!"#$%&"'$ ($ !")$*&+,-". .!'",+/"& * 0% &)"$!

)$&
$1234 (21526734 &'")

!"#"
!"#$%&"$ '()! (*&$%+! ,$ &+&!$ -$

tel-00408745, version 1 - 3 Aug 2009

!$%&'&($(&)* + ,&-&./- 0/1 2/34/-34/1

84 39 :;<46=;5> 84 !;14 &2?@;7 .:5;?23;=
.'"/+0,+&" 1 2%3(!40&+5)$
'0!

!"#$ %&'&(&)

#)'5(&)*1 /63$3/1 7)5- '$ 89(-)').&/

0/ ':;*(/-*/(
.()&$%)$ ')*,+5)$4$%& ,$ 66 70%8+$! 699: -$80%& ,$ ;)!< /(4'(#" -$ 1
=>
=>
=>
=>
=>
=>

=+/?$,
G0,+B0J+4
M)<
0&!+/P
M)+,,0)4$

@+8$+,,
H0**()#
CK(),4+%$
N$-)/
.$%0/
A!8(<QR$,,$!

AB.CDEB@.D2F.
2B@2C
A> L8!<
A> N+OK$
2.CL
L)!$/(4

TU9 1F9 Q 2B@2CV

!"#+-$%&
I)&$)!
@0''(!&$)!
@0''(!&$)!
@0''(!&$)!
LS04+%0&$)!

tel-00408745, version 1 - 3 Aug 2009

!"!#$%!"!&'(

tel-00408745, version 1 - 3 Aug 2009

!""! "#$%! !%" &! '()*" +,)- &.-/ "(010*& 2)* 0 !) &*!) 3(*-4*30&!5!-" 6 &,7897: ;.3#*0 :-"*<
3.&*% 0) %!*- +! &,=2)*3! >&0-$"! !" &.(% +! 5!% %=?.)(% 6 &,@>AB B0)%0--! !" 6 7-"!& 9!%!0(4#
05C(*+/!D E!% 3(!5*!(% (!5!(4*!5!-"% -! 3!)1!-" 0&&!( 2),6 5!% 4.&&$/)!% +0-% 4!% "(.*% &0C.<
(0".*(!% !" 0)F =2)*3!% +! %)33.(" 0+5*-*%"(0"*' !" *-'.(50"*2)! 2)* 5,.-" 044.530/-=D ;0-% !)F
4!""! "#$%! -,0)(0*" 30% 1) &!% ?.)(%D G,!F3(*5! 0)%%* 50 /(0"*")+! 6 5!% =")+*0-"% 2)* .-" ="=
".)" 0) &.-/ +!% 4!% +!(-*$(!% 0--=!% 5.- 3(!5*!( %)33.(" !" 50 /(0-+! ?.*!D
G! (!5!(4*! 4#0&!)(!)%!5!-" &!% 5!5C(!% +! 5.- ?)(H 2)* .-" 044!3"= +! (033.("!( %)( 4!""!
"#$%! !" +! %! +=3&04!( 3.)( &0 3(=%!-"0"*.-D B!)( 1!-)! 6 ;.3#*0 :-"*3.&*% !%" )- /(0-+ #.--!)(
3.)( 5.*D
A*-0&!5!-"I 4,!%" 6 50 '!55! G.)50-0 2)! ?! +=+*! 4! "(010*&D ;0 3(=%!-4! 6 5!% 4J"=% !"
%!% !-4.)(0/!5!-"% .-" ".)?.)(% ="= 50 3(!5*$(! %.)(4! +,*-%3*(0"*.-D

tel-00408745, version 1 - 3 Aug 2009

!"#$ %$& '!()*+$&

tel-00408745, version 1 - 3 Aug 2009

!"#$%&'(#)%"
!

"#$#%&'()#* *+% ',-$).%$.) ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! !

!/

0& 1#)%2'23(. .) *.* ($)#%4)* ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! !

!5

0.* 6(78+')#* 6. '& 1#)%2'23(. ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! !

!9

0. :&**&3. ; ',#8<.''. 6. '& 1#)%2'23(. ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! !

* +,-.%/'#)%" &' &%01)"2 &2 /1 0-#$%/%3)2

/!

0.* 1#8&$(*1.* 6. >&*. 6. 1.*+%. ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! !

/!/

0,#@2'+)(2$ 6. '& 1#)%2'23(. :&**(@. ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! !

/!/!

A#)<26.* :'+* .78&8.* :2+% '& 8&:)+%. 6.* :&B+.)*

! ! ! ! ! ! ! ! ! ! ! !

/!/!/

-$C#%.$8. 6.* :.%C2%1&$8.* %#*.&+ ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! !

/!/!5

E3%#3&)(2$ 6.* ($C2%1&)(2$* *+% '. )%&F8 ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! !

/!/!9

0,#8<&$)(''2$$&3. 6&$* '& 1#)%2'23(. ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! !

/!/!D

E$&'H*. .) 126#'(*&)(2$ 6+ )%&F8 82''.8)# ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! !

0,#@2'+)(2$ 6. '& 1#)%2'23(. &8)(@. ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! !

/J

/!5!

A#)%2'23(. 6,+$ 8<.1($ 6. ',-$).%$.) 6,+$. 1&$(K%. &8)(@. ! ! ! ! ! ! ! ! !

/J

/!5!/

A#)%2'23(. 6,+$ %#*.&+ 6,+$. 1&$(K%. &8)(@. ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! !

/=

/!5

/!9

L2$8'+*(2$* ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! !

5 +1 0-#$%/%3)2 6177).2

55

5!

M8<&$)(''2$$&3. ).1:2%.' .) *:&)(&' 6+ %#*.&+ ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! !

59

5!/

0,#8<&$)(''2$$&3. 8211. +$ :%2>'K1. N2($) ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! !

5D

5!5

O.%C2%1&$8.* 6. ',2:)(1(*&)(2$ N2($). ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! !

5G

5!9

L2$8'+*(2$* ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! !

5?

8 +2 9$')# )"#$%&')# 61$ /,-(:1"#)//%""132 ; &-#2(#)%" 2# (/177202"# &27 <%#7

.%/'0)"2'=
8
9!

A#)<262'23(. ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! !

9/

9!/

P#*+')&)* $+1#%(B+.* ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! !

95

!"#$ %$& '!()*+$&

!"#

$%&'())( *+,-%.( .( /)011(*(2,3.+,(/,4%2 501+( 1&6 .(1 427%6*0,4%21 86%,%/%)046(1 !9

!"!

:%2/)&14%21 " " " " " " " " " " " " " " " " " " " " " " " " " " " " " " " " " " " " " " "

!;

!" #$%%&'"()$* (+,-$%+''+ ". /+%0)#+ 1+ '2&#3"*()''$**"4+

56

9"<

=1,4*0,4%2 .( )0 ,04))( .( >602.(1 8%8&)0,4%21 " " " " " " " " " " " " " " " " " " " "

!?

9"@

A%.B)( (, C20)D1(

" " " " " " " " " " " " " " " " " " " " " " " " " " " " " " " " " " "

9E

9"#

F0)4.0,4%2 (, 088)4/0,4%2 " " " " " " " " " " " " " " " " " " " " " " " " " " " " " " " "

9#

9"!

:%2/)&14%21 " " " " " " " " " " " " " " " " " " " " " " " " " " " " " " " " " " " " " " "

9!

tel-00408745, version 1 - 3 Aug 2009

7 !" ,&(%$'$4)+ "#()0+

;"<

G( /%2,6H)( .( /%2>(1,4%2 (, )( /%2,6H)( .I(66(&61 .021 J :K " " " " " " " " " " " "

9;

;"@

K06/%&61 .(1 *0/-42(1 .021 J :K

" " " " " " " " " " " " " " " " " " " " " " " " " "

9?

;"#

:%2/)&14%21 " " " " " " " " " " " " " " " " " " " " " " " " " " " " " " " " " " " " " " "

;<

6 8$*#'./)$*/

79

:);')$4%"-3)+

8< 1&(")''&

69

!)/(+ 1+ -.;')#"()$*/

6=

>**+?+/
> @+A$%,.'"()*4 (3+ B$*)($% C'"#+,+*( C%$;'+, D E-(),"' F+(G$%HIJ)1+ K",I
-')*4
: @"*H)*4 L'$G/ A%$, K",-'+1 M%"N#
8 O/(),"()*4 B+,;+%/3)- )* " B.'()#"/( K+//)$*
P O?-+%)+*#+/ $* +*3"*#)*4 1"(" #$''+#()$* )* '"%4+ *+(G$%H/

!"#$%&' (
!"#$%&'"($!

tel-00408745, version 1 - 3 Aug 2009

"#$#%&'()#* *+% ',-$).%$.)

!"#$%&#%$ ' ($( )&(( )*++% ,# &(-%', -.+/0%1 *,2%&$1 3%4.50%1 *6 0% -%,0 -%&2.)% *7%&$ ',4
,$.0.-'$%,&- %-$ 0% 8&*,$'9% ', +.%,48 :%- /';,%$- <%# '#90'.- =%-$ >7*&$?@ A% )B*.4 ' ($( :(0.5(&(
/*,& 0' -.+/0% &'.-*# :% C').0.$%& 0!.#$%&)*##%4.*# :%- &(-%',41 0' )&('$.*# :%- '//0.)'$.*#- %$
0!'D*,$ :%- -%&2.)%- E 0!"#$%&#%$@ F# :%- /&.#)./%- C*#:'+%#$',4 :% 0!"#$%&#%$ %-$ ;,% 0% &(-%', :*.$
&%-$%& 0% /0,- -.+/0% /*--.50% %$ ;,% 0!.#$%00.9%#)% :*.$ -% $&*,2%& :'#- 0%- +')B.#%- :%- ,$.0.-'$%,&;,. %7%)$,%#$ %00%- 0%- */(&'$.*#- )0'--.;,%- :% $G/% &%)*,2&%+%#$ :%- %&&%,&-1 ':'/$'$.*# :%
:(5.$1 0*)'0.-'$.*# :%- &%--*,&)%-1 %$ '.#-. :% -,.$%@ A!%-$ 0% C'+%,4 /&.#)./% :% 5*,$H%#H5*,$ :%
0!"#$%&#%$ <%# '#90'.- %#:H$*H%#:?@ A% /&.#)./% /%&+%$ E $*,$% '//0.)'$.*# :% 0!"#$%&#%$ :(-.&'#$
)*++,#.;,%&1 :!%#2*G%& :%- /';,%$- E #!.+/*&$% ;,%00% ':&%--% "I %$ #!.+/*&$% ;,%0 #,+(&*
:% /*&$@ % &(-%', C'.$ %#-,.$% :% -*# +.%,4 /*,& ')B%+.#%& )% /';,%$ E 0' 5*##% :%-$.#'$.*#@
J. 0' :%-$.#'$.*# #!%-$ /'- $&*,2(%1 0' -*,&)% %-$ #*$.K(%1 -.#*# 0% /';,%$ %-$ 0.2&( /*,&2, ;,!.0
#!' /'- ($( &%D%$( *, )*&&*+/, 0% 0*#9 :, )B%+.#@ L*,$ )%). -% C'.$ -'#- 0% 5%-*.# :% &(-%&2%&
:%- &%--*,&)%- :'#- 0% &(-%', *, :% :%+'#:%& 0!',$*&.-'$.*# :% )%&$'.#%- %#$.$(- ':+.#.-$&'$.2%)*++% )% ;,. -% C'.$ :'#- 0%- &(-%',4 $(0(/B*#.;,%-@ A%$$% /*--.5.0.$( :% )*++,#.;,%& %# $%+/&(%0 '2%) /0,-.%,&- %#$.$(- &(-%',4 :.-$&.5,(%- /'&$*,$ :'#- 0% +*#:% ' ($( ,# C')$%,& .+/*&$'#$
/*,& 0' )&('$.*# :!.#$(&%--'#$- -%&2.)%- )*++% 0% M>=1 0% +'.01 0% I%%&H$*HI%%&1 0%- *2%&0'G-1 %$
'.#-. :% -,.$%@ >$ )!%-$ /&.#)./'0%+%#$ E %00% ;,% &%2.%#$ 0% 9&'#: -,))N- ;,!' )*##, 0!"#$%&#%$@
O'0B%,&%,-%+%#$1 &.%# #!%-$ 9&'$,.$ :'#- 0' 2.%@ !*,2%&$,&% :, &(-%', ', $&'K) )*+/0.;,%
0' $P)B% :% 0!':+.#.-$&'$.*# :% 0!"#$%&#%$ %$ :% -*# .#9(#.%&.% %$ )*#:,.$ E :%- )*+/*&$%+%#$#(C'-$%- E 0' C*.- /*,& 0%- ,$.0.-'$%,&- %$ /*,& 0%- */(&'$%,&- &(-%',4@ F# ,$.0.-'$%,& :% 0!"#$%&#%$
/%,$ %#2*G%& ',$'#$ :% /';,%$- ;,% -*# .#$%&C')% &(-%', %$ 0' 2.$%--% :% -' +')B.#% 0% /%&+%$$%#$@
F#% /'&$.% :% )% $&'K) %-$ :(-.&'50%1 +'.- 0% &%-$% /%,$ )*#-.-$%& %# ,# $&'K) )*+/0N$%+%#$
.#:(-.&'50% )*++% 0%- '$$';,%- ;,. 2.-%#$ E C'.&% $*+5%& :%- -%&2%,&- *, 5.%# 0%- %+'.0- :, $G/%
JIQO ;,. C*#$ :(5*&:%& 0%- 5*R$%- ',4 0%$$&%-@ "0 #!G ' ',),# +()'#.-+% &(-%', ;,. %+/S)B%
0%- ,$.0.-'$%,&- :% 0!"#$%&#%$ :!%#2*G%& :, $&'K) .#:(-.&'50% T "0- *#$ $*,- 0' /*--.5.0.$( :!,$.0.-%&
', +'4.+,+ 0%,&- 5'#:%- /'--'#$%- :.-/*#.50%-1 %$ 0%- &*,$%,&- :, &(-%', C*#$ :% 0%,& +.%,4
U

!"#$%&' () *+%&,-./%$,+
!"#$ $%&'(%$ &%#$) *$'+,)- .%) "!/$'*%#$) $/)%'#0 1% &%#$ ,2*/ 3"4%5* #5 6#0 1% !'7#%*) '$$43'5*
8 ,9',#5 1% &%#$) $"#*%#$) %* *:,9%5* 1% &% $%&'(%$ 8 *$'3%$) &%#$ $/)%'# '#0 $/)%'#0 3"4)45)- ;&)
%0%$,%5* '#,#5 ,"5*$2&% 8 !$4"$4 )#$ ,% *$'+, %* 5% !%#3%5* $4%5 '!!$%51$% 1%) "<)%$3'*4"5) ='4*%)
!'$ &>#5 %* &>'#*$%- ?5 "!/$'*%#$ $/)%'# 1"4* '@4$ *"#* )%#& %5 )% <')'5* )#$ ,% 7#>4& "<)%$3% 1'5)
)"5 $/)%'#A %* &% B'04B#B 7#>4& !%#* ='4$% !"#$ &4B4*%$ &% 1/@:* %5 ,') 1% ,"B!"$*%B%5* '5"$B'&
C%-@A ="$*% %* )#<4*% ,$"4))'5,% 1# *$'+,D %)* 1% &% <&"7#%$ 8 &' !/$4!9/$4% 1% )"5 $/)%'#- E'4)
!"#$ &% <&"7#%$ 4& ='#* !"#3"4$ &% B%)#$%$A &% ,"B!$%51$%A %* &>4)"&%$A 1%) *:,9%) 7#4 !%#3%5* F*$%
=',4&%) 1'5) #5 !%*4* $/)%'#A B'4) 7#'51 4& )>'@4* 1>#5 @$'51 $/)%'# 1% *$'5)4* "G 1%) ,%5*'45%)
1% E<!) ,4$,#&%5*A ,%) *:,9%) 1%34%55%5* 14H,4&%) 8 $/'&4)%$ 3"4$% 4B!"))4<&%)- I,*#%&&%B%5*A &%)
"!/$'*%#$) 1# ,"%#$ 1% &>;5*%$5%* )% ,"5*%5*%5* &' !&#!'$* 1% *%B!) 1>#5 )4B!&% ="$J'$145@ 1%)
!'7#%*) '#0 $/)%'#0 3"4)45) %* &'4))%5* &' *:,9% 1% 1/*%,*4"5 %* <&",'@% 1# *$'+, 451/)4$'<&% '#0
$/)%'#0 1%)*45'*'4$%)- .",'&%B%5*A 1%) 1/,4)4"5) )4B!&%) )"5* !$4)%) ,"BB% &% ,9"40 1%) )"$*4%) 7#4

tel-00408745, version 1 - 3 Aug 2009

!%$B%**%5* 1% B4%#0 $/!'$*4$ &% *$'+,A %* &% $%$"#*'@% 1>#5% !'$*4% 1# *$'+, %5 ,') 1% )#$,9'$@%
"# 1% $#!*#$% 1>#5 1%) &4%5).%) ,"5)/7#%5,%) 5'*#$%&&%) 1% &>'<)%5,% 1# ,"5*$2&% +5 1'5) &% ,"%#$ 1# $/)%'# %* 1% &'
,"&&'<"$'*4"5 %5*$% &%) "!/$'*%#$) C'#*$% 7#% &>/,9'5@% 1>45="$B'*4"5) )#$ &% $"#*'@%D )"5* K
L .' !$"&4=/$'*4"5 1%) '**'7#%) 1'5) &>;5*%$5%* C!'$ %0%B!&% ,%&&%) 1% *(!% MM"N 7#4 34)%5* 8
*"B<%$ #5 )%$3%#$A '45)4 7#% &%) 3%$) %* &%) 34$#)D %* 1# *$'+, 451/)4$'<&% C&% B'4& NOIED 7#%
)%#&%) &%) B',945%) %* $/)%'#0 1%)*45'*'4$%) !%#3%5* *$'4*%$ %* /&4B45%$- P% *$'+, 451/)4$'<&%
%)* ,94Q$/ 1% 5") R"#$) 8 #5% 14S'45% 1% !"#$,%5*) 1# 3"&#B% *"*'& 1# *$'+, ;5*%$5%* '3%,
#5 %Q%* 5/=')*% !"#3'5* 1/!'))%$ &% $'*4" 1% ,% *$'+,L P"BB% &%) '**'7#%) %* &% *$'+, 451/)4$'<&% 5% )"5* !') <&"7#/) 8 &%#$) )"#$,%) )#4*% 8
&>'<)%5,% 1% ,"&&'<"$'*4"5 %5*$% &%) "!/$'*%#$)A &%) !'7#%*) ,"$$%)!"51'5* *$'3%$)%5* &%
$/)%'# !"#$ F*$% $%R%*/) 8 &' +5A ,'#)'5* '45)4 #5% !%$*% 1% $%))"#$,%) %* #5% 1/@$'1'*4"5
1%) )%$34,%)L .>'<)%5,% 1% ,"&&'<"$'*4"5 %5*$% &%) "!/$'*%#$) C'#*$% 7#% &>/,9'5@% 1%) 45="$B'*4"5) )#$ &%
$"#*'@%D $%51 14H,4&% &% 14'@5")*4, 1%) !$"<&TB%) $/)%'#0- ?5 "!/$'*%#$ !%#* 8 &' &4B4*% %*
)>4& 14)!")% 1% B/*9"1%) %H,',%) ,"55'U*$% ,% 7#4 )% !'))% 1'5) )"5 $/)%'#A B'4) 4& ' '#,#5
B"(%5 1% )'3"4$ ,% 7#4 )% !'))% 1'5) &%) '#*$%) $/)%'#0 54 &' !$"3%5'5,% 1%) !$"<&TB%)- .%)
%0%B!&%) *(!47#%) )"5* &' 1/@$'1'*4"5 1%) !%$="$B'5,%) )#4*% 8 &>'#@B%5*'*4"5 1# 1/&'4
%* &' ,9#*% 1% &' <'51% !'))'5*% 1"5* &>"!/$'*%#$ $/)%'# 5% !%#* !') ,"55'U*$% &>"$4@45%
%0',*% %5 1%9"$) 1% )"5 $/)%'# CP>%)* 1/R8 #5% *:,9% 14H,4&% 8 $/'&4)%$ 1'5) )"5 $/)%'#DP% !$"<&TB% 1% B'57#% 1% 34)4"5 @&"<'&% )>'!!&47#% <4%5 5'*#$%&&%B%5* '#0 #*4&4)'*%#$)
7#4 %#0 '#))4 5>"5* !') &' !"))4<4&4*/ 1% ,"55'U*$% &>"$4@45% 1%) !$"<&TB%) 1'5) &>;5*%$5%*L .' !$"&4=/$'*4"5 1%) V3%$&'() C$/)%'#0 1% $%,"#3$%B%5*D 7#4 ,9%$,9%5* 8 $/)"#1$% &%) !$"W
<&TB%) 1% &>;5*%$5%* %* 8 ,"B<&%$ &% B'57#% %5 )%$34,%) ="#$54)- XVY Z[\] %)* #5 %0%B!&%
*(!47#% 1>#5 $/)%'# V3%$&'( 7#4 $%14$4@% &% *$'+, 8 *$'3%$) 1%) !$"0() '+5 1% ,"5*"#$5%$
&%) !$"<&TB%) 1% $"#*'@% 1'5) &>;5*%$5%*- .%) )()*TB%) 1% ,""$1"55/%) ,"BB% ^43'&14 Z_`]

!"! #$ %&'()*)+,- -' .-. ,/'&(0'.

!"# $#$ %&!%!'$' %!(& )*##&* + ,- ./'%!'/#/!" .*' -%%,/0-#/!"' *# .*' *"#/#$' &$'*-(1 ("
)!2*" %!(& '* ,!0-,/'*& #!%!,!3/4(*)*"# ,*' ("*' %-& &-%%!&# -(1 -(#&*'5 6*' %,-#*7!&)*'
.* )*'(&* 0!))* 89/##*& .* :;<=; > ?@ '!"# A*"(*' 0!)B,*& ,* )-"4(* .* A/'/!" 3,!B-,*
*# 0!,,-B!&-#/!" *"#&* !%$&-#*(&'5 C,,*' %*&)*##*"# -A*0 .*' )*'(&*' .* B!(# *" B!(# *# .*'
/"7!&)-#/!"' '(& ,*' #-B,*' .* &!(#-3* 0!,,*0#$*' *" 0*&#-/"' &!(#*(&' !(A*&#' .* ,D<"#*&"*#
.* .!""*& + #!(# ,* )!".* ("* /.$* .* 0!))*"# ,* &$'*-( '* %!&#* + (" /"'#-"# .!""$ *"
#*&)*' .* %*&7!&)-"0*' *# #!%!,!3/*5

!" #$ %&'()*)+,- -' .-. ,/'&(0'.

6*' !%$&-#*(&' .* ,D<"#*&"*# *# ,*' (#/,/'-#*(&' '* #&!(A*"# -,!&' + *(1 #!(' '*(,' %!(& 'D-.-%#*&

tel-00408745, version 1 - 3 Aug 2009

-(1 A-&/-#/!"' *# -(1 -"!)-,/*' .* 0* 3&-". &$'*-( 4(/ 'D-%%*,,* ,D<"#*&"*# !E ,*' '*(,' '*&A/0*'
4(/ *1/'#*"# '!"# ,* &!(#-3* %-4(*# %-& %-4(*# *# ,* 0!"#&F,* .* 0!"3*'#/!" .* B!(# *" B!(#
7!(&"/ %-& G:H 4(* )-,I*(&*('*)*"# B*-(0!(% .D(#/,/'-#*(&' %*(A*"# -&&J#*& %!(& ,*' B*'!/"'
.* ,*(&' -%%,/0-#/!"'5 G!(# ,* &*'#* *'# /"0!""( *# + .$0!(A&/& K ,- #!%!,!3/* .*' &$'*-(1 *# ,)-"/L&* -A*0 ,-4(*,,* /,' '!"# 0!""*0#$'M ,*' 0-&-0#$&/'#/4(*' .*' 0I*)/"' *# ,/*"' .* ,D<"#*&"*#M
,*' %!,/#/4(*' .* &!(#-3* -.!%#$*' %-& ,*' -(#&*' !%$&-#*(&' &$'*-(1M ,*' -%%,/0-#/!"' 4(* ,*'
(#/,/'-#*(&' (#/,/'*"#M ,D/"#*"#/!" .*' (#/,/'-#*(&'M ,* 0!)%!&#*)*"# .*' (#/,/'-#*(&'M ,D!&/3/"* .*'
%&!B,L)*' &$'*-(1M *# -/"'/ .* '(/#*5 G!(# 0*0/ "* %*(# %-' '* 7-/&* '-"' ("* )$#&!,!3/* .( &$'*-(
4(* 0- '!/# %-& ,*' !%$&-#*(&' !( %-& ,*' (#/,/'-#*(&'5 =* "!' N!(&'M ,- %&-#/4(* ,- %,(' (#/,/'$*
*'# 4(* 0I-0(" %&!0L.* + '- )-"/L&* %!(& 0!)%&*".&* -( )/*(1 0* 4(/ '* %-''*5 O" !%$&-#*(&
)*'(&* '!" &$'*-( *# *''-/* .* 0!)%&*".&* ,* )/*(1 %!''/B,* '(& 0* 4(/ '* %-''* .*.-"' *# '(& 0*
4(/ -&&/A* .* ,D*1#$&/*(&5 6*' (#/,/'-#*(&' *P*0#(*"# + ,*(& #!(& .*' )*'(&*' .* B!(# *" B!(# %!(&
0!)%&*".&* ,*' 0-&-0#$&/'#/4(*' .*' ./P$&*"#' &$'*-(1 .* ,D<"#*&"*# .-"' ,- ,/)/#* .( %!''/B,* Q("
!%$&-#*(& %*(# #&L' B/*" B,!4(*& ,* #&-R0 .* )*'(&* .D(" (#/,/'-#*(& R"-,S5 C# R"-,*)*"#M /, 2
- ,- 0!))("-(#$ '0/*"#/R4(* 4(/ 'D*P!&0* .D*P*0#(*& .*' !%$&-#/!"' .* )*'(&* + 3&-".* $0I*,,*
Q0!))* 89/##*& .* :;<=;M ;TH .* U6;UVM <HTWU .* 8%&/"#,-B'M *# .*' %&!N*#' .* &*0I*&0I*
0!))* T*#&!%!,/'M CGWT<:M *# TWTCS *# .* )*##&* ,*' &$'(,#-#' *# ,*' !B'*&A-#/!"' *P*0#($*'
+ ,- ./'%!"/B/,/#$ .*' (#/,/'-#*(&' *# .*' !%$&-#*(&'5
H-&)/ ,*' /"#$&J#' .* ,- )$#&!,!3/* .*' &$'*-(1M !" %*(# 0/#*& ,*' '(/A-"#' K
X 6- )$#&!,!3/* '*&# + .$#*0#*& ,*' %&!B,L)*' .* 7!"0#/!""*)*"# Q%-& *1*)%,* ,*' -##-4(*'M
,*' &(%#(&*' .*' ,/*"'M ,*' B-/''*' .* .$B/#S *# %-& 0!"'$4(*"0* + 2 &*)$./*& .-"' ,- )*'(&*
.( %!''/B,*5
X 6- )$#&!,!3/* '*&# + )/*(1 0!)%&*".&* ,D<"#*&"*# Q*535M '!" #&-R0M '- #!%!,!3/*S *# + %&!%!'*&
.*' )!.L,*' #I$!&/4(*' 4(/ %*&)*##*"# .* %&$./&* '!" $A!,(#/!" *# .* &*%&!.(/&* ,*' &$'(,#-#'
.-"' .*' '/)(,-#/!"' *# .*' %,-#*7!&)*' *1%$&/)*"#-,*'5
X 6- )$#&!,!3/* '*&# + 0-&-0#$&/'*& ,*' -%%,/0-#/!"' *# ,*' 0!)%!&#*)*"#' .*' (#/,/'-#*(&'5
:*0/ %*&)*# .D$A-,(*& ,D/)%-0# .* #!(#* "!(A*,,* -%%,/0-#/!" Q*535M ,* H-/&Y+YH-/&S *# .*

!"#$%&' () *+%&,-./%$,+
!"#! $"#%&'# (")*"+!&)&$! ,&-.-/ 0& 1234 5#+ 0& +65&'# &! 5"$ !+'7(8 9' )6!+"0".:& *&+)&! '#; "*6+'!&#+5 <& ):&#; <:)&$5:"$$&+ 0&#+5 +65&'#; &! 0&5 '<):$:5=
!+&+ ,("$$&(!:%:!6/ :$.6$:&+:& <& !+'7(/ 70!+'.&/ >'(!#+'!:"$4 &! '#; #!:0:5'!&#+5 <& ):&#;
("$(&%":+ 0&#+5 '**0:('!:"$5 ,0&5 +&$<+& *'+ &;&)*0& '<'*!'!:%&5 ? 0' !"*"0".:& <# +65&'#
&! 5&5 *&+>"+)'$(&54-

!" #$% &'()*+,-% &$ +. /-,01+12'$

@0 &5! 6%:<&$! A#B&$ 0B'C5&$(& <B#$ *0'$ <& ("$!+D0& <'$5 0B@$!&+$&! A#: *&+)&! '#; #!:0:5'!&#+5
&! '#; "*6+'!&#+5 <B'.:+ &$5&)C0& *"#+ ):&#; "*6+&+ 0& +65&'#/ <&5 5"0#!:"$5 *"#+ 0' )6!+"0".:&
5"$! ? <6%&0"**&+ *"#+ +6'0:5&+ 0&5 >"$(!:"$5 (:!6&5 (:=<&55#5- E&5 5"0#!:"$5 <":%&$! 5& C'5&+ 5#+

tel-00408745, version 1 - 3 Aug 2009

0& *&# <& >"$(!:"$5 <:5*"$:C0&5 <'$5 0& ("&#+ <# +65&'# ("))& 0&5 *"0:!:A#&5 <& 5(F&<#0:$. G@GH
<'$5 0&5 +"#!&#+5/ 0&5 )&55'.&5 @EI2 JKLM ,*+"C'C0&)&$!4 +&$%"N65 *'+ 0&5 +"#!&#+5 &! )'(F:$&5
&$ ('5 <& +&O&! <& *'A#&!5/ &! 0&5 !'C0&5 <& +"#!'.& ,5#+!"#! 0&5 !'C0&5 PQ2 JKRM4- SB#$ (D!6 0'
)6!+"0".:& *&#! 5& >':+& &$ ("00&(!'$! 0& !+'7( 5#+ 0&5 0:&$5 <& 0B@$!&+$&! &! &$ 0B'$'0N5'$! '*+T5E&(: &5! 5"#%&$! '**&06 0B'**+"(F& *'55:%& *"#+ 0' )6!+"0".:& ,*'5 <B:$O&(!:"$ <& !+'7(4- S& 0B'#!+&
(D!6/ 0' )6!+"0".:& *&#! 5& >':+& &$ &$%"N'$! <&5 *'A#&!5 &! &$ )&5#+'$! 0' +6'(!:"$ <# +65&'# ?
(&5 *'A#&!5- E&(: &5! 5"#%&$! '**&06& 0B'**+"(F& '(!:%& <'$5 0' )6!+"0".:&- 9&5 *'A#&!5 *&#%&$!
U!+& <6<:65 ? 0' )&5#+& ("))& :05 *&#%&$! U!+& <&5 *'A#&!5 <& <"$$6&5- 9' )&5#+& &00& )U)& 5&
>':! <& C"#! &$ C"#! &$!+& <&5 )'(F:$&5 "+<:$':+&5 "# <&5 5"$<&5 <6<:6&5- V! A#& (& 5":! <'$5 0&
('5 *'55:> "# '(!:>/ #$& >":5 0&5 )&5#+&5 &W&(!#6&5/ :0 >'#! *'55&+ *'+ #$& '#!+& *F'5& A#: &5! (&00&
<& 0B'$'0N5& <&5 +65#0!'!5 &! <& 0B:$>6+&$(& <& (& A#: 5& *'55& <'$5 0& +65&'#- S'$5 0& ('5 *'55:>/ :0
5B'.:! <& (")*+&$<+& (& A#: 5& *'55& '# *":$! <& )&5#+& &! ':00&#+5- S'$5 0& ('5 '(!:>/ :0 5B'.:! <&
(")*+&$<+& (& A#: 5& *'55& !"#! '# 0"$. <# (F&):$ !+'%&+56 *'+ 0&5 *'A#&!5 <& )&5#+&- E& A#B"$
)&5#+& <& C"#! &$ C"#! $B&5! A#& 0B'((#)#0'!:"$ <& !"#! (& A#& +&$("$!+&$! 0&5 *'A#&!5 <'$5
0&5 <:W6+&$!5 0:&$5 &! +"#!&#+5 <# +65&'#- 9&5 )&5#+&5 *&#%&$! U!+& &W&(!#6&5 &$ *0#5:&#+5 *":$!5
&$5#:!& ("++606&5 &$5&)C0& *"#+ "C!&$:+ *0#5 <& +65#0!'!5- V00&5 *&#%&$! '#55: U!+& +&*+"<#:!&5
<'$5 0& !&)*5 *6+:"<:A#&)&$! "# 5&0"$ #$ *+"(&55#5 5!"(F'5!:A#& *"#+ *&+)&!!+& #$& 6!#<& <&
0B6%"0#!:"$ !&)*"+&00& <# +65&'# &! 0' <6!&(!:"$ <& !"#! (F'$.&)&$! <'$5 0& (")*"+!&)&$! &! 0&5
*&+>"+)'$(&59B:$!6+U! <& 0' )6!+"0".:& &5! (0':+ )':5 0&5 <:X(#0!65 A#: 0B'((")*'.$&$! 5"$! $")C+&#5&5E&5 <:X(#0!65 *+"%:&$$&$! *+:$(:*'0&)&$! <& 0' .+'$<& !':00& <& 0B@$!&+$&! &! <& 5"$ "#%&+!#+&/
&! <# .+'$< %"0#)& <& !+'7( A#: N (:+(#0&- 9&5 ("$56A#&$(&5 <& (&00&5=(: 5#+ 0&5 <&#; '**+"(F&5
<& )6!+"0".:& 5"$! <:W6+&$!&5 )':5 &00&5 5& +&O":.$&$! 5#+ 0' <:X(#0!6 <& 0' !Y(F& Z
8 2"#+ 0&5 )&5#+&5 '(!:%&5/ #$ .+'$< +65&'# %&#! 5:)*0&)&$! <:+& #$& *F'5& <B:$>6+&$(&
(")*0:A#6& *#:5A#& <:%&+5 *F6$")T$&5 *&#%&$! '**'+'[!+& !"#! 0& 0"$. <# (F&):$ 5#:%: *'+
0&5 *'A#&!5- 2"#+ (F'A#& )6!+:A#& ,&-.-/ !'#; <& *&+!&5/ C'$<& *'55'$!&4/ :0 <&%:&$! '0"+5
:)*"+!'$! <& C:&$ (F":5:+ 0' )'$:T+& '%&( 0'A#&00& 0&5 *'A#&!5 5"$! &$%"N65 '7$ <B"C!&$:+

!"! #$% &'()*+,-% &$ +. /-,01+12'$

!" #"$!!"%& &'(%!)*) +,(($-!". /*!0"%&"%("#"1) 2"))" ,+)$#$(*)$,1 13"() +*( (,%4"1) 5*2$!"
6 5*$&". 7" 8%$ "() -,1 (%& 2"&)*$1( 20"#$1( +"%) (3*4'&"& #*%4*$( (%& 93*%)&"( 20"#$1(.
:1 ;&*19 &'("*% 4"%) *%(($ 9$&" %1 ;&*19 4,!%#" 9" +*8%")( (,19"( 6 "14,<"& ") +!%( 9"
#*20$1"( 6 5*$&" $1)"&4"1$& +,%& *))"$19&" %1" -,11" 2,%4"&)%&" 9% &'("*%.
= >,%& !"( #"(%&"( +*(($4"(? !" +&,-!@#" "() +&$12$+*!"#"1) 9*1( !" ;&*19 4,!%#" 9% )&*A2
6 2,!!"2)"& (%& !"( !$"1( 9% &'("*% +,%& *4,$& %1" 4$($,1 2!*$&" 9" 2" 8%$ (" +*((". B) ($ ,1
9%+!$8%" 2"( +,$1)( 9" #"(%&" +*(($4"? !" 4,!%#" 9" )&*A2 6 2,!!"2)"& *%;#"1)" *%(($. C%&
!" !$"1 93*22@( 93%1" #*20$1"? 8%"!8%3%1 +"%) (" 2,1)"1)"& 9" 2,!!"2)"& ),%) !" )&*A2 ") 9"
!3*1*!<("&. D" 4,!%#" 9" 2" )&*A2 4* &"()"& 9*1( !* !$#$)" 9% &*$(,11*-!". /*!0"%&"%("#"1)
(%& 9"( !$"1( &'("*%E -$"1 20*&;'( F".;.? 6 9"( 2"1)*$1"( 9" /-+( +*& "E"#+!"G? 2"2$ 9"4$"1)
$#+,(($-!". H! (3*;$) *!,&( 9" )&,%4"& 9"( *!;,&$)0#"( 93*;&';*)$,1 9% )&*A2 8%$ &'9%$("1)
!" 4,!%#" 9" 9,11'"( 6 &"#,1)"& *+&@( 6 !3%1$)' 2"1)&*!" ,I !" )&*$)"#"1) ("&* "J"2)%'.

tel-00408745, version 1 - 3 Aug 2009

7!*$&"#"1) 2"( *!;,&$)0#"( (,1) 9'+"19*1)( 9" !3*++!$2*)$,1 K %1 *!;,&$)0#" 93*;&';*)$,1

;'1'&*! "() 2"!%$ 8%$ +"&#") *%E 9,11'"( *;&';'"( 93L)&" %)$!$('"( +*& !" +!%( ;&*19 1,#-&"
93*++!$2*)$,1( F!3*;&';*)$,1 "() %1" ,+'&*)$,1 $&&'4"&($-!" M MG. B) 2"2$ 13"() +*( !" ("%! +&,N
-!@#". >*&2" 8%" #L#" ($ %1 *!;,&$)0#" 9" 2,#+&"(($,1 "O2*2" "() )&,%4'? %1" %1$)' 9"
#"(%&" +*(($4" +"%) L)&" 9*1( !3$12*+*2$)' 9" )&*$)"& ),%( !"( +*8%")( "1 )"#+( &'"! ") 93*2N
2'9"& 6 !* #'#,$&" +,%& 20*2%1 9" 2"( +*8%")( *A1 9" #"))&" 6 P,%& !"( 20*#+( ") &";$()&"(
2,&&"(+,19*1)(. H! (3*;$) *!,&( 9" )&,%4"& 9"( *!;,&$)0#"( 8%$ +"&#"))"1) 9" 2,1()&%$&" %1"
$#*;" 9% )&*A2 6 +*&)$& 93%1 (,%(N"1("#-!" 9" +*8%")(? 2" 8%$ 1,%( &*#@1" 6 ),%) %1 !*&;"
9,#*$1" 8%$ "() 2"!%$ 9" !3'20*1)$!!,11*;" 9% )&*A2 ") 9" (,1 $#+*2) (%& !* #')&,!,;$".
= D" ),%) 1*)%&"! +&,-!@#" 9" !3$1)"&+&')*)$,1 9"( &'(%!)*)( 8%$ +&,4$"1) 9" !* 2,#+!"E$)'
9" !3H1)"&1") ") 9% ;&*19 1,#-&" 9" 5*2)"%&( 8%$ 9'2$9"1) 9" 2" 8%$ (" +*((". 7"&)*$1"(
#')&$8%"( (,1) 5*2$!"( 6 2*!2%!"& 2,##" +*& "E"#+!" !" 4,!%#" 9% )&*A2? #*$( 93*%)&"(
2,##" !" 1,#-&" 9"( *))*8%"( ,% !* -*19" +*((*1)" 9$(+,1$-!" (,1) -"*%2,%+ +!%( 9$O2$!"(
6 $15'&"&. 7"( 9"&1$@&"( #')&$8%"( &"8%$@&"1) %1 2"&)*$1 #,9@!" +,%& !3H1)"&1") F".;.? #,9@!"
+,%& !" )&*A2? #,9@!" +,%& !"( 20"#$1(G 2" 8%$ "() #*!0"%&"%("#"1) 9$O2$!" 6 ,-)"1$&.
= D"( #'2*1$(#"( 8%" !"( %)$!$(*)"%&( ") ,+'&*)"%&( #"))"1) "1 +!*2" +,%& (" +&,)';"& !"( %1(
9"( *%)&"(. 7"( #'2*1$(#"( &"19"1) 9$O2$!" !* #')&,!,;$" 9" !3H1)"&1") 4,$&" $#+,(($-!".
7,##" "E"#+!" 9" 2"( #'2*1$(#"( ,1 )&,%4" !" 2&<+)*;" 9% )&*A2 ") !* 9'(*2)$4*)$,1 9"
!3"14,$ 9"( #"((*;"( 9" 2,1)&Q!" H7/> 8%" 2"&)*$1( ,+'&*)"%&( +&*)$8%"1). D* (,!%)$,1 6 2"(
+&,-!@#"( +*(("1) (,%4"1) +*& 9"( #')0,9"( ()*)$()$8%"( ") 9"( 0"%&$()$8%"( 8%$ "((*$"1)
93*++&,20"& !* &'*!$)' "1 +*&)*1) 9% +"% 8%$ * +% L)&" #"(%&'.
D"( 9$O2%!)'( 9*1( !* #')&,!,;$" (" &'(%#"1) *!,&( *%E +,$1)( (%$4*1)( R
= D" +!*2"#"1) 9"( +,$1)( 9" #"(%&"? *++"!' (,%4"1) '20*1)$!!,11*;" (+*)$*!. >,%& %1 ,+'N
&*)"%&? 2"2$ &"4$"1) 6 20,$($& !"( +,$1)( 9" (,1 &'("*% ,I !"( #"(%&"( (,1) 6 "J"2)%"&. >,%&
!"( %)$!$(*)"%&(? 2"2$ &"4$"1) 6 )&,%4"& !"( #*20$1"( 4"&( !"(8%"!!"( "14,<"& !"( +*8%")( 9"
#"(%&" ") !"( #*20$1"( (,%&2"(. /*!0"%&"%("#"1) ") 2" 8%$ "() !" 2*( *2)%"!!"#"1)? !"( ,+'N

!"#$%&' () *+%&,-./%$,+
!"#$%!& $# %#'('&"#$%!& )$ *+((",+!$)# -"& -+%! (" !."('&"#'+) /$ *$##$ #0*1$ )' -+%! (2.*1")3$
/$& !.&%(#"#& +,#$)%& -"! *1"*%) /2$)#!$ $%45
6 72"3!.3"#'+) /% #!"8* -+%! !./%'!$ ($ 9+(%:$ /% #!"8* ; #!"'#$!5 <$##$ "3!.3"#'+) /+'# =#!$
>"'#$ /2%)$ :")'?!$ ')#$(('3$)#$ "8) /$ -!.&$!9$! ($ :"4':%: /2')>+!:"#'+)& @%' -$%9$)#
&$!9'! "% :"4':%: /2"--('*"#'+)&5
6 72.*1")#'((+))"3$ A#$:-+!$( -+%! ($ /'&#')3%$! /% &-"#'"(B @%' -$!:$# /$ !./%'!$ ($ 9+(%:$
/% #!"8* ; *+(($*#$!5 C+%# *+::$ (2"3!.3"#'+)D (2.*1")#'((+))"3$ )$ /+'# -"& #!+- ')/%'!$ $)
$!!$%! ($& !.&%(#"#&5 E+%! *1"@%$ "--('*"#'+) /$ :$&%!$D %)$ .#%/$ $&# "(+!& ; :$)$! -+%!
#!+%9$! (2':-"*# /$ (2.*1")#'((+))"3$ $# '/$)#'8$! (" :$'(($%!$ :.#1+/$ -+%! ')>.!$! ($ #!"8*
+!'3)"( ; -"!#'! /% #!"8* .*1")#'((+)).5 <$##$ /$!)'?!$ -!+*./%!$ $&# "--$(.$

$+0'&1$,+5

6 7" *+(($*#$ /$& !.&%(#"#& /$ :$&%!$ /+'# &$ >"'!$ ($ -(%& !"-'/$:$)# -+&&',($ &")& *+)3$&F
#'+))$! ($ !.&$"%5 7" :$&%!$ -$%# =#!$ (" *+(($*#$ /% #!"8* &%! %) ('$) *+::$ $(($ -$%# =#!$

tel-00408745, version 1 - 3 Aug 2009

%)$ !$@%=#$ $)9+G.$ ; %)$ $)#'#. !.&$"% A+% %)$ :"*1')$ /'&#")#$B -+%! +,#$)'! *$!#"')$&
')>+!:"#'+)&D -"! $4$:-($ ($ /.,'# &%! ($ ('$) +% (2"/!$&&$ HE /$ (2$)#'#.5 I)$ /$& *"!"*F
#.!'&#'@%$& /$ *$##$ *+(($*#$ $&# ($ 3!")/ )+:,!$ /2$)#'#.& !.&$"%4 @%' -$%9$)# ')#$!9$)'!5
I)$ "%#!$ *"!"*#.!'&#'@%$ *2$&# (" 8",'('#. @%' -$%# 9"!'$! $)#!$ %)$ $)#'?!$ 8",'('#. A#+%#
*+(($*#$!B $# %)$ 8",'('#. -"!#'$(($ A*"& /2%)$ ')>+!:"#'+) "3!.3.$ *+::$ -"! $4$:-($ ($
/.,'# :+G$) &%! ($& ('$)&D ($ )+:,!$ /$ J+#& &%! %) ('$)D ($ )+:,!$ /$ -"@%$#& &%! %) ('$)D
($ )+:,!$ /$ :"*1')$& *+))$*#.$& ; %) !.&$"%D $# "')&' /$ &%'#$B5
6 7" :")'?!$ "9$* ("@%$(($ $)9+G$! ($& -"@%$#& &+)/$& -+%! *"!"*#.!'&$! %) *1$:') /% !.&$"%5
7$ 3!")/ )+:,!$ /$ >"*#$%!& @%' /.*'/$)# *$ @%' &$ -"&&$ &%! %) *1$:') A-$)&$K ; #+%& ($&
-"@%$#& @%' -"&&$)# -"! *$ *1$:') $# ($%! -!+9$)")*$B !$)/ *$##$ *"!"*#.!'&"#'+) /'L*'($
9+'!$ ':-+&&',($5 <1"@%$ :.#!'@%$ ).*$&&'#$ %)$ :")'?!$ &-.*'8@%$ -+%! $)9+G$! ($& &+)/$&5
<$ @%2'( >"%# >"'!$ -+%! :$&%!$! ($ /.("' :+G$) $&# /'M.!$)# /$ *$ @%2'( >"%# >"'!$ -+%!
:$&%!$! (" ,")/$ -"&&")#$ /'&-+)',($ +% (" ,")/$ -"&&")#$ /% ('$) ($ -(%& ($)# #+%# "% (+)3
/% *1$:')5
6 N$& &+(%#'+)& -+%! *+)#+%!)$! ($& ,"!!'?!$& :'&$& $) -("*$ -"! ($& %#'('&"#$%!& $# ($& +-.!"F
#$%!& /$ (2H)#$!)$# "8) /$ &$ -!+#.3$! ($& %)& /$& "%#!$&5 I)$ :.#1+/$ /$ :$&%!$ $L*"*$
/+'# =#!$ *"-",($ /$ >+)*#'+))$! /")& #+%& ($& &*.)"!'+&5 I)$ &+(%#'+) -"! $4$:-($ @%' >+)*F
#'+))$ @%$ &%! /% #!"8* )+)F*!G-#. +% @%' &%--+&$ @%$ ($& !+%#$%!& /% !.&$"% *++-?!$)#
$) !$)9+G")# /$& :$&&"3$& H<OE $&# :+')& $L*"*$ @%2%)$ "%#!$ &+(%#'+) )2$4'3$")# -"&
*$& *+)/'#'+)& :=:$ &' (" /$!)'?!$ $&# %) -$% :+')& $L*"*$ $) #$!:$& /$ -$!>+!:")*$5

!" #$ %&''&($ ) *+,-.$**$ /$ *& 0,123*3(4$

O"(3!. #+%& ($& #!"9"%4 $M$*#%.& /")& (" ('##.!"#%!$D /+)# %) "-$!P% 3.).!"( &$!" -!.&$)#. /")&
($ *1"-'#!$ &%'9")#D ,$"%*+%- /$& -+')#& /%!& /")& (" :.#!+(+3'$ !$&#$)# &")& &+(%#'+) *("'!$ $#
#!?& -!+,",($:$)# 9+)# ($ !$&#$! 9%$ (" *+:-($4'#. /$ (2H)#$!)$# $# &+) *+)&#")#$ .9+(%#'+)5 E"!:'

!"! #$ %&''&($ ) *+,-.$**$ /$ *& 0,123*3(4$

!"# $%&'()*"# +,-%" $%./!.$0("*"/1 (" $0##0," 2 (34!5"((" 6" (0 *41%&(&,."7 8/ "9"1: !"%10./# 6"#
$&./1# 6-%# !.14# 60/# (0 #"!1.&/ $%4!46"/1" /4!"##.1"/1 #&-;"/1 6"# #&(-1.&/# $%&1&!&(0.%"# <-.
/" #&-9%"/1 $0# #. (0 10.((" 6- %4#"0- 0-,*"/1"7 =/ 1%&-;" 60/# !"11" !014,&%." $0% ">"*$(" ("#
#&(-1.&/# #101.#1.<-"# <-. !(0##"/1 (" 1%0+! !%?$14 6" (3@/1"%/"1 "1 !"(("# <-. !5"%!5"/1 (0 *".(("-%"
*0/.)%" 0;"! (0<-"((" "/;&?"% ("# $0<-"1# $&-% *"#-%"% -/" !"%10./" *41%.<-" %"(01.;" 2 -/ !5"*./
6" (3@/1"%/"17 A" ,"/%" 6" $%&'()*"# /3"#1 $0# -/ 50/6.!0$ "/ !0# 630-,*"/101.&/ 6" (0 10.(("
6- %4#"0-7 B0% !&/1%" ("# 0-1%"# $&./1# 6-%# 60/# (0 *41%&(&,." &/1 ("-% 6.C!-(14 6.%"!1"*"/1
(.4" 2 (0 10.((" 6- %4#"0- !" <-. %"/6 6.C!.(" (" $0##0," 2 (34!5"(("7 D3">"*$(" 1?$.<-" "#1 !"(-.
6" (3&'#"%;01.&/ 6" !" <-. #" $0##" 60/# -/ ,%0/6 %4#"0- <-" !" #&.1 $0% 6"# *"#-%"# $0##.;"# &6"# *"#-%"# 0!1.;"#7 A" ,"/%" 6" $%&'()*"# /4!"##.1" '"0-!&-$ 63"9&%1 6" (0 $0%1 6" !5"%!5"-%#
$&-% 0%%.;"% 2 6"# #&(-1.&/# <-. $0##"/1 2 (34!5"((" #0/# 1%&$ .*$0!1"% (0 $%4!.#.&/ 6"# *"#-%"#7
E0/# !" *0/-#!%.1: /&-# F&!0(.#&/# #-% ("# 0#$"!1# 6" *41%&(&,." <-. #&-9%"/1 6" $%&'()*"#

tel-00408745, version 1 - 3 Aug 2009

6" $0##0," 2 (34!5"(("7 G&-# 0;&/# 0-##. *"/4 6"# 0!1.;.14# #-% (30-1%" !014,&%." 6" $%&'()*"#:
*0.# "(("# /" #"%&/1 $0# $%4#"/14"# 60/# !" *0/-#!%.1 $0% #&-!. 635&*&,4/4.14 6" (0 $%4#"/101.&/7
G&-# %",%&-$&/# !"# 0#$"!1# 6" (0 *41%&(&,." "/ 6"-> !014,&%."# H
I E3-/ !J14 #" 1%&-;" (" $%&'()*" 6- ,%0/6 ;&(-*" 6" 6&//4"# 2 !&(("!1"% $&-% %40(.#"% ("#
*"#-%"#7 A"!. ;0 6- ,%0/6 ;&(-*" 6- 1%0+! 2 (0 ,%0/6" 10.((" 6- %4#"0-7 G&# !&/1%.'-1.&/#
60/# !"11" 6.%"!1.&/ ;.#"/1 2 %46-.%" (" ;&(-*" 6" 6&//4"# 2 1%0.1"% $0% (3./1"%*46.0.%" 6"
(34!50/1.((&//0," 2 (0 F&.# #$01.0( "1 1"*$&%"(7 8/ $0%0(()(": -/" 41-6" #"%0 *"/4" $&-%
4;0(-"% (3.*$0!1 6" (34!50/1.((&//0," #-% (0 <-0(.14 6" (0 *"#-%"7
I E" (30-1%" !J14 #" 1%&-;" (" $%&'()*" 6- ,%0/6 ;&(-*" 6" 6&//4"# 2 "/;&?"% #-% (" %4#"0-7
A"# 6&//4"# $"-;"/1 K1%" ("# %4#-(101# 6" *"#-%" $0##.;" !&**" .(# $"-;"/1 K1%" ("# $0L
<-"1# #&/6"# "1 %"<-K1"# "/;&?4# 2 6"# "/1.14# %4#"0-> "1 ("# *0!5./"# 6"# -1.(.#01"-%#7 A"
$%&'()*" 0 414 #&-;"/1 .,/&%4 60/# (0 (.114%01-%"7 G&-# #&-(";&/# !" $%&'()*" 60/# !" *0L
/-#!%.1 "1 /&-# $%&$&#&/# -/" #&(-1.&/ $%&1&!&(0.%" "C!0!" <-. $"%*"1 63"/;&?"% -/ ,%0/6
;&(-*" 6" 6&//4"# 2 $0%1.% 63-/ ,%0/6 /&*'%" 6" #&-%!"# #0/# !&/,"#1.&//"% (" %4#"0-7
M !50!-/ 6" !"# $%&'()*"# /&-# $%4#"/1&/# 6"# #&(-1.&/# "1 /&-# ("# ;0(.6&/# $0% *&64(.#01.&/
0/0(?1.<-": #.*-(01.&/# "1 ">$4%.*"/101.&/#7 G&# #&(-1.&/#: $0% ("-% "C!0!.14 "1 ("-% #.*$(.!.14:
$%4#"/1"/1 6"# %4$&/#"# 2 !"%10./# 6"# $&./1# 6-%# 6" (0 *41%&(&,." "1 !&/#1.1-"/1 $0% !&/#4<-"/!"
-/" ./14%"##0/1" 0;0/!4 60/# !" 6&*0./"7 8(("# &/1 (0 !0%0!14%.#1.<-" 630$$&%1"% -/" /&-;"(("
;.#.&/ #-% !"%10./# 6"# 0#$"!1# 6" (0 *41%&(&,." "1 6" F&/!1.&//"% "/ !0# 6" ,%0/6# %4#"0-> "1 6"
,%0/6"# <-0/1.14# 6" 6&//4"# 2 1%0.1"%7
D"# !50$.1%"# #-.;0/1# $%4#"/1"/1 -/" 6"#!%.$1.&/ 6" /&# !&/1%.'-1.&/#7 A50$.1%" N 1%0.1" ("
$%&'()*" 6" (0 *41%&(&,." $0##.;" 6"# ,%0/6# %4#"0-> $0% (34!50/1.((&//0," 2 (0 F&.# #$01.0( "1
1"*$&%"(7 D34!50/1.((&//0," /3"#1 $0# ,%01-.1 "1 0 -/ !&O17 A50$.1%" P $%"/6 -/" *41%.<-" <-. "#1
!"((" 6" (0 641"!1.&/ 6"# -1.(.#01"-%# ("# $(-# ,&-%*0/6# "/ 1"%*"# 6" 1%0+! "1 *&/1%" (3.*$0!1
6" (34!50/1.((&//0," 6"##-#7 @( #30,.1 63-/ ">"*$(" 63-/" 41-6" <-" /&-# 6";&/# *"/"% !50<-"
F&.# <-3.( #30,.1 634!50/1.((&//0,"7 A50$.1%" Q F0.1 -/" 410$" #-$$(4*"/10.%" "/ *&/1%0/1 #-% -/

!"#$%&' () *+%&,-./%$,+
!"#$% %&%'()% *+''%,# +, (%"# $-."/$% )% 0$"/# /,#$+."/# (!$ )1-*2!,#/))+,,!3% %, *+,4/.-$!,# )!
*+$$-)!#/+, #%'(+$%))% .!,4 )%4 5!$/!#/+,4 .% )! #!/))% .% )! (+(")!#/+, -*2!,#/))+,,-%6 7%*/ (%$'%#
0/%, -5/.%''%,# .%4 #!"& .1-*2!,#/))+,,!3% %,*+$% ()"4 $-."/#46 72!(/#$% 8 #$!/#% )% ($+0)9'% .%
)! '-#$+)+3/% !*#/5% %# .% *+))%*#% .1/,:+$'!#/+,46 ;) ($-4%,#% ,+#$% 4+)"#/+, <;7= ><7=?:$/%,.)@
;,:+$'!#/+, 7+))%*#/+, =$+#+*+)A B"/ (%$'%# .1%,5+@%$ .%4 .+,,-%4 %# .%4 $%B"C#%4 .!,4 ", 3$!,.
$-4%!" 4!,4 )% *+,3%4#/+,,%$ %# 4!,4 ,"/$% !" $%4#% ." #$!D* >,+#!''%,# )% #$!D* <7= .1+E )%
#%$'% <7=?:$/%,.)@ .!,4 )% #/#$%A6
F5!,# .% *+''%,*%$ G .-*$/$% ,+4 4+)"#/+,4 %# (+"$ '/%"& )%4 '+#/5%$H ,+4 $-4"'+,4 .!,4
)% *2!(/#$% 4"/5!,# )1-5+)"#/+, .% )! '-#$+)+3/% .% )1;,#%$,%# %# )%4 ($/,*/(!"& #$!5!"& .!,4 )!
)/##-$!#"$%6 I% '!,"4*$/# 4%$! *+,*)" .!,4 )% 72!(/#$% J (!$ ,+#$% (+/,# .% 5"% 4"$ )% .+'!/,%

tel-00408745, version 1 - 3 Aug 2009

%# 4+, -5+)"#/+, !/,4/ B"% )%4 (/4#%4 B"% ,+"4 *+'(#+,4 4"/5$% (!$ )! 4"/#%6

!"#$%&' (

tel-00408745, version 1 - 3 Aug 2009

!"#$%&'($) *& *$+,()- *- %,

+"'.$%$/(!"#$%&#%$ ' ($( )&(( *'#* +#,&'*$&-)$-&% .% /%*-&% *$'#.'&. 0-+ 1%&/%$ '-2 31(&'$%-&* %$
'-2 -$+4+*'$%-&* .!35*%&6%& )% 0-+ *% 1'**% %$ .!()7'#8%& 4%* &(*-4$'$* %#$&% %-29 :3//% .+$ .'#*
4% )7'1+$&% 1&()(.%#$; 4% )3%-& .% 4!"#$%&#%$ ,3-&#+$ -# *%&6+)% *+/14% )3#*+*$'#$ %# -# &3-$'8%
'- /+%-2 .%* 1'0-%$*; %$ $3-$% 4!+#$%44+8%#)% %*$ &%13-**(% < 4' 1(&+17(&+% .- &(*%'- =.'#* 4%*
/')7+#%*>9 ?%&*3##% < -# /3/%#$ .3##( #!' -#% +.(% .% )3//%#$ %*$ ,3&/( 4!"#$%&#%$ #+ 0-%44%*
*3#$ 4%* )3//-#+)'$+3#* %# )3-&*9 @#% /')7+#% 1%-$ %#63A%& %$ &%)%63+& .%* 1'0-%$* '-$'#$
0-% 4' 5'#.% 1'**'#$% 4% 1%&/%$; %$ 4%* 31(&'$%-&* &%4'+%#$ 4%* 1'0-%$* %# 1&36%#'#)% .%* '-$&%*
31(&'$%-&* 4' 14-1'&$ .% $%/1* *-+6'#$ .%* 134+$+0-%* 43)'4%* 0-+ #% $+%##%#$ 1'* )3/1$% .%*
1%&,3&/'#)%* 8435'4%*9 ?'& )3#*(0-%#)%; )7'0-% 31(&'$%-& %$ )7'0-% -$+4+*'$%-& =3- 8&3-1%*
.!-$+4+*'$%-&*> .(143+%#$ 4%-&* 1&31&%* *34-$+3#* .% /%*-&% 'B# .!31$+/+*%& 4% &(*%'- .% 4%-& 13+#$
.% 6-%9
C('#/3+#*; )%&$'+#* /()'#+*/%* 3#$ '))3/1'8#( 4!"#$%&#%$ .%1-+* *' )&('$+3# %$ 3#$ ,3&/(
-#% 53##% 5'*% .% .(1'&$ 13-& 4' /($&3438+%9 C3-* )3//%#D3#* 1'& .()&+&% )%* /()'#+*/%*;
'1&E* #3-* 1&(*%#$3#* 4%* *34-$+3#* 0-+ 3#$ '11'&- .%1-+* %# 4%* )4'**'#$ %# ,3#)$+3# .% 4%-&
#'$-&% F !""#$% 3- !&'#$%9 C3-* &'11%43#* 0-!-#% *34-$+3# 1'**+6% %*$ -#% *34-$+3# 0-+ #!+#G%)$%
1'* .% $&'B) .'#* 4% &(*%'- /'+* *%-4%/%#$ *% )3#$%#$% .!35*%&6%& 4% $&'B) 0-+ )+&)-4% *-& -# 314-*+%-&* 4+%#* .% 4!"#$%&#%$9 H 4!3113*(; *% $&3-6%#$ 4%* *34-$+3#* ')$+6%* 0-+ )3#*+*$%#$ < %#63A%&
.%* 1'0-%$* .'#* 4% &(*%'- .(.+(* < 4' /%*-&% %$ < 4!35*%&6'$+3# .% 4' &(')$+3# .- &(*%'- < )%*
1'0-%$* *3#.%*9 ' &(')$+3# .- &(*%'- < -# 1'0-%$ *% /'#+,%*$% 1'& *3# $%/1* .% 4+6&'+*3#; (8'4
< 4!+#B#+ 0-'#. 4% 1'0-%$ %*$ &%G%$(9

!" #$% &'()*+%&$% ,$ -)%$ ,$ &$%./$

I'48&( 4' *+/14+)+$( .% 4' )3-)7% &(*%'- .% 4!"#$%&#%$; )%&$'+#* /()'#+*/%* 3#$ ($( +#$&3.-+$*
13-& '**-&%& -# /+#+/-/ .% ,3#)$+3#*9 H 4' $J$% .% )%* /()'#+*/%* *% $&3-6%#$ 4%* /%**'8%*
":I? ="#$%&#%$ :3#$&34 I%**'8% ?&3$3)34> KLMN %#63A(* 1'& 4%* &3-$%-&* .% 4!"#$%&#%$ %$ *%*
L

!"#$%&' () *+,-./0%$.1 20 2.3"$1' 2' /" 3,%&./.4$'

"#$%&'() *+#', ,() -#*+(.) /(0+) -/1)('.('. ,() #'2"#3&()4 5' (6(.7 *+#', +' -#*+(. #//&8#'.
9 3# $2+$%( :; ,<+'( "#$%&'( 2+ ,<+' /2+.(+/ -/1)('.( +'( #'2"#3&(7 +' "())#=( :>?; (). ('82@1
A2+ -3+.B. ,(8/#&. 3<C./(D 9 3# )2+/$( ,+ "())#=(7 E&(' 18&,(""('. )& 3<('.C.( :; ,+ "())#=( ().
F+=1 '( -#) $2'.('&/ ,() (//(+/)7 -#/$( *+( )&'2' 3( "())#=( :>?; /&)*+( ,<C./( ('82@1 9 3#
"#+8#&)( )2+/$(4 >() "())#=() :>?; $%#'=('. (' G2'$.&2' ,( 3# '#.+/( ,( 3<#'2"#3&(4 H' $2,(
(). #))2$&1 9 $%#*+( #'2"#3&( (. (). &'$3+) ,#') 3<('.C.( ,+ "())#=( :>?;4 5' 3&)#'. $( $2,(7
3# )2+/$( /1$(-./&$( ,+ "())#=( :>?; (). #+ $2+/#'. ,+ -/2E3I"( *+&7 2' 3( /#--(33(7 $2'$(/'(
)(+3("('. 3( -#*+(. $2//()-2',#'.4 5. -2+/ ,#8#'.#=( ,<&'G2/"#.&2') 9 3# )2+/$(7 +'( -#/.&( ,+
-#*+(. #'2/"#3 (). $2-&1( ,#') 3( "())#=( :>?;4
J( -/2.2$23( :>?; (). 3( )(+3 "2@(' *+& -(/"(. ,<#82&/ +' /(.2+/ ,( 3# -#/. ,() /2+.(+/) ,(
3<:'.(/'(.4 K& 3<#,/())( :; ,+ /2+.(+/ (). $2''+(7 :>?; -(/"(. ,( 3<&'.(//2=(/ ,&/($.("('. -#/ +'
"())#=( ,( .@-(

'5!. -2+/ #82&/ +' /(.2+/ ,( .@-( '5!. &'#/64 L( $(..( G#02'7 +' #,"&'&)./#.(+/

tel-00408745, version 1 - 3 Aug 2009

2+ +' +.&3&)#.(+/ -(+8('. )#82&/ )& 3( /2+.(+/ (). E&(' 2-1/#.&2''(3 2+ -#)4 >(..( 2-1/#.&2'
-(+. #+))& )( G#&/( #8($ .2+.( "#$%&'( (. )(/8(+/ ,( 3<:'.(/'(. &"-31"('.#'. 3( -/2.2$23( :>?;4
?#&'.('#'.7 )& 3<#,/())( :; ,+ /2+.(+/ '<(). -#) $2''+(7 3( )(+3 "2@(' *+& /().( -2+/ #82&/ +'
/(.2+/ $<(). ,<('82@(/ +' -#*+(. (. ,( G#&/( ,( 3# )2/.( *+( 3( -#*+(. )2&. #'2/"#3 (' #//&8#'.
#+ /2+.(+/7 (. $<(). 39 *+( 3( /(.2+/ /1-2', -2+/ )&='#3(/ 3<#'2"#3&(4 L#') )# /1-2')( 3( /2+.(+/
"(. 3<#,/())( :; ,( )2' &'.(/G#$( ,( )2/.&(4
J() #'2"#3&() #--#/#&))('. )2+8('. *+#', 3( /2+.(+/ )( ./2+8( 2E3&=1 ,( /(F(.(/ 3( -#*+(.4
><(). (' $( "2"('. *+<&3 '2.&M( 3# )2+/$( ,( $( /(F(.4 5. 3() /#&)2') -2+/ /(F(.(/ +' -#*+(. )2'.
,&8(/)()7 $2""( -#/ (N("-3(7 ,().&'#.&2' &'$2''+(7 3( $%#"- OOJ # #..(&'. P1/27 (. 3( -#*+(.
(). ./2- =/#', -2+/ *+( 3( /2+.(+/ -+&))( 3( ./#')"(../(4 H'( /I=3( ,( E#)( ,#') :>?; $<().
*+<+' "())#=( :>?; /(F(.1 -#/ +' /2+.(+/ '( =1'I/( -#) +' #+./( "())#=( :>?; #M' ,<18&.(/
+'( #8#3#'$%( ,( "())#=() :>?;4 J() /2+.(+/) /(3#&('. 3() "())#=() :>?; 3() +' ,() #+./()
$2""( ,() -#*+(.) :; /1=+3&(/) ,( .(33( )2/.( *+<+' G((,E#$Q (). -2))&E3( 9 2E.('&/ ,( .2+.
/2+.(+/ (. "#$%&'( ,( 3<:'.(/'(.4 ?#3%(+/(+)("('. ,#') 3# -/#.&*+( 3() $%2)() '( )2'. -#) )&
F2@(+)() *+( $# RS T4 >(/.#&') /2+.(+/) ,( 3<:'.(/'(. ,1)#$.&8('. $2"-3I.("('. 3# G2'$.&2''#3&.1
:>?;4 L<#+./() 3&"&.('. 3( ,1E&. ,() "())#=() :>?;4 5. &3 @ # $(/.#&') /2+.(+/) *+& E32*+('. 3()
"())#=() :>?; =1'1/1) -#/ 3() #+./() /2+.(+/)4 O2+. $($& $2"-3&*+( 3() .U$%() ,( "()+/( E#)1()
)+/ 3() "1$#'&)"() ,( :>?;4
:>?; # &')-&/1 -3+)&(+/) $%(/$%(+/) 9 ,18(32--(/ ,() 2+.&3) -2+/ 3( ,&#='2).&$ (. 3# "1./232=&(
,( 3<:'.(/'(.4 >() 2+.&3) -(/"(..#'. -2+/ 3# -3+-#/. ,( .("-) 3( ,&#='2).&$ ,<+' $%("&' ('./(
,(+N "#$%&'() 2+ ('./( +'( "#$%&'( (. +' /2+.(+/7 2'. 1.1 #-/I) +.&3&)1) -2+/ +'( "1./232=&(
9 =/#',( 1$%(33( ,( 3<:'.(/'(.4 ;#/"& 3() 2+.&3) 9 E#)( ,( :>?; 2' ./2+8( A,() 8#/&#'.() 9 $()
2+.&3) (N&).('. )2+) ,<#+./() #--(33#.&2')D V
W

#$14 +.&3&)( 3() "())#=() :>?; '5!. (. '5!. &'#/6 -2+/ 81/&M(/ )& +'( #,/())( :; (). #$.&8(
2+ -#)4

%&"5'&.0%' ('82&( ,() -#*+(.) HL; #8($ +' OOJ 3&"&.1 (. #..(', 3( /(.2+/ :>?; ,+ /2+.(+/

!"! #$% &'()*+%&$% ,$ -)%$ ,$ &$%./$

!"# $%&%''% (% )%**+,% !"+-. */- 001 %23#$%4 5/"$ "- 001 6,+( 7 n +" .63+$' %' 3/"$
"- n 3("* 3%'#' !"% (% -/)8$% .% *+"'* (% (/-, ." 9:%)#-; 0/)($/1.0$ 9/((%9'% (<+.$%**%
=5 ." $/"'%"$ !"# *% '$/">% +" n?)% *+"' +>%9 (% .6(+# +((%$@$%'/"$ %-'$% (+ */"$9% %' 9%
$/"'%"$4 A- 9/))%-B+-' .%

%' %- +",)%-'+-' n .%

7 9:+!"% C/#*; 0/)($/1.0$ 3%"' .%

9%''% )+-#?$% 9/((%9'%$ (%* +.$%**%* =5 .% '/"* (%* $/"'%"$* (% (/-, .<"- 9:%)#- +#-*# !"%
(%* .6(+#* +((%$@$%'/"$ 9")"(+'#C*4 0/)($/1.0$ +$$D'% .<+",)%-'%$ n !"+-. "- $%'/"$ .% (+
)+9:#-% .%*'#-+'$#9% %*' -/'6 %' 9%9# %*' $%-." 3/**#8(% 3+$ (<"'#(#*+'#/- .<"- -")6$/ .%
3/$' *"$ (+ )+9:#-% !"# -<%*' 3+* +((/"6 9% !"# 3$/>/!"% "- $%'/"$ .% 'E3% 21/0 .*/$)(3)-4$4
1<#.6% .% 0/)($/1.0$ .% 9/-'$F(%$ (% 001 + 6'6 %-*"#'% $%3$#*% 3+$ .#>%$* /"'#(* !"# 9:%$9:%-'
7 +>/#$ "- C%%.8+9G .%* $/"'%"$* (% (/-, .<"- 9:%)#- )+#* !"# 3%$)%''%-' .% 9/--+H'$% "3%" 3("* !"% 0/)($/1.0$4 I- '$/">% 3+$)# 9%* /"'#(* 7 8+*% .% '$+9%$/"'% J 2)03(3)/ KLMN
3/"$ (% .68#' .%* (#%-*; 2)03*$(5 KLON 3/"$ (+ (/9+(#*+'#/- ." ,/"(/' .<6'$+-,(%)%-'; 2)/+%

tel-00408745, version 1 - 3 Aug 2009

0/)($/1.0$ K PN 3/"$ (+ .6'%9'#/- .%* $63+$'#'%"$* .% '$+Q9 (% (/-, .<"- 9:%)#-; %' 67
0/)($/1.0$ 3/"$ (% $%36$+,% .%* ./)+#-%* (% (/-, .<"- 9:%)#- %' 3+* *%"(%)%-' .%* $/"'%"$*4
R 2)03 89: ,+%(1;$/< KSTN !"# 3%$)%' .% .6'%9'%$ (% U0V WU+2#)") 0$+-*)#**#/- V-#'X
(% 3("* 3%'#' (% (/-, .<"- 9:%)#-4 Y%9# *% C+#' %- %**+E+-' 3("*#%"$* '+#((%* .% 3+!"%' %9/))%-B+-' 3+$ .%* ,$+-.* 3+!"%'* %' %- .6*+9'#>+-' 9:+!"% C/#* (+ C$+,)%-'+'#/- W8#' Z[
.+-* (<%-'D'% =5 )#* 7 X4 1% $%'/"$ .<"- )%**+,% =YU5 .% 'E3% \[$+,)%-'+'#/- -%%.%.
+-. Z[ *%'\ %*' "-% #-.#9+'#/- !"</- -<+ 3+* %-9/$% +''%#-' (+ '+#((% .% U0V (+ 3("* 3%'#'%4
=YU5 -% $63/-. 3+* 7 '/"* (%* 8%*/#-* .%* /36$+'%"$* !"# >%"(%-' 9/--+H'$% (%* 9+$+9'6$#*@
'#!"%* ." '$+Q9 *"$ (%"$* (#%-* +Q- .% 3/">/#$ $6+,#$ 7 '/"'% +",)%-'+'#/- .% '$+Q9; .6'%9'%$
.%* +-/)+(#%*; %' 3(+-#Q%$ .%* C"'"$%* )#*%* 7 &/"$ .% (%"$ $6*%+"4 5/"$ 9%(+ #( E + %" (+ 9$6+'#/." 3$/'/9/(% ]^U5 !"# 3%$)%' 7 "- /36$+'%"$ .<#-'%$$/,%$ (%* $/"'%"$* .% */- $6*%+" %' .% (#$%
7 .#*'+-9% 3+$ ]^U5 (% 9/-'%-" .% 9%$'+#-* $%,#*'$%* .+-* 9:+!"% $/"'%"$4 Z+-* 9%* $%,#*'$%*;
+33%(6* U=_ WU+-+,%)%-' =-C/$)+'#/- Z+'+8+*%X *"#>+-' (+ '%$)#-/(/,#% ]^U5; */-' *'/9G6%*
.%* #-C/$)+'#/-* +,$6,6%* *"$ (% $/"'%"$ %' *"$ (% '$+Q9 !"# 3+**% 7 '$+>%$*4 1%* $/"'%"$* 9+(9"(%-'
.%* *'+'#*'#!"%* +,$6,6%* *"$ (% '$+Q9 9/))% (% .68#' )/E%- %' (% '+"2 .%* 3%$'%* *"$ (<#-'%$C+9%
.% */$'#%; %' (</36$+'%"$ 9/((%9'% 9%* *'+'#*'#!"%* 7 .#*'+-9% +>%9 (% 3$/'/9/(% ]^U5; 9% !"# ("#
3%$)%' .<+>/#$ "-% #.6% .% :+"' -#>%+" .% (<6'+' .% */- $6*%+" %' .% *%* (#%-*4 U+(:%"$%"*%)%-'
+>%9 ]^U5 #( -<%*' 3+* 3/**#8(% .<+>/#$ "-% #-C/$)+'#/- 3("* Q-% *"$ (% '$+Q9 -# "-% #-C/$)+'#/*"$ (%* $6*%+"2 .%* +"'$%* /36$+'%"$*4
]^U5 3%$)%' (+ )%*"$% 3+**#>% .% :+"' -#>%+" .%* .#`6$%-'* $/"'%"$* %' (#%-* .<"- $6*%+"4
U+#* (%* "'#(#*+'%"$* $6,"(#%$* .% (<=-'%$-%' /-' +"**# 8%*/#- .% 9/--+H'$% (% '$+Q9 !"# 9#$9"(%
*"$ (%"$* #-'%$C+9%* $6*%+"2 3/"$ .#>%$*%* +33(#9+'#/-* !"# '/"9:%-' +"2 3%$C/$)+-9%* %' 7 (+
*69"$#'64 1+ )6':/.% (+ 3("* 9(+**#!"% .% (% C+#$% 9<%*' .<+996.%$ 7 (+ 9+$'% $6*%+" %' .% (+ )%''$%
%- )/.% \3$/)#*9"/"*\; (% )/.% !"# 3%$)%' (+ (%9'"$% .% '/"* (%* 3+!"%'* !"# 3+$'%-' .% (+
)+9:#-% %' !"# (<+''%#,-%-' )D)% 9%"2 !"# -% */-' 3+* +.$%**6* 7 (+ )+9:#-% )D)%4 V-% C/#*
(+ 9+$'% .+-* 9% )/.%; #( *<+,#' .% 9/3#%$ (%* 3+!"%'* +" -#>%+" ." I] WI3%$+'#-, ]E*'%)X;

!"#$%&' () *+,-./0%$.1 20 2.3"$1' 2' /" 3,%&./.4$'

"# $#% &$'(#() #' "# $#% (#*+,'#( -. ,/0#-. -11$/2-'/3 1+.( ., '(-/'#*#,' .$'4(/#.(5 67.'/$/%-'#.(
1#.' #8#2'.#( 2#''# /,'#(0#,'/+, $./ *9*#) *-/% /$ #:/%'# -.%%/ "#% $/;(-/(/#% %'-,"-("% 1+.( $#
3-/(# 2+**# $#% $/;(-/(/#%

/$5#6"#

<= > #' ?@A <BC>5 D,# 3+/% -. ,/0#-. -11$/2-'/3) $#% 1-E.#'%

1#.0#,' 9'(# -F2G4% "7.,# *-,/H(# -11(+1(/4# #' %'+2I4% 1+.( "-0-,'-J# "7-,-$K%#5 L/0#(% +.'/$%
%'-,"-("% +,' 4'4 1(+1+%4% 1+.( 2#' +;M#2'/3) $#% 1$.% 2+,,.% 4'-,'

%6#203# #' '%!'&'"/5 N# J#,(#

"7+.'/$ 1#(*#' .,# 2-1'.(# "# '+.' $# '(-&2 1(4%#,' %.( $7/,'#(3-2# (4%#-. "7.,# *-2G/,#) *-/%
*-$G#.(#.%#*#,' #%' %#.$#*#,' +(/#,'4 0#(% $#% %'-'/+,% "# '(-0-/$ #' +("/,-'#.(% #' ,+, 1-% $#%
(+.'#.(%5 N#% "#(,/#(% 1+%%H"#,' .,# -(2G/'#2'.(# "/84(#,'# #' ,42#%%/'#,' -$+(% ., -.'(# 'K1#
"7+.'/$%5 O, 1$.%) $# '(-&2 E./ '(-0#(%# ., (+.'#.( -''#/,' "#% 0-$#.(% ;#-.2+.1 1$.% 4$#04#%
E.# %.( $#% /,'#(3-2#% "#% *-2G/,#%) 2# E./ ,42#%%/'# "#% '#2G,/E.#% "# 2-1'.(#) "7-J(4J-'/+,) #'
"# &$'(-J# 1$.% %+1G/%'/E.4#%5

7'%8/.9 <BP>) $7+.'/$ %'-,"-(" 1+.( $# 2-1'.(# "# '(-&2 "-,% $#%

(+.'#.(% N/%2+ #' $- (434(#,2# "-,% 2# "+*-/,#) %#(- #:1+%4 "-,% $- %#2'/+, $/4# Q $740+$.'/+, "#

tel-00408745, version 1 - 3 Aug 2009

$- *4'(+$+J/# 1-%%/0# "#% (4%#-.:5

D,# -.'(# /,3+(*-'/+, "# 0-$#.( "-,% $7R,'#(,#' 27#%' $# 2+,'#,. "#% '-;$#% "# (+.'-J#) %.('+.'
$#% '-;$#% ?S@ <T=>5 L-,% $#% (+.'#.(% /,'#(,#% -.: "+*-/,#%) +, '(+.0# 1+.( 2G-E.# 1(4&:#
-11-('#,-,' -. "+*-/,# $7/,'#(3-2# "# %+('/# #' $- "/%'-,2# Q 1-(2+.(/( 1+.( -''#/,"(# 2# 1(4&:#5
D, 1(4&:# #%' ., #,%#*;$# "7-"(#%%#% R@ 2+,'/J.U% (#1(4%#,'4 1-( $#% ;/'% "# 1+/"% $# 1$.% 3+('
2+**.,% Q 2#% -"(#%%#% V1-( #:#*1$# $# 1(4&:# WB5TX5C5CY X 1+.( '+.'#% $#% -"(#%%#% E./ +,'
$#% *9*#% 1(#*/#(% X ;/'% E.# WB5TX5C5CZ5 @+.( $#% 1(4&:#% #:'4(/#.(% -. "+*-/,#) "#% %+('/#%
1-( "43-.' -. ,/0#-. "# 2G-E.# (+.'#.( /,'#(,# 1#.0#,' 9'(# -%%+2/4#%5 D,# -.'(# 1+%%/;/$/'4 #%'
"7-%%+2/#( Q 2#% 1(4&:#% #:'#(,#% .,# "/%'-,2# 4J-$# -. ,+*;(# "# %-.'% E.7/$ 3-.' 1+.( %+('/(
". "+*-/,# 1$.% .,# 0-$#.( /,"/E.-,' $- $+,J.#.( ". 2G#*/, (#%'-,' .,# 3+/% $# 1-E.#' E./''# $#
"+*-/,#5 N#('-/,% 1(+'+2+$#% "# (+.'-J# /,'(-["+*-/,# 2+**# \]@A "/8.%#,' $- 2-('# #,'/H(#
". "+*-/,# Q '+.% $#% (+.'#.(%) "7-.'(#% 2+**# ^R@ "/8.%#,' %#.$#*#,' $- $+,J.#.( ". 2G#*/,
$# 1$.% 2+.(' 1+.( 2G-E.# 1(4&:#5 _.# 2# %+/' 1+.( 2# J#,(# "# 1(+'+2+$# /,'(-["+*-/,# +.
$7-.'(#) $#% '-;$#% "#% (+.'#.(% Q $7/,'4(/#.( "7., "+*-/,# 3+.(,/%%#,' .,# /,3+(*-'/+, /*1+('-,'#
Q $7+14(-'#.( %.( $74'-' "# %+, (4%#-.5 `-$G#.(#.%#*#,' 2#''# /,3+(*-'/+, ,# 1#(*#' 1-% "7-0+/(
.,# /"4# %.( 2# E./ %# 1-%%# #, "#G+(% ". (4%#-.5 R$ #%' Q -M+.'#( E.# $#% '-;$#% "# (+.'-J# "#%
(+.'#.(% /,'#(,#% Q ., "+*-/,# ,# %+,' $/%/;$#% E.# 1-( $7+14(-'#.( $./[*9*#5
@+.( -0+/( .,# /"4# %.( $# (+.'-J# "-,% '+.' $7R,'#(,#' Q ., *+*#,' "+,,4) /$ 3-.' (#2+.(/(
-.: '-;$#% "# (+.'-J# ?S@5 O, #8#') "-,% ., (+.'#.( ?S@) %# '(+.0# #, J4,4(-$ #' Q '+.'
*+*#,' $7#,%#*;$# "# 1(4&:#% "-,% $7R,'#(,#') #' 1+.( 2G-E.# 1(4&:# $#% "+*-/,#% Q 1-(2+.(/(
1+.( K -((/0#(5 N7#%' .,# %+('# "7-(;(# "# 2G#*/,% $#% 1$.% 2+.('% 1+.( -$$#( "7., "+*-/,#
0#(% '+.% $#% -.'(#% "+*-/,#%) -0#2 $- (41-('/'/+, "#% 1(4&:#% %.( $#% "/84(#,'% "+*-/,#%5 O,
2+*;/,-,' $#% '-;$#% "# 1$.%/#.(% (+.'#.(% ?S@) .,# 2-('# "# $7R,'#(,#' -. ,/0#-. "+*-/,# 1#.'
9'(# +;'#,.#5 6- E.#%'/+, /,'./'/0# E./ %# 1+%# #%' %/ 2#% '-;$#% "# (+.'-J# %+,' (#,".#% 1.;$/E.#%5
N#('-/,% +14(-'#.(% +,' 2G+/%/ "7+.0(/( $#.(% (+.'#.(% ?S@) #' 2+**# ., (+.'#.( ?S@ -22#1'#
"#% 2+,,#:/+,% Q "/%'-,2# "#% -.'(#% (+.'#.(% ?S@) /$ %.F' -$+(% "# %# 2+,,#2'#( Q ., "# 2#%

tel-00408745, version 1 - 3 Aug 2009

!"! #$% &'()*+%&$% ,$ -)%$ ,$ &$%./$

!"# "#

$ %&'() *) +,-.()'.)( &/ .01)&/ *23&0.) ')(2/'.4) 5&' '2/()10)67

'2/()/'7 2/1)'(78 *) 9&0') :233) 7,0+ 7,&;0( *,/. '2/()/' <=> 12070. )( *) +/0 *)3&.*)' +& +07()
*)7 5'4?@)7 )( '2/()7 A/,0+ *4(0).(# %)'(&0.)7 0.0(0&(01)7 2.( 4(4 +&.:4)7 *&.7 :)(() *0'):(02. )(
2.( &B2/(0 C /.) 72'() *) B&7) *) *2..4)7 52/' +)7 5'4?@)7 *) +,-.()'.)(8 +) 3&550.; ).(') +)7
5'4?@)7 )( +)7 *23&0.)78 )( +& 3&.0D') &1): +)7A/)++)7 +)7 *23&0.)7 *) +,-.()'.)( 72.( :2..):(47#
E,0.0(0&(01) +& 5+/7 :2../) )7( /0.1$2+$3% F G H# >&'30 +)7 &55+0:&(02.7 92/'.0)7 5&' '2/()10)67
/.) &55+)( I&1& A/0 ('&:) 52/' :J&A/) &*')77) ->8 +& :&'() *) +,-.()'.)( &/ .01)&/ *23&0.) 1/)
5&' :)(() &*')77)# K0;/') "# 32.(') /. )@)35+) *) :)(() :&'() ')+&(01) C /.) 3&:J0.) C +,-LM-N
O25J0&PN.(052+07#
E& :23B0.&072. *)7 (&B+)7 *) '2/(&;) <=> *2..) /.) :)'(&0.) 03&;) *) +& :&'() *) +,-.()'.)(
&/ .01)&/ *23&0.)# Q. +0). ).(') *)/@ *23&0.)7 &55&'&R( 7/' :)(() :&'() 70 :) +0). ?;/') 7/'
/.) '2/() :J2070) 5&' /. *)7 '2/()/'7 <=> :2.70*4'47# Q. +0). A/0 .,&55&'&R( 7/' &/:/.) '2/()
.) 5)/( 5&7 S(') *4():(4 5&' :)(() 34(J2*)# -.(/0(01)3).(8 +& 5'4:0702. *) +& :&'() &/;3).()
&1): +) .23B') *)7 (&B+)7 :2++):(4)7 )( :23B0.4)7 ).7)3B+)# -+ & 4(4 32.('4 *&.7 F TH A/,/.)
*0U&0.) *) (&B+)7 :2++):(4)7 C *)7 ).*'20(7 *0V4').(7 )7( 7/W7&.( 52/' &120' /.) B2..) 5'4:0702.
*) +,-.()'.)(# N/ *)+C *) :) 7)/0+ +) ;&0. ). :2.70*4'&.( 5+/7 *) (&B+)7 )7( 3&';0.&+#
E)7 (&B+)7 *) '2/(&;) <=> 5)/1).( &/770 5'47).()' /. &/(') &1&.(&;) A/&.* 2B7)'14)7 *&.7
+) ()357 )( .2. 5&7 C /. 323).( *2..4# X. )V)(8 A/&.* /. +0). *) +,-.()'.)( (23B) ). 5&..)8
(2/7 +)7 5'4?@)7 *2.( +)7 '2/()7 5&77).( 5&' :) +0). *075&'&077).( 52/' 5)/( S(') '4&55&'&R(') 5+/7
(&'* 70 /.) &/(') '2/() )@07() Y2/ +) +0). )7( '4(&B+0Z# E& :2''4+&(02. ).(') +)7 ()357 *) *075&'0(02.
)( *,&55&'0(02. *)7 5'4?@)7 )7( *2.: /.) B2..) 0.*0:&(02. A/,0+7 7) ('2/1).( 5'2:J)7 +)7 /.
*)7 &/(')7# %)(() :2''4+&(02. & 4(4 2B7)'14) *&.7 F

H 7/' *)7 ;'&.*7 0.()'1&++)7 *) ()3578 52/'

:2.:+/') A/) :,)7( /. 2/(0+ )W:&:) 52/' *4:0*)' */ ');'2/5)3).( *)7 5'4?@)7 *&.7 +,-.()'.)(#

!"#$%&' () *+,-./0%$.1 20 2.3"$1' 2' /" 3,%&./.4$'

"#$%&'()*&+ ,- '. /$)0&'&1*- 2.33*%"# $%&'()(*+, -#..+/, 01 '%.,#1 23# -#. 4,..% 03%/()1,' 0,-1+. )# 4'%#&+(2 0, )352&,'2,&6 ",.

&'#/#17 0#2. 4,&&, 0+',4&+(2 -,1/,2& 8&', *'(1-%. ,2 4+29 4)#..,. .,)(2 ),1' (:;,4&+< =
> ?44%)%',' )# /+&,.., 0, 4#-&1', ,& .&(4@#*, ,2 $%$(+', 0,. -#91,&.6
> ? -#'&+' 01 &'#A4 4()),4&% ,2 12 (1 -)1.+,1'. -(+2&.B +2<%',' 4, 91+ ., -#.., #+)),1'. 0#2. ),
'%.,#16
> C, 2(1/,)),. $%&D(0,. 03#*'%*#&+(2 91+ -,'$,&&,2& 0, 4($-',..,' ), &'#A4 4()),4&% .#2.
-,'0', &'(- 03+2<('$#&+(2.6
> C,. $%&D(0,. 03%4D#2&+))(22#*, 91+ -,'$,&&,2& 0, '%01+', ), /()1$, 0, &'#A4 E 4()),4&,'
,& E &'#+&,' &(1& ,2 ',.&#2& 4#-#:), 03+2<%',' ),. 4#'#4&%'+.&+91,. 01 &'#A4 ('+*+2#)6
> ?2#)F., ,& $(0%)+.#&+(2 01 &'#A4 4()),4&%6

tel-00408745, version 1 - 3 Aug 2009

",. .,4&+(2. .1+/#2&,. 0%4'+/,2& :'+G/,$,2& ),. &'#/#17 ,H,4&1%. 0#2. 4D#412, 0,. 4+29 4)#..,.6

! !"

#$%&'()* +,-* )./0/)* +'-1 ,0 /0+%-1) ()* +02-)%*

", -'(:)G$, 0,. (1&+). 4($$,

%5#203#

,&

'%!'&'"/

,.& 913+). 2%4,..+&,2& -(1' 4D#91, -#I

91,& 4#-&1'%B 0, ), ',$(2&,' E &'#/,'. ), .F.&G$, 03,7-)(+&#&+(2 -(1' -(1/(+' ), .&(4@,' 0#2. )#
$%$(+',6 J,&&, -'(4%01', ,.& 4(1&,1., 4, 91+ '#),2&+& )# /+&,.., 0, 4#-&1', ,& 4#1., 0,. -,'&,.
0, -#91,&.6 K(1' #44%)%',' 4,&&, -'(4%01',B )# &,4D2+91, 91+ # %&% #0(-&%, -#' -)1.+,1'. 4#'&,.
.-%4+#)+.%,. 0#2. ),. $,.1',. 4($$, )# 4#'&, C?L 03M20#4, NOPQ # %&% 03#44%0,' 0+',4&,$,2&
E )# $%$(+', -(1' )3%4'+&1', 0,. -#91,&. .#2. 0,/(+' -#..,' -#' ), .F.&G$, 03,7-)(+&#&+(26 J,4+
# 4(2.&+&1% 12, +$-('&#2&, #/#24% 0#2. 4, 0($#+2, ,2 -,'$,&&#2& )# 4#-&1', 0,. -#91,&. E )#
/+&,.., 91, ), :1. 0, )# .&#&+(2 0, &'#/#+) ,& )# $%$(+', )# -,'$,&&,2&6
R2, #1&', .()1&+(2 1&+)+.%, # %&% 4,)), 0, )# 4#-&1', 0312, -#'&+, 0,. -#91,&. 4($$, -#'
,7,$-), )3,2&8&, SJKT5K -)1. ),. 91,)91,. -',$+,'. (4&,&. 0, 0(22%,.6 "# -)#&,<('$, 0, $,.1',
-#..+/, 5KUVW 0, X-'+2&"#:. N!YQ # #0(-&% 4,&&, .()1&+(26 ", <#+& 91, ),. -#91,&. 4#-&1'%. .(2&
'#$,2%. E 12, -)1. -,&+&, &#+)), # -,'$+. 0, ),. %4'+', ,2 $%$(+', E 12, /+&,.., -)1. %),/%,6
K(1' )# -)1-#'& 0,. #--)+4#&+(2.B 4(22#Z&', )3,2&8&, ,.& .1[.#2&6 U#)D,1',1.,$,2&B .+ -(1' 12,
#--)+4#&+(2 91,)4(291, 4($$, -#' ,7,$-), )3+0,2&+A4#&+(2 01 &'#A4 -#' .(2 4(2&,21B )31&+)+.#&,1'
(1 )3(-%'#&,1' # :,.(+2 0, -)1. 0, 0(22%,.B ),. (:&,2+' 4# .,'# +$-(..+:), %&#2& 0(22% ), 4#'#4&G',
+''%/,'.+:), 0, )3#--'(4D,6
"3+2&'(014&+(2 0,. <(24&+(2. 0, D#4D#*, N\OQ -(1' )# 4#-&1', ,& )3#2#)F., 0,. -#91,&. # #1..+
-,'$+. 0, *#*2,' 4(2.+0%'#:),$,2& ,2 -,'<('$#24,6 ?/,4 ),. <(24&+(2. 0, D#4D#*, #--)+91%,. E
4,'&#+2. 4D#$-. 0#2. ),. -#91,&.B +) ,.& 0,/,21 <#4+), 03#44%0,' 0+',4&,$,2& 0#2. )# $%$(+', E )#
4#., 91+ 4(2&+,2& ),. +2<('$#&+(2. .1' 4D#91, -#91,&6 J,4+ # -,'$+. 0, 4#)41),' <#4+),$,2& .1' 12
&'#A4 0,. $%&'+91,. 4($$, ), 2($:', 0, -#91,&. -#' #0',.., 5K (1 -#' '%.,#16 C#2. ), 4#. (] )#
/+&,.., 03#44G. E )# $%$(+', ,.& 4(2.+0%'%, &(1;(1'. +2#44,-&#:),B 0,. A)&',. 0, &F-, ^)(($ N_OQ
-,1/,2& 8&', 1&+)+.%. -(1' ,H,4&1,' 12 -',$+,' A)&'#*, '#-+0, 0,. +2<('$#&+(2.B .1+/+ -#' 12 #44G.

! ! "#$%&'()*&+ ,- '. /$)0&'&1*- 2.33*%-

" #$ %&%'()*+ ,-*.*%/#* #* /#01 23/(40* *12 5*#0( 6* #$ %*10)* 6*1 $6)*11*1 78 1'0)5*1 6$91 09
2)$:5+ ;$91 09 /)*%(*) 12$6* 09 :#2)* <#''% *12 02(#(1& /'0) 1$='() 1( 09* $6)*11* 78 1'0)5* $
6&>" &2$(2 =0*? *2 1( 5* 9-*12 /$1 #* 5$1? #$ %&%'()* *12 $55&6&* /'0) #-*9)*@(12)*%*92 6* 5*22*
$6)*11*+ ,*1 2)$=$0. A$1&1 10) #*1 B'952('91 6* C$5C$@* *2 #*1 :#2)*1 <#''% 1'92 1'0=*92 /)&1*92&1
6$91 #$ #(22&)$20)* 1'01 #* 2(2)* D;$2$ E2)*$%(9@D+

! !

"#$%&'#(' )'* +'&$,&-.#('* &%*'./

,$ %&2)'#'@(* /$11(=* 60 )&1*$0 5'91(12* /$) 6&:9(2('9 *9 #$ 5$/20)* 6*1 6'99&*1 *9 5*)2$(91
/'(921 60 )&1*$0+ F*5( /*)%*2 B$5(#*%*92 #* 5$#50# 6*1 %&2)(40*1 #(&*1 " 5*1 /'(921 5'%%* /$)
*.*%/#* #$ )&/$)2(2('9 60 2)$:5 *92)* #*1 6(G&)*92*1 $//#(5$2('91? #* ='#0%* 60 2)$:5? #* 9'%A)*
6* H'21? #$ 6(12)(A02('9 6* #$ 2$(##* 6*1 /$40*21? *25+ 8$) 5'92)*? (# *12 6(I5(#* 6* 5$#50#*) 6*1

tel-00408745, version 1 - 3 Aug 2009

%&2)(40*1 6* 9$20)* A'02J*9JA'02? 5-*12J"J6()* 6*1 %&2)(40*1 #(&*1 " 09 5C*%(9 6* #-792*)9*2 /#02K2
40-" 09 '0 /#01(*0)1 /'(921 6* %*10)*+ ,-(9B&)*95* 6*1 5$)$52&)(12(40*1 6-09 5C*%(9 9&5*11(2* #*
6&/#'(*%*92 6*1 /'(921 6* %*10)* /$11(=* /$)2'02 6$91 #* )&1*$0? %$(1 %$#C*0)*01*%*92 5*5(
9-*12 /$1 B$(1$A#*? '0 $0 %'(91? (# *12 (%/'11(A#* 6-(%$@(9*) 40* 5*1 /'(921 6* %*10)*1 %L%*
1( 6&/#'3&1 /$) #*1 6(G&)*921 '/&)$2*0)1? /0(11*92 /$)2$@*) #*0)1 %*10)*1+ 8'0) 6*1 )$(1'91 6*
1&50)(2&? #*1 '/&)$2*0)1 9* /*)%*22*92 /$1 " 6*1 *.2&)(*0)1 6-*G*520*) 6*1 %*10)*1 10) #*0) )&1*$0?
*2 9-$55*/2*92 /$1 6* /$)2$@*) #*1 )&10#2$21 6* #*0)1 %*10)*1+
M$#@)& 5*1 'A12$5#*1? 6*1 *G')21 '92 &2& *92)*/)(1 /'0) (9B&)*) 5*)2$(9*1 5$)$52&)(12(40*1 6*1
5C*%(91 " /$)2() 6*1 %*10)*1 /$11(=*1+ ;$91 N!OP? 09* &206* $ &2& %*9&* /'0) (9B&)*) #*1 5$)$52&J
)(12(40*1 6-09* 5'99*.('9 QF8 " /$)2() 6*1 %*10)*1 *G*520&*1 *9 09 1*0# /'(92+ ;*1 12$2(12(40*1
5'%%* #* 6&A(2? #$ =*)1('9 QF8? #-&='#02('9 6* #$ B*9L2)* *2 60 RQQ SR'096JQ)(/ Q(%*T? '92
&2& 5$#50#&*1+ F*5( *12 )*960 /'11(A#* /$) #$ 9$20)* A(6()*52('99*##* 6* QF8? 'U #-'A1*)=$2('9 1(J
%0#2$9&* 6*1 /$40*21 6* 6'99&*1 *2 6*1 $540(22*%*921 /*)%*2 6* 5'%/)*96)* 5* 40( 1* /$11* $0
9(=*$0 6* #$ 1'0)5* *2 6* #$ 6*12(9$2('9+ NV!P $ *11$3& 6* 5$#50#*) #*1 6&A(21 /C31(40*1 6*1 #(*91 #*
#'9@ 6-09 5C*%(9 6* #-792*)9*2 " /$)2() 6* #-C(12'@)$%%* 60 2*%/1 40( 1&/$)* #*1 /$40*21 6-09*
5'99*.('9 QF8 2)$=*)1$92 5* 5C*%(9+ N WOP 02(#(1* #*1 /$40*21 6-09* 5'99*.('9 QF8 6$91 #*1
6*0. 1*91 /'0) (9B&)*) #$ A$96* /$11$92* 6(1/'9(A#* #* #'9@ 6-09 5C*%(9+ X(9$#*%*92? /#01(*0)1
2)$=$0. S*+@+? N WVPT '92 *11$3& 6-(9B&)*) 5* 40-'9 $//*##* #$ %$2)(5* 6* 2)$:5 6-09 )&1*$0 S#* 2)$:5
*92)* 5C$40* /'(92 6-*92)&* *2 5C$40* /'(92 6* 1')2(*T " /$)2() 6*1 %*10)*1 /$11(=*1 *G*520&*1 "
#-(92&)(*0) 60 )&1*$0+ ,* /'(92 5'%%09 " 2'01 5*1 2)$=$0. *12 #$ 6(I50#2& 6* #$ /)'5&60)* *2 #*
@)$96 A)0(2 6$91 #* )&10#2$2 'A2*90+
,* /#$5*%*92 6*1 /'(921 6* %*10)*1 /$11(=*1 *12 09 /)'A#Y%* *9 1'(+ ;-09 5K2& 5*1 /'(921
'92 09 5*)2$(9 5'Z2 *2 5'91'%%*92 6*1 )*11'0)5*1 40* 5* 1'(2 /'0) 5'##*52*) #* 2)$:5 'U /'0)
#* 2)$91/')2*)+ ;* #-$02)* 5K2&? 5*)2$(91 6* 5*1 /'(921 6* %*10)* /*0=*92 L2)* 5'%/#Y2*%*92
)*6'96$921 $=*5 6-$02)*1 '0 5'%/#Y2*%*92 (902(#*1 /'0) #$ %&2)(40* 5'91(6&)&*+ 8$) *.*%/#*?
/'0) 5$#50#*) #* 2)$:5 40( 2)$91(2* *92)* 6*0. )&1*$0.? (# 10I2 6* %*22)* 09 /'(92 6* %*10)* 10)
#* #(*9 6* 2)$91(2 $0 #(*0 6* %*10)*) 2'01 #*1 #(*91 60 )&1*$0 *2 (6*92(:*) #* 2)$:5 40( 2)$91(2* /$) 5*

!"#$%&' () *+,-./0%$.1 20 2.3"$1' 2' /" 3,%&./.4$'

"#$%& '$()*$* ($)"$+$%, "$ "#$% -$ ,*.%(#, $(, /.* 01%,*$ #%()2(.%, /1)* 01%%.3,*$ ". *4/.*,#,#1%
-$ ". 05.*6$ 7 "8#%,4*#$)* -) *4($.)&
9#:$*( ,*.:.); 1%, *$6.*-4 "$ /*1<"=+$ -$ /".0$+$%,> (1):$%, -) /1#%, -$ :)$ -) 0."0)"
-$ ". 05.*6$ ()* "$ *4($.) ?$&6&> @ ABCD& E. F)$(,#1% /1(4$ 4,.#, 01++$%, /".0$* "$ +#%#+)+ -$
+1%#,$)*( /1)* 0./,)*$* .) +1#%( )%$ G1#( ,1)( "$( /.F)$,( F)# ,*.:$*($%, )% *4($.)& H" (8.6#, -8)%
/*1<"=+$ 0".((#F)$ -$ ". ,541*#$ -$( 6*./5$(> F)# /$), I,*$ *4(1") .:$0 -$( 5$)*#(,#F)$( :1#*$ )%$
*$05$*05$ $;5.)(,#:$ -.%( "$ 0.( -$ /$,#,( *4($.);& 9$ %1,*$ 0J,4> %1)( .:1%( /*1/1(4 -.%( @ C
)%$ (1"),#1% 64%4*."$ .-./,.,#:$ F)# /*1K,$ -) G.#, F)$ "$( *1),$)*( -$ "8H%,$*%$, -#(/1($%, -4L7
-81),#"( -$ +$()*$ 01++$ "$( 'HM( -$ NO'P $, O$,Q"1R @BSC& O1)( .:1%( <T,# ()* 0$( 1),#"( /1)*
"$ -4:$"1//$+$%, -8)%$ (1"),#1% .-./,#:$ F)#> $% G1%0,#1% -$ ". ,T05$ -$ +$()*$ 7 $U$0,)$*> G.#,
.0,#:$* "$( <1%( *1),$)*( /1)* ". 01""$0,$ -8#%G1*+.,#1%( ()* "$ ,*.K0 -.%( "$ *4($.)& E$ V5./#,*$
()#:.%, -40*#, $% -4,.#" %1,*$ .//*105$& W% /")( -$ ". (#+/"$ .0,#:.,#1% 1) -4(.0,#:.,#1% -8)%

tel-00408745, version 1 - 3 Aug 2009

*1),$)*> %1,*$ .//*105$ ,#$%, $% 01+/,$ "$ G.#, F)$ "$( *1),$)*( /$):$%, /.*,#0#/$* -#U4*$++$%, 7
)%$ ,T05$ -$ +$()*$ $% 01""$0,.%, "$ ,*.K0 7 -$( :#,$(($( -#U4*$%,$( ?0$ F)# 01**$(/1%- 7 -$( ,.);
-8405.%,#""1%%.6$ -#U4*$%,(D& V$0# /$*+$, )% -$6*4 -$ "#<$*,4 -$ /")( 0$ F)# .+4"#1*$ -.:.%,.6$
"$( +$()*$( /.((#:$( $, "$( *$%- /")( .-./,.,#:$( 7 ". ,T05$ -$ +$()*$ F)8)% 1/4*.,$)* -4(#*$
$U$0,)$*&

! !"

#$%&$'()*+ ,-. )+/*%0'()*+. .1% 2- (%'34

V1""$0,$* $, *$/1*,$* ,1)( "$( /.F)$,( $(, )%$ 1/4*.,#1% ,*=( 01),$)($ F)# %$ /.(($ /.( 7
"8405$""$& P1)* 0$". #" $(, #+/1*,.%, -$ ,*1):$* -$( +4,51-$( -8.6*46.,#1% -) ,*.K0 F)# /$*+$,,$%,
-$ *4-)#*$ "$ :1")+$ -$( -1%%4$( 7 *$/1*,$* $, 7 ,*.#,$* (.%( /$*-*$ ,*1/ -8#%G1*+.,#1%(& V$ F)#
/1)(($ $% G.:$)* -$ 0$( +4,51-$( 08$(, F)$ "$( .//"#0.,#1%( -.%( ". +4,*1"16#$ %81%, /.( <$(1#% -$
01%%.3,*$ ,1), ()* "$ ,*.K0& P.* $;$+/"$> "$( -1%%4$( .//"#0.,#:$( -.%( "$( /.F)$,( (1%, *.*$+$%,
),#"#(4$(& P.*$#" /1)* 0$*,.#%( 05.+/( 01++$ "$ 05$0X()+> ". :$*(#1% -$ HP> "$( 1/,#1%(> "$ %)+4*1
-$ (4F)$%0$ YVP> $,0& Z%$ +4,51-$ -8.6*46.,#1% F)# *$6*1)/$ "$( /.F)$,( $% G1%0,#1% -$ 0$*,.#%(
0*#,=*$( $, F)# ,*.%(+$, F)$ -$( #%G1*+.,#1%( .6*464$( 01++$ "$ %1+<*$ -810,$,( $, "$ %1+<*$
-$ /.F)$,(> $(, (1):$%, ()2(.%,$& P1)* "$( .//"#0.,#1%( F)# -$+.%-$%, /")(> +."5$)*$)($+$%,
#" ($*. ,*1/ ,.*- $, #" G.)-*. *$G.#*$ ". +$()*$ +.#( 0$,,$ G1#([0# .:$0 )%$ .6*46.,#1% -#U4*$%,$&
E. +4,51-$ -8.6*46.,#1% ". /")( 01%%)$> -#,$ ". +4,51-$ -$ <.($ 1) ". +4,51-$ ". /")( K%$>
$(, 0$""$ F)# *$6*1)/$ "$( /.F)$,( $% ),#"#(.%, "$( 0#%F 05.+/( ()#:.%,( -.%( "8$%,I,$ \ .-*$(($ HP
(1)*0$> .-*$(($ HP -$(,#%.,#1%> %)+4*1 -$ /1*, (1)*0$> %)+4*1 -$ /1*, -$(,#%.,#1%> $, /*1,101"$
-$ ,*.%(/1*,& E8$%($+<"$ -$( /.F)$,( 6*1)/4( F)# $% *4()",$%, 01**$(/1%- 7 )%$ 01%%$;#1% YVP
1) <#$% 7 )% ]1, Z9P& V$, $%($+<"$ $(, (1):$%, .//$"4 ]1, S[,)/"$ 1) ]1, -4K%# 7 <.($ -$
F)#%,)/"$& P1)* 05.F)$ ]1, S[,)/"$> )% $%($+<"$ -8#%G1*+.,#1%( /$):$%, ^ I,*$ .((10#4 01++$ "$
%1+<*$ -810,$,(> "$ %1+<*$ -$ /.F)$,(> "$ ,$+/( -$ -4<), $, -$ K%> "$( <#,( -) 05.+/ 5
"$( <#,( -) 05.+/

6 7"48>

%9#' .: 8'&-$;'> $,0& _:$0 0$,,$ +4,51-$> )% 1/4*.,$)* 01""$0,$ -1%0 -$( ]1,(

.:$0 -$( #%G1*+.,#1%( ()* 05.F)$ ]1, $, $((.#$ -8),#"#($* 0$ *$,1)* /1)* 01+/*$%-*$ 0$ F)# ($ /.(($

! ! "#$%&'()*&+ ,- '. /$)0&'&1*- 2.33*%-

"#$% %&$ '(%)#*+ ,)% %-#-.%-./*)% 0&11) 23*-.2.%#-)*' 2) 42*% 5&*'1#$"6 2)% '(%)#*7 5($('#$- 2)
42*% ") -'#806 2) $&19') ") ')/*:-)% ;<= )$>&?()% @ *$ %)'>)*' &* @ *$ '(%)#*6 )-06 4)*>)$:-') A#0.2)1)$- 0#20*2()% @ 4#'-.' ")% B&-% CD-*42)%+ E2 %3#5.- %&*>)$- "3*$) #5'(5#-.&$ ")% B&-%
CD-*42)% @ *$ $.>)#* 42*% (2)>(+ =#' )7)142)6 *$) #5'(5#-.&$ 4 5 .,0-33- 3&(06- )%- (/*.>#2)$-) @
1)--') )$%)192) -&*% 2)% B&-% CD-*42)% "&$- 2)% #"')%%)% E= %&*'0)% 4#'-#5)$- 2)% 1:1)% FG 9.-%
)$ 4#'-#$- ") 5#*0H)+
I$ 1&- .14&'-#$- @ ".') @ 4'&4&% ")% B&-% )%- 2)*' "*'()+ J 23)70)4-.&$ ")% B&-% ;<= "&$2) "(9*- )- 2# 8$ %&$- 9.)$ "(8$.% 4#' 2)% 4#/*)-% KLM )- NEM6 *$ B&- )$ 5($('#2 )%- #%%&0.( @ *$
0)'-#.$ .$-)'>#22) ") -)14%+ J* 9&*- ") 0)- .$-)'>#22) ") -)14%6 2) B&- )%- 0&$%."('( 0&11) (-#$8$. )- *$ #*-') B&- )%- 0'(( 1:1) %. ")% 4#/*)-% 0&$-.$*)$- @ -'#$%.-)'+ <)0. )%- A#.- 4&*' (>.-)'
/*) 0)'-#.$% B&-% $) "*')$- 4#% .$"(8$.1)$- )- 4&*' 2.1.-)' 2) -)14% ") ')-&*' ")% 1)%*')%+
O3#5'(5#-.&$ )$ B&-% )%- %*44&'-() 4#' 2)% $&*>)#*7 '&*-)*'% 4&*' 4)'1)--') #*7 &4('#-)*'%

tel-00408745, version 1 - 3 Aug 2009

*$) 0&$$#.%%#$0) 42*% 8$) /*) 0)22) ") KMP= ") 0) /*. %) 4#%%) "#$% 2)*' '(%)#*+ O3)7)142) 2)
42*% -?4./*) )%- M)-N2&Q RSCT ") <.%0& /*. )%- %&'-. )$ 42*%.)*'% >)'%.&$%+ O# >)'%.&$ C 4)'1)- *$)
#5'(5#-.&$ @ 9#%) ") /*.$-*42) U 2# >)'%.&$ S 4)'1)- 42*% ") 0&$-'V2) )$ 4'&4&%#$-

$.>)#*7

"3#5'(5#-.&$ U )- 2# >)'%.&$ W 4)'1)- *$) #5'(5#-.&$ B)7.92) )- )7-)$%.92) 4&*' %*44&'-)' ")
$&*>)#*7 0H#14% )- -)0H$&2&5.)%+ M)-N2&Q ".%4&%) "3*$ 0)'-#.$ 0#0H) 4&*' 2) %-&0X#5) ")% B&-%+
I$ B&- )%- ')4&'-( /*#$" .2 )%- "(02#'( 8$.+ O) "(9&'")1)$- "* 0#0H) 4'&>&/*) #*%%. 2) '#44&'-#5)
")% B&-% 2)% 42*% Y5(%+

! !"

#$%&'()*+,,-))(./ 0()1 ,( 2%*3-,-.+/

E2 %3#5.- "3*$ "&1#.$) "3*$ 5'#$" .$-(':- "#$% 2# '("*0-.&$ "* >&2*1) "* -'#80 @ 0&22)0-)'
)- @ -'#.-)'+ <) "&1#.$) # 5#5$( 9)#*0&*4 "3.14&'-#$0) "#$% 2)% ")'$.Z')% #$$()% )- # (-(
1&-.>( 4#' 2) A#.- /*) 0)'-#.$)% 1(-'./*)% "#$% 2# 1(-'&2&5.) $) $(0)%%.-)$- 4#% 2# 0#4-*') ")
-&*% 2)% 4#/*)-%+ =#' 0&$%(/*)$0)6 .2 # (-( 4'&4&%( ") '("*.') 2) >&2*1) "* -'#80 0&22)0-( 4&*'
0)% 1(-'./*)% #8$ ") '("*.') 2# 0H#'5) %*' 2)% '&*-)*'% )- 4)'1)--') #*7 %?%-Z1)% ") 1)%*')
"3&4(')' @ ")% "(9.-% )$0&') 42*% (2)>(%+ O) 0#20*2 %) A#.- #2&'% %*' 2# 4#'-.) 0&22)0-() "* -'#80+
<2#.')1)$-6 )- 0&11) "#$% 0H#/*) #442.0#-.&$ &[ 23(0H#$-.22&$$#5) )%- 4'#-./*(6 .2 A#*- -'&*>)'
*$) 1(-H&") 4&*' .$>)'%)' 2)% &9%)'>#-.&$% )- 0#20*2)' 2# >#2)*' ") 2# 1(-'./*) ')2#-.>) #* -'#80
&'.5.$#2 @ 4#'-.' "* -'#80 (0H#$-.22&$$(+ E2 A#*- #*%%. -'&*>)' ")% 1(-H&")% .$-)22.5)$-)% 4&*'
)\)0-*)' 23(0H#$-.22&$$#5) %#$% -'&4 .$"*.') )$ )'')*' 2) '(%*2-#-+ ,#$% 0)'-#.$% 0#% &[ 2# 1(-'./*)
)%- ')2#-.>) @ *$) 5'#$") 4&4*2#-.&$ 0&11) 4#' )7)142) 2) >&2*1) 1&?)$ "* -'#806 2) 0#20*2
)%- %.142) )- ")% -#*7 "3(0H#$-.22&$$#5) '("*.-% 4)*>)$- :-') 4'#-./*(%+ =#' 0&$-')6 "#$% 2)%
#*-')% 0#% &[ 2# 1(-'./*) )%- "(4)$"#$-) ") 4&4*2#-.&$% ") 4)-.-)% -#.22)% 0&11) 2)% 4)-.-% B&-%6
23(0H#$-.22&$$#5) )- 23.$>)'%.&$ ")1#$")$- *$ )\&'- $&$ $(52.5)#92) %&*>)$- #>)0 23#.") "3&*-.2%
1#-H(1#-./*)% )- %-&0H#%-./*)%+
O) 0H&.7 ")% 4&.$-% ") 1)%*') 4)*- #*%%. :-') >* 0&11) *$ 4'&92Z1) "3(0H#$-.22&$$#5)+
<3)%- 42*% ") 23(0H#$-.22&$$#5) %4#-.#2 /*) ") 23(0H#$-.22&$$#5) -)14&')2+ J* 2.)* ") '(4&$"') @ 2#

!"#$%&' () *+,-./0%$.1 20 2.3"$1' 2' /" 3,%&./.4$'

"#$%&'() *(+,'$) -$ ./"#$&% '0 1/#& .2$)-2$ .(#2 */0*#0$2 #)$ *$2&/')$ +3&2'"#$4 053*6/)&'00())/7$
%./&'/0 %5')&32$%%$ 8 23.()-2$ 8 0/ "#$%&'() -5(9 '0 1/#& */.&#2$2 0$% ./"#$&%: ;/)% 0/ 0'&&32/&#2$
*$% -$#< /%.$*&% -$ 053*6/)&'00())/7$ =%./&'/0 $& &$+.(2$0> ()& 3&3 /,(2-3% ')-3.$)-/++$)& 05#)
-$ 05/#&2$: ?(&2$ &2/@/'0 / *()%'%&3 $) #)$ 72/)-$ ./2&'$ 8 0$% *(+,')$2 $)%$+,0$ .(#2 (,&$)'2 #)$
/2*6'&$*&#2$ "#' .$2+$&4 $) 1()*&'() -$ 0/ &A*6$ -$ +$%#2$ -3%'23$4 -$ *6('%'2 (9 '0 1/#& *(00$*&$2
0$% ./"#$&% $& 8 "#$0 &/#<4 /@$* 0/ .(%%','0'&3 -5/@('2 -$% &/#< -53*6/)&'00())/7$ -'B32$)&% -/)%
0$% -'B32$)&% 2(#&$#2%: ?(#% /@()% @(#0# "#$ )(&2$ /2*6'&$*&#2$ %('& */./,0$ -$ %5/-/.&$2 /#<
@/2'/&'()% -# &2/C* $& 8 0/ -'@$2%'&3 -$% &A*6$% -$ +$%#2$ "#' .$#@$)& D&2$ $)&2$.2'%$% -/)% #)
23%$/#: ?(&2$ /2*6'&$*&#2$4 $) *(+,')/)& 0$% -$#< /%.$*&%4 $%& */./,0$ -5/&&$')-2$ -$ +$'00$#2$%
.$21(2+/)*$%:
E53*6/)&'00())/7$ $%& #&'0'%3 ./2 0$% (.32/&$#2% +/'% %(#@$)& %/)% #)$ 3&#-$ %(0'-$ -$22'F2$
"#' ')-'"#$ *(++$)& '0 1/#& 1/'2$ .(#2 /@('2 -$% 23%#0&/&% ./% &2F% $22()3%: G$2&/')% (.32/&$#2%

tel-00408745, version 1 - 3 Aug 2009

#&'0'%$)& -$% &/#< &2F% ,/% "#' .$#@$)& *()@$)'2 "#58 #) $)%$+,0$ 2$%&2$')& -$ +3&2'"#$%: G'%*(
.2(.(%$ -5/H(#&$2 053*6/)&'00())/7$ 8 ?$&I0(J /@$* #)$ +3&6(-$ -5')@$2%'() %'+.0$ K .(#2 #)
&/#< -53*6/)&'00())/7$ pL =%$#0$+$)& pL -$% ./"#$&% %()& *(00$*&3%>4 0$% M(&% %()& .2$+'F2$N
+$)& 1(2+3% $) ./2&/)& -# &2/C* 3*6/)&'00())34 /.2F% *6/"#$ M(& / %() @(0#+$ +#0&'.0'3 ./2

100/p .(#2 ./2@$)'2 8 #)$ $%&'+/&'() -$ %() @(0#+$ (2'7')/0 /@/)& 053*6/)&'00())/7$: G$% .2/N
&'"#$% %()& 0(') -$ 05(.&'+/0 ='0 1/#& .$)%$2 /# */% -$% .$&'&% M(&%> $& -$ 0/ 2$*6$2*6$ $%& $)*(2$
)3*$%%/'2$ .(#2 +'$#< *(+.2$)-2$ 05'+./*& -$ 053*6/)&'00())/7$ $& *(++$)& 05(.&'+'%$2: E$%
&2('% *6/.'&2$% %#'@/)&% %()& *()%/*23% 8 )(% *()&2',#&'()% -/)% *$ -(+/')$: ;/)% G6/.'&2$ O
=-3&/'003 -/)% P))$<$ P>4 )(#% .23%$)&$2()% )(&2$ /2*6'&$*&#2$ "#' (.&'+'%$ 053*6/)&'00())/7$
-/)% #) 23%$/# 3&/)& -())3$ #)$ &A*6$ -$ +$%#2$: G6/.'&2$% Q $& R =-3&/'003% -/)% P))$<$% S
$& G 2$%.$*&'@$+$)&> $<.0'"#$)& %#2 -$#< $<$+.0$% *()*2$&% 05'+./*& -$ 053*6/)&'00())/7$ %#2
0$% 23%#0&/&% -$ +$%#2$ $& *(++$)& 1/'2$ .(#2 ')@$2%$2 0$% ')1(2+/&'()% 3*6/)&'00())3$%: G$&&$
3&#-$ $%& 0(') -5D&2$ $<6/#%&'@$4 *$.$)-/)& 0/ +3&6(-(0(7'$ #&'0'%3$ -$-/)% .$#& D&2$ /..0'"#3$
8 -5/#&2$% .2(,0F+$% *(++$ 0/ -3&$*&'() -$% /)(+/0'$% $& 0/ .23-'*&'() -# &2/C*:

! !"

#$%&'() )* +,-.&/(%*/,$ -0 *1%23 3,&&)3*.

E/ *(00$*&$ -# &2/C* / -'@$2%$% /..0'*/&'()%: E$% (.32/&$#2% %5$) %$2@$)& .(#2 *())/T&2$ 053&/&
-$ 0$#2% 23%$/#<: G$*' .$#& %$ 1/'2$ $) &$2+$% -# @(0#+$ -# &2/C* %#2 0$% -'B32$)&% 0'$)% $)
,'&%U% (# ,'$) $) &$2+$% -$% /..0'*/&'()% $& -$% '-$)&'C/)&% -$% #&'0'%/&$#2% "#' ./2&'*'.$)& 8 0/
73)32/&'() -# &2/C*: G$% (,%$2@/&'()% %$2@$)& 8 0/ 1('% 8 +'$#< /-+')'%&2$2 0$ 23%$/# $& 8 -3&$*&$2
$& 23/7'2 8 -$% /)(+/0'$% $& -$% %'&#/&'()% -$ *2'%$% =*$ -(+/')$ %5/..$00$ 0$ V2/C* W)7')$$2')7>:
X)$ /#&2$ /..0'*/&'() $%& 0/ .0/)'C*/&'() -$% +'%$% 8 H(#2 -# 23%$/# $) 1()*&'() -$ 053@(0#&'() -$
0/ *6/27$: Y/06$#2$#%$+$)&4 *$2&/')$% -$ %$% (,%$2@/&'()% %()& -'Z*'0$% 8 (,&$)'2 "#$ *$ %('& ./2*$
"#$ 0$% #&'0'%/&$#2% *2[.&$)& 0$#2% ./"#$&% $& -()* $+.D*6$)& 05(.32/&$#2 /# +'0'$# -5/**3-$2 8
0$#2 *()&$)# \]^_4 (# ,'$) ./2*$ "#$ *$2&/')$% -$ *$% /)(+/0'$% $& /&&/"#$% %()& %#,&'0$% "#$ 0$%
(.32/&$#2% )$ .$#@$)& ./% 0$% -3&$*&$2 %/)% #)$ 3&#-$ +')#&'$#%$ -# &2/C* \`a_:

! ! "#$%&'()*&+ ,- '. /$)0&'&1*- 2.33*%-

"#$ $%&#'(&)*+#$ , -#+. (/+. $/'( &'(0.#$$0$ 12. -32'2-4$# 5+ (.2)% 5# -36'(#.'#( 52'$ -# 7+( 5#
%/81.#'5.# $2 %/81/$&(&/' #( $/' 09/-+(&/': #( 5# 1./1/$#. 5#$ 8/5;-#$ 2'2-4(&*+#$ *+& 1#+9#'(
$#.9&. , -2 </&$ 1/+. 1.05&.# -# (.2)% #( 1/+. -# =0'0.#. $4'(>0(&*+#8#'( 52'$ 5#$ $&8+-2(&/'$
#( 5#$ #?10.&8#'(2(&/'$ %/'(.@-0#$A B# 5#.'&#. 1/&'( #$( (.;$ &81/.(2'( 1/+. -2 92-&52(&/' 5#$
1./(/%/-#$ 5# -36'(#.'#( #( 5# (/+(# 1./1/$&(&/' 9&$2'( , -#$ 280-&/.#.A B3#$( 2+$$& &81/.(2'( 1/+.
-2 92-&52(&/' 5# (/+( '/+9#2+ $4$(;8# 52'$ 5#$ %/'5&(&/'$ *+& 1#+9#'( $# .#1./5+&.# #( *+& $/'(
-#$ 1-+$ .02-&$(#$ 1/$$&7-#$A "2 92-&52(&/' 5#$ '/+9#--#$ 9#.$&/'$ 5# CBD #( 5#$ 80%2'&$8#$ 5#
%/'(.@-# 5# %/'=#$(&/' #' #$( +'# 7/''# &--+$(.2(&/'A

"2 8/50-&$2(&/' 5+ (.2)% 5# -36'(#.'#( 12$$# 5327/.5 12. +'# 0(+5# $(2(&$(&*+# 1/+. %/8E

tel-00408745, version 1 - 3 Aug 2009

1.#'5.# #( *+2'(&)#. -#$ %2.2%(0.&$(&*+#$ 1.&'%&12-#$ 5+ (.2)%A F'$+&(# 5#$ 8/5;-#$ 2'2-4(&*+#$
$/'( 1./1/$0$ 1/+. %21(+.#. %#$ %2.2%(0.&$(&*+#$A "#$ 1.#8&#.$ (.292+? 52'$ %#((# 5&.#%(&/' <2&E
$2&#'( 5#$ >41/(>;$#$ $&81-#$ $+. -# (.2)% 5# -36'(#.'#( G5#$ 53>41/(>;$#$ 5+ $/.(# H2.I/9&#''#$J:
21.;$ &- $3#$( 290.0 *+# -# (.2)% 5# -36'(#.'#( 1.0$#'(# +' %/81/.(#8#'( -/'=+# 808/&.# G211#-0
#' 2'=-2&$ -/'= .2'=# 5#1#'5#'(J *+& 5#82'5# 5#$ '/+9#2+? 8/5;-#$ KLMNA "2 5&O0.#'%# #'(.#
+' (.2)% -/'=+# 808/&.# #( +' (.2)% H2.I/9&#' #$( *+# -2 808/&.# 5+ 1.#8&#. 50%./P( 1/-4'/E
8&2-#8#'( 29#% -# (#81$ (2'( 5&$ *+# -2 808/&.# 5+ 5#+?&;8# 50%./P( #?1/'#'(&#--#8#'(A B#((#
1./1.&0(0 -/'=+# 808/&.# 2 0(0 2((.&7+0# , -2 1./1.&0(0 *+#+# -/+.5# G/+ >#294 (2&- #' 2'=-2&$J
5# -2 5+.0#: 9/-+8# #( '/87.# 5#$ )%>&#.$ #( Q/($ 52'$ -36'(#.'#( G1./1.&0(0 (.;$ %/''+# #' &'</.E
82(&*+# 12. .211/.( , (/+( %# *+& #$( >+82&' #( *+& 1.0$#'(# +' %/81/.(#8#'( 1-+(@( H2.I/9&#'
/+ #?1/'#'(&#-J KRRNA D2. %/'$0*+#'%# -#$ 8/5;-#$ *+& $/'( 1./1/$0$ 1/+. -# (.2)% 5# -36'(#.'#(
$/'( #' =.2'5# 12.(&# >&0.2.%>&*+#$ *+& 1./%;5#'( 12. -2 $+1#.1/$&(&/' 5#$ Q/($ , 1./1.&0(0 *+#+#
-/+.5# #A=A K S: LT:

NA 6- #?&$(# %#1#'52'( 532+(.#$ 8/5;-#$ 72$0$ $+. 5#$ 0*+2(&/'$ 5# .0%+.E

.#'%# $(/%>2$(&*+# #( *+& 1#.8#((#'( , 12.(&. 5#$ 92-#+.$ 5# (.2)% 52'$ -# 12$$0 #( 5#$ 92-#+.$
532+(.#$ 92.&27-#$ 2-02(/&.#$ #( 1./%#$$+$ $(/%>2$(&*+#$: 5# (./+9#. +'# 92-#+. 1/+. -# (.2)% ,
-3&'$(2'( 2%(+#- 5# (#--# $/.(# *+# -# (.2)% (/(2- =0'0.0 $2(&$<2&( %#.(2&'#$ %/'5&(&/'$ 7&#' 50)'&#$
#' (#.8#$ 5# 507&( 8/4#': 92.&27&-&(0 #( %/..0-2(&/'A "# 8/+9#8#'( 7./U'&#' <.2%(&/''2&.# #( -#$
1./%#$$+$ VWHV #( VW6HV #' $/'( +' #?#81-#A

X2'$ K SN '/+$ 29/'$ %/'(.&7+0 , %# 5/82&'# #' 1./1/$2'( +' 8/5;-# , 72$# 5# Q/($ 1/+.
-# (.2)% 5# -36'(#.'#(A Y/+$ 29/'$ -&0 -#$ 1./1.&0(0$ 5+ (.2)% (/(2- , -2 92.&27&-&(0 5+ 507&( 5#
%>2*+# Q/(A X2'$ +' (.292&- #' %/+.$: '/+$ $/88#$ #' (.2&' 532--#. 1-+$ -/&' 1/+. 1./1/$#.
5#$ 8/5;-#$ $(2(&$(&*+#$ 5#$ Q/($ =0'0.0$ 12. -#$ 5&O0.#'(#$ 211-&%2(&/'$ 5# -36'(#.'#( 52'$ -#
7+( 5# -#$ %-2$$#. +-(0.&#+.#8#'( #' +(&-&$2'( $#+-#8#'( 5#$ &'</.82(&/'$ $+. -2 (2&--# #( &'$(2'($
532..&90# 5#$ 12*+#($ $2'$ 5#9/&. $# )#. $#+-#8#'( 2+? &'</.82(&/'$ 52'$ -#$ #'(Z(#$A B#%& 2 +'
=.2'5 &'(0.Z( #' %2$ 5# (.2)% %.41(0A

!"#$%&' () *+,-./0%$.1 20 2.3"$1' 2' /" 3,%&./.4$'

!" #$%&'()*+', -. (/ 0%*1'('2+. /3*+&.

"#$ %&'()*#+($ ,# -./0*#(0#* &#+1#0* 2#$+(#( *%+$ -#$ -3#0$ ,# -#+($ ('$#)+4 ,.+0# 2)035(#
&)$$31# &%+( 6%2&(#0,(# 6# 7+3 $# &)$$#8 6#&#0,)0* 3-$ $%0* 306)&)9-#$ ,.#4'6+*#( -) 2:2#
;%06*3%0 ,)0$ -#$ ('$#)+4 ,#$ )+*(#$ %&'()*#+($< =+)0* )+4 +*3-3$)*#+($ #* )+4 ),2303$*()*#+($
,#$ ('$#)+4 ,# (#6%+1(#2#0*8 *%+* 6# 7+.3-$ &#+1#0* #4'6+*#( 6.#$* +0# ;%06*3%0 ,# 6)&*+(# ,+
*()>6 $+( -.30*#(;)6# ('$#)+ ,# -#+($ 2)6?30#$< @#**# ,#(035(# 6)&*+(# ,%00# +0# 30;%(2)*3%0 *(5$
&#+ +*3-# &%+( -) 6%2&('?#0$3%0 ,# -.'*)* ,+ ('$#)+ #* -#$ &#(;%(2)06#$ ,#$ )&&-36)*3%0$< /- A )
,%06 +0 ;%(* 9#$%30 ,# 6%--#6*#( ,#$ 30;%(2)*3%0$ $+( -.'*)* ,+ ('$#)+ )+ ,#-B ,# -.30*#(;)6# ('$#)+
#* 6#63 0# &#+* $# ;)3(# 7+# &)( +0# )&&(%6?# )6*31# &%+( 6#+4 7+3 %0* &)$ )665$ )+4 -3#0$ ,+
('$#)+< C.363 13#0* -.30*'(:* ,# 6#**# )&&(%6?#< D%0 E()0, &(%9-52# 6#&#0,)0* #$* -# 2)07+# ,#
&('63$3%0 '*)0* ,%00'$ -) 6%2&-#43*' ,+ ('$#)+ #* -# E()0, 0%29(# ,.)&&-36)*3%0$ #* +*3-3$)*#+($<

tel-00408745, version 1 - 3 Aug 2009

C.)&(5$ -# &(3063&# ,# -) 2'*(%-%E3# )6*31#8 ,#$ &)7+#*$ $%0* #01%A'$ ,)0$ -# ('$#)+ #* 6.#$*
#0 ;%06*3%0 ,# -) ('&%0$# ,+ ('$#)+ B 6#$ &)7+#*$ F3<#<8 ,'-)3 7+.3-$ &(#00#0*8 ,'93*8 &#(*#$G 7+#
7+#-7+.+0 &#+* 30;'(#( 6# 7+3 $# &)$$# B -.30*'(3#+( ,+ ('$#)+ #* $+( -# 6?#230 #0*(# -) 2)6?30#
7+3 #01%3# -#$ &)7+#*$ ,# 2#$+(# #* 6#--# 7+3 -#$ (#H%3* #* 7+3 (#01%3# -#$ )67+3**#2#0*$ $3 I)2)3$<
J%*#K 363 7+# -) 2#$+(# )6*31# &#+* $# ;)3(# $)0$ )67+3**#2#0* $3 6.#$* -) 2)6?30# ,#$*30)*(36# 7+3
2#$+(# #* $3 -#$ ?%(-%E#$ $%0* $A06?(%03$'#$< L 0%*#( )+$$3 7+# -#$ &)7+#*$ ,# 2#$+(# &#+1#0*
:*(# ,#$ &)7+#*$ ,',3'$ )1#6 ,#$ ,%00'#$ ,# 9%+(()E# F,+22A &)6M#*$G %+ 93#0 ,#$ 1()3$ &)7+#*$
,# ,%00'#$ 6%22# 6#+4 ,#$ *()0$;#(*$ ,# >6?3#($<
"#$ *()1)+4 $+( -#$ 2#$+(#$ )6*31#$ &#+1#0* :*(# 6-)$$'$ #0 ,#+4 6)*'E%(3#$8 -) ,#+4352#
6)*'E%(3# $# 9)$# #0 E()0,# &)(*3# $+( ,#$ %+*3-$ ,# -) &(#235(# 6)*'E%(3# N
O P()1)+4 13$)0* B '1)-+#( -#$ &#(;%(2)06#$ ,.+0 6?#230 #0*(# ,#+4 2)6?30#$<
O P()1)+4 13$)0* B 2#$+(#( *%+* +0 ('$#)+ #0 6%(('-)0* &-+$3#+($ 2#$+(#$ )6*31#$< "# ('$#)+
B 2#$+(#( &#+* :*(# -./0*#(0#* -+3 2:2# )+ 031#)+ ,# $#$ (%+*#+($ #* ,%2)30#$8 6%22# H)
&#+* :*(# )+ 031#)+ ,.+0 ('$#)+ Q1#(-)A F('$#)+ )+ 031#)+ )&&-36)*3; %+ ('$#)+ $+&#(&%$'G<
J%+$ ('$+2%0$ ,)0$ -) $+3*# -#$ &(3063&)+4 *()1)+4 ,)0$ -) -3**'()*+(# ,)0$ 6?)6+0# ,# 6#$ ,#+4
6)*'E%(3#$ )30$3 7+# 0%$ 6%0*(39+*3%0$ ,)0$ -# ,%2)30#<

!"!#

$%&'()(*+, -./0 12,3+0 -, ).40&,'0,& -./0, 350+6', 51&+7,

"#$ %+*3-$ ,# 9)$# ,# *A&# &30E %+ *()6#(%+*# &#(2#**#0* )+4 )&&-36)*3%0$ ,# *#$*#( -) ,3$R
&%0393-3*' ,.+0 6?#230 ,# -./0*#(0#*< @#&#0,)0* -#$ )&&-36)*3%0$ $%0* )+$$3 $#0$39-#$ B ,.)+*(#$
&)()25*(#$8 *A&37+#2#0* -# ,'-)38 -) 9)0,# &)$$)0*# #* -# *)+4 ,# &#(*#$ ,#$ &)7+#*$< "# *)+4 ,#
$#0$393-3*' ,'&#0, ,# -) 0)*+(# ,# -.)&&-36)*3%0 7+3 &#+* $# *(),+3(# $%+$ -) ;%(2# ,.+0# ;%06*3%0
,.+*3-3*' -3)0* -) 7+)-3*' &#(H+# &)( -.+*3-3$)*#+( ,# -.)&&-36)*3%0 B -.#0$#29-# ,#$ 2'*(37+#$ 6)()6R
*'(3$)0* -# 6?#230 ('$#)+ #2&(+0*' &)( -#$ &)7+#*$ ,# -.)&&-36)*3%0< S%+( -#$ )&&-36)*3%0$ ,# *A&#
*()0$;#(* ,# >6?3#($ F,3*#$ '-)$*37+#$ TUVWG8 -) ;%06*3%0 ,.+*3-3*' 1)(3# &(3063&)-#2#0* #0 ;%06*3%0
,# -) 9)0,# &)$$)0*# ,3$&%039-#8 2)3$ &%+( -#$ 6%+(*$ *()0$;#(*$ #* B 6)+$# ,# -) &?)$# D-%X D*)(*

!"! #$%&'()*+', -. (/ 0%*1'('2+. /3*+&.

"# $%&' #((# #)* +,--.# +,//# 0*1-* 2,-+*3,- 1.))3 ". "0(13 1((#454#*,.4 6789: ;- #<#*' 1. "0=.*
">.- *41-)2#4* "# ?+@3#4' 3( 21.* .- +#4*13- *#/A) A,.4 1**#3-"4# (1 A@1)# )*1*3,--134# #* +# *#/A)
1.B/#-*# (3-0134#/#-* 1C#+ (# "0(13: D-# 2,3) (1 A@1)# )*1*3,--134# 1**#3-*#' 3( #)* (,B3E.# "# "34#
E.# $%& #))13# ">.*3(3)#4 A(#3-#/#-* (1 =1-"# A1))1-*# "3)A,-3=(# F)3 (# 40+#A*#.4 (# A#4/#*G' #*
",-+ (# *#/A) 4#)*1-* "1-) (1 C3# ">.- *41-)2#4* #)* A43-+3A1(#/#-* (# 41AA,4* C,(./# "# ",--0#)
4#)*1-* ).4 =1-"# A1))1-*#:
&,.4 (#) 1AA(3+1*3,-) /.(*3/0"31 E.3 .*3(3)#-* ),.C#-* (# A4,*,+,(# DH&' (1 )3*.1*3,- #)*
"3<04#-*#: H>.- +I*0 .-# 1AA(3+1*3,- /.(*3/0"31 ->1 A1) *,.J,.4) =#),3- "# *,.*# (1 =1-"#
A1))1-*# "3)A,-3=(#: H# (>1.*4# +I*0' #* #- A(.) "# (1 =1-"# A1))1-*#' .-# 1AA(3+1*3,- /.(*3/0"31
1 )1 E.1(3*0 "0A#-"1-*# ". "0(13 1((#454#*,.4 F).4*,.* )>3( )>1B3* ">.-# 1AA(3+1*3,- *#/A) 40#(G #*
". *1.K "# A#4*#) H# A1E.#*) 6 !9: &,.4 40".34# (# *1.K "# A#4*#) #* -# A1) L*4# 1B4#))32' 3( 1 0*0
2,4*#/#-* 4#+,//1-"0 E.# (#) 1AA(3+1*3,-) /.(*3/0"31 *41-)/#**#-* 1. /L/# "0=3* E.# $%&

tel-00408745, version 1 - 3 Aug 2009

F(# 21/#.K A43-+3A# "# $%&5243#-"(3-#)) 6MM9G: N1 +,--13))1-+# "# +# "0=3* #)* *4O) 3/A,4*1-*#
A,.4 (>1AA(3+1*3,- 1?- E.>#((# A.3))# +@,3)34 ),- +,"#+ #* (1 E.1-*3*0 "# 4#",-"1-+# P 1J,.*#4 1.
Q.K A,.4 +,443B#4 *,.*# A#4*# 0C#-*.#((# "# A1E.#*):
%#**# "0A#-"1-+# "# (1 E.1(3*0 1AA(3+1*3C# #- (>0*1* ". 40)#1. 1 /,*3C0 .- B41-" -,/=4# "#
4#+@#4+@# ).4 "#) ,.*3() "# /#).4# #R+1+#) A,.4 (1 /0*4,(,B3# "#) +@#/3-) "# (>S-*#4-#*: T,.4-34
.-# (3)*# #K@1.)*3C# "# +#) ,.*3() "# /#).4# #)* .-# *U+@# 3/A,))3=(# 0*1-* ",--0 (#.4 B41-"
-,/=4#: %#4*13-) )3*#) +,//# +#(.3 "# %VSHV 67!9 A40)#-*#-* .- "#)+43A*32 "#) A43-+3A1.K ,.*3()
"# /#).4# E.>3( +,-C3#-* "# +,-).(*#4 1C1-* (# "0C#(,AA#/#-* "# ),- A4,A4# ,.*3(: H1-) +#**#
)#+*3,- +#A#-"1-* -,.) 1((,-) -,.) +,-*#-*#4 "# +3*#4 (#) A43-+3A1(#) 1AA4,+@#) "1-) (1 /0*4,(,B3#
">.- +@#/3- "# (>S-*#4-#*:

!"#$! %# %&'()
N1 /#).4# ". "0(13 ">.- +@#/3- "# (>S-*#4-#* 1 =#1.+,.A 3-*04#))0 (#) +@#4+@#.4): W1(5
@#.4#.)#/#-* #* #- (>1=)#-+# ">@,4(,B#) A14213*#/#-* )X-+@4,-3)0#)' (1 /#).4# #K1+*# ". "0(13
.-3"34#+*3,--#( #)* .-# *U+@# "3R+3(# )3 +# ->#)* A1) 3/A,))3=(#: Y#1.+,.A "# A4,*,+,(#) +,//#
$%& 6Z[9 #* $T\% 6MZ9 )# +,-*#-*#-* "# (1 /#).4# ". "0(13 1((#454#*,.4: %#**# 3-2,4/1*3,- A#.*
L*4# ,=*#-.# #K1+*#/#-* )3 (#) 40+#A*#.4) 4#-C,3#-* 41A3"#/#-* (#.4) 1+E.3**#/#-*) ,. 1. /,3-)
3-2,4/#-* (#) ),.4+#) ". *#/A) A1))0 1C1-* (# 4#-C,3 "#) 1+E.3**#/#-*): H1-) (# +1) "#) +@#/3-)
A14213*#/#-* )X/0*43E.#)' (# "0(13 .-3"34#+*3,--#( A#.* L*4# +,-)3"040 +,//# (1 /,3*30 ". "0(13
1((#454#*,.4' +# E.3 ->#)* A1) /1(@#.4#.)#/#-* (# +1) "1-) (>S-*#4-#* ,] .-# B41-"# A14*3# "#)
+@#/3-) ),.<4#-* ">.-# 1)X/0*43#: N1 )#.(# ),(.*3,- 4#)*1-*# )#413* "# )X-+@4,-3)#4 (#) @,4(,B#)
1.*1-* E.# A,))3=(# 1C#+ "#) A4,*,+,(#) +,//# ^$& C,34# ". _&`' #* #-).3*# ">#)*1/A3((#4 (#)
A1E.#*) P (>#-C,3 #* P (1 40+#A*3,-: %#4*13-) *41C1.K "1-) (1 (3**041*.4# ,-* 2,+1(3)0 ).4 (>0(3/3-15
*3,- "# *,.*# "3<04#-+# "1-) (#) 240E.#-+#) "#) @,4(,B#) E.3 A#.* +1.)#4 .- "0+1(1B# "#) @,4(,B#)
1C#+ (# *#/A) #:B: 6aM9:

!"#$%&' () *+,-./0%$.1 20 2.3"$1' 2' /" 3,%&./.4$'

!"#$! %# &'#( %! )!$&!"

!" #$%#&% '& ($&) '" *"+(", '" *$-&"(, %" %./0 '1&/ #2"34/ $ $&,,4 4/(5+",,5 %", #2"+#2"&+, ,&+6
(.&( %", '57"%.**"&+, '", $**%4#$(4./, 3&%(435'4$ -&4 ./( 8 .*(434,"+ %$ -&$/(4(5 '" +"'./'$/#"
9:;<= 8 $>.&("+ $&) *$-&"(, '$/, %" ?&( '" *.&7.4+ %", +"#./,(+&4+" ,$/, '"7.4+ %", +"(+$/,6
3"((+"@ A% ",( #.//& -&" %" ($&) '" +"'./'$/#" "( %$ 3$/4B+" './( "%%" ",( #.'5" ,./( C.+("3"/(
'5*"/'$/(, '" #.33"/( %", *"+(", ," *+.'&4,"/( D

E@ !" ($&) '" *"+(", ",( ,.&7"/( #$%#&%5 ,&+

&/" #"+($4/" C"/F(+" *.&+ 8 %$ C.4, './/"+ &/" 7$%"&+ ,40/4G#$(47" "( *.&+ -&" #"((" 7$%"&+ *&4,,"
$'$*("+ 8 (.&( #2$/0"3"/( '$/, %15($( '& +5,"$&@ !" #2.4) '" #"((" C"/F(+" ",( ,.&7"/( &/ #.36
*+.34, "/(+" *+5#4,4./ "( +5$#(474(5 DHIE@ ;/ *%&, '& ($&) 3.J"/ '" *"+(",K %" *+.G%" '", *"+(",
'" *$-&"(, ",( $&,,4 '1&/" #"+($4/" 43*.+($/#"@ !" C$4( -&" %", *$-&"(, ," *"+'"/( ,.&7"/( "/
+$C$%", '$/, %1A/("+/"( D

E $ 3.(475 %" '57"%.**"3"/( '", 35#$/4,3", :;< ,*5#4G-&", DLK

E@

tel-00408745, version 1 - 3 Aug 2009

!1$/$%J," '& *+.G%" '" *"+(", *$,,"/( ,.&7"/( *$+ &/" #.%%"#(" '" ,($(4,(4-&", ,&+ %", *$-&"(,
$7"# %", 4'"/(4G$/(, '" #"&) +"M&, "( '" #"&) *"+'&,K "( "/,&4(" *$+ %" #$%4?+$0" '1&/ 3.'B%"
,($(4,(4-&" 8 #" *+.G%"@ N/ ")"3*%" '1.&(4% 8 &(4%4,"+ ",( %$ #2$O/" '" P$+Q.7 8 '"&) .& *%&,4"&+,
5($(, #.33" /.&, %1$7./, C$4( '$/, DRE@

!"#$! %! *' +',%! )'""',&!

<1",( %$ 35(+4-&" '" #2"34/ -&4 $ ,&,#4(5 %" *%&, '" +"#2"+#2" *$+#" -&" G/$%"3"/( %$ ?$/'"
*$,,$/(" ",( #" -&4 4/(5+",," %" *%&, %1&(4%4,$("&+ "( %" C.&+/4,,"&+ '" ,"+74#"@ S& '5?&(K %$ ?$/'"
*$,,$/(" '1&/ #2"34/ $ 5(5 '5G/4" #.33" 5($/( %" '5?4( *2J,4-&" '& %4"/ %" *%&, %"/( %" %./0
'1&/ #2"34/ '" %1A/("+/"(@ ;( #1",( '$/, #"((" *"+,*"#(47" -&" '", .&(4%, #.33"

#"%!5!"& DITE "(

#"56'%7#"$& DUUE ./( 5(5 '57"%.**5,@ !" *+"34"+ &(4%4," %" '5%$4 *.&+ 4/C5+"+ %$ ?$/'" *$,,$/(" "(
%" '"&)4B3" &(4%4," %1",*$#"3"/( "/(+" *$-&"(, *.&+ %" 3F3" .?>"#(4C@ :4/$%"3"/(K ,"&% %" '5%$4
*"&( F(+" 3",&+5 '" ?.&( "/ ?.&( "( (.&(" %14/("%%40"/#" +"74"/( 8 (+.&7"+ %" %4"/ "/(+" %$ ?$/'"
*$,,$/(" '$/, %" #."&+ '& +5,"$& "( %" '5%$4 '", *$-&"(,K $*+B, 4% ,1$04( '1&/ ,43*%" *+.?%B3"
'14/7"+,4./ "( G%(+$0" *.&+ (+.&7"+ %$ 7$%"&+ %$ *%&, *+.?$?%" *.&+ %$ ?$/'" *$,,$/("@
!" '5?4( '& %4"/ %" *%&, %"/( 9$**"%5 ,.&7"/( %" 0.&%.( '15(+$/0%"3"/(= ",( &/" 35(+4-&"
43*.+($/(" *$+#" -&1"%%" './/" &/" 4'5" ,&+ %$ ?.+/" ,&*5+4"&+" -&" %" '5?4( *"&( $(("4/'+"@

#"%!5!"& ",( 3F3" #$*$?%" '" %.#$%4,"+ %1"3*%$#"3"/( '" #" %4"/@ <"*"/'$/( %" #2"34/ "3*+&/(5

*$+ %", *$-&"(, ",( (+B, *+.?$?%"3"/( #2$+05 *$+ %" (+$G# '1$&(+", &(4%4,$("&+ "( './# %$ ?$/'"
*$,,$/(" -&1&/" $**%4#$(4./ *"&( +5$%4,"+ *"&( F(+" ?4"/ 4/C5+4"&+" 8 #"%%" '& '5?4( '& %4"/ %" *%&,
%"/(@ !$ /.(4./ '" /"

8"12' #"99"1%' 2$9#.1$8/' $ $%.+, 5(5 4/(+.'&4(" "( $ 5(5 '5G/4" #.33" 5($/(

%$ *$+(4" '" %$ ?$/'" *$,,$/(" /./ &(4%4,5" %" %./0 '1&/ #2"34/ 8 &/ 4/,($/( './/5@ <1",( %" '5?4(
3$)43&3 -&1&/" $**%4#$(4./ *"&( +5$%4,"+ %" %./0 '" #" #2"34/ 8 #"( 4/,($/(K "/ /" #./,4'5+$/(
*$, ?4"/ ,V+ %" #$, '1&/" $**%4#$(4./ $0+",,47" -&4 ?.3?$+'" %" +5,"$& $7"# '& (+$G# NWX "( -&4
*.&,," %" +",(" '& (+$G# Y<X 8 +"#&%"+@ S %$ '4Z5+"/#" '& '5?4( '& 0.&%.( '15(+$/0%"3"/(K %$
?$/'" *$,,$/(" '4,*./4?%" ",( &/" -&$/(4(5 -&4 7$+4" $7"# %" ("3*,@

;/ *%&,K %$ ?$/'" *$,,$/("

!"#$ %! &'%()$*$(# +$ ,-!#$. /$ )-!/-# +01#,%()/$*$(# 2$!# %!33" 4%,"$, *%"3 &$&" %,,"4$ 5 !($ 6,17!$(&$

!"! #$%&'()*+', -. (/ 0%*1'('2+. /3*+&.

!"#

" # $%&'%( )*+(,-./'010.( 2345 0( 6-.)0 7-88-.(0 )987%.96'0 2: 8&, 3 5

)987%.96'0 70&( ;(,0 917%8+0 7-, &. -&(,0 '90. <&0 =0'&9 )& /%&'%( )*+(,-./'010.( =%110 =*08(

tel-00408745, version 1 - 3 Aug 2009

'0 =-8 7-, 0>017'0 )-.8 ?9/&,0 " "

@. /,-.) .%16,0 )*%&(9'8 )0 108&,0 )0 6-.)0 7-88-.(0 )987%.96'0 - +(+ 7,%7%8+ )-.8 ''9((+,-(&,0 0"/" A! B C!B CDB D B 4EFG" 308 %&(9'8 -77-,(90..0.( H )0&> ='-8808" I- 7,019J,0 ='-880
08( =0''0 <&9 =%.898(0 H 08(910, '*9.(0.89(+ )& (,-K= ,+80-&B 0( =%..-988-.( '0 )+69( )& /%&'%(
)*+(,-./'010.( 2%& '0 )+69( 7LM89<&0 )& '90.5B 0''0 =-'=&'0 '- 6-.)0 7-88-.(0 )987%.96'0 =%110
+(-.( '- )9N+,0.=0 0.(,0 '0 )+69( 7LM89<&0 0( '*9.(0.89(+ )& (,-K=" I0 )+69( 7LM89<&0 )& '90. 70&(
;(,0 %6(0.& 7-, &.0 1+(L%)0 H 6-80 )0 7-=O0(P7-9, =%110 0>7'9<&+ =9P)088&8" I0 =-'=&' )& (,-K=
,+80-& )08 -&(,08 &(9'98-(0&,8 80 Q-9( 8%&R0.( 7-, '*0.R%9 )08 7-9,8 )0 7-<&0(8 7,%=L08 '08 &.8
)08 -&(,08 8-.8 <&*9'8 8%90.( =%''+8B 0( 0.8&9(0 7-, '0 =-'=&' )0 '*087-=010.( )0 =08 7-<&0(8 H
'- 8%,(90 )& ,+80-&B &. 087-=010.( <&9 .0 70&( ,+8&'(0, <&0 )08 7-<&0(8 )08 -&(,08 &(9'98-(0&,8
<&9 R90..0.( 8*9.8+,0, 0.(,0 0&>" :R0= )08 (0=L.9<&08 )0 K'(,-/0 )0 6,&9(B '0 )+69( 1%M0. )&
(,-K= 70&( -'%,8 ;(,0 =-'=&'+" S08 %&(9'8 =%110 AF B C!B 4EFG 0.(,0.( )-.8 =0((0 =-(+/%,90" @. )08
-R-.(-/08 )0 =0((0 -77,%=L0 08( <&*0''0 .0 )01-.)0 7-8 '*0.R%9 )0 60-&=%&7 )0 7-<&0(8 0( )%.=
7-, =%.8+<&0.=0 0''0 .0 70,(&,60 7-8 '0 ,+80-&B 8%. 7,%6'J10 =070.)-.( 08( <&*0''0 .+=0889(0 '=%..-988-.=0 )& )+69( 7LM89<&0 )& '90. '919(-.( '- 6-.)0 7-88-.(0 )987%.96'0"
I*-&(,0 ='-880 )*%&(9'8 08( 6-8+0 8&, '0 7,9.=970 )*0.R%9 )0 (,-9. )0 7-<&0(8 H &. )+69( +'0R+
<&9 =,+0 &.0 =%./08(9%. 1%10.(-.+ )-.8 '0 ,+80-&B 0( 0.8&9(0 )0 108&,0, '0 )+69( )0 8%,(90 )0
=08 7-<&0(8" I0 7,0190, (,-R-9' )-.8 =0((0 )9,0=(9%. - +(+ 341'5. A! G <&9 - Q-9( '*LM7%(LJ80 <&0 '0
)+69( )0 8%,(90 )08 7-<&0(8 )& ,+80-& 08( =0'&9 )0 '- 6-.)0 7-88-.(0 )987%.96'0" :&(,010.( )9(B
8&, &. =L019. <&9 - &.0 6-.)0 7-88-.(0 AB (%&( (,-9. )0 7-<&0(8 0.R%M+ H &. )+69( 7'&8 /,-.)
<&0 A <&9((0 '0 ,+80-& H &. )+69( +/-' H A -R0= 690. +R9)010.( &. =0,(-9. 6,&9( H 0.'0R0," T9 '0
)+69( )08 7-<&0(8 8%.)08 08( 9.Q+,90&, H AB 9'8 <&9((0,%.( '0 ,+80-& -& 1;10 )+69( )*0.(,+0" I0
7,%6'J10 -R0= &. (0' %&(9' 08( <&*9' 8&77%80 <&0 '0 (,-K= ,+80-& .0 ,0=&'0 7-8 Q-=0 -&> 7-<&0(8
8%.)08" 30=9 - &. 80.8 <&-.) '08 7-<&0(8 8%.)08 8%.( 0.R%M+8 H &. )+69( '+/J,010.( 8&7+,90&,
H '- 6-.)0 7-88-.(0 )987%.96'0B 1-98 <&-.) 9'8 8%.( 0.R%M+8 H &. )+69( (,J8 +'0R+B 9'8 R%.( K.9,
7-, 1%.%7%'980, '08 '90.8 )& ,+80-& 0( 7-880.( )0R-.( '0 ,08(0 )& (,-K= =0 <&9 Q-9( <&*9'8 70&R0.(
!"#$%#& '%()* ()+,-(!#-! .#! /" ")0! &"**")1! 0(*&%)( /!2

!"#$%&' () *+,-./0%$.1 20 2.3"$1' 2' /" 3,%&./.4$'

"#$%%&' (& ')*&+# , #- .)/$% /&+#01#2 2(#* 3'+-. "#& A &- 41-0%$1- .& (&#' +3'&**$5$%)6 7&0$ 5$&-%
.# 4+$% "#& (& ')*&+# %'+$%& %1#* (&* 2+"#&%* .& (+ 898& 8+-$:'& &% -;$*1(& 2+* (& %'+<0 ')*&+#
.&* 2+"#&%* *1-.&*6 =& 0+* ($8$%& 0;&*% (;&-51$ .&* 2+"#&%* *1-.&* 01(()* "#$ +''$5&-% 01(()* +#
31#(1% .;)%'+-3(&8&-%> "#$ +%%&-.&-% .+-* (& /#?&' @A@B C@$'*%DA- @$'*%DB#%E .# '1#%&#' (&* #-*
.&''$:'& (&* +#%'&*> &% "#$ *1'%&-% +2':* .# 31#(1% 01(()* &*2+0)* 2+' (& .)/$% 2FG*$"#& .# 31#(1%6
H+-* 0& 0+*> (& .)/$% 8&*#') &*% 0&(#$ .# .)/$% 2FG*$"#& .# ($&- &% -1- 2+* (+ /+-.& 2+**+-%&
.$*21-$/(&6
I#&("#&* +--)&* +2':*

5#&.6' (+ -1%$1- .& 8&*#'& .& /+-.& 2+**+-%& + )%) '&5#& 21#' %&-$'

0182%& .& (;+3'&**$5$%) .&* 2+"#&%* *1-.&* JKL> L M6 N#221*1-* "#& (& ($&- ($8$%+-% (+ /+-.&
2+**+-%& + #- .)/$% 2FG*$"#& C &% "#& (+ /+-.& 2+**+-%& .$*21-$/(& &*% A6 =& %'+<0 ')*&+# &*%
.1-0 &- 81G&--& )3+( , C A6 O'&-1-* R (& .)/$% .;)8$**$1- .&* 2+"#&%* *1-.&* &% )%#.$1-*

tel-00408745, version 1 - 3 Aug 2009

0188&-% (& .)/$% .& ')0&2%$1- .&* 2+"#&%* *1-.&* 5+'$& +5&0 R6 7188& 1- (& 51$% .+-* @$3#'& 6P>
%+-% "#& (& .)/$% R -& .)2+**& 2+* A> (& ')*&+# -;&*% 2+* 01-3&*%$1--) &% (&* 2+"#&%* *1-.&*
"#$%%&-% &- 2'$-0$2& +# 898& .)/$% .;)8$**$1-6 =&* 2+"#&%* *1-.&* 01-3&*%$1--& (& ')*&+# "#+-.
$(* *1-% &-51G)* , 2(#* "#& A6 7;&*% , 2+'%$' .& 0& 21$-% "#& (& ')*&+# .&5$&-% $-0+2+/(& .& *&'5$'
(&* 2+"#&%* *1-.&* , (&#' .)/$% .;)8$**$1- &% *& %'1#5& 1/($3) .& (&* '+(&-%$'6 A( &*% 01--# "#;#'1#%&#' $82()8&-%+-% #-& 21($%$"#& @A@B '&2+'%$% (& %'+<0 , *+ *1'%$& 2'121'%$1--&((&8&-% +#
.)/$% .;&-%')&6 O+' 01-*)"#&-0& (&* 2+"#&%* *1-.&* *& %'1#5&-% *&'5$* , #- .)/$% CR/(C A + R)
%+-% .$* "#& (& '&*%& .# %'+<0 ')*&+# *& %'1#5& *&'5$ , #- .)/$% C(C A)/C A + R) .& %&((&
*1'%& "#& (+ *188& .&* .&#Q .)/$%* &*% )3+(& , (+ 0+2+0$%) .# ($&- C 6 R- +#38&-%+-% R +# .&(,
.& A> (&* 2+"#&%* *1-.&* .&5$&--&-% .& 2(#* &- 2(#* +3'&**$4* &% 2+**&-% .& 2(#* &- 2(#* .&5+-%
(&* 2+"#&%* .# %'+<06 =&#' .)/$% .& *1'%$& 2+' 01-*)"#&-0& +#38&-%& *1#* ($-)+$'&8&-% S#*"#;,
+%%&$-.'& C 21#' #- .)/$% R %':* 3'+-. C(& 0+* .& 2+"#&%* 01(()*E6 =& .)/$% .& ')0&2%$1- .&*
2+"#&%* *1-.&* &*% +(1'* )3+( , A .+-* (& *&#( 0+* 1T R &*% )3+( , A> 0& "#;#- 1#%$( .& 8&*#'&
0F&'0F& , 0+(0#(&' .;1T (+ .$U0#(%) .# 2'1/(:8&6
H;+2':* 0&%%& +22'10F&> (+ /+-.& 2+**+-%& .$*21-$/(& A 2&#% *& 8&*#'&' .& .&#Q 8+-$:'&*6
O'&8$:'&8&-% &- $-5&'*+-% (+ 01#'/& .+-* (+ @$3#'& 6P &636 JL M6 =+ .$U0#(%) ')*$.& .+-* (& 4+$%
"#& 0&%%& 01#'/& .&8+-.& (+ 01--+$**+-0& .& C 6 H+-* (& 0+* 1T (& ($&- 31#(1% .;)%'+-3(&8&-% &*%
(& 898& "#& 0&(#$ ($8$%+-% (+ /+-.& 2+**+-%&> 0&%%& $-41'8+%$1- 2&#% 9%'& 1/%&-#& +5&0 #- 0&'%+$&?1'% '+$*1--+/(&> 2+' 01-%'& "#+-. (&* .&#Q ($&-* *1-% .$?)'&-%*> (+ 01--+$**+-0& .& C &*% (1$.;9%'& )5$.&-%& 2#$*"#;&((& .&8+-.& (+ 01--+$**+-0& .& (+ /+-.& 2+**+-%& .$*21-$/(& 0& "#$ &*%
(+ "#+-%$%) , '&0F&'0F&'6 =;+#%'& *1(#%$1- 21#' 0+(0#(&' (+ /+-.& 2+**+-%& .$*21-$/(& *&'+$% .&
'&2)'&' (& 21$-% .& .)/#% .& (+ V1-& .& 01-3&*%$1- C21$-% 1T R &*% )3+( , AE6 7;&*% &- 0& 21$-% (,
"#& (& .)/$% .& ')0&2%$1- .&* *1-.&* &*% )3+( , 0&(#$ .& (+ /+-.& 2+**+-%& .$*21-$/(&6 =+ .)%&0%$1.& 0& 21$-% 2&#% *& 4+$'& &- /+(+G+-% (;&*2+0& .& .)/$% .;)8$**$1- R &% &- %'+"#+-% (;)51(#%$1.# .)/$% .& ')0&2%$1- .&* *1-.&*6 =& 2'1/(:8& 0;&*% "#& , 0+#*& .# /'#$% ')*&+# &% .& (;&''&#' .&
8&*#'&> (+ .)%&0%$1- .# 0F+-3&8&-% .& 01821'%&8&-% .# ($-)+$'& +# *1#*D($-)+$'& 2&#% 8+-"#&'
.& 2')0$*$1-6 =+ 8)%F1.& (+ 2(#* &U0+0& &% "#$ + )%) +.12%)& 2+'

#"%!/."2 &*% (+ .)%&0%$1- .#

!"! #$%&'()*+', -. (/ 0%*1'('2+. /3*+&.

!"#

"# $ %&'() *+&,(--(./ *0- 12340)- -./*0- 0) '2/*0 12--2/)0 *(-1./('50

*&'4) *0 52 6./0 *0 7./80-)(./ 129 5+248,0/)2)(./ *4 *&52(" :/ 0;0)< *=- 340 R *&12--0 52 '2/*0

tel-00408745, version 1 - 3 Aug 2009

12--2/)0 *(-1./('50< 50 '4;09 *2/- 50 9.4)049 7.,,0/70 > -0 90,15(9 0) 52 ?50 *+2))0/)0 > -0
@.9,09 70 34( 19.A.340 4/0 248,0/)2)(./ 7./)(/40 *4 *&52( *0- 12340)- *0 19.'0" B5 -4C) 25.9*0 *&)07)09 70))0 248,0/)2)(./ 1.49 7./75490 340 /.4- /.4- )9.4A./- > *9.()0 *4 1.(/) *&-(9&
D6./0 *0 7./80-)(./E" F 8247G0 *4 1.(/) *&-(9& 70))0 248,0/)2)(./ *4 *&52( *0- 12340)- -./*0/0 *.() 12- 2A.(9 5(04" FA07 4/0 )0550 ,&)G.*0 1.49 *&)07)09 -( /.4- /.4- )9.4A./- > *9.()0 .4
> 8247G0 *4 1.(/) *&-(9& 0) 4/0 ,&)G.*0 *(7G.).,(340 1.49 5+2*21)2)(./ *4 *&'() R< (5 *0A(0/)
@27(50 *0 5.725(-09 921(*0,0/) 0) 19&7(-&,0/) 50 1.(/) *&'4) *0 7./80-)(./" H/0 -(,150 ,0-490 *4
*&'() *+&,(--(./ .4 *0 9&701)(./ 0/ 70 1.(/) D50- *04I &824IE 0-) -4C-2/) 1.49 *.//09 52 A25049
*0 52 '2/*0 12--2/)0 *(-1./('50"
%2/- /.)90 )92A2(5 -49 52 '2/*0 12--2/)0 50- *09/(=90- 2//&0- /.4- /+2A./- 12- 7G097G& >
)9.4A09 *0- ,0(55049- .4)(5- ,2(- 154)J) > ,(04I (/)0919&)09 50- 9&-45)2)- .')0/4- 129 *0- .4)(50I(-)2/)-" F 129)(9 *0- ,0-490- *0 '2/*0 12--2/)0 0;07)4&0- 0/)90 4/ 709)2(/ /.,'90 *0 ,27G(/0-<
/.4- 2A./- A.454 7.,190/*90 -( 4/0 7.99&52)(./ 45/*+/(. 0I(-)0 0/)90 50- *(;&90/)0- ,0-490- 0)
-+(5 K 2 ,.K0/ *0 )92709 4/0 729)0 *0 '2/*0 12--2/)0 *0 5+B/)09/0) 34( 109,0) *+(/@&909 52 '2/*0
12--2/)0 0/)90 *04I ,27G(/0- > ).4) (/-)2/) -2/- *0A.(9 52 ,0-4909 *(907)0,0/) LMN< MMO" P.42A./- A.454 A&9(?09 -( 52 7.99&52)(./ -12)(250 34( 0I(-)0 1.49 4/0 ,&)9(340 7.,,0 50 *&52( 0I(-)0
24--( 1.49 52 '2/*0 12--2/)0" Q25G04904-0,0/)< 707( -+0-) 2A&9& *(C7(50 > 2))0(/*90< > ,.(/34+4/ 892/* /.,'90 *0 19.I(0- *0 ,0-490 *0 52 '2/*0 12--2/)0 0-) *&15.K& *2/- 50 9&-024 LMN< MMO"
P.- 0I1&9(,0/)2)(./- -49 R52/0)S2' LTUO ./) ,./)9& 340 1.49 (/@&909 52 '2/*0 12--2/)0 0/)90
*04I ,27G(/0-< (5 @24) 192)(340,0/) *&15.K09 4/ 19.IK > 5+0/)9&0 *4 9&-024 5.725 *0 7G274/0 *0(/)4()(./- 129)0/2(90- > R52/0)S2'" S2 ,0-490 *0 '2/*0 12--2/)0 0/)90 70- 19.I(0- 104) 25.9-09A(9 > 2119.I(,09 52 '2/*0 12--2/)0 0/)90 50- *04I ,27G(/0- 0/ 340-)(./" S0- 9&-45)2)- 1.49 50
*&52( -./) 7.,15=)0,0/) *(;&90/)- LN#O" S0 @2() 340 50 *&52( @.9,0 4/0 ,&)9(340 *2/- 4/ 0-1270
:475(*(0/ 109,0) 4/ 15./80,0/) 0C7270 *0- *&52(- *0- 7G0,(/- *0 5+B/)09/0) > 129)(9 *+4/ -.40/-0,'50 *0 ,0-490-" V0 15./80,0/) 104) 25.9- -09A(9 1.49 5+(/@&90/70 *0- *&52(- 0/)90 ,27G(/0129 50 -(,150 725745 *0 52 *(-)2/70 :475(*(0//0 0/)90 70- ,27G(/0- *2/- 5+0-1270 :475(*(0/" S0
15./80,0/) *0 52 '2/*0 12--2/)0 /&70--()0 192)(340,0/) ).4)0 52 ,2)9(70 *0 ,0-490- 0) 0-) 129

!"#$%&' () *+,-./0%$.1 20 2.3"$1' 2' /" 3,%&./.4$'

!
"#$%&'()$") *)( )+",")-

!"!

#$%&'(')*+ ,-./ &$0+1. ,-./+ 21/*3&+ 14%*5+

.$ *,/,0010) 2 0, ",/,"3&/4%,34#$ 56($ "7)84$ 5) 069$3)/$)3 "#$$)"3,$3 ($) *,4/) 5) 8,"74$)%:

($ ;/,$5 $#8</) 5) 3/,=,(> , ?#",04%& %(/ 0, "#//&0,34#$ 5)% 8)%(/)% ,"34=)% )@)"3(&)% *,/
($ )$%)8<0) 5) 8,"74$)% *#(/ 0, ",/,"3&/4%,34#$ 5( /&%),( 0)% "#$$)"3,$3- A)33) "#//&0,34#$ 5)%
#<%)/=,34#$% , ($ ;/,$5 ;,4$ *(4%'(6)00) *)/8)3 5) 34/)/ 5)% "#$"0(%4#$% %(/ 0, 3#*#0#;4) 5( /&%),(
)3 %(/ 0, 0#",04%,34#$ 5)% 4$"45)$3%- B,/ )>)8*0): ($) 8,"74$) '(4 *,/0) ,=)" ($ ;/,$5 $#8</) 5)
8,"74$)% )3 '(4 "#$%3,3) '() 0) 5&0,4 )%3 *0(% ;/,$5 '() 0, $#/8,0) =)/% 3#(3)% ")% 8,"74$)% *)(3
"#$"0(/) '() 0) */#<018) =4)$3 5) %#$ 04)$ 56,""1%- C) 8D8): ($) 8,"74$) '(4 %#$5) ($ ")/3,4$
$#8</) 5) 8,"74$)% )3 '(4 "#$%3,3) '() *#(/ 5)(> 5)% 8,"74$)% 0, "#//&0,34#$ 5)% *)/3)% 5)

tel-00408745, version 1 - 3 Aug 2009

*,'()3% )%3 *0(% ?#/3) '() *#(/ 0)% ,(3/)% *)(3 "#$"0(/) '() ")% 5)(> 8,"74$)% %) 3/#(=)$3 5,$%
0) 8D8) =#4%4$,;) 5,$% 0) /&%),(- C,$% 0, 0433&/,3(/): 0)% 3/,=,(> 5,$% ")33) 54/)"34#$ %#$3 %#(=)$3
*/&%)$3&% %#(% 0) 343/) E3#8#;/,*74) 5) 069$3)/$)3E #( <4)$ *,/?#4% E4$?&/)$") 5) 0, 3#*#0#;4) 5)
069$3)/$)3E-

!"!#$%"&'!$&() *+' ,&+)' *-.) "%'+!.

90 %6,;43 564$?&/)/ 0)% */#*/4&3&% 5)% 04)$% 5( /&%),( 2 *,/34/ 5) 8)%(/)% 5) <#(3 )$ <#(3 %,$%
?,4/) 4$3)/=)$4/ 0)% $#)(5% 4$3)/8&54,4/)%- B,/84 0)% */#*/4&3&% 2 4$?&/)/ %) 3/#(=)$3 0) 5&0,4 )3
0) 3,(> 5) *)/3)% 5) *,'()3%- A6)%3 ($ */#<018) '(4 $#/8,0)8)$3 $6, *,% 5) %#0(34#$ %(/ ($
%)(0 "7)84$- .$ )@)3: ") '(6#$ 8)%(/) 5) <#(3 )$ <#(3 %(/ ($ "7)84$ "6)%3 0, "#$=#0(34#$ 5)%
54%3/4<(34#$% 5) 0, 8&3/4'() F)-;- 5&0,4G %(/ 0)% 54@&/)$3% 04)$% 5( "7)84$: )3 0, 5&"#$=#0(34#$
5,$% ") ",% )%3 48*#%%4<0) "#88) 40 H , *0(%4)(/% 54%3/4<(34#$% 04&)% *,/ ($) %)(0) &'(,34#$- B,/
"#$3/) 51% '(6#$ *,/0) 5( /&%),(: 0, 5&"#$=#0(34#$ 5)=4)$3 *#%%4<0) ;/I") 2 0, "#//&0,34#$ '(4
)>4%3) )$3/) 0)% 8)%(/)%- J) 3/,=,40 0) *0(% "#$$( 5,$% ") 5#8,4$) )3 '(4 "7)/"7) 2 ",/,"3&/4%)/
0)% 04)$% 56($ /&%),( %6,**)00) K9LA FK(034",%3 9$?)/)$") #? L)3M#/N A7,/,"3)/4%34"%G O P- J,
*/)841/) =)/%4#$ 5) K9LA , =4%& 2 ",/,"3&/4%)/ 0) 5&0,4 )3 0) 3,(> 5) *)/3)% 5)% 04)$% 0) 0#$; 56($
,/</) 5) 54@(%4#$ 8(034",%3- J)% $#)(5% 5) 06,/</) %#$3 0)% /#(3)(/% 8(034",%3 #( 5)% */#>4)%
8(034",%3 ,( $4=),( ,**04",34?- Q *,/34/ 5)% /,**#/3% /)$=#H&% *,/ 0)% /&")*3)(/% %(/ 0)% *,'()3%
/)R(% )3 %(/ 0)(/ 5&0,4 5) /&")*34#$: 0, %#(/") )%3 ",*,<0) ,=)" K9LA 5) ",/,"3&/4%)/ 0)% 04)$%
5) 06,/</) 5) <#(3S)$S<#(3 %,$% '() 0)% $#)(5% 5) 06,/</) F/#(3)(/% #( */#>4)%G ,4)$3 <)%#4$ 5)
%) %#$5)/ )$3/) )(>- A)"4 , ($ ;/,$5 4$3&/D3 *(4%'(6($) %#(/") 5) 54@(%4#$ 8(034",%3 *)(3 ,0#/%
45)$34T)/ 5) <#(3S)$S<#(3 0)% 04)$% 5( /&%),( '(4 */&%)$3)$3 ($ */#<018) )3 /&,;4/ )$ ?#$"34#$
*,/ )>)8*0) )$ /&#/;,$4%,$3 06,/</) 5) 54@(%4#$ #( <4)$ )$ %&*,/,$3 0)% /&")*3)(/% )$ *0(%4)(/%
;/#(*)% )3 )$ )$=#H,$3 5( "#$3)$( 54@&/)$3 2 "7,"($ 5)% ;/#(*)%B/)$#$% 06)>)8*0) %(4=,$3 *#(/ "#8*/)$5/) 0, */#<0&8,34'() 5) K9LA )3 064$3&/D3 5) 0,
"#//&0,34#$ 5)% #<%)/=,34#$%- C,$% 0, U4;(/) -V: ($) %#(/") 54@(%) )$ 8(034",%3 2 3/,=)/% ($
$#)(5 4$3)/8&54,4/) F/#(3)(/ #( */#>HG- p: p1 )3 p2 %#$3 /)%*)"34=)8)$3 0) 3,(> 5) *)/3)% 5)%

!"! #$%&'()*+', -. (/ 0%*1'('2+. /3*+&.

!"#

"# $ %&'()*' +,()*' (-./01./ *2,./304/ 5' *1 6-003*1/,-. 5'+ -7+'081/,-.+ *-0+ 5' *1 (3/0-9
*-:,' 52;. 03+'1; )10 *21))0-6<' 16/,8'
)1=;'/+ +;0 *' *,'. 6-((;.> *' *,'. )0-)0' 1; 036')/';0 ? '/ *' *,'. )0-)0' 1; 036')/';0 " @'
/1;& 5' )'0/'+ ('+;03 )10 036')/';0 ? +';* '+/ 3:1* A 1 (1 p)(1 p1 ) '/ )10 036')/';0

tel-00408745, version 1 - 3 Aug 2009

+';* '+/ 3:1* A 1 (1 p)(1 p2 )" B10 6-.+3=;'.6' *'+ -7+'081/,-.+ ,.5,8,5;'**'+ .2-./ 1;6;.'
6<1.6' 5' 61*6;*'0 *'+ /1;& 5' )'0/'+ 5'+ /0-,+ *,'.+ +3)103('./" C* (1.=;' ;.' 3=;1/,-. )-;0
)-;8-,0 ,.8'0+'0 *' )0-7*D('" E2'+/ *A =;' *1 6-003*1/,-. 5'+ -7+'081/,-.+ 5'8,'./ ,./30'++1./'"
F. )1=;'/ 7,'. 0'G; '. ? '/ )'05; '.

H-; ,.8'0+'('./I '&6*;/ 1;/-(1/,=;'('./ *1 )'0/' 5;

)1=;'/ +;0 *' *,'. '. 6-((;.> 6' =;, '+/ ;.' ,.J-0(1/,-. )036,';+'" E'6, )';/ +' /015;,0' +-;+
*1 J-0(' 52;. .-;8'1; /1;& 5' )'0/'+ =;, +21))'**' *' /1;& 5' )'0/'+ K-,./ H)1=;'/ )'05; )10 *'+
5';& 036')/';0+ '. (4(' /'()+I '/ =;, 81;/ L p + (1 p)p1 p2 " M-;+ 18-.+ 1*-0+ /0-,+ 3=;1/,-.+
=;' .-;+ )-;8-.+ J16,*'('./ ,.8'0+'0 )-;0 61*6;*'0 *'+ /0-,+ =;1./,/3+ (1.=;1./'+"
@' )0,.6,)' 5' NCME )';/ J16,*'('./ +23/'.50' A 5'+ 03+'1;& )*;+ 6-()*,=;3+" C* )';/ 1;++,
+21))*,=;'0 1; 61+ 5' *2,.J30'.6' 5' *1 5,+/0,7;/,-. 5; 53*1, +;0 6<16;. 5'+ *,'.+" E'0/1,.+ /0181;&
-./ '++1O3 5' *23/'.50' 1; 61+ ;.,61+/ '":" PQR> #ST" U; *,'; 52'.8-O'0 ;. )1=;'/ (;*/,61+/ =;, '+/
5;)*,=;3 )10 *' 0-;/';0 ,./'0(35,1,0'> 5';& )1=;'/+ 6-**3+ +-./ '.8-O3+ ;. 1; )0'(,'0 036')/';0
'/ ;. 1;/0' 1; 5';&,D(' 036')/';0" @' J1,/ =;2,*+ +-,'./ 6-**3+ 6'+ 5';& )1=;'/+ 8-./ 8,80' /0D+
)0-717*'('./ *1 (4(' '&)30,'.6' 1; .,8'1; 5; 0-;/';0 ,./'0(35,1,0'" %. 'V'/> +, *' 0-;/';0
'+/ 51.+ ;. 3/1/ 5' 6-.:'+/,-.> *1 )'0/' 5'+ 5';& )1=;'/+ '+/ /0D+ )0-717*'" C* O 1 /-;K-;0+ *1
)-++,7,*,/3 =;2;. 5'+ 5';& )1=;'/+ +' )'05 '/ )1+ *21;/0' (1,+ *1 )0-717,*,/3 5' 6'/ 383.'('./
'+/ 6-.+,5303' 6-((' )*;+ )'/,/' =;' *1 )0'(,D0'"

!"!#$%"&'!$&() *+ ,! $(-(,(.&+ */ "%'+!/

@2C./'0.'/ '+/ 6-.+/0;,/ 1;/-;0 5' =;'*=;'+ :01.5+ -)301/';0+ 5' /01.+,/ HW,'09?I =;, +2,./'09
6-..'6/'./ './0' ';& '/ =;, 1++;0'./ *1 6-..'6/,8,/3 5'+ 1;/0'+ -)301/';0+ 5' .,8'1;

H.1/,-.1*

-; 03:,-.1*I" E'+ 5'0.,'0+ A *';0 /-;0 -V0'./ *1 6-..'6/,8,/3 1;& './0')0,+'+ '/ J-;0.,++';0+ 5'
+'08,6'" @'+ -)301/';0+ 5' .2,()-0/' =;'* .,8'1; )';8'./ A *';0 /-;0 '/ A /-;/ (-('./ +,.:'0
5'+ 166-05+ './0' ';& )-;0 36<1.:'0 *';0 )0-)0' /01X6 +1.+ )1++'0 )10 *'+ -)301/';0+ 5' .,8'1;
+;)30,';0 H)1++1:' +-;8'./ )1O1./I 8-,0' -V0,0 *'+ ;.+ 1;& 1;/0'+ 5'+ +-*;/,-.+ 5' 716Y;) '.
61+ 5' )0-7*D('" E'6, 0'.5 /0D+ 6-()*,=;3 '/ /0D+ ,()038,+,7*' *1 /-)-*-:,' 5; 03+'1; :*-71* =;'

!"#$%&' () *+,-./0%$.1 20 2.3"$1' 2' /" 3,%&./.4$'

"#$%&''# '# (&')$*+# #) ,-.)$.%#/ 0#% &"1$-)#2$% (&''-.%%#') 3.#' +#2$% 4&,-.'#%5 6 72. .+% %&')
(&''#()1% #) 6 72. .+% &8$#') +# %#$9.(#5 ,-.% -2 4#+6 4# :- )&2) (# 72;.+% (&''-.%%#') %2$ +# $#%)#
4# +;<')#$'#) (# %&') +#% .'=&$,-).&'% >?@ 72. .'4.72#') +- +.%)# 4#% 4&,-.'#% 6 "-$(&2$.$ "&2$
-))#.'4$# 2' (#$)-.' "$1AB# $1%#-2/ C#))# +.%)# "#2) ,D,# D)$# .'(&,"+E)# 4-'% +# (-% &F 2'
&"1$-)#2$ .')#$,14.-.$# 41(.4# 4# '# "-% -''&'(#$ +#% '&,% 4#% 4&,-.'#% 72. +# %2.9#')/
C# ,-'72# 4;.'=&$,-).&'% %2$ +- )&"&+&G.# 42 $1%#-2 - %2%(.)1 +- (2$.&%.)1 4#% (H#$(H#2$%
72. &') #%%-I1 4# 419#+&""#$ 4#% %&+2).&'% "&2$ .'=1$#$ +- )&"&+&G.# 4# +;<')#$'#) 4-'% +# 32)
4# (&''-.)$# +- (&''#().9.)1 G+&3-+# #) 4# +&(-+.%#$ +#% -4$#%%#% <@ +#% 2'#% "-$ $-""&$) -2B
-2)$#%/ C#(. - 4.9#$%#% -""+.(-).&'% 72# (# %&.) "&2$ +#% 2).+.%-)#2$% &2 "&2$ +#% =&2$'.%%#2$% 4#%
%#$9.(#% &2 ,D,# "&2$ +#% &"1$-)#2$%/ J2#+72# %&.) +# K&2#2$5 +- (&''-.%%-'(# 4# +- )&"&+&G.# 4#
+;<')#$'#) "#2) -.4#$ 6 &").,.%#$ +#% -""+.(-).&'% #' (H&.%.%%-') ,.#2B "-$ #B#,"+# +# %#$9#2$ &2
+# "-.$ 6 (&')-()#$ &2 +# $1%#-2 6 (&''#()#$/ C#(. "#2) -2%%. %#$9.$ 6 4.-G'&%).72#$ +# $1%#-2 #)

tel-00408745, version 1 - 3 Aug 2009

6 (&,"$#'4$# +# (&,"&$)#,#') 4#% &"1$-)#2$%/ L.'-+#,#') +- (&''-.%%-'(# 4# +- )&"&+&G.# "#2)

%#$9.$ 6 419#+&""#$ 4#% ,&4E+#% )H1&$.72#% 72. "#$,#))#') 4# G1'1$#$ 4#% )&"&+&G.#% 4#% $1%#-2B
72. $#%%#,3+#') +# "+2% "&%%.3+# -2B )&"&+&G.#% $1#++#%/ C#(. #%) )$E% .,"&$)-') "&2$ +- "$1(.%.&'
4#% %.,2+-).&'% #) "&2$ +- 9-+.4-).&' 4#% "$&)&(&+#%/
M-'% +- %2.)# '&2% (&,,#')&'% +#% "$.'(."-2B )$-9-2B 4-'% (#))# 4.$#().&'/

!"!#$%"&'!$&() *+ ,"!-./ */ #())/#$&0&$%

C;#%) +;-""$&(H# +- "+2% (+-%%.72#/ N++# (&'%.%)#

6 )$-(#$ 2'# (-$)# "&2$ +;<')#$'#) &F +#% "&.')% %&') +#% $&2)#2$% &2 +#% 4&,-.'#% #) &F +#% -$$D)#%
42 G$-"H# %&') +#% +.#'% +#% .')#$(&''#()-')/ @&2$ +- (-$)# -2 '.9#-2 $&2)#2$%5 #++# "#2) D)$#
&3)#'2# "-$ 4#% )$-(#$&2)#% #'9&I1% "-$ 2' #'%#,3+# 4# ,-(H.'#% 9#$% 4;-2)$#% ,-(H.'#%5 #'%2.)#
"-$ 2'# $#"$1%#')-).&' 4#% (H#,.'% &3)#'2% #' 2' %#2+ G$-"H# #/G/ OPQ5 RS5 TP5 PUV5 PUSW/ @&2$
+- (-$)# -2 '.9#-2 4&,-.'#%5 #++# "#2) D)$# &3)#'2# 4# 4#2B ,-'.E$#% X #' -G$1G#-') +#% )-3+#%
4# $&2)-G# >?@ "$.%#% 4-'% 4.81$#')% $&2)#2$% 4# +;<')#$'#) OPUPW5 &2 #' "-$)-') 4;2'# (-$)#
-2 '.9#-2 $&2)#2$% #) #' =&$,-') +#% 4&,-.'#% 6 "-$).$ 4#% .'=&$,-).&'% %2$ +#% -%%&(.-).&'%
Y"$1AB# <@5 '&, 42 4&,-.'#Z #/G/ OQ[W/ \-.% 72#+72# %&.) +- '-)2$# 4# +- (-$)#5 4.9#$%#% 4.](2+)1%
+;-((&,"-G'#') (&,,# "-$ #B#,"+# +# '&,3$# 4# "&.')% 4# ,#%2$# Y)&2) #' %-(H-') 72;&' '#
"#2) "-% #' "$#'4$# -2)-') 72;&' #' 9#2)Z5 +- +&(-+.%-).&' 4# (#% "&.')% 4# ,#%2$#5 +;&$4$# "-$
+#72#+ +#% ,#%2$#% %&') 6 =-.$#5 +- $142().&' 4# +- $#4&'4-'(#5 +- 41)#().&' 4#% -+.-% Y(-% 4;2'
4&,-.'# &2 4;2' $&2)#2$ -9#( "+2%.#2$% '&,% #) .')#$=-(#%Z5 #)(/
^' 4#% $1%2+)-)% +#% "+2% .,"&$)-')% 4# (#% )$-9-2B - 1)1 +;&3%#$9-).&' =-.)# %2$ +- 4.%)$.32).&'
42 4#G$1 4# (&''#().9.)1 4;2' '&#24 72# (# %&.) $&2)#2$ &2 4&,-.'#/ N' #8#)5 K2%72;6 +;-''1#
P[[[5 (#))# 4.%)$.32).&' - )&2K&2$% 1)1 %2""&%1# %2.9$# 2'# +&. #B"&'#').#++# -2 ,&.'% 4-'% %72#2#/ N' P[[[5 2' )$-9-.+ ".&''.#$ "-$ +#% )$&.% =$E$#% L-+&2%)-% OVPW - (&'(+2 72# (#))# 4.%)$._
32).&' %2.) 2'# +&. 72#2#_+&2$4# #) "-$ (&'%172#'(# )&2% +#% ,&4E+#% 4# G$-"H#% 2).+.%1% 6 (#))#
1"&72# "&2$ +- ,&41+.%-).&' 4#% $1%#-2B (&,,# +# ,&4E+# N`5 4# a-B,-' &2 ?b_<b\5 &')
1)1 (&'%.41$1% .'%2]%-')% "&2$ $#c1)#$ (#))# "$&"$.1)1/ N' #8#)5 +- "$&"$.1)1 72#2#_+&2$4# 9#2)

!"! #$%&'()*+', -. (/ 0%*1'('2+. /3*+&.

"#$" %&'()*'*+" ,&-* .$* /*-"0&+% +#*$,% ,* )12+"*-+*" %#+" 3#-"*'*+" /#++*/"4% (0- -0((#-"
5 )0 '#6*++* /* .$* )*% '#,7)*% 5 .$*$* *8(#+*+"&*))* +10--&9* (0% 5 -*(-#,$&-*: ;*/& 0 4"4 5
)1#-&<&+* ,1$+ <-0+, +#'=-* ,* "-090$8 %$- )1*8()&/0"&#+ ,* )0 %#$-/* ,* /* (>4+#'7+* ?@AB *" %$)0 (-#(#%&"&#+ ,*% +#$9*0$8 '#,7)*% ">4#-&.$*% (#$- )0 "#(#)#<&* ,$ -4%*0$ /#''* (0- *8*'()*
)* '#,7)* 4'5.1 (/5 1/,-'0 21/46 ?AB *" )* '#,7)* 2.,.1/( (+,./1 41.7.1.,3. ? CB: D#$9*+"E )0
">4#-&* ,$ 7'1* -.&+.,* 4()8 7'1* *%" ,#++4* /#''* *8()&/0"&#+ 5 /* (>4+#'7+*:

!"#$%&%#' (% !)*#'%+#%' (,#- .# %-/,0% 12+'.%! (% 0""+("##3%-

F#/0)&%*- )*% '0/>&+*%

)*% $+*% (0- -0((#-" 0$8 0$"-*% 5 (0-"&- ,1$+* /0-"* +1*%" (0% )0 %#)$"&#+ )0 ()$% *G/0/*: H1$+ /I"4
&) *%" ,&G/&)* ,1#="*+&- /*""* /0-"* *" ,* )14"*+,-* 5 ,* +#$9*))*% 0,-*%%*% #$ ,* )0 '*""-* 5 J#$-E *"
,* )10$"-* /I"4 )0 /0-"* *%" ,1$+ <-0+, 9#)$'* .$1&) *%" ,&G/&)* ,* '0+&($)*-: K#$- /*)0 )*% *L#-"%

tel-00408745, version 1 - 3 Aug 2009

%* %#+" /#+/*+"-4% %$- )0 '#,4)&%0"&#+ ,* )12+"*-+*" (0- $+ *%(0/* ,* /##-,#++4*%E %#$9*+" $+

*%(0/* M$/)&,&*+ 5 ()$%&*$-% ,&'*+%&#+%E *" )0 '#,4)&%0"&#+ ,*% '0/>&+*% *" ,*% -#$"*$-% (0- ,*%
(#&+"% ,0+% /*" *%(0/*: F*% /##-,#++4*% ,*% '0/>&+*% %#+" />#&%&*% ,* "*))* %#-"* .$* )0 ,&%"0+/*
<4#'4"-&.$* *+"-* *))*% ,0+% )1*%(0/* *%" $+* 0((-#8&'0"&#+ ,$ ,4)0& -4%*0$ *+"-* )*% '0/>&+*%:
H* /*""* '0+&7-* &) %$G" ,* /#++0&"-* )*% /##-,#++4*% ,* ,*$8 '0/>&+*% (#$- (#$9#&- *%"&'*- )*
,4)0& *+"-* *))*% %0+% ,*9#&- 0))*- %$- $+* ,*% '0/>&+*% *" %#+,*- )10$"-*: F0 <-0+,* .$*%"&#+ *%"
%0+% ,#$"* %& /*""* &,4* *%" 30&%0=)* *" %& #+ (*$" #="*+&- )*% /##-,#++4*% ,1$+* '0+&7-* %&'()* *"
*G/0/*: ;#''* )* ,4)0& %0"&%30&" *+ (-&+/&(* )*% (-#(-&4"4% ,1$+* '4"-&.$*E &) 6 0 3#-"* />0+/* .$*
/0 %#&" )* /0%: ;* (-#=)7'* *%" /#++$ ,0+% )0 )&""4-0"$-* /#''* 4"0+" )* (-#=)7'* ,* ()#+<*'*+"
,*% ,4)0&% ,* )12+"*-+*" #$ =&*+ )* (-#=)7'* ,* /0)/$) ,*% /##-,#++4*% 9&-"$*))*%:
K)$%&*$-% 0((-#/>*% *8&%"*+": F* (-*'&*- "-090&) 0 /#+%&%"4 *+ )* ,4()#&*'*+" ,1$+* &+3-0%"-$/N
"$-* ,* =0)&%*% 0((*)4*% (/,-0/198 ?OPB: F*% =0)&%*% '*%$-*+" )* ,4)0& *+"-* *))*% *" 3#-'$)* $+
(-#=)7'* ,1#("&'&%0"&#+ .$1*))*% -4%#)9*+" (#$- "-#$9*- )*% '*&))*$-*% /##-,#++4*% (#$- *))*% .$&
'&+&'&%*+" )1*--*$- .$0,-0"&.$* '#6*+ -4%$)"0+" ,$ ()#+<*'*+" ,* )*$-% ,4)0&%: M+%$&"*E *" 4"0+"
,#++4*% )*% /##-,#++4*% ,*% =0)&%*%E $+* '0/>&+* .$*)/#+.$* )*% %#+,* *" 3#-'$)* 5 %#+ "#$$+ (-#=)7'* ,1#("&'&%0"&#+ +#+N)&+40&-* (#$- "-#$9*- %*% '*&))*$-*% /##-,#++4*% .$& '&+&'&%*+"
)1*--*$- .$0,-0"&.$* '#6*+ -4%$)"0+" ,$ ()#+<*'*+" ,* %*% ,4)0&% ,0+% )1*%(0/*: ;*""* ,*$8&7'*
"Q/>* *%" -40)&%4* (0-0))7)*'*+" *" &+,4(*+,0''*+" (0- )*% '0/>&+*%: M+ (-#/4,0+" /#''* /*/&E
"#$"*% )*% '0/>&+*% -4$%%&%%*+" 5 09#&- )*$-% /##-,#++4*% *" (*$9*+" *+%$&"* )*% $"&)&%*- ,0+% )*%
0(()&/0"&#+%: R+* $"&)&%0"&#+ "6(&.$* %*-0&" (0- *8*'()* ,1*+9#6*- $+* -*.$S"* 5 $+ %*-9*$- HTD
*+ ,*'0+,0+" )1&+9*-%&#+ ,1$+* RUF: F* %*-9*$- HTD 0$ )&*$ ,1*+9#6*- $+* )&%"* ,10,-*%%*% 2KE
&) *L*/"$* $+* *%"&'0"&#+ ,$ ,4)0& *+"-* )0 '0/>&+* ,*'0+,*$%* *" )*% ,&L4-*+"*% 0,-*%%*% 2K ,*
)0 'S'* RUF *" -*+9#&* (0- *8*'()* )10,-*%%* 2K )0 ()$% (-#/>*:
R+ ,*% <-0+,% 090+"0<* ,*% 0((-#/>*% 5 =0%* ,* =0)&%*% /#''* ?OPE !CB *%" )*$- 0%(*/" +#+
&+"-$%&3: V )1*8/*("&#+ ,*% =0)&%*%E )*% '0/>&+*% +* %* 3#+" (0% %#+,*- (0- ,10$"-*% '0/>&+*%
,0+% )* =$" ,* /0)/$)*- )*% /##-,#++4*%: M+ ()$%E )*% "-090$8 #+" '#+"-4 .$1$+ (*"&" +#'=-* ,*
=0)&%*% *%" +4/*%%0&-* (#$- ()#+<*- )*% ,4)0&% ,* )12+"*-+*" 09*/ $+* =#++* (-4/&%&#+ /* .$& -4,$&"

!"#$%&' () *+,-./0%$.1 20 2.3"$1' 2' /" 3,%&./.4$'

"#$%&'()*+,-.-$/ ,* "0*)1- "#.2*)(- 3 4$- *22)#"0- '- /52- "*)*"/()&%*/&#$ /#/*,- '- ,* "*)/'4 )(%-*46 7-2-$'*$/8 ,* 2)("&%&#$ '4 2,#$1-.-$/ 2-4/ %#49)&) '- "-)/*&$-% &))(14,*)&/(% '*$% ,)#4/*1- :$/-)$-/ ;4& $- 2-4<-$/ 2*% =/)- "*2/4)(-% '*$% 4$ -%2*"- '- "##)'#$$(-%6 >*) -?-.2,-8
4$- '-% 2)#2)&(/(% '@4$- '&%/*$"- '*$% 4$ -%2*"- -%/ ;4- ,* %#..- '-% '&%/*$"-% -$/)- $#-4'
A -/ $#-4' B -/ $#-4' B -/ $#-4' 7 '#&/ =/)- %42()&-4)- 3 ,* '&%/*$"- -$/)- A -/ 76 C*$%
,@:$/-)$-/ .*,0-4)-4%-.-$/ $#4% 2#4<#$% /)#4<-) '-% "*% #D "-//- )E1,- $@-%/ 2*% %*/&%F*&/- -/ #D
,- '(,*& AG7 -%/ %42()&-4) 3 ,* %#..- '-% '(,*&% '- AGB -/ BG76 7@-%/ ,- F*.-4? 2)#+,E.- '<&#,*/&#$ '- ,@&$(1*,&/( /)&*$14,*&)- HI!8 JJ!K6 L* %#,4/&#$ 3 "- 2)#+,E.- -%/ ,#&$ '@=/)- (<&'-$/-6
M$- 2#%%&+&,&/( %-)*&/ '@&$/)#'4&)- '-% $#-4'% &$/-).('&*&)-% .#'(,&%*$/ ,-% )#4/-4)% .*&% "-//%#,4/&#$ )&%;4- '@*41.-$/-) "#$%&'()*+,-.-$/ ,* "#.2,-?&/( '- ,@#2()*/&#$6 N(*$.#&$%8 .=.-$ ,@*+%-$"- '- "-//- %#,4/&#$ 3 ,@&$(1*,&/( /)&*$14,*&)-8 ,-% %5%/E.-% '- "##)'#$$(-% )-%/-$/ 4$&'(- &$/()-%%*$/- 24&%;4- 3 4$ "#O/ )-,*/&<-.-$/ F*&+,- &,% 2-4<-$/ '#$$-) 4$- +#$$- &'(- %4) ,*

tel-00408745, version 1 - 3 Aug 2009

/#2#,#1&- '4 )(%-*4 -/ %4) ,* 2)#?&.&/( '-% .*"0&$-%6

L-% "##)'#$$(-% 2-4<-$/ *4%%& %- "*,"4,-) 2*) 4$- *22)#"0- "#.2,E/-.-$/ '&%/)&+4(- %*$%
,- +-%#&$ '-% +*,&%-%6 7-//- &'(- * (/( &$/)#'4&/- 2*) P&<*,'& H IK8 4$ %5%/E.- '- "##)'#$$(-%
'&%/)&+4( '5$*.&;4- /)E% &$/()-%%*$/ ;4& /)#4<- %#$ "0-.&$ *4 '(2,#&-.-$/ '*$% '&<-)%-% *22,&G
"*/&#$% "#..- ,- ",&-$/ B&/Q#))-$/ '- AR4)-4% #4 ,* 2,*/-F#).- S$-L*+8 ,* <-)%&#$ -4)#2(-$$'- >,*$-/L*+6 A4 ,&-4 '- "*,"4,-) ,-% "##)'#$$(-% 3 2*)/&) '-% .-%4)-% '- '(,*& <-)% ,-% +*G
,&%-%8 P&<*,'& 4/&,&%- '-% .-%4)-% '- '(,*& -$/)- ,-% .*"0&$-% -,,-% .=.-% 2#4) ,@*T4%/-.-$/ '-%
"##)'#$$(-%6 70*;4- .*"0&$- "#..-$"- 2*) 4$- %422#%&/&#$ %4) ,* <*,-4) '- %-% "##)'#$$(-%8
-$%4&/- 3 "0*;4- .-%4)- '- '(,*& ;4@-,,- -9-"/4- <-)% 4$- *4/)- .*"0&$-8 -,,- "#.2*)- ,* '&%/*$"U4",&'&-$$- *4 '(,*& .-%4)(6 V& ,* '&%/*$"- -%/ 2,4% 1)*$'- ;4- ,- '(,*& .-%4)(8 -,,- %@*22)#"0'- ,@*4/)- .*"0&$-6 C*$% ,- "*% "#$/)*&)- -,,- %@(,#&1$- '@-,,-8 -/ 3 (1*,&/( -,,- )-%/- 3 %* 2,*"-6
L@*4/)- .*"0&$- -9-"/4- ,* .=.- #2()*/&#$ '- %#$ "W/(6 U/ "-//- #2()*/&#$ -%/ )(2(/(- <-)%
'@*4/)-% .*"0&$-%6 >#4) /)#4<-) ,- .-&,,-4) *T4%/-.-$/ '-% "##)'#$$(-% -/ 2)#4<-) ,* "#$<-)G
1-$"- '4 %5%/E.-8 P&<*,'& * F*&/ ,@*$*,#1&- -$/)- 4$ %5%/E.- '- "##)'#$$(-% -/ 4$ %5%/E.- ')-%%#)/% #D ,-% .*"0&$-% %#$/ ,-% 2#&$/% '@*//*"0-.-$/ '-% )-%%#)/% -/ #D ,-% "0-.&$% -$/)- .*G
"0&$-% "#))-%2#$'-$/ *4? )-%%#)/% .=.-%6 L* ,#$14-4) '@4$ )-%%#)/ *4 )-2#% )-2)(%-$/- ,- '(,*&
)(%-*4 '4 "0-.&$6 A"")#"0(% '*$% ,@-%2*"- *4? 2#&$/% ;4& )-2)(%-$/-$/ ,-% .*"0&$-% -/ ,X"0(%8 &,
-%/ "#$$4 ;4- ,-% )-%%#)/% <#$/ #%"&,,-) '*$% /#4/-% ,-% '&)-"/&#$% T4%;4@3 "#$<-)1-) <-)% 4$ )(1&.%/*/&#$$*&)- ;4& .&$&.&%- ,@($-)1&- /#/*,-6 L@($-)1&- /#/*,- $@-%/ *4/)- ;4- ,* %#..- '- ,@-))-4)
;4*')*/&;4- -$/)- '(,*&% -/ '&%/*$"-% U4",&'&-$$-%6 >*) "#$%(;4-$"-8 &, %4Y/ ;4- ,-% .*"0&$-%
*T4%/-$/ ,-4)% "##)'#$$(-% 3 ,* .*$&E)- ,-% )-%%#)/% ,- F#$/ 2#4) "#$<-)1-) <-)% 4$ (/*/ #2/&.*, -$
/-).-% '@-))-4) ;4*')*/&;4-6 C*$% P&<*,'&8 4$- .*"0&$- *<*$"- #4 %@(,#&1$- '@4$- *4/)- .*"0&$*<-" 4$ 2*% 2)#2#)/&#$$-, 3 ,* '&9()-$"- -$/)- '(,*& .-%4)( -/ '&%/*$"- "*,"4,(-6
P&<*,'& -%/ 4$ %5%/E.- /)E% &$/()-%%*$/ ;4& * /#4/-% ,-% "0*$"-% '@=/)- '(2,#5( 4$ T#4) "#..%-)<&"- 4$&<-)%-,6 L-% .-%4)-% '- '(,*& 2-4<-$/ %- F*&)- 0#)%G+*$'- '*$% 4$ S<-),*5 '('&(8 #4 +&-$
'*$% ,* +*$'- -$ -?2,#&/*$/ ,-% "#..4$&"*/&#$% -$ "#4)%6 N#4% '#$$#$% +-*4"#42 '@&$/()=/ 3 "-

!"! #$%&'()*$%)

"#"$%&' ()*" *+$,' ,'-.',-.'/ 0)*" 1* ,2-'*$ $,)3)45 67 89 *+1" )3+*" 2$1(42 5) "2-1,4$2 (1 "#"$%&'
:), 5;23)51)$4+* ('" (2<=$" >1' ('" &)-.4*'" &)53'455)*$'" :'13'*$ -)1"', '* )**+*?)*$ (' @)1""'"
-++,(+**2'" +1 A4'* '* *' :)" )B1"$)*$ 5'1," -++,(+**2'" -+,,'-$'&'*$/ C+1" )3+*" :,+:+"2 1*'
"+51$4+* D A)"' (' E1,3'#+," F&)-.4*'" (' -+*G)*-'H :+1, 32,4G', 5' A+* @+*-$4+**'&'*$ (1
"#"$%&'/ 0)*" 1* )1$,' $,)3)45 '* -+1," *+1" 2$1(4+*" -+&&'*$ 1* "#"$%&' -+&&' I43)5(4 :'1$
"',34, ()*" 5) (2$'-$4+* (' $+1$ -.)*<'&'*$ $+:+5+<4>1' ()*" 5' ,2"')1 '$ ()*" 5;4('*$4G-)$4+*
('" ,)4"+*" (' -' -.)*<'&'*$/

!" #$%&'()*$%)
C+1" )3+*" 3+151 (+**2 ()*" -' -.):4$,' 1* ):',?1 <2*2,)5 (' 5;23+51$4+* (1 (+&)4*' ('
&2$,+5+<4' )1 -+1," ('" (',*4%,'" )**2'"/ J;'"$ 1* (+&)4*' >14 ) )$$4,2 A')1-+1: (' ,'-.',-.'"

tel-00408745, version 1 - 3 Aug 2009

'$ 3) -+*$4*1', D '* )$$4,', "1,$+1$ )3'- 5) :,+54@2,)$4+* ('" )$$)>1'" '$ (1 $,)G- 4*(2"4,)A5'
'$ 5;)1<&'*$)$4+* (' 5) $)455' '$ 5;.2$2,+<2*24$2 (1 ,2"')1/ K' <,)*( (2G :+1, 5) -+&&1*)1$2
"-4'*$4G>1' 3) -',$)4*'&'*$ L$,' 5;1$454")$4+* ('" ,2"15$)$" '$ ('" +A"',3)$4+*" (2BD @)4$'" ()*" 5'
:,+-'""1" (' -+*"$,1-$4+* (;1* M*$',*'$ (' N1$1, :51" "O,9 :51" )():$)$4@9 '$ :51" 'P-)-' >1'
-'514 (' *+" B+1,"/ Q* <,)*( +AB'-$4@ :+1, 5) :,+-.)4*' :2,4+(' '"$ -',$)4*'&'*$ 5) -+*"$,1-$4+*
(;1*' ),-.4$'-$1,' (4"$,4A12' :',&'$$)*$ 5;+A"',3)$4+* (' 5;2$)$ (1 ,2"')1 '$ 5' (4)<*+"$4- (' "'"
:,+A5%&'"/ J'$$' ),-.4$'-$1,' :'1$ $,%" A4'* @)4,' :),$4' (' 5;M*$',*'$ (1 N1$1, '* @+1,*4"")*$ 5'"
4*@+,&)$4+*" *2-'"")4,'" )1 A+* @+*-$4+**'&'*$ (1 ,+1$)<'9 (1 $,)*":+,$ '$ ('" )::54-)$4+*"/
0)*" -' (+&)4*' *+1" )3+*" :,4*-4:)5'&'*$ -+*$,4A12 ()*" ('" "+51$4+*" 'P-)-'" >14 :)""'*$
D 5;2-.'55' '$ >14 :',&'$$'*$ 1*' &'455'1,' -+&:,2.'*"4+* ('" :',@+,&)*-'" ('" ,2"')1R/ K'" -.)S
:4$,'" "143)*$" (+**',+*$ 1*' 4(2' (' *+" -+*$,4A1$4+*" "1, 5;)::,+-.' :)""43'9 "1, 5' :,+A5%&'
(;2-.)*$455+**)<' '$ "1, 5' -+*$,T5' (' -+*<'"$4+* :+1, 5'" "+*('" )-$43'"/ U+1, 5'" )1$,'" -+*$,4S
A1$4+*" >1' *+1" *' -4$+*" :)" :), &)*>1' (;'":)-'9 '55'" "+*$ '* <,)*(' :),$4' :+,$2'" "1, 5)
&'"1,' ('" :',@+,&)*-'" '$ 5) &+(254")$4+* (1 :,+$+-+5' VJU '/</ 6W9 !789 5) &2$,+5+<4' 31' :),
5'" )::54-)$4+*" '/</ 6WW89 5) &2$,+5+<4' ('" ,2"')1R ")*" G5 '/</ 6!X9 YZ9 W[8 '$ 5) &2$,+5+<4' ('"
)::54-)$4+*" &15$4&2(4)" '/</ 6Z9 Z\8/

tel-00408745, version 1 - 3 Aug 2009

!"#$%&' () *+,-./0%$.1 20 2.3"$1' 2' /" 3,%&./.4$'

!"#$%&' (
! "#$%&'&()* +!,,)-*
tel-00408745, version 1 - 3 Aug 2009

!"#$%%! #&' (&&)*) (

!"#$%&'(&) *" %!+),"-)", ") ,"-."' *" ,/(%%" 0)&.1-" *" ./23()"'4 -&5,"5-' ", 5,(%('/,"5-'6 ",
*" ,-/72 08&%5." *" $/95",'4 )&.1-" *" :&,'4 )&.1-" *!/$$%(2/,(&)'6 $&'" $&5- %"' &$;-/,"5-'
5) ';-("5# $-&1%<." *!&1'"-8/,(&) *" 2" -;'"/5= >(',&-(95"."),4 %"' &1'"-8/,(&)' ;,/("), %(.(,;"'
? 5)" '(.$%" 2&%%"2," *" ,-/72 ") 1(,' $/- '"2&)*" '5- %"' *(@;-"),' %(")' *5 -;'"/5 $/- %" $-&,&A
2&%" *" B"',(&) *"' -;'"/5# CDEF GHIJ= K"$5('4 ", $&5- *(8"-'"' -/('&)' 2&.." %/ ,/-(72/,(&)4 %/
*;,"2,(&) *"' /)&./%("'4 ", %!()B;)("-(" *" ,-/724 (% L / "5 M&-, 1"'&() *!5)" &1'"-8/,(&) 5) $"5
$%5' 7)" *5 -;'"/5= F/- "#".$%"4 (% "', *"8")5 );2"''/(-" *" 2&))/N,-" %" 2&),")5 *5 ,-/72 ",
'/ -;$/-,(,(&) "),-" %"' *(@;-"),"' /$$%(2/,(&)' ", %"' *(@;-"),' 5,(%('/,"5-' /()'( 95" *"' ',/,('A
,(95"' 95( $"-.",,"), *" .("5# 2&.$-")*-" %"' $"-M&-./)2"' *5 -;'"/5 ", *"' /$$%(2/,(&)' *"'
5,(%('/,"5-'= O)" &1'"-8/,(&) 7)"4 ", /5 2&),-/(-" *!5)" &1'"-8/,(&) *" 3/5, )(8"/5 ? %/ CDEF4
*"./)*" *" 2/$,5-"- %"' $/95",'4 *" -"B/-*"- %"' 8/%"5-' *"' 23/.$' */)' %"5-' "),P,"' ", *"
.",,-" ? Q&5- *"' ,/1%"' ',&2R;"' */)' %/ .;.&(-" ", 2&),")/), %"' ',/,(',(95"' '5- %" ,-/72= O)"
,"%%" &$;-/,(&) $"5, '" *;-&5%"- '/)' $-&1%<." $&5- *"' 8&%5."' *" ,-/72 $/' ,-<' (.$&-,/),'4
./(' 95/)* %" 8&%5." *" ,-/72 /,,"(), *" B-/)*"' 8/%"5-' *" %!&-*-" *" 2"),/()"' *" E1$'4 2",,"
&1'"-8/,(&) *;,/(%%;" *"' $/95",' ", *" %"5- 2&),")5 *"8("))"), ,-<' *(S2(%" 8&(-" (.$&''(1%" $&5*"5# -/('&)' $-()2($/%"' GTU4 TV4 VW4 XI4 YZJ [
\ / *(S25%,; *!5) /22<' -/$(*" ? %/ .;.&(-" $&5- .",,-" ? Q&5- %"' *(@;-"),' 23/.$' -"%/,(M'
? %/ ."'5-" *5 ,-/72=
\ " B-/)* 8&%5." *!()M&-./,(&) B;);-; $/- 5)" ."'5-" 2&.$%<," *5 ,-/72=
] 2"2( '!/Q&5," 5) /5,-" *;7, %(; /5 23&(# *"' $&(),' &^ "@"2,5"- %"' ."'5-"' GXI4 YI4 ZX4 H_4
Y_I4 IHJ= F-")&)' 2&.." "#".$%" %/ '(.$%" ,`23" *" 2&))/N,-" ? 5) ()',/), *&)); ,&5, %" ,-/72
95( 8/ *5 '(," a /5 '(," b= C( %"' $/95",' '5(8"), 5) '"5% 23".()4 (% '5S, /%&-' *" ."'5-"- %"
,-/72 ") 5) '"5% $&(), *" 2" 23".()= F&5- ;8(,"- *" 2&%%"2,"- *!/5,-"' ,-/72' )&) (),;-"''/),'4
2" $&(), "', ? 23&('(- *" ,"%%" ./)(<-" 95" %" ,-/72 "),-" a ", b '&(, ./Q&-(,/(-"= +% "', ;8(*"),
95" 2" 23&(# *" %!".$%/2"."), *"' $&(),' *" ."'5-" *"8("), *" $%5' ") $%5' 2&.$%(95; /8"2 %/
2&.$%"#(,; *" %/ ,`23" ? "@"2,5"- ", %/ ,/(%%" *5 -;'"/5= K/)' %/ $-/,(95" $%5'("5-' ,`23"' *"
."'5-" $"58"), P,-" (./B();"' 2&.." $/- "#".$%" ."'5-"- ,&5, %" ,-/72 95( 95(,," 5) '(,"4 ,&5,
TT

!"#$%&' () *" +,%&-.-/$' #"00$1'

"# $%&'( )#*$+,- . /, *+$#0 $1/$ "# $%&'( )2/,# &33"+(&$+1, )1,,-#0 #$ (144# (&* #5$%64# $1/$ "#
$%&'( )/ %-*#&/7

!" #$%&'()**+''&,- (-./+0-* -( 1/&()&* 23 041-&3

8# 3%19":4# ;/2/, 13-%&$#/% (1,<%1,$# )&,* "& 4#*/%# )# *1, %-*#&/ #*$ (144#,$ #=#($/#%
/,# $>(?# )# 4#*/%# )# "& 4&,+:%# "& 3"/* 3%-(+*# 31**+9"# -$&,$ )1,,- "# @%&,) A1"/4# )/ $%&'(
;/+ B (+%(/"# CA1"/4# #, 1($#$*0 A1"/4# #, 1($#$*D*#(E7 F2&3%:* "2G,$%1)/($+1,0 +" #*$ ("&+% ;/2/,#
19*#%A&$+1, )/ %-*#&/ . (#%$&+,* 31+,$* 9+#, (?1+*+* %-)/+$ (1,*+)-%&9"#4#,$ "# A1"/4# )/ $%&'(
. (1""#($#%0 . $%&+$#% #$ . *$1(H#%7 G" %#*$# (#3#,)&,$ "# 3%19":4# )# (1""#($# )# 3&;/#$* &/5 31+,$*
)# 4#*/%# *-"#($+1,,-* #$ "2&)&3$&$+1, )# (#$$# (1""#($# &/ )-9+$ -"#A- )/ $%&'(7 8# (?1+5 13$+4&"

tel-00408745, version 1 - 3 Aug 2009

)#* 31+,$* )# 4#*/%# ,2#*$ 4&"?#/%#/*#4#,$ 3&* */I*&,$ 31/% %-*1/)%# (# 3%19":4#7
J', )# 31/A1+% @-%#% "# )-9+$ -"#A- )/ $%&'( #, /, 31+,$ )# 4#*/%#0 /,# )+%#($+1, +431%$&,$#
. */+A%#0 ;/+ #*$ )2&+""#/%* *1/A#,$ &)13$-# 3&% "#* 13-%&$#/%* #$ */331%$-# 3&% "#* (1,*$%/($#/%*0
#*$ (#""# )# "2-(?&,$+""1,,&@# $#431%#" K L0 M0 M!N7 O"/*+#/%* <1%4#* )2-(?&,$+""1,,&@# $#431%#"
3#/A#,$ 6$%# #,A+*&@-#*7 82+)-# )# 9&*# )# "2-(?&,$+""1,,&@# $#431%#""# #*$ )# $%&+$#% /,# 3&%$+#
)#* 3&;/#$* )/ $%&'( &/ "+#/ )# $1/* "#* 3&;/#$*7 8# %&$+1 )#* 3&;/#$* (1""#($-* #$ $%&+$-* #*$
(# ;/21, &33#""# "# $&/5 )2-(?&,$+""1,,&@#0 *1/A#,$ #53%+4- #, 31/%(#,$7 O1/% /, $&/5 )2-(?&,P
$+""1,,&@# )# 100.pQ0 /,# $#(?,+;/# *+43"# (1,*+*$# . (1""#($#% /, 3&;/#$ &A#( /,# 3%19&9+"+$- p
#$ )# "2+@,1%#% CA+*P.PA+* )# "& 4-$%1"1@+#E &A#( /,# 3%19&9+"+$- 1 p0 #$ )1,( (2#*$ *#/"#4#,$ 31/%

100.pQ )#* 3&;/#$* ;/# "& 4-41+%# #*$ &((-)-# 31/% "& 4+*# . R1/% )#* (?&43* (1,$#,&,$ )#*
*$&$+*$+;/#* */% "# $%&'(7 S#(+ %-)/+$ 3&% /, <&($#/% )# 1/p "# ,149%# )2&((:* 4-41+%#0 #$ %-)/+$
&/**+ "# A1"/4# )#* %-*/"$&$* )# 4#*/%# . *$1(H#% #$ . $%&,*31%$#% . $%&A#%* "# %-*#&/ &/5 /,+$-*
(#,$%&"#* (?&%@-#* )/ $%&+$#4#,$7 T,# &/$%# 4&,+:%# )# <&+%# "2-(?&,$+""1,,&@# #*$ )# (1""#($#%
#$ $%&+$#% /, 3&;/#$ $1/* "#* 1/p 3&;/#$*7 G" & -$- )-41,$%- ;/# 31/% /, $%&'( +431%$&,$ *$&P
$+1,,&+%#0 $1/$#* "#* $#(?,+;/#* )1,,#,$ 3%&$+;/#4#,$ "# 464# %-*/"$&$ K LN7 8& )+=-%#,(# #,$%#
/,# $#(?,+;/# #$ /,# &/$%# #*$ )&,* *& (&3&(+$- . )-$#($#% )#* 3?-,14:,#* $%&,*+$1+%#* #$ )#*
&($+A+$-* 3-%+1)+;/#* &+,*+ ;/# )&,* "& <&(+"+$- )2+43"-4#,$&$+1,7
T,# ;/#*$+1, ;/+ *# 31*# #*$ (144#,$ /,# /,+$- )# 4#*/%# #*$ (&3&9"# )2&33"+;/#% "2-(?&,P
$+""1,,&@# . $1/* "#* 3&;/#$* )2/, $%&'( A1"/4+,#/5 *&,* ;/2#""# *1+$ (&3&9"# )# $%&+$#% $1/*
(#* 3&;/#$*7 U, #=#$0 "2-(?&,$+""1,,&@# *# <&+$ *1/A#,$ &/ ,+A#&/ 3?B*+;/# C?&%)V&%#E 3&% )#*
<1,($+1,* )# ?&(?&@# ;/+ *1,$ +,)-3#,)&,$#* )/ (1,$#,/ )#* 3&;/#$*7 S#$$# 13-%&$+1, #*$ %&3+)#
#$ ,2#*$ 3&* 9"1;/&,$# 31/% /,# /,+$- )# 4#*/%#7 8& 3%1(-)/%# 9"1;/&,$# #*$ 3&% (1,$%# "2&((:*
. "& 4-41+%# C#7@70 K MNE ;/+ #, (&* )2-(?&,$+""1,,&@# ,2#*$ <&+$ ;/# 31/% "#* 3&;/#$* *-"#($+1,,-*
)21W "# @&+, +431%$&,$ #, A+$#**# C$%1+* 1%)%#* )# @%&,)#/% 31/% /, $&/5 )2-(?&,$+""1,,&@# )#
X7XXYQE7
82-(?&,$+""1,,&@# #*$ /,# $#(?,+;/# 9+#, (1,,/# )&,* )2&/$%#* )14&+,#* (144# 3&% #5#43"#
"& $?-1%+# )/ *+@,&" #$ "& $?-1%+# )/ *1,)&@#7 J (?&;/# 13-%&$+1, )2-(?&,$+""1,,&@# (1%%#*31,) /,#

!"! #$%&'()*+,,-))(./ &-00/ 1) 23-4,50/ 6-+)*

"#$%& '()%"$*'+ ,*$& ,-*+.&%/*'+ 0!123 4+ &5&$6 /&#7/ 7&/ ("8#&$/ )9:"+$*77'++)/ /'+$ $%"*$)/ ('#%
7& ;#$ ,& 7" <)$%'7'=*& &$ 7&/ "#$%&/ ("8#&$/ /'+$ /&#7&<&+$ %&7">)/ /"+/ $%"*$&<&+$ ("% 7-#+*$) ,&
<&/#%&3 ?&/ /$"$*/$*8#&/ /#% 7& $%"@9 /'+$ &+/#*$& 9"79#7)&/ A ("%$*% ,&/ ("8#&$/ )9:"+$*77'++)/3
?" 8#&/$*'+ B'+,"<&+$"7& &/$ 9'<;*&+ 9&/ /$"$*/$*8#&/ /'+$ @,C7&/ "# $%"@9 '%*=*+"7 &$ 9'<<&+$
('#% #+& $&9:+*8#& ,-)9:"+$*77'++"=& ,'++)& &$ #+ $"#D ,-)9:"+$*77'++"=& ,'++)6 *+B)%&% 7&/
/$"$*/$*8#&/ ,# $%"@9 '%*=*+"7 A ("%$*% ,& 9&77&/ ,# $%"@9 )9:"+$*77'++)3 ?" ,&%+*C%& '()%"$*'+ &/$
9&77& ,*$& E *+.&%/*'+ E &$ 7-&%%&#% 8#* 7-"99'<("=+& &/$ 7-&%%&#% 9"#/)& ("% 7-)9:"+$*77'++"=&3
F'#% 9&%$"*+&/ /$"$*/$*8#&/ 7-*+.&%/*'+ &/$ B"9*7& &$ 7-&%%&#% &/$ B"*;7& .'*%& +)=7*=&";7& 0 G6 !123
H&9* &/$ 7& 9"/ ,&/ /$"$*/$*8#&/ 9"79#7)&/ /#% #+ =%"+, +'<;%& ,& ("8#&$/3 I+ &D&<(7& $>(*8#&
&/$ ,& 9"79#7&% 7& .'7#<& ,# $%"@9 &+ '9$&$/ '# &+ ("8#&$/6 7& +'<;%& ,& ("8#&$/ ("% (%'$'9'7&
,& $%"+/('%$6 7& +'<;%& ,& ("8#&$/ ("% "((7*9"$*'+6 &$93 H-&/$ 7& %)/#7$"$ *+$#*$*B ,& 7" 7'* ,&/
=%"+,/ +'<;%&/ 8#* (&%<&$ ,-*+B)%&% 7& .'7#<& ,-#+& =%"+,& ('(#7"$*'+ A ("%$*% ,# .'7#<&

tel-00408745, version 1 - 3 Aug 2009

,-#+& ("%$*& ,& 9&$$& ('(#7"$*'+ $*%)& "# :"/"%,3 J7 > " 9&(&+,"+$ ,&/ /$"$*/$*8#&/ ('#% 7&/8#&77&/
7-'()%"$*'+ ,-*+.&%/*'+ &/$ ,*K9*7& .'*%& *<('//*;7& A $"#D ,-)9:"+$*77'++"=& %),#*$3 L+ $%'#.&
,"+/ 9&$$& 9"$)='%*& $'#$&/ 7&/ /$"$*/$*8#&/ 8#* 9'+9&%+&+$ #+ (&$*$ .'7#<& ,& ("8#&$/ 9'<<& ("%
&D&<(7& 7& .'7#<& ,&/ M'$/6 7" ,#%)& ,&/ M'$/6 7&/ (:)+'<C+&/ $%"+/*$'*%&/ &$ "+'<"7*&/6 &$93 ?"
%&9:&%9:& ,"+/ 9& ,'<"*+& &/$ =%"+,& '#.&%$& &$ 9*;7& ,& (7#/ &+ (7#/ ,& /$"$*/$*8#&/ ".&9 7&#%/
<)$:',&/ ,-)9:"+$*77'++"=& &$ ,-*+.&%/*'+ '($*<*/)&/ N&3=36 0 G6 OP6 1!6 !12Q3 R"+/ 7& 9:"(*$%&
/#*."+$ +'#/ *77#/$%'+/ 7" ,*K9#7$) ,-*+.&%/*'+ /#% #+& <)$%*8#& /*<(7& <"*/ ,-#+ =%"+, *+$)%S$
('#% 7&/ '()%"$&#%/ 8#* &/$ 9&77& ,& 7" ,)$&9$*'+ ,&/ M'$/ 7&/ (7#/ ='#%<"+,/ &$ ,& 7&#% 97"//&<&+$
7&/ #+/ ("% %"(('%$ "#D "#$%&/ 01!23 T'#/ <'+$%'+/ ("% ,&/ 9"79#7/ ,& (%';";*7*$) ."7*,)/ /#%
,&/ $%"9&/ %)&77&/ 8#& 9&$$& *+.&%/*'+ +)9&//*$& #+ $"#D ,-)9:"+$*77'++"=& )7&.)6 "# <'*+/ (7#/
)7&.) 8#& 7&/ $"#D ,-)9:"+$*77'++"=& "9$#&77&<&+$ (%"$*8#)/ ("% 7&/ '()%"$&#%/3 R"+/ 7& 9:"(*$%&
,-"(%C/6 +'#/ <'+$%'+/ /#% #+& "#$%& <)$%*8#& 9'<<&+$ 7-*<("9$ ,& 7-)9:"+$*77'++"=& (&#$ S$%&
%),#*$ &+ $&+"+$ 9'<($& ,& 7-"#$'U9'%%&7"$*'+ 8#* (#*//& &D*/$&% ,"+/ 7" 8#"+$*$) <&/#%)&3

!" #$%&'()*+,,-))(./ &-00/ 1) 23-4,50/ 6-+)*

I+ '()%"$&#% ,)/*%"+$ ';/&%.&% /'+ %)/&"# ('//C,& ,'+9 ,&#D <"%=&/ ,& <"+'&#.%& ('#%
<*+*<*/&% 7& .'7#<& ,& ("8#&$/ A 9'77&9$&% &$ A $%"*$&%3 R-#+ 9V$) *7 (&#$ B"*%& ,& 7-)9:"+$*77'+U
+"=& /("$*"7 ('#% <*&#D 9:'*/*% 7&/ ('*+$/ ,& <&/#%&3 R& 7-"#$%& 9V$) *7 (&#$ 9'+@=#%&% 7& $"#D
,-)9:"+$*77'++"=& $&<('%&7 &+ B'+9$*'+ ,& 7" 9"("9*$) ,& /&/ #+*$)/ ,& <&/#%& &$ ,& 7" (%)9*/*'+
,)/*%)&3 H&/ ,&#D ,*<&+/*'+/ (&#.&+$ S$%& 9'+/*,)%)&/ /)("%)<&+$ 7-#+& ,& 7-"#$%& W <*&#D (7"9&%
7&/ ('*+$/ ,& <&/#%& &+/#*$& <*&#D 9:'*/*% 7& $"#D ,-)9:"+$*77'++"=&3 477&/ (&#.&+$ "#//* S$%& 9'<U
;*+)&/ &+/&<;7& ,& $&77& B"X'+ A B'%<&% #+ (%';7C<& ,-'($*<*/"$*'+ =7';"7& 8#* 9:&%9:& 9'<<&
/'7#$*'+ 7& <&*77&#% (7"9&<&+$ ,&/ ('*+$/ ,& <&/#%& &$ 7& <&*77&#% $"#D ,-)9:"+$*77'++"=& <*+*U
<*/"+$ 7-&%%&#% ,-)9:"+$*77'++"=&6 )$"+$ ,'++)&/ ;*&+ /Y% 7&/ 9'+$%"*+$&/ ,# />/$C<& &+ $&%<&/
,& 9"("9*$) ,&/ ('*+$/ ,& <&/#%& &$ ,& .'7#<& <"D*<#< ,# $%"@9 A 9'77&9$&%3 ?-'($*<*/"$*'+

!"#$%&' () *" +,%&-.-/$' #"00$1'

"#$%&' (') ('*+ ,-#./01') 2 /3242%&25' (3$('%&$6'- ($4'-) ,#$%&) (' 1')*-' 72$./'1'%& 892-5:);
2* /$'* (3*% %#1.-' 72$./' (' ,#$%&) (' 1')*-' <*$ )#%& )*-892-5:) ,2- /') #,:-2&$#%) (' 1')*-'=
>/2$-'1'%&; ,#*- <*' 8'&&' #,&$1$)2&$#% "#$%&' &-#*4' )#% $%&:-?&; /3#,:-2&'*- (#$& ($),#)'- (3*%
5-2%( %#1.-' (' ,#$%&) (' 1')*-' @ 28&$4'- '& @ ,-#5-211'- 24'8 *% &2*+ (3:892%&$//#%%25'
#,&$12/ <*$ 8#--'),#%( @ /2 &A89' (' 1')*-' (:)$-:'= >3')& *%' 9B,#&90)' <*$ ')& &#*& @ 72$& -:2C
/$)&' :&2%& (#%%: <*328&*'//'1'%& /2 ,/*,2-& )$ 8' %3')& ,2) /2 &#&2/$&: (') -#*&'*-) (* -:)'2* (3*%
#,:-2&'*- )#%& :<*$,:) ,2- (') #*&$/) (' 1')*-' <*$ :892%&$//#%%'%& /' &-268 '& /' -2,,#-&'%& 2*+
*%$&:) 8'%&-2/') D/3'+'1,/' &B,$<*' ')& /') -#*&'*-) >EF>G :<*$,:) ,2- /3#*&$/ (' 1')*-' ,2))$4'
H'&I/#J KLMNO=
P2%) /2 )*$&' (' 8'&&' )'8&$#%; %#*) 2//#%) 8#%)$(:-'- *%' &A89' (' 1')*-' '& %#*) 2//#%)
82/8*/'- )2 ,-:8$)$#% @ ,2-&$- (* &-268 :892%&$//#%%: &'1,#-'//'1'%& @ &#*) /') ,#$%&) (' 1')*-'
($),#%$./') (2%) /' -:)'2*= H#&#%) 8'&&' ,-:8$)$#% ,2- M (p); #Q p ')& /' 4'8&'*- (') &2*+ (3:892%C

tel-00408745, version 1 - 3 Aug 2009

&$//#%%25' (2%) /') ($R:-'%&) ,#$%&) (' 1')*-' D#* -#*&'*-)O= S' ,-#./01' -'4$'%& 2/#-) @ &-#*4'/') 1'$//'*-') 42/'*-) (' p <*$ 12+$1$)'%& M (p)=
S') 8#%&-2$%&') )*- /' ,-#./01' (3#,&$1$)2&$#% )#%& /') )*$42%&')= T-'1$0-'1'%&; p ')& *%
4'8&'*- @ :/:1'%&) ,#)$&$7)= U% :/:1'%& (' p :52/ @ V:-# 4'*& &#*& )$1,/'1'%& ($-' <*' /' ,#$%&
(' 1')*-' '% <*')&$#% %3')& ,2) 28&$4: ,#*- 8'&&' &A89' (' 1')*-'= P'*+$01'1'%&; /') &2*+
(3:892%&$//#%%25' (2%) /') ,#$%&) (' 1')*-' D-#*&'*-)O %' (#$4'%& ,2) (:,2))'- *%' 8'-&2$%'
42/'*- (:8$(:' ,2- /' 12&:-$'/= >'&&' .#-%' )*,:-$'*-' ,'*& &9:#-$<*'1'%& 2//'- "*)<*3@ WXXY=
Z-#$)$01'1'%&; /' 4#/*1' (* &-268 5/#.2/'1'%& 8#//'8&: %' (#$& ,2) '+8:('- *%' 8'-&2$%' 42/'*,-:(:6%$' ,2- /3#,:-2&'*- '% 7#%8&$#% (' )') 82,28$&:) (' )&#8[25' '& (' &-2%),#-&= S' 4#/*1' (*
&-268 5/#.2/'1'%& 8#//'8&: ')& )$1,/'1'%& /2 )#11' (* &-268 8#//'8&: (2%) /') ($R:-'%&) ,#$%&)
(' 1')*-'= H#*) ,#)):(#%) 2/#-) *% ,-#./01' (3#,&$1$)2&$#% 8/2))$<*' <*3*% #,:-2&'*- ,'*&
-:)#*(-' ,#*- &-#*4'- /' 1'$//'*- 4'8&'*- p :&2%& (#%%:' /2 &A89' (' 1')*-' @ 'R'8&*'-= F$ ,2'+'1,/' /2 &A89' (' 1')*-' ')& /3')&$12&$#% (* &-268 '%&-' /' )$&' \ '& /' )$&' ]; /3#,:-2&'*- ,'*& ^
_ ,2-&$- (3*%' ')&$12&$#% (' 8' &-268 #-$5$%2/;
_ 82/8*/'- ,#*- *% 4'8&'*- p <*'/8#%<*' /2 ,-:8$)$#% (' /3$%4'-)$#%;
_ &-#*4'- /2 1'$//'*-' 42/'*- (' p;
_ 8#%65*-'- /') ,#$%&) (' 1')*-' 24'8 8'&&' 42/'*- (' p '& 8#//'8&'- (* &-268 ,'%(2%& *%
8'-&2$% &'1,);
_ 1')*-'- /' 4#/*1' (* &-268 :892%&$//#%%: '%&-' \ '& ];
_ '%)*$&' &-#*4'- *%' 1'$//'*-' ')&$12&$#% ,#*- /' 4#/*1' (* &-268 #-$5$%2/ '%&-' \ '& ]
:&2%& (#%%: /' 4#/*1' (* &-268 8#//'8&:;
_ -'&-#*4'- *%' 2*&-' 42/'*- #,&$12/' ,#*- p '% *&$/$)2%& /2 %#*4'//' ')&$12&$#% (* 4#/*1'
(* &-268 #-$5$%2/ '%&-' \ '& ];
_ '& 2$%)$ (' )*$&' "*)<*3@ 8' <*' /' ,-#./01' 8#%4'-5'=
H#) '+,:-$1'%&2&$#%) #%& 1#%&-: <*' &-0) ,'* (3$&:-2&$#%) )#%& )*`)2%&') ,#*- <*' /' )B)&01'
8#%4'-5' 4'-) /2 42/'*- (:)$-:'=

! ! "#$%&$'()*#+ ,# -.&/01'1+(01&) 2&1)0#

tel-00408745, version 1 - 3 Aug 2009

!"#

"# $ %& '()(*(+,- ./ 012-&/ -/0()1-3 45678

"#$%&$'()*#+ ,# -.&/01'1+(01&) 2&1)0#

7(/2 &9(32 &))*,:/1 *& ;1'<(.- .=()',;,2&',(3 >(,3'- &/ '0&?@ @(**-@'1 2/0 *- 012-&/ -/0()1-3

45678 .(3' *& '()(*(+,- -2' ,**/2'01- .&32 A,+/0- "#" 7('0- 'B@<- @(32,2'&,' C -2',;-0 '(/' *'0&?@ (0,+,3&,0- ./D.-2',31 &/ 012-&/ .- 0-@<-0@<- E6758 &/ F(G&/;- H3," I- '0&?@ -2' *&
2(;;- .- '(/2 *-2 '0&?@2 1@<&3+12 -3'0- *- 0(G&/;- /3, -' @<&@/3 .-2 &/'0-2 )&G2 -/0()1-32"
I<&@/3 .- @-2 '0&?@2 J(0;- @- :/=(3 &))-**- /3 KL )&,0 M(0+,3N.-2',3&',(3 )&,0O ,"-" *-2 *,+3-2
0(/+-2 .&32 A,+/0- "#" P&0 -Q-;)*-R HSNAF 0-)012-3'- *- '0&?@ 1@<&3+1 -3'0- *- 0(G&/;- /3, -'
*& A0&3@-" L&32 @- @&2R *& J(3@',(3 M (p) .- 3('0- )0(T*U;- .=()',;,2&',(3 3=-2' :/=/3- 2(;;2/0 *& )01@,2,(3 .=-2',;&',(3 ./ 9(*/;- .- @<&@/3 .- @-2 '0&?@2 KL" P(/0 19&*/-0 *& )01@,2,(3
3(/2 &9(32 @(32,.101 *=-00-/0 :/&.0&',:/- ;(G-33-R 9(,0 633-Q- 6 )(/0 )*/2 .- .1'&,*2" 7(/2
&9(32 -32/,'- '(/031 3('0- &*+(0,'<;- )(/0 '0(/9-0 *-2 ;-,**-/02 '&/Q .=1@<&3',**(33&+- .&32 *-2
.,V10-3'2 0(/'-/02 ./ 012-&/ 45678R -' 3(/2 &9(32 @&*@/*1 *& )01@,2,(3 (T'-3/- )(/0 @<&@/3 .-2
KL )&,02" 8(/' @-@, & 1'1 J&,' .&32 /3 2,;/*&'-/0 &*,;-3'1 )&0 *- 90&, '0&?@ ./ 012-&/ 45678
&9-@ *& 90&,- '()(*(+,- ./ 012-&/" 8(/2 *-2 0(/'-/02 ./ 012-&/ 1'&,-3' 2/))(212 ,;)*1;-3'-0 .-2
(/',*2 .- ;-2/0- C '&/Q .=1@<&3',**(33&+- @(3?+/0&T*-"
P(/0 T,-3 @(;)0-3.0- *-2 )-0J(0;&3@-2 .- 3('0- ;1'<(.-R 3(;;1- ;1'<(.-N+*(T&*-R 3(/2
*=&9(32 @(;)&01 C /3- 9-02,(3 2,;)*,?1-" 6/ @(3'0&,0- .- *& ;1'<(.-N+*(T&*- (W 3(/2 J&,2(32
/3 1@<&3',**(33&+- C '(/2 *-2 0(/'-/02 ./ 012-&/ 45678R @-''- 9-02,(3 2,;)*,?1- @(32,2'- 2,;N
)*-;-3' C (T2-09-0 *- '0&?@ &/Q ,3'-0J&@-2 ./ 0(/'-/0 .=&@@U2 ./ F(G&/;- H3, M&9-@ /3 '&/Q
.=1@<&3',**(33&+- ()',;,21 *(@&*-;-3'O" 7('(32 @-''- 9-02,(3 2,;)*,?1- )&0 ;1'<(.-N*(@&*-" %-2
012/*'&'2 (T'-3/2 2(3' 012/;12 .&32 *& A,+/0- "X"
%& @(/0T- Y69-0&+- (9-0 &** KL )&,02Y ;(3'0- *& )01@,2,(3 ;(G-33- 2/0 '(/2 *-2 KL )&,02
)(/0 *& ;1'<(.-N+*(T&*-" %- YZ-2' KL )&,0Y -' Y[(02' KL )&,0Y ;(3'0-3' 0-2)-@',9-;-3' *&
;-,**-/0- -' *& )*/2 ;&/9&,2- )01@,2,(3R &/22, )(/0 *& ;1'<(.-N+*(T&*-" %-2 '0(,2 &/'0-2 @(/0T-2
;(3'0-3' *& ;\;- @<(2- ;&,2 @-@, )(/0 *& ;1'<(.-N*(@&*-R *& 9-02,(3 2,;)*,?1- .- *& ;1'<(.-N

!"#$%&' () *" +,%&-.-/$' #"00$1'

tel-00408745, version 1 - 3 Aug 2009

!"#

"# $ %&' ()*+,',-.' /0-1&..&2 (,)& &3 0&,44&5)&6 -73&.5&' '5) 3-5' 4&' 89 (:,)'

;4-7:4&" <-3&= >5& 4?:@& A&' @ A:.' 4: B;5)& +-))&'(-.A C 4: +-.3):,.3& '5) 4& D-450& A& 3):B+
+-44&+3* &. (:>5&3'E'&+"
<-5' D-1-.' +4:,)&0&.3 A:.' 4: B;5)& +-00&.3 .-3)& 0*3F-A&G;4-7:4& A-..& &. 0-1&..&
A?&@+&44&.3&' (&)H-)0:.+&' (-5) A,I*)&.3&' D:4&5)' A& 4: +-.3):,.3& '5) 4& 3):B+ +-44&+3*" J&33&
()*+,',-. 0-1&..& &'3 4*;K)&0&.3 '5(*),&5)& C +&44& -73&.5& (:) 4: 0*3F-A&G4-+:4& -L 4?-7'&)G
D:3,-. &'3 H:,3& '&54&0&.3 :5@ ,.3&)H:+&' A5 )-53&5) A?:++K' MN" %&' A&5@ 0*3F-A&' ;4-7:4& &3
4-+:4& A-..&.3 ():3,>5&0&.3 4&' 0O0&' (&)H-)0:.+&' (-5) 4&' 89 (:,)' :1:.3 4: 0&,44&5)& ()*G
+,',-." J&(&.A:.3 4&' )*'543:3' '-.3 A,I*)&.3' (-5) 4&' 89 (:,)' :1:.3 4&' (45' H:,74&' ()*+,',-.'"
P. &I&32 5. 89 (:,) :D&+ 5.& H:,74& ()*+,',-. &'3 H-)0* A?5. 3):B+ H:,74& &. D-450& (*.:4,'*
(:) 5.& 0*3F-A& A& 0&'5)& ',0(4,B*& '& 4,0,3:.3 C 3)K' (&5 A& (-,.3' A& 0&'5)&" P. 4?-7'&)G
D:.3 C A&' &.A)-,3' +F:);*' /+-00& 4&' ,.3&)H:+&' A5 )-53&5) A?:++K' A5 )-1:50& 5.,62 4& 3:5@
A?*+F:.3,44-..:;& +F-,', &'3 7:' /(:)+& >5& 4& 3):B+ &'3 ;):.A6 &3 A-.+ +& (&3,3 89 (:,) '-5I)&
*.-)0*0&.3" %&' (&3,3' 89 (:,)' -.3 7&'-,. A?O3)& -7'&)D*' '5) A&' 4,&.' -L ,4' '-.3 '&54' -5
:D&+ 3)K' (&5 A?:53)&' 3):B+' (-5) (-5D-,) :5;0&.3&) 4& 3:5@ A?*+F:.3,44-..:;& &3 :0*4,-)&) 4:
()*+,',-. A& 4&5) 0*3)-4-;,&2 >5&4>5& +F-'& >5& .-3)& 0*3F-A&G;4-7:4& 4& (&)0&3"
Q,;5)& "# 0-.3)& :5'', 4& )*'543:3 ,.35,3,H >5& 4: ()*+,',-. A& .-3)& 0*3F-A&G;4-7:4& '?:0*4,-)&
&. )&4:@:.3 4: +-.3):,.3& '5) 4& D-450& A5 3):B+ +-44&+3*"
J-.+&).:.3 4: +-.D&);&.+& A& .-3)& :4;-),3F0& &. 4?:7'&.+& A?5.& +-..:,'':.+& ()*+,'& '5) 4&
D-450& A5 3):B+ -),;,.:42 .-' &@(*),0&.3:3,-.' -.3 0-.3)* >5?,4 '5R3 A?5. 3)K' (&5 .-07)& A?,3*G
):3,-.' (-5) :33&,.A)& 4&' 7-..&' D:4&5)' (-5) 4&' 3:5@ A?*+F:.3,44-..:;& &3 (:) +-.'*>5&.+& 4&'
0&,44&5)&' &'3,0:3,-.' (-5) 4& 3):B+ )*'&:5" 9:.' Q,;5)& " .-5' 0-.3)-.' +-00&.3 +&33& ()*G
+,',-. *D:45& &. H-.+3,-. A5 .-07)& A?,3*):3,-.'" %?:@& A&' @ A:.' 4: B;5)& 0-.3)& 4&' A,I*)&.3&'
(-'',7,4,3*' (-5) 4& 3:5@ A?*+F:.3,44-..:;& :5 A*(:)3 &3 4?:@& A&' 1 4: ()*+,',-. A& 4?&'3,0:3,-. A5

tel-00408745, version 1 - 3 Aug 2009

!"! #$%&'()*$%)

!"#

" # $% &'()*+*,- .- /,-)0*,- 12 -,34'. 15*0('%0*,-+

6,723. 1. 7589 &%*'. 7. &72+ 1*:)*7. .0 7% &'()*+*,- 3,;.--(. +2' 0,2+ 7.+ 89 &%*'+" <7 .+0 )7%*'
=25> &%'0*' 1. 1.2? *0('%0*,-+ -,2+ %00.*@-,-+ 2-. &'()*+*,- *-1(&.-1%-0. 12 0%2? 15()A%-0*77,-B
-%@. &'*+ %2 1(&%'0" <7 .+0 &',4%47. =2. ). -,34'. 15*0('%0*,-+ )A%-@. 152- ),-0.?0. > 75%20'. .0
152-. 3(0'*=2. > 75%20'. 3%*+ -,2+ )',;,-+ =25*7 '.+0. 0,2C,2'+ /%*47." $. &',26.' 0A(,'*=2.3.-0
/%*0 &%'0*. 1. -,+ 0'%6%2? /202'+ 1%-+ ).00. 1*'.)0*,-"

!" #$%&'()*$%)
9%-+ ). )A%&0*'.D -,2+ %6,-+ &'(+.-0( -,0'. %')A*0.)02'. &,2' 7. )A,*? 12 0%2? 15()A%-0*77,-B
-%@. 1%-+ 7.+ ',20.2'+ 152- '(+.%2 %E- 15,&0*3*+.' 2-. ).'0%*-. 0F)A. 1. 3.+2'." $5%')A*0.)02'.
% 3,-0'( +,- *-0('G0 1%-+ 75.+0*3%0*,- 12 6,723. 1.+ 89 &%*'+D 3G3. ).2? 7.+ &72+ /%*47.+D .0
).)* > 6,723. 12 0'%E) ),77.)0( (@%7" H72+ 1. 1(0%*7+ +2' ). 0'%6%*7 +. 0',26.-0 1%-+ I--.?. I"
J,2+ &,2'+2*6,-+ -,+ '.)A.')A.+ 1%-+ ).00. 1*'.)0*,- .0 -,2+ +,33.+ .- &%'0*)27*.' *-0('.++(+
&%' &',26.' 75.:)%)*0( 1. -,0'. %')A*0.)02'. &,2' 15%20'.+ 0F)A.+ 1. 3.+2'. .0 &%' 7% ),-6.'@.-).
1. 75%7@,'*0A3. .- )%+ 15*-/,'3%0*,-+ 3%-=2%-0.+" K.+ 0'%6%2? +. /.',-0 1%-+ 7. )%1'. 12 &',C.0
.2',&(.- LK89L =2* 1(3%''.'% .- M.&0.34'. NOOP"

tel-00408745, version 1 - 3 Aug 2009

!"#$%&' () *" +,%&-.-/$' #"00$1'

!"#$%&' (

tel-00408745, version 1 - 3 Aug 2009

! "#$%& %'&#()$%& *+#

,-./0+'&%,,(''+1! 2 ).&!/&%(' !&
/,+33!4!'& )!3 5(&3 6(,$4%'!$7
!"#$%%! #&' (&&)*) +
!"#$ %#&'("$&")#$ *!)# +% ,-"#!+!.($ *%//(0$ $" +$ &'!(1 2$/ "%)1 23-&'%4"(++!44%.$ #$*!/$ /)#
)4$ 5!4&"(!4 (,*!#"%4"$6 &$++$ 2$ +3$##$)# 7)( %&&!,*%.4$ +3(40$#/(!4 23)4 "#%8& -&'%4"(++!44-9
:%4/ +$ &'%*("#$ *#-&-2%4" 4!)/ %0!4/ ,!4"#- +3(4"-#;" 2$ 4!"#$ %**#!&'$ /)# )4$ ,-"#(7)$
/(,*+$ < (40$#/$#6 &$++$ 2) 0!+),$ 2) "#%8& $4"#$ #!)"$)#/ /!)#&$/ $" #!)"$)#/ 2$/"(4%"(!4/9 =4
$>$"6 +% ,$(++$)#$ $/"(,%"(!4 2) 0!+),$ 23)4 "#%8& -&'%4"(++!44- *$)" ;"#$ !?"$4)$ *%# )4$
/(,*+$ 2(0(/(!4 *%# +$ "%)1 23-&'%4"(++!44%.$ $" +3$##$)# &!,,(/$ $/" +(,("-$ $" *$)" 5%&(+$,$4"
;"#$ &'(>#-$9 @$*$42%4" *!)# 23%)"#$/ ,-"#(7)$/ +3$1$#&(&$ $/" +!(4 23;"#$ 5%&(+$ /)#"!)" 7)%42 &$&(
"!)&'$ < 2$/ A!"/ 2$ 5%(?+$ 0!+),$ &!,,$ *%# $1$,*+$ +!#/ 2$ +3$/"(,%"(!4 2) 0!+),$ ,!B$4 2$/
A!"/6 2) 4!,?#$ 2$ A!"/ !) 2$ +% 2(/"#(?)"(!4 2$/ 0!+),$/ 2$/ A!"/ CDE6 FG6 HHI9 J% 2(K&)+"- $/"
2%4/ +$ 5%(" 7)$ ?$%)&!)* 2$ A!"/ *%//$4" (4%*$#L)/ < &%)/$ 2$ +$)# 5%(?+$ 0!+),$9 M%# $1$,*+$
*!)# )4 "%)1 23-&'%4"(++!44%.$ 2$ GNGOO $" )4 A!" 2$ GO *%7)$"/ P&$ 7)( $/" +% "%(++$ ,!B$44$ 23)4
A!" 2%4/ +3Q4"$#4$"R6 +% &'%4&$ 7)$ &$ A!" /!(" &%*")#- $/" 23$40(#!4 GOS &$ 7)( 0$)" 2(#$ 7)$ HOS
2$ &$/ *$"("/ A!"/ *%//$4" (4%*$#L)/9 J% 2(K&)+"- < &!44%("#$ +$)# 4!,?#$ #-/(2$ 2%4/ +$ 5%(" 7)$
&$&( 4-&$//("$ +% &!44%(//%4&$ 2$/ 0!+),$/ !#(.(4%)1 2$/ A!"/ *!)# #-&!,*$4/$# +% *$#"$ &%)/-$
*%# +3-&'%4"(++!44%.$ &$ 7)( $/" /!)0$4" +3(45!#,%"(!4 < #$&'$#&'$#9 :$/ /!+)"(!4/ *#!"!&!+%(#$/
&!,,$ *%# $1$,*+$ +3(40$/"(.%"(!4 2$/ $4";"$/ 2$/ *%7)$"/ < +% #$&'$#&'$ 2$/ *%7)$"/ TU
7)( %44!4&$4" +$ 2-?)" 2$/ A!"/ *$)" 2%4/ &$#"%(4$/ &%/ %,-+(!#$# +% *#-&(/(!49 J% 2-"$&"(!4
2$/ %4!,%+($/ /!)>#$4" %)//( 2$ +3-&'%4"(++!44%.$9 V +3$1&$*"(!4 2$/ %4!,%+($/ &%)/-$/ *%# )4$
%).,$4"%"(!4 ?#)/7)$ 2) "#%8&6 +$/ %4!,%+($/ $4 .-4-#%+ /!4" &%)/-$/ *%# )4 .#%42 4!,?#$ 2$
*$"("/ A!"/ 7)( 0(/$4" < /%")#$# )4 /$#0$)# !) < &'$#&'$# )4 *!#" 2(/*!4(?+$ /)# )4$ ,%&'(4$9
J% 2-"$&"(!4 /$ 5%(" /!)0$4" *%# )4$ (2$4"(8&%"(!4 23%).,$4"%"(!4 2%4/ +$ 4!,?#$ 2$ A!"/9 W%(/
&!,,$ (+/ /3%.(" /!)0$4" 2$ A!"/ 2$ 5%(?+$ 0!+),$6 )4 .#%42 4!,?#$ 23$4"#$ $)1 *%//$4" /%4/
;"#$ 0)/ $" +% &%*%&("- < 2-"$&"$# +3%4!,%+($ $/" %+!#/ #-2)("$ CXY6 EFI9 J$/ ,-"'!2$/ ?%/-$/ /)# +$
YG

!"#$%&' () *' +&,$% $-%&./,$% #"& 0123!"-%$00.--"4' 5 /2%'3%$.- '% 30"66'7'-% /'6 8.%6 9.0,7$-',:

"#$"%$ &'()*+,-.( (* -#/ $( ),01+( &( 2,*/ ,)* 0,)*+3 %)( 0(.$$(%+( +3/./*#)*( 4 "(* 35#+& 6! 78
9#)/ "( ":#-.*+( ),%/ .$$%/*+,)/ $'.0-#"* &( $'3":#)*.$$,))#5( &% *+#;" /%+ %) (<(0-$( ",)"+(*=
"($%. &( $# &3*("*.,) (* "$#//(0()* &(/ 2,*/ >,$%0.)(%< &#)/ %) +3/(#% ?#--($3/ /,%>()* !"#$

%&&!'(

! "#$ %&#'()#*'$ $+,-).'#$$#-) / 0#$ 1%)$ &%*' 2,3#'$#$ '(,$%-$ 4 0%5&',$ 6+,-7.-,#',# 2#

)'(80 9#- (-76(,$

&'")* !+,%+!!'%+,

: 6( ;(0)*'(),%- #) 6( $.0*',).! <- 1%) &#*) =)'# 2.8-, 0%55#

.)(-) *- $,5&6# )'(-$;#') 0('(0).',$. &(' 6# ;(5#*> ?*,-)*&6# 9(2'#$$# @A $%*'0#: &%') $%*'0#:
(2'#$$# @A 2#$),-(),%-: &%') 2#$),-(),%-: &'%)%0%6#

0%55# ,6 &#*) =)'# 2.8-, &(' *- &'.8>#

'.$#(* $%*'0# %* 2#$),-(),%- 9&(' #>#5&6# )%*$ 6#$ &(?*#)$ #- &'%3#-(-0# 2# )#6 &'.8># BCD %*
2#$),-.$ / )#6 &'.8># BCD ! E+#$) *-# (&&6,0(),%- ?*, -+( &($ .). 0%*3#') 2(-$ 6( 6,)).'()*'# 0%55#
)#66# &*,$?*# 6#$ %&#'()#*'$ 0%-$,2F'#-) $%*3#-) ?*# 6#$ 1%)$ $%-) 3%6*5,-#*> #) &(' 0%-$.?*#-0#
6+.0G(-),66%--(7# -+( &($ *- 7'(-2 ,5&(0)! H%*$ (66%-$ 5%-)'#' ?*# 0%-)'(,'#5#-) / 0#))# &#-$.#:
6# ;(,) ?*# 6#$ 1%)$ $%,#-) 3%6*5,-#*> -+#>06*) &($ 6#$ #''#*'$ 2# 2.)#0),%- #) 2# 06($$#5#-)! "#

tel-00408745, version 1 - 3 Aug 2009

3%6*5# 2+*- 1%) 3%6*5,-#*> #$) #$),5. (3#0 *-# #''#*' '#6(),3# &6*$ ;(,I6# 5(,$ #- )#'5#$ 2#
3(6#*' (I$%6*#: 6+#''#*' #$) 7'(-2# #) &(' 0%-$.?*#-0# 6# 1%) &#*) ;(0,6#5#-) &($$#' &%*' *- (*)'#
1%) 3%6*5,-#*> $, 6# )(*> 2+.0G(-),66%--(7# -+#$) &($ 0G%,$, 0%-3#-(I6#5#-)! J(6G#*'#*$#5#-)
-%*$ )'%*3%-$ ?*+*- )(*> 2+.0G(-),66%--(7# .6#3. 9KLM 3%,'# &6*$ #$) -.0#$$(,'# &%*' 2.)#0)#'
#) 06($$#' 6#$ 1%)$ 3%6*5,-#*>! E# )(*> I(,$$# 2+*- %'2'# 2# 7'(-2#*' $, $#*6#5#-) ,6 $+(7,)
2+*-# 2.)#0),%- 2#$ 1%)$ 3%6*5,-#*> $(-$ 06($$#5#-) %* $, 2+(*)'#$ ,-;%'5(),%-$ $%-) 2,$&%-,I6#$
0%55# &(' #>#5&6# 6# -*5.'% 2# $.?*#-0# 2(-$ 6#$ #-)=)#$ 2#$ &(?*#)$!

!" #$%&'(')'*+,
"+.)*2# 2# 6+.0G(-),66%--(7# -.0#$$,)# $%*3#-) 2#$ %*),6$ )G.%',?*#$ &%*' &%*3%,' #$),5#' 6+#'N
'#*' 5%4#--# 6%'$ 2# 6+,-3#'$,%- #) 6# 6,#' (* )(*> 2+.0G(-),66%--(7#! "#$ $,5*6(),%-$ &#*3#-)
(,2#' 2(-$ 0#))# .)*2# 5(,$ .)(-) 2%--. 6# 7'(-2 3%6*5# 2# 2%--.#$: #66#$ 2#3,#--#-) )'F$ 0%5N
&6,?*.#$ / 5#))'# #- %#*3'#! O66#$ $%-) &6*)P) *),6,$.#$ &%*' 3(6,2#' 6#$ '.$*6)()$ )G.%',?*#$! A%*'
-%)'# (&&6,0(),%- -%*$ $%55#$ &('),$ 2#$ G4&%)GF$#$ $*,3(-)#$ Q
R <- 7'(-2 -%5I'# 2# 1%)$ 2%-) 6# 3%6*5# $*,) *-# 0#')(,-# 6%, 2# &'%I(I,6,).! A%*' 6#$
3(6#*'$ -*5.',?*#$: -%*$ (3%-$ &',$ 6( 6%, 2# A('#)% (3#0 *- 3%6*5# 5%4#- 2# D!S TI4)#$
&%*' 6#$ 1%)$ UN)*&6# #) KV!V TI4)#$ &%*' 6#$ 1%)$ 2.8-,$ / I($# 2* &'.8># 2#$),-(),%- BCD!
"( 6%, 2# A('#)% #$) 0%--*# &%*' (*)(-) *-# I%--# '#&'.$#-)(),%- 2* 3%6*5# 2#$ 1%)$ 2(-$
6+@-)#'-#) #) 2(-$ 6+,-;%'5(),?*# #- 7.-.'(6 WXXY!
R "+.0G(-),66%--(7# #$) $*&&%$. *-,;%'5# ,-2.&#-2(-) 2* 0%-)#-* 2#$ &(?*#)$ #) 2+*- )(*>
5%4#-

p!

R <- %&#'()#*' #$) ,-).'#$$. &(' 2.)#0)#' #) 06($$#' 6#$

1%)$ 6#$ &6*$ 3%6*5,-#*>! E#))#

G4&%)GF$# $#'( '#6(>.# *6).',#*'#5#-)!

A%*' .3(6*#' 6+,5&(0) 2# 6+.0G(-),66%--(7#: ,6 -%*$ ;(*) *-# 5#$*'# 2# &#';%'5(-0#! H%*$ 6(
2.8-,$$%-$ 0%55# 0#0,! H%*$ ;%'5%-$ )%*)#$ 6#$ &(,'#$ 2# 1%)$ 2%-) 6# &'#5,#' #$) *- 2#$

1%)$

!"! #$%&'()(% *&+$,-.&/%

"#$%&'()%* )+ $) ,)%*'-&) %( .#+ /%)$0#(/%)1 2 0340%() ,) 0)5 64'7)58 (#%5 455#0'#(5 %( 6#',5
9 5' $)5 ,)%* .#+5 5#(+ &4$ 0$455:5 $;%( 647 7466#7+ < $;4%+7) 467-5 $;:034(+'$$#((4=) )+ %( 6#',5
> 5;'$5 5#(+ ?')( 0$455:51 @#+7) &)5%7) ,) 6)7A#7&4(0) )5+ 4$#75 +#%+ 5'&6$)&)(+ $4 5#&&) ,)5
6#',5 ,) +#%+)5 $)5 64'7)5 4'(5' A#7&:)51 B' $4 ,:+)0+'#( )+ 0$455)&)(+ 5) 6455)(+ 54(5 67#?$-&)8
4$#75 $4 &)5%7) ,) 6)7A#7&4(0) )5+ :=4$) < C:7#8 5'(#( 54 "4$)%7 ,#(() %() ',:) 5%7 $;'&6#7+4(0)
,% ,:5#7,7) ,4(5 $4 $'5+) ,)5 t .#+5 04%5: 647 $;:034(+'$$#((4=)1 D47 )*)&6$) %() =74(,) "4$)%7
")%+ 5'&6$)&)(+ ,'7) /%;%( ,)5 .#+5 "#$%&'()%* )5+ ,)")(% +7#6 6)+'+ 5%'+) < $;:034(+'$$#((4=)1
E)++) &)5%7) ,) 6)7A#7&4(0) 4 $) =74(, 4"4(+4=) /%;)$$) 6)%+ 5) 04$0%$)7 4(4$F+'/%)&)(+1
@#7&4$)&)(+ 034/%) 7:4$'54+'#( ,% 67#0)55%5 ,;:034(+'$$#((4=) ,#(() $')% < %() "4$)%7 6#%7
0)++) &:+7'/%)1 @#%5 5#&&)5 '(+:7)55:5 647 54 "4$)%7 &#F)(() 5%7 +#%+)5 $)5 7:4$'54+'#(51 D#%7
0)$4 (#%5 67#0:,#(5 0#&&) 0)0' G
H @#%5 04$0%$#(5 ,;4?#7, $4 67#?4?'$'+: /%;%( .#+ ,) +4'$$) /%)$0#(/%) 5#'+ ,4(5 $4 $'5+) ,)5

tel-00408745, version 1 - 3 Aug 2009

t .#+5 $)5 6$%5 "#$%&'()%*1

H I(5%'+) (#%5 04$0%$#(5 $4 67#?4?'$'+: /%;%( .#+ ,4(5 $4 $'5+) ,)5 t .#+5 5#'+ &4$ 0$455: 647
7466#7+ < %( 4%+7) .#+ ,) +4'$$) /%)$0#(/%)1
H J$ 5%K+ 4$#75 ,) &%$+'6$')7 0)++) 67#?4?'$'+: 647 $) (#&?7) +#+4$ ,) 64'7)5 6#%7 +7#%")7 $4
&)5%7) /%;#( 03)703)1 E) ,)7(')7 (#&?7) )5+ 5'&6$)&)(+ %() A#(0+'#( ,) t )+ ,% (#&?7)
,) +#+4$ ,)5 .#+5 ,'56#('?$)5 ,4(5 $4 +740)1 D#%7 %( (#&?7) +#+4$ ,) .#+5 :=4$ < N 8 $)
(#&?7) ,) 64'7)5 )5+ :=4$ < (2N t 1)t/21
L4 ,:+)0+'#( ,)5 t .#+5 $)5 6$%5 "#$%&'()%* 54(5 $) ?)5#'( ,) $)5 0$455)7 $)5 %(5 647 7466#7+
4%* 4%+7)5 5) A4'+ 674+'/%)&)(+ ,) $4 &M&) &4('-7)1 J$ A4%+ A4'7) 4++)(+'#( < 0) /%) $) (#&?7)
,) 64'7)5 647 $)/%)$ '$ A4%+ &%$+'6$')7 )5+ &4'(+)(4(+ N.t )+ ,#(0 )5+ '(A:7')%7 4% 045 67:0:,)(+1

!" #$%&'()(% *&+$,-.&/%

D#%7 $4 "4$',4+'#(8 (#%5 5#&&)5 647+'5 ,)5 +740)5 7:)$$)5 )+ (#%5 4"#(5 A#7&: $)5 .#+5 ,),4(51
J$ 5;4='+ ,)5 +740)5 4% ('")4% 64/%)+ 0#$$)0+:)5 7)56)0+'")&)(+ 5%7 %( $')( ,% 7:5)4% 4&:7'04'(
2NJLI@I O9P )+ %( $')( ,;400-5 ,% 7:5)4% A74(Q4'5 RI@2SIR OTUP1 D#%7 034/%) +740)8 (#%5
4"#(5 04$0%$: $) (#&?7) ,) .#+58 $)%7 "#$%&) &#F)( )+ $) 6474&-+7) ,) $4 $#' D47)+# /%' ,:07'+ $4
/%)%) V6474&-+7) W1 E)5 "4$)%75 #(+ :+: 647 $4 5%'+) %+'$'5:)5 ,4(5 (#+7) &#,-$) 6#%7 +7#%")7 $)
(#&?7) &#F)( ,) 64'7)5 ,) .#+5 &4$ 0$455:58 0) /%' 0#(5+'+%) (#+7) &)5%7) ,) 6)7A#7&4(0) 6#%7
$4 67:0'5'#( ,) $4 ,:+)0+'#( )+ ,% 0$455)&)(+ 467-5 :034(+'$$#((4=)1 E)++) #6:74+'#( 4 :+: 7:6:+:)
6#%7 6$%5')%75 "4$)%75 ,% +4%* ,;:034(+'$$#((4=) )+ 6#%7 6$%5')%75 "4$)%75 ,) t V$) (#&?7) ,)
3)4"F 3'++)75 )( /%)5+'#(W1 I( 6474$$-$)8 (#%5 4"#(5 466$'/%: $;:034(+'$$#((4=) ,'7)0+)&)(+ 5%7
$4 +740) 4% ('")4% 64/%)+ )+ (#%5 4"#(5 04$0%$: (#+7) &:+7'/%) ,'7)0+)&)(+ 54(5 6455)7 647 $)
&#,-$)1 S#%5 $)5 7:5%$+4+5 5#(+ ,:07'+5 ,4(5 2(()*) N ,4(5 ,)5 X=%7)5 /%' &#(+7)(+ $;)77)%7
&#F)(() )( A#(0+'#( ,% +4%* ,;:034(+'$$#((4=) 5%7 %() :03)$$) $#=Y$#=1 Z4(5 0) +74"4'$8 (#%5
4"#(5 0#(5',:7: %( =74(, (#&?7) ,) 0#&?'(4'5#(5 G 0$455)&)(+ "51 ,:+)0+'#( 5)%$)8 .#+5 SED

!"#$%&' () *' +&,$% $-%&./,$% #"& 0123!"-%$00.--"4' 5 /2%'3%$.- '% 30"66'7'-% /'6 8.%6 9.0,7$-',:
Abilene trace, N = 487K, beta = 2, blind ranking

Jussieu trace, N = 103K, beta = 1.5, blind ranking

1e+06

t = 25
t = 10
t=5
t=2
t=1

1e+06
100000
10000
1000
100
10
1
0.1
0.01

Average number of misranked flow pairs

Average number of misranked flow pairs

1e+07

0.001

t = 25
t = 10
t=5
t=2
t=1

100000
10000
1000
100
10
1
0.1
0.01
0.001

0.1

!"#

1
10
Packet sampling rate (%)

0.1

!" # $%%&'% ()*&++& ,& ,-.&/.0)+ &. /1233&(&+. ,&3 4).3 5)1'(0+&'6
Abilene trace, N = 487K, beta = 2, blind detection

Jussieu trace, N = 103K, beta = 1.5, blind detection

1e+06

t = 25
t = 10
t=5
t=2
t=1

1e+06
100000
10000
1000
100
10
1
0.1
0.01

Average number of misranked flow pairs

Average number of misranked flow pairs

1e+07

tel-00408745, version 1 - 3 Aug 2009

1
10
Packet sampling rate (%)

t = 25
t = 10
t=5
t=2
t=1

100000
10000
1000
100
10
1
0.1
0.01
0.001

0.1

10

Packet sampling rate (%)

!"#

0.1

10

Packet sampling rate (%)

!7 # $%%&'% ()*&++& ,& ,-.&/.0)+ 3&'1& ,&3 4).3 5)1'(0+&'6

53! 4).3 87 9 ,0:-%&+.3 +)(;%&3 ,& 4).39 ,0:-%&+.&3 521&'%3 <)'% 1& <2%2(=.%& ,& 12 1)0 >2%&.)9
&./! ?0@'%& !" ()+.%& '+ &6&(<1& ,& /&3 %-3'1.2.3 <)'% 1& /23 ,& 4).3 AB> &. ,-.&/.0)+ ,&3 4).3
5)1'(0+&'6! C2 D@'%& E @2'/F& /)%%&3<)+, E 12 .%2/& G;01&+& &. /&11& E ,%)0.& E 12 .%2/& H&+2.&%!
?0@'%& !7 ()+.%& 1& (I(& &6&(<1& (203 ,2+3 1& /23 ,J'+& ,-.&/.0)+ 3&'1& 32+3 /1233&(&+.! K)3
/)+/1'30)+3 <&'5&+. 3& %-3'(&% /)((& 3'0. L
# M+ .2'6 ,J-/F2+.011)++2@& -1&5- &3. 3)'5&+. %&N'03 &. /& .2'6 2'@(&+.& 25&/ 1& +)(;%& ,&
4).3 E ,-.&/.&% &.8)' /1233&%! >)'% 1& 4). )' 1&3 ,&'6 4).3 1&3 <1'3 5)1'(0+&'6 '+ .2'6
,J-/F2+.011)++2@& ,& 1J)%,%& ,& "O <&'. 3'P%& E 1&3 ,-.&/.&% &. 1&3 /1233&%! Q=3 N'& /&
+)(;%& 2'@(&+.&9 1& .2'6 ,J-/F2+.011)++2@& %&N'03 2'@(&+.& ,J'+& R2S)+ &6<)+&+.0&11&!
>2% &6&(<1& 12 ,-.&/.0)+ &. /1233&(&+. ,&3 ,06 4).3 1&3 <1'3 5)1'(0+&'6 +-/&330.& '+ .2'6
,J-/F2+.011)++2@& ,& 1J)%,%& ,& TUO 5)0%& <1'3! B&/0 5&'. ,0%& .)'. 30(<1&(&+. 1J0+&P/2/0.,& 1J-/F2+.011)++2@& <)'% /& @&+%& ,J2<<10/2.0)+3!
# C2 ,-.&/.0)+ 3&'1& %-,'0. 1& .2'6 ,J-/F2+.011)++2@& %&N'03 ,J'+ )%,%& ,& @%2+,&'% <)'% 1&
(I(& +)(;%& ,& 4).3 <2% %2<<)%. E 12 ,-.&/.0)+ &. /1233&(&+.! V203 /& .2'6 %&3.& N'2+,
(I(& .%=3 -1&5-!
# G'@(&+.&% 1& +)(;%& ,& 4).3 2'@(&+.& &+ @-+-%21 12 <%-/030)+ <'03N'& 12 /F2+/& ,J25)0%

!"! #$%&'((' )*+,$-' -' .(/00')'1+2-*+'.+3$1 4/0*' 0%5 -'0 316$5)/+3$10 75$+$.$(/35'0
Abilene trace, N = 487K, beta = 2, protocol-aware ranking

Jussieu trace, N = 103K, beta = 1.5, protocol-aware ranking

1e+06

t = 25
t = 10
t=5
t=2
t=1

1e+06
100000
10000
1000
100
10
1
0.1
0.01
0.001

Average number of misranked flow pairs

1e+07
Average number of misranked flow pairs

t = 25
t = 10
t=5
t=2
t=1

100000
10000
1000
100
10
1
0.1
0.01
0.001
0.0001

0.1

1
10
Packet sampling rate (%)

0.1

1
10
Packet sampling rate (%)

!"#

"# $ %&&'(& )*+',,' -' ./011')',2 -'1 3*21 4*/()5,'(6 ', (25/510,2 /' ,()7&* -' 178(',.'
-' /0 .*(.9' 2&0,1:*&2

tel-00408745, version 1 - 3 Aug 2009

-'1 3*21 ',.*&' :/(1 4*/()5,'(6 0(;)',2'" <&0258(')',2 2*(2'1 /'1 /*51 -' -512&5=(25*, -'
4*/()' 10251>*,2 .'22' :&*:&5727" ?'1 /*51 8(5 ,' 10251>*,2 :01 .'22' :&*:&5727 1*,2 .'//'1 8(5
2&*,8(',2 /' 4*/()' -'1 3*21 4*/()5,'(6"
$ ?0 :&7.515*, 0(;)',2' 8(0,- /0 8('(' -' /0 /*5 -' -512&5=(25*, -( 4*/()' -'1 3*21 -'@
45',2 :/(1 /*(&-'" A'.5 '12 (,' .*,178(',.' ,02(&'//' -( >052 8(' /'1 3*21 -'45',,',2 :/(1
4*/()5,'(6 '2 :/(1 '1:0.71 ', 2'&)'1 -' 4*/()'"

!" #$%&'((' )*+,$-' -' .(/00')'1+2-*+'.+3$1 4/0*' 0%5 -'0 316

7$5)/+3$10 85$+$.$(/35'0
?B(25/51025*, -'1 5,>*&)025*,1 :&*2*.*/05&'1 0 2*(C*(&1 727 .*,15-7&7' .*))' (, )*+', 'D.0.'
:*(& &7-(5&' /B'&&'(& .0(17' :0& /B7.90,25//*,,0;'" <0& '6'):/' /' >052 -' .*):2'& /'1 :08('21 EA<
8(5 :*&2',2 /' FGH 30; '2 -' 1(::*1'& 8(' .908(' :08('2 FGH .*&&'1:*,- I (, ,*(4'0( 3*2
:'&)'2 (,' '125)025*, &0:5-' '2 'D.0.' -( ,*)=&' -' 3*21 -0,1 /' &71'0(" J0,1 .'22' )K)'
-5&'.25*, ,*(1 04*,1 &';0&-7 15 /0 .*,,05110,.' -( ,()7&* -' 178(',.' -51:*,5=/' -0,1 /B',2K2'
2&0,1:*&2 :'(2 05-'& ,*2&' 0::/5.025*," %, 1*(12&0+0,2 /' ,()7&* -' 178(',.' -( :&')5'& :08('2
7.90,25//*,,7 -' .'/(5 -( -'&,5'& :08('2L *, 2&*(4' /' ,*)=&' -'1 :08('21 ',4*+71 ',2&' /'1 -'(6
'2 *, (25/51' .' ,*)=&' :*(& /' ./011')',2 '2 /0 -72'.25*," ?B'&&'(& -0,1 .'22' '125)025*, 45',2 -'1
:08('21 ,*, 4(1 :0& /B7.90,25//*,,0;' 0( -7=(2 '2 I /0 M, -( 3*2" N:&'1 0C(12')',2 -( )*-O/' I
.'22' ,*(4'//' )729*-' ,*(1 04*,1 4( 8(B'//' 72052 .0:0=/' -' &7-(5&' /' 20(6 -B7.90,25//*,,0;'
-B(, *&-&' -' ;&0,-'(& .*):0&7 I /0 )729*-' :&7.7-',2'" ?0 P5;(&' "# /' )*,2&' :*(& /'1 -'(6
2&0.'1 '2 :*(& /' .01 -72'.25*, '2 ./011')',2 -'1 3*21 4*/()5,'(6"
Q0/;&7 1*, 5,27&K2L .'22' )729*-' 1*(R&' )0/9'(&'(1')',2 -' :/(15'(&1 :&*=/O)'1 8(5 :'(4',2
/0 )'22&' ', .0(1'" JB(, .S27L '//' &'8(5'&2 /B0..O1 I /B',2K2' 2&0,1:*&2 .' 8(5 45*/' /' :&5,.5:' -'1
.*(.9'1 '2 :0& .*,178(',.' /B'):K.9' -' >*,.25*,,'& ', .01 -' :08('21 .&+:271" J' /B0(2&' .S27L
'//' :&71',2' -'1 :&*=/O)'1 -' >*,.25*,,')',2 ', .01 -' &'2&0,1)5115*,1L )0(4051 *&-*,,0,.'@

!"#$%&' () *' +&,$% $-%&./,$% #"& 0123!"-%$00.--"4' 5 /2%'3%$.- '% 30"66'7'-% /'6 8.%6 9.0,7$-',:

"#$% &#' ()*+#%' &)$' ,# -.'#)+ #% -.+%/,/')%/0$ &+ $+".-0 &# '.*+#$1#2 30""# 0$ ,# "0$%-#
&)$' ,45$$#6# 78 &#' ),90-/%:"#' (#+;#$% <%-# /$%-0&+/%' (0+- &.%#1%#- 1#' &#-$/#-' (-0=,>"#'
#% -.&+/-# ,#+- /"()1%2

"#$%&'()#$(
?)$' 1# 1:)(/%-# $0+' );0$' "0$%-. '+- +$ #6#"(,# 10$1-#% ,4/"()1% &# ,4.1:)$%/,,0$$)9#

'+- ,) *+),/%. &4+$# %@1:# &# "#'+-# /"(0-%)$%# *+/ #'% 1#,,# &# ,) &.%#1%/0$ #% 1,)''#"#$% &#'
A0%' ;0,+"/$#+62 B0+' );0$' "#$. +$# .%+&# )$),C%/*+# )110"()9$.# &# '/"+,)%/0$' ),/"#$%.#'
()- &#' %-)1#' -.#,,#'2 3#%%# .%+&# D)/%# '+- +$ ,/#$ &+ -.'#)+ (#+% '4.%#$&-# E ,4/$%.9-),/%. &+
-.'#)+ #$ (-0F%)$% &+ 1)&-# &# %-);)/, /$%-0&+/% &)$' ,# 1:)(/%-# (-.1.&)$%2 34#'% +$ &# $0'
0=G#1%/D' (0+- ,# D+%+-2 H$ )+%-# 0=G#1%/D #'% ,4.%+&# &# ,4/"()1% &# ,4.1:)$%/,,0$$)9# '+- &4)+%-#'

tel-00408745, version 1 - 3 Aug 2009

)((,/1)%/0$' 10""# ,) &.%#1%/0$ &4)$0"),/#' 0+ ,) 1,)''/F1)%/0$ &#' )((,/1)%/0$' E ()-%/- &#'

&0$$.#' '%)%/'%/*+#' '+- ,# %-)F12

!"#$%&' (
! "#$$%&!'(#) '*+,#$*&&* !- .*$/("* 0*
&1%"2!)'(&&#))!3*
tel-00408745, version 1 - 3 Aug 2009

!"#$%%! #&' (&&)*) +

! "#$$%&!'(#) *+' ,)* -#))%* (./#$'!)'* -!)+ &! .%'$#&#0(*1 2&&* /*$.*' -* '($*$ -*+ "#)"&,3
+(#)+ (./#$'!)'*+ *) "#./!$!)' -*+ .*+,$*+ *4*"',%*+ +,$ -*+ 5,!)'('%+ "#$$%&%*+1 6*''* "#.3
/!$!(+#) /*,' -#))*$ -*+ ()7#$.!'(#)+ +,$ &! "#./#+!)'* &(!)' "*+ 5,!)'('%+ "#$$%&%*+8 "#..*
*&&* /*,' -#))*$ ,)* (-%* +,$ &! /$%"(+(#) -*+ .*+,$*+ *4*"',%*+1 *+ !//&("!'(#)+ +#)' -(9*$+*+8
)#,+ "('#)+ &*+ +,(9!)'*+ :
; ! "#$%&'()*+,#* -) ./'0!0 -1 &'+)!1 2 %!&0,& -) $)+1&)+ !"0,3)+4 2) *4*'8 *' "#..* -('
-!)+ &*+ "<!/('$*+ /$%"%-*)'+8 &*+ +#)-*+ !"'(9*+ .*+,$*)' -*+ "<*.()+ *)'(*$+1 2' "#..*
&*+ "<*.()+ -* &=>)'*$)*' +* "<*9!,"<*)'8 (& ? ! &=!//!$('(#) -=,)* "#$$%&!'(#) 5,( /*,' +*$9($
@ .(*,A *A/&(5,*$ &*+ $%+,&'!'+ -*+ .*+,$*+1 B!$ *A*./&* C>D6 EFG ,'(&(+* "*''* /$#/$(%'%
/#,$ "!$!"'%$(+*$ &*+ &(*)+ -=,) $%+*!, @ /!$'($ -*+ .*+,$*+ !"'(9*+1 H) /*,' /!$&*$ -!)+ "*
"!+ -* "#$$%&!'(#) +/!'(!&*1 * $*+'* -*+ !//&("!'(#)+ *A/&#('*)' &! "#$$%&!'(#) '*./#$*&&*1
; ! %&'-,"0,#* 5,( /*$.*' @ /!$'($ -*+ .*+,$*+ !"',*&&*+ *' -*+ ()7#$.!'(#)+ +,$ &! "#$$%&!'(#)
-=*+'(.*$ *' -* /$%-($* &* 7,',$1 6*"( /*$.*' -=!)'("(/*$ !9*" &*+ $%!"'(#)+ "#)9*)!I&*+1 *
0&!5" )*6,*))&,*6 J&=()0%)(*$(* -, '$!K"L *4*"',% /!$ &*+ #/%$!'*,$+ /*,' /$#K'*$ -* "*''*
!//&("!'(#) *) /!$'!)' -*+ .#-M&*+ I(*) "#)),+ /#,$ &* '$!K" *101 ENO8 PQG1
; ! -'0)"0,#* -/!*#$!.,)+ 5,( ,'(&(+* -*+ '*"<)(5,*+ +(.(&!($*+ @ "*&&* -* &! /$%-("'(#)1 B#,$
&! /$%-("'(#)8 (& +=!0(' -* '$#,9*$ &! 9!&*,$ &! /&,+ /$#I!I&* -!)+ &* 7,',$1 B#,$ &! -%'*"'(#)
-=!)#.!&(*+8 (& +=!0(' -* '$#,9*$ &! 9!&*,$ &! /&,+ /$#I!I&* /#,$ &* /$%+*)' *' -* &! "#./!$*$
@ &! 9!&*,$ .*+,$%*1 R9*" ,)* (-%* +,$ &* '!,A -=*$$*,$ .#?*) "!&",&% /*)-!)' ,)* /<!+*
-* "!&(I$!0*8 (& -*9(*)' !&#$+ 7!"(&* -* "#./!$*$ &! 9!&*,$ .*+,$%* @ &! 9!&*,$ *+'(.%* *'
-* "#)"&,$* +=(& +=!0(' -=,)* 0$!)-* -%9(!'(#) #, /!+1 S)* 0$!)-* -%9(!'(#) *+' !++(.(&%*
@ ,)* !)#.!&(*1 6*''* .%'<#-* /*,' +=!//&(5,*$ !, '$!K" -* &=>)'*$)*' /#,$ -%'*"'*$ &*+
!''!5,*+ "#)'$* &*+ .!"<()*+ EFT8 UVG "#..* *&&* /*,' +=!//&(5,*$ @ '#,' !,'$* +?+'M.*
/#,9!)' /$%+*)'*$ ,) "#./#$'*.*)' !)#$.!&1 S) *A*./&* '?/(5,* *+' )#'$* "#)'$(I,'(#)
$%"*)'* +,$ &! -%'*"'(#) -*+ !''!5,*+ "#)'$* &*+ $%+*!,A -* "##$-#))%*+ 9($',*&&*+ EUWG1
XP

!"#$%&' () *" +,&&-."%$,/ %'0#,&'..' "1 2'&3$+' 4' .5-+!"/%$..,//"6'

!
"

*" +,&&'+%$,/ 4'2 0'21&'2) #$ %&'($ )$ *+&,+- $.' &/$0$,1&$ '21345$ )/5-$ '$&&$ +11&36+7
'38- 9: ;< =>+45$ 81?(+'38- )$ ,$.5($ $.' @?-?(+&$,$-' +668,1+@-?$ )/5-$ 6$('+3-$ $(($5(<
=$''$ $(($5( 1$5' A'($ 6+5.?$ 1+( 5- B(53' $0'?(3$5( 68,,$ 6/$.' &$ 6+. &8(. )$ &+ ,$.5($
)$ &+ B+-)$ 1+..+-'$< C+3. &/$(($5( 1$5' +5..3 D$-3( )5 E+3' 45/8- 8B.$(D$ 5-$ 1+('3$ )5
.2.'F,$ $' 45/8- $..+3$ )/3-E?($( &$ @&8B+& G 1+('3( )/8B.$(D+'38-. 1+('3$&&$.< #/?6>+-'3&&8-7
-+@$ )5 '(+%6 +11+('3$-' G 6$''$ )$503F,$ 6+'?@8(3$< H85( ?&3,3-$( &$ B(53' $' 68((3@$(
&$. ,$.5($.I )$50 6&+..$. )$ ,?'>8)$. $03.'$-'< #+ 1($,3F($ 68-.3.'$ G (?1?'$( &+ ,A,$
,$.5($ 1&5.3$5(. E83. $' G 1(86?)$( 1+( )$. ,?'>8)$. .'+'3.'345$. 185( -$ @+()$( 45$ &$.
D+&$5(. &$. 1&5. 1(8B+B&$. 85 &$. ,82$--$.< =$''$ ,?'>8)$ ./+11&345$ .85D$-' G )$. 18157
&+'38-. 453 -$ D+(3$-' 1+. $- E8-6'38- )$ '$,1.< #/+5'($ ,?'>8)$I 453 $.' 6$&&$ +)81'?$ 1+(
&$ %&'($ )$ *+&,+- $' &$. +5'($. %&'($. &3-?+3($. 85 -8-7&3-?+3($.I 68-.3.'$ G 1(8%'$( )$ &+
68((?&+'38- )5 .2.'F,$ G ,$.5($( 185( 68((3@$( &$. ,$.5($.< J &+ )3K?($-6$ )$ &+ 1($,3F($

tel-00408745, version 1 - 3 Aug 2009

,?'>8)$I &+ )$503F,$ ,?'>8)$ 1(8%'$ )$. ,$.5($. 1(?6?)$-'$. 185( +,?&38($( &+ 1(?63.38)$ &+ ,$.5($ +6'5$&&$< #/3)?$ $.' .3,1&$< L5+-) 5-$ ,?'(345$ $.' 68((?&?$ +D$6 $&&$ ,A,$
M+5'8768(($&+'38-N 85 +D$6 5-$ +5'($ ,?'(345$I .+ D+&$5( $.' $- 1+('3$ )?63)?$ 1+( &$. D+&$5(.
1+..?$.I $' 6$63 )?1$-) )$ &/3,18('+-6$ )$ &+ 68((?&+'38-< H+( $0$,1&$ )+-. &$ 6+. )/5-$
68((?&+'38- 18.3'3D$I .3 &$ 1+..? $.' @(+-) 3& 2 + E8('$ 6>+-6$ 45$ &$ 1(?.$-' .83' +5..3 @(+-)
$' &/3-D$(.$ $.' +5..3 D(+3< H+( 68-.?45$-6$I &$ E+3' )$ 68--+O'($ &$. D+&$5(. 1+..?$. +3)$ G
,3$50 $.'3,$( &+ D+&$5( )$ &+ ,?'(345$ )+-. &$ 1(?.$-' 68,1+(? +5 6+. 8P 8- -$ 68--+O'
(3$- .5( &$ 1+..? $' 8P &$ .$5& ,82$- )3.18-3B&$ $.' )/$.'3,$( &+ ,?'(345$ 1+( .+ D+&$5(
,82$--$ 85 .+ )3.'(3B5'38- ,+(@3-+&$< Q85D$-' &$. 85'3&. 1(86F)$-' 68,,$ .53' R
" J 1+('3( )$. ,$.5($. $' )$. $.'3,+'38-. 1(?6?)$-'$. $' )$ &+ 68((?&+'38-I $.'3,$( 5-$
D+&$5( 185( &$ 1(?.$-'<
" C$.5($( &$ 1(?.$-' $' (+S-$( &/$.'3,+'38- $- 1($-+-' 5-$ 1+('3$ )5 1(?.$-' $' 5-$ 1+('3$
)5 1+..?< #$ 183). $-'($ &$ 1+..? $' &$ 1(?.$-' $.' 1+(E83. +11$&? &$ @+3- )$ &/$.'3,+'$5(
$' )?1$-) )$ &/8BT$6'3E G +''$3-)($ M,3-3,3.$( &/$(($5( 45+)(+'345$ ,82$--$I ?&3,3-$(
6$('+3-$. E(?45$-6$. $' $- @+()$( )/+5'($.I $'6N<
" JT85'$( &+ ,$.5($ $' $.'3,+'38- )5 1(?.$-' G &/>3.'8(345$ $' ($D$-3( +5 183-' U<
H85( ,3$50 68,1($-)($ 6$''$ ,?'>8)$I 1($-8-. &$ 6+. .3,1&$ )$ &/$.'3,+'38- )5 )?&+37+&&$(
($'85( ,82$- )+-. V=H M.85D$-' +11$&? WVV 185( W85-)7V(31 V3,$N< #$ WVV $.' 5-$
,?'(345$ 453 D+(3$ G )$50 ?6>$&&$. )$ '$,1. R X-$ 1$'3'$ ?6>$&&$ )$ '$,1. 8P &+ D+(3+'381$5' A'($ 68-.3)?(?$ 68,,$ +&?+'83($ -8- 68((?&?$ 6+5.?$ 1+( &+ -+'5($ 1+45$' )$. (?.$+50<
Y' 5-$ +5'($ @(+-)$ ?6>$&&$ )$ '$,1. )?63)?$ 1+( &/?D8&5'38- )$ &+ 68-@$.'38- )+-. &$ (?.$+5
$' 453 3-'(8)53' 5-$ 6$('+3-$ 68((?&+'38- )+-. &$ 1(86$..5.< V=H D3.$ G 6+&65&$( 5-$ D+&$5(
,82$--$ )5 WVV 453 .53' &+ 1+('3$ 68((?&?$< H85( 6$&+ 3& 1(86F)$ 1+( &/5'3&3.+'38- )/5%&'($ 1+..$ B+. )$ '21$ ,82$--$ @&3..+-'$ +D$6 183). $018-$-'3$& MY018-$-'3+&&2 Z$3@>'$)
C8D3-@ JD$(+@$N 9[:;< #8(. )$ 6>+45$ -85D$&&$ ,$.5($ )$ WVV &/$.'3,+'$5( )5 WVV ,82$E8(,$ 5-$ -85D$&&$ $.'3,+'38- $- 1($-+-' &$ \]! )$ &/+-63$--$ $.'3,+'38- $' &$ U]! )$ &+

!"! #$%&'(%&)* +, -( %(&--, +, ./(*+,$ 0)01-(%&)*$

"#$%&# '()%#**# #) #+ ('*(%*'+) #+$%,)# *' $-""#. /#(, 0#&"#) 123*,",+#& *#$ -$(,**'),-+$
1# (-%&)#$ 03&,-1#$ #) 1# 4'&1#& (#**#$ 1# *-+4%#$ 03&,-1#$. 5, (#))# (-&&3*'),-+ +2#6,$)',)
0'$7 8/9 '%&',) :+, 0'& ('*(%*#& %+# ;'*#%& "-<#++# $,"0*# =%, +# (>'+4# 0'$ ?#'%(-%0
#+ @-+(),-+ 1# )#"0$ #) =%# $, #**# (>'+4#7 $-+ (>'+4#"#+) #$) ?#'%(-%0 "-,+$ $,4+,:('),@.
A-)#B ,(, =%# *' "#,**#%&# #$),"'),-+ 12%+ 0&-(#$$%$ +-+ (-&&3*3 #$) $' ;'*#%& "-<#++#7
%+# ;'*#%& =%, 0#%) C)&# ?,#+ *-,+ 1#$ ;'*#%&$ "#$%&3#$ $%&)-%) $, *# 0&-(#$$%$ 0&3$#+)# %+#
@-&)# ;'&,'?,*,)3. D#$ 0&-(#$$%$ (-&&3*3$ $# 0&C)#+) ",#%6 E *2#$),"'),-+.
F* $2'4,) 12%+ 1-"',+# )&G$ ,"0-&)'+) =%, $# $,)%# E *' *,",)# 1# *' )>3-&,# #) 1# *' 0&'),=%#. H#$
-%),*$ $-0>,$),=%3$ #6,$)#+) 1-+) *' )>3-&,# 1% $,4+'*7 1#$ 0&-(#$$%$ $)-(>'$),=%#$7 1% (-+)&I*#
#) 12'%)&#$7 #) =%, 0#%;#+) ';-,& %+ 4&'+1 ,+)3&C) *-&$ 1# *' (-+(#0),-+ #) -0),",$'),-+ 1#$
$-*%),-+$ #J('(#$ 0-%& *' "3)&-*-4,#. H# +-)&# (-)3 +-%$ ';-+$ (-+)&,?%3 1'+$ (# 1-"',+# E
)&';#&$ )&-,$ '00*,('),-+$7 =%# +-%$ +-%$ (-+)#+)-+$ 1'+$ (# (>'0,)&# 12#+ #60*,=%#& %+#. D#$

tel-00408745, version 1 - 3 Aug 2009

1#%6 '%)&#$ '00*,('),-+$ )&',)#+) &#$0#(),;#"#+) 1# *2#$),"'),-+ 1% 13*', 1'+$ *#$ 0*'<-%) ?%K#&$
1#$ '00*,('),-+$ '%1,- L!MN #) 1# *2#$),"'),-+ 1#$ (--&1-++3#$ +-",+'*#$ 12%+# "'(>,+# 1'+$ %+
&3$#'% 1# (--&1-++3#$ $-%",$ E 1#$ '))'=%#$ LOPN. D#$ (--&1-++3#$ +-",+'*#$ 0#&"#))#+) '0&G$
1# 13)#()#& $, 1#$ "'(>,+#$ ,+)#+),-++#**#"#+) '++-+(#+) 1#$ (--&1-++3#$ )&G$ 1,$)'+)#$ 1# (#
=%2#**#$ 1#;&',#+) @',&#.

!" #$%&'(%&)* +, -( %(&--, +, ./(*+,$ 0)01-(%&)*$

A-%$ ';-+$ 3)%1,3 (# 0&-?*G"# 1'+$ *# ('1&# 12'00*,('),-+$ 1# 1,K%$,-+ =%, 13$,&#+) (-++'Q)&#
E )-%) ,+$)'+) *# +-"?&# 1# 0#&$-++#$ =%, *#$ &#R-,;#+). /2#$) *# ('$ 0'& #6#"0*# 12%+ $#&;#%&
1,K%$'+) 1# *' ;,13- -% 1# *' )3*3 #+ "%*),('$) ';#( 1% S89 E %+ 4&'+1 +-"?&# 1# 0'&),(,0'+)$.
D' 1,K%$,-+ 0#%) $# @',&# $%& 1#$ $')#**,)#$ -% $%& 1#$ *,4+#$ TH5D. /#))# (-++',$$'+(# 1% +-"?&#
1# 0'&),(,0'+)$ #+ )#"0$ &3#* 0#%) $#&;,& E 3;'*%#& *' 0-0%*'&,)3 1#$ 1,K3&#+)$ ('+'%6 1# 1,K%$,-+
U#.4. (>',+#$ 8VW #) E ",#%6 1,"#+$,-++#& *# &3$#'% #) *#$ $#&;,(#$ 0'& (-+$3=%#+(#. H'+$ (#&)',+$
('$ #**# 0#%) '%$$, $#&;,& E 3=%,*,?&#& &#1-+1'+(# #) &#)&'+$",$$,-+$ 0-%& '$$%&#& *' :'?,*,)3 1#
*' 1,K%$,-+ U0*%$ 1# 0'&),(,0'+)$ (2#$) 0*%$ 1# &#)&'+$",$$,-+$ #) 0'& (-+$3=%#+(# 0*%$ 1# (>'&4#
(# =%, 0-%&&',) X%$),:#& *2%),*,$'),-+ 1# *' &#1-+1'+(#W. /#0#+1'+) 12'%)&#$ '00*,('),-+$ 0#%;#+)
'%$$, 0&-:)#& 1% ('*(%* 1# *' )',**# 1# 4&'+1#$ 0-0%*'),-+$. A-%$ (,)-+$ 0'& #6#"0*# *# ('*(%*
1% +-"?&# 1# 0',&$ =%, 0'&),(,0#+) E %+ 0'&)'4# 1# (-+)#+% ';#( Y,)8-&&#+) U#.4. *# ('$ 1#$
%),*,$')#%&$ "#))'+) E X-%& *#%& ;#&$,-+ 1# Z5 [,+1-\$W7 *# ('*(%* 1% +-"?&# 1# ]-)$ $%& %+
*,#+ -% 1'+$ %+ &3$#'%7 *# ('*(%* 1% +-"?&# 1# "'(>,+#$ =%, $-+) &3;#,**3#$ 1'+$ %+# 4&'+1#
0*')#@-&"#7 #)(.
H'+$ +-)&# )&';',* +-%$ 3),-+$ (-+@&-+)3$ '% 4&'+1 ;-*%"# 1# 1-++3#$ =%, 0#%) &3$%*)#& 12%+
('*(%* #6'() 1# *' )',**# 12%+# 4&'+1# 0-0%*'),-+. ^+ ('*(%* #6'() +3(#$$,)# =%# (>'=%# "'(>,+#
#+;-,# 03&,-1,=%#"#+) %+ "#$$'4# _>#**-_ E *' $-%&(# #) =%# *' $-%&(# @'$$# *' $-""# 1# )-%$ (#$
"#$$'4#$. /#(, 0#%) ('%$#& %+# ,"0*-$,-+ E *' $-%&(# =%'+1 *#$ &3(#0)#%&$ $-+) +-"?&#%6. ^+#

!"#$%&' () *" +,&&-."%$,/ %'0#,&'..' "1 2'&3$+' 4' .5-+!"/%$..,//"6'

!"#

3)0(&
video1
video2
video3
video4

"# $ %&'()*+,*-. /&' ,)0(&' 1*/2-'

4-.56&6) /& 70 ,)0(&
1/(s)
d
h
m
s
3 13 33 20
89": #;<#=
11d 1h 46m 8s
#9"# #=9:=
d
h
m
s
50 22 13 20
;"#
==;><
29d 16h 43m 13s
#:"8 ;<<8!

'-76,*-. +-6))0*, ?,)& 7@&'+0(&A&., /&' A&''05&' /0.' 7& ,&A+' A0*' (&(* (-./6*, B 6. A0.C6&
/& +)2(*'*-. (06'2 +0) 7@21-76,*-. ,&A+-)&77& /& 70 +-+670,*-." DE. /& )2'-6/)& (& +)-F7GA&H
.-6' '-AA&' +0''2' +0) 6.& '-76,*-. +)-F0F*7*',& /0.' 70C6&77& 7&' )2(&+,&6)' '-., /&A0./2'
/@&.1-I&) +2)*-/*C6&A&., 7&6)' A&''05&' JKL44MJ A0*' 01&( 6.& (&),0*.& +)-F0F*7*,2 C6* +&6,

tel-00408745, version 1 - 3 Aug 2009

?,)& (N-*'*& &. O-.(,*-. /& 70 +)2(*'*-. 1-676& &, /& 70 ,0*77& /& 70 +-+670,*-." D++&7-.' (&,,&
+)-F0F*7*,2 p &, 0++&7-.' S 70 +2)*-/& /@&.1-*" D7-)' ,-6,&' 7&' S '&(-./&'H (N0C6& )2(&+,&6) ,*)&
/& '-. (-,2 6. .-AF)& 0720,-*)& &.,)& ! &, # &, &.1-*& '-. A&''05& JKL44MJ '* (& .-AF)& ,*)2
&', *.O2)*&6) B p P(&(* &', 2C6*107&., B (& C6@-. 0++&77& &.1-* 01&( +)-F0F*7*,2 pQ" L. +)-(2/0.,
(-AA& (&(*H 7& .-AF)& /& A&''05&' &.1-I2' &', )2/6*, +0) 6. O0(,&6) /& 1/p" 40 C6&',*-. B (-AR
+)&./)& &', C6&7 *A+0(, 0 (&,,& )2/6(,*-. '6) 70 +)2(*'*-. /& 7@&',*A0,*-. &, (-AA&., O0*)& +-6)
*.1&)'&) 7& +)-F7GA& 01&( 7& A*.*A6A /@&))&6)" S-,)& ,)010*7 +6F7*2 /0.' T H =U &, /2,0*772 /0.'
D..&V& W &''0*& /& )2+-./)& B (&,,& C6&',*-." X7 '@05*, /@6. +)-F7GA& ,I+*C6& /@2(N0.,*77-..05&
/-., 7@*.1&)'*-. +&6, +)-E,&) +7&*.&A&., /& 7@06,-R(-))270,*-. /0.' 7&' 10)*0,*-.' /& 70 ,0*77& /& 70
+-+670,*-." %0.' (& (0' 7@06,-R(-))270,*-. 1*&., '*A+7&A&., /6 O0*, C6& 7&' +0),*(*+0.,' )&',&.,
6. ,&A+' .-. .257*5&0F7& /0.' 70 '&''*-. &, +0) (-.'2C6&.(& (&6V C6* '-., +)2'&.,' B 6. *.',0.,
/-..2 0Y&(,&., (&),0*.&A&., 7& .-AF)& ,-,07 /& +0),*(*+0.,' B ,-6, 06,)& *.',0., +)2(2/0., 7&6)
/2+0),"

!" #$%&'( () *+,'-.(

D10., /& +)2'&.,&) 7& A-/G7& +0)7-.' /&' ,)0(&' C6& .-6' 01-.' 6,*7*'2&' +-6) 70 107*/0,*-." X7
'@05*, /& C60,)& ,)0(&' 1*/2-' (-77&(,2' '6) 7& Z[-.& T9U &.,)& D-\, &, ]&+,&AF)& >!!#" 4&' ,)0(&'
-., /&' /6)2&' 0770., /& C6&7C6&' ^-6)' +-6) 7@6.& B 6.& 1*.5,0*.& /& ^-6)' +-6) 7@06,)&" S-6'
(-..0*''-.' +-6) (N0C6& ,)0(& &V0(,&A&., 7& .-AF)& /& +0),*(*+0.,'H 7&6)' *.',0.,' /@0))*12& &,
7&6)' *.',0.,' /& /2+0)," D++&7-.' 7& ,06V A-I&. /@0))*12& /&' +0),*(*+0.,' +0) ,)0(&H 1/ 7&
,&A+' A-I&. C6@*7 +0''& /0.' 70 '&''*-. +0) ,)0(&H &, = / 70 (N0)5&" X7 &', (-..6 /@0+)G' 7&
,N2-)GA& /& 4*,,7& T##>U C6& )&+)2'&.,& 7& .-AF)& A-I&. /& +0),*(*+0.,' +)2'&.,' '*A67,0.2A&.,
/0.' (N0C6& '&''*-." %0.' 70 '6*,& 0++&7-.' N (t) 7& .-AF)& /& +0),*(*+0.,' B 6. *.',0., /-..2 t
P-6 70 ,0*77& /& 70 +-+670,*-.Q" 4& 30F7&06 "# /-..& 7&' 107&6)' /& (&' /*Y2)&.,' +0)0AG,)&' C6&
.-6' 01-.' (07(672&' +-6) 7&' /*Y2)&.,&' ,)0(&'"
4& +)-F7GA& +-'2 &', 07-)' (-AA&., &',*A&) 70 ,0*77& /& 70 +-+670,*-. N (t) B +0),*) /6 '-6'R

!"! #$%&'( () *+,'-.(

"#$"%&'" ("$ %"$$)*"$ +",-$. /+"#0#$ -# 1#$2)#2 tn = nS "2 )33"'0#$ Yn '" #0%&+" (" %"$$)*"$
4567784 +",-$ 9 :"2 1#$2)#2. ;)33"'0#$ <-" :=)<-" +>:"32"-+ "#?01" -# %"$$)*" 0- #0# )?": -#"
3+0&)&1'12> p. 7) $0'-210# ') 3'-$ $1%3'" $"+)12 (" (1?1$"+ '" #0%&+" (" %"$$)*"$ +",-$ 3)+ p "2 ("
3+"#(+" N (tn ) = Yn /p. @A"$2 ') $0'-210# <-1 #" 21"#2 3)$ :0%32" (" ') :0++>')210#. 6''" B0#:210##"C
?01+" %D%" "''" "$2 ') %"1''"-+"C <-)#( ') 303-')210# :=)#*" 2"''"%"#2 ?12" <-" ') :0++>')210#
(1$3)+)1$$" "#2+" '"$ 1#$2)#2$ (" %"$-+". /)+ :0#2+" "# 3+>$"#:" (" ') :0++>')210#C "''" "$2 '01#
(AD2+" 'A0321%)'" 3-1$<-" <-"'<-A-# 3"-2 3+0E2"+ ("$ "$21%)210#$ B)12"$ ()#$ '" 3)$$> 30-+ '" :)':-'
(A-#" %"1''"-+" "$21%)210# ()#$ '" 3+>$"#2. 7" *)1# $"+)12 9 ') B01$ -#" %"1''"-+" 3+>:1$10# "2 -#"
3'-$ 3"212" ?)'"-+ (" p. F#" %)#1G+" 30-+ 2"#1+ :0%32" (" ') :0++>')210# $"+)12 'A-21'1$)210# ("
:" <-A0# )33"''" -# E'2+" (" %0H"##" *'1$$)#2" )?": 301($ "I30#"#21"' J6I30#"#21)''H K"1*=2"(
L0?1#* M?"+)*" 6KLMN. @A"$2 -# E'2+" 3)$$"O&)$ 2+G$ :0##- -21'1$> ()#$ (1?"+$ 3+020:0'"$
(" 'AP#2"+#"2 3)+21:-'1G+"%"#2 '" 3+020:0'" Q@/ R ST 30-+ '1$$"+ ("$ <-)#212>$ ?)+1)&'"$. 7A1(>"

tel-00408745, version 1 - 3 Aug 2009

$"+)12 (A"$21%"+ N (tn ) "# 3)+21" "# -21'1$)#2 Yn /p "2 "# -#" )-2+" 3)+21" "# -21'1$)#2 'A)#:1"##"
"$21%)210# N (tn1 ). F# 3)+)%G2+" [0, 1] "$2 -21'1$> 30-+ (>E#1+ 'A1%30+2)#:" (A-#" 3)+21"
(tn1 ) + (1 )Yn /p.
(tn ) = N
$-+ 'A)-2+". @":1 (0##" '1"- 9 -# "$21%)2"-+ (" :"22" B0+%" U N
6# )-*%"#2)#2 ') ?)'"-+ (" (" V 9 !C :"2 "$21%)2"-+ (0##" 3'-$ (A1%30+2)#:" )- 3)$$> )(>2"+%1#)#2 (- 3+>$"#2. W1"# >?1("%"#2 1' H ) -#" ?)'"-+ 0321%)'" (" "2 :"22" ?)'"-+ 0321%)'"
(>3"#( (" 'A1%30+2)#:" (" ') :0++>')210#. /0-+ -# 3+0:"$$-$ B0+2"%"#2 :0++>'> 1' B)-2 3+"#(+" -#"
?)'"-+ (" 3+0:=" (" 1C :"3"#()#2 30-+ -# 3+0:"$$-$ #0# :0++>'>C (012 D2+" 3+0:=" (" V. X)#$
:" ("+#1"+ :)$ 1' H #A) )-:-# *)1# ()#$ ') :0#$1(>+)210# (" 'A=1$20+1<-" (" ') 2)1''" (" ') 303-')210#.
Y1*-+" .! %0#2+" $-+ -#" 2+):" :0''":2>" '0+$ (A-#" $"$$10# )-(10 :0%%"#2 'A"$21%)2"-+ 6KLM
$" :0%30+2" 30-+ (1Z>+"#2"$ ?)'"-+$ (" . @" *"#+" (A"$21%)2"-+ "$2 2+G$ 1#2>+"$$)#2 %)1$ 1' )
("-I 3+0&'G%"$ %)["-+$ U (i) (A-# :02> 1' "$2 1#:)3)&'" (A"$21%"+ ("$ $H$2G%"$ <-1 3+>$"#2"#2
-#" (H#)%1<-" )-2+" <-" '1#>)1+" )-20O+"*+"$$1?" (A0+(+" ! J1.". $"-'" (>3"#()#:" (" 'A1#$2)#2
3+>:>()#2NC (ii) "2 (" 'A)-2+" :\2> 1' ) '" 3+0&'G%" (- :=01I (" C -# :=01I <-1 (>3"#( (" ')
?)+1)&1'12> "2 (" 'A)-20O:0++>')210# (" ') 2)1''" (" ') 303-')210#.
X)#$ :" 2+)?)1'C #0-$ $0%%"$ )''>$ 3'-$ '01# "# :="+:=)#2 '" %"1''"-+ "$21%)2"-+ '1#>)1+" <-1
21"#2 :0%32" (" ') :0++>')210# ()#$ '" 3+0:"$$-$ (" ') 2)1''" (" ') 303-')210#. F# "$21%)2"-+
'1#>)1+" ) ') 3+03+1>2> <-" 'A"$21%)210# (" ') 2)1''" (" ') 303-')210# 9 'A1#$2)#2 ):2-"' "$2 -#"
B0#:210# '1#>)1+" (" ') %"$-+" ):2-"''"C ("$ "$21%)210#$ ()#$ '" 3)$$> "2 ("$ %"$-+"$ ()#$ '" 3)$$>.
KP6]6; ".*. R VT ) (>?"'033> -#" 2=>0+1" 30-+ '" :)':-' (" :"2 "$21%)2"-+ 0321%)'. 7A0321%)'12>
"$2 "# 2"+%"$ (A"++"-+ <-)(+)21<-" %0H"##". 7" +>$-'2)2 %)["-+ (" ') 2=>0+1" (" KP6]6; "$2 <-"
1' $-^2 (" :0##)_2+" 'A)-20O:0++>')210# (- 3+0:"$$-$ 9 "$21%"+C 1.".C E[Nn Nnk ]C "2 ') :0?)+1)#:"
"#2+" '"$ %"$-+"$ "2 '" 3+0:"$$-$ 9 "$21%"+C 1.".C E[Nn Ynk ]C 30-+ 30-?01+ 2+0-?"+ '" %"1''"-+
"$21%)2"-+ '1#>)1+". 7" :)':-' :"3"#()#2 "$2 :0%3'1<-> "2 #>:"$$12" '" 3'-$ $0-?"#2 ("$ =H302=G$"$
B0+2"$ 30-+ 30-?01+ 3)+?"#1+ 9 ("$ "I3+"$$10#$ "I3'1:12"$. ]0-$ )?0#$ B)12 30-+ :"') 'A=H302=G$"
<-" '"$ %"%&+"$ )++1?"#2 $-1?)#2 -# 3+0:"$$-$ /01$$0# (A1#2"#$12> :0#$2)#2" "2 <-" :=):-# +"$2"
()#$ '" $H$2G%" -# 2"%3$ "I30#"#21"''"%"#2 (1$2+1&-> (" ?)'"-+ %0H"##" 1/. @"$ =H302=G$"$

!"#$%&' () *" +,&&-."%$,/ %'0#,&'..' "1 2'&3$+' 4' .5-+!"/%$..,//"6'

Membership vs. time and EWMA estimation: p = 0.01, S = 1s

30

60

90

= 0.95
= 0.99
= 0.999
Audio trace

tel-00408745, version 1 - 3 Aug 2009

!"#

40000

(s)

80000

120000

"# $ %&'()*+', -. (/ */+((. -. (/ 0'0)(/*+', .* .1*+2/*+', %345

,. 1',* 0/1 1')&.,* &/(+-.16 2/+1 2/(789 :.(/6 .((.1 0.82.**.,* -;/&'+8 ), .1*+2/*.)8 1')1 <'82.
.=0(+:+*. .* ),. 78/,-. 089:+1+', -;.1*+2/*+', :'22. ,')1 /((',1 (. &'+8 1)8 :.8*/+,.1 *8/:.1"
>( 8.1*. ?);@ :/(:)(.8 (;/)*'A:'889(/*+', -) 08':.11)1 -. (/ */+((. -. (/ 0'0)(/*+', ?)+ .1* B+.,
:',,). -/,1 (/ *C9'8+. -.1 D(.1 -;/**.,*. E Cov(Nn , Nnk ) = (/)ekS " F/ :'889(/*+', .,*8. (.
08':.11)1 -.1 2.1)8.1 .* (. 08':.11)1 @ .1*+2.8 ?)/,* @ .((. .* 1')1 ,'1 CG0'*CH1.1 -;),+<'82+*9 .*
-;+,-90.,-/,:. -/,1 (.1 890',1.16 0.)* 1;9:8+8. :'22. 1)+* E Cov(Nn , Ynk ) = p2 Cov(Nn , Nnk )
0')8 k 6= 0 .* Cov(Nn , Ynk ) = pCov(Nn , Nnk ) 0')8 k = 0" I, 1+20(. 0(',7.2.,* -. :.1
891)(*/*1 -/,1 (.1 <'82)(.1 -98+&9.1 0/8 3>%J%K L&'+8 5,,.=. M 0')8 0()1 -. -9*/+(1N 0.82.*
-. *8')&.8 (/ 2.+((.)8. <'82. (+,9/+8. 0')8 ,'*8. .1*+2/*.)8 -. (/ */+((. -. (/ 0'0)(/*+', /+,1+
?). (.1 2.+((.)81 :'.O:+.,*1 E

n = AN
n1 + BYn + (/)(1 A pB).
N

A .* B 1',* -.)= :'.O:+.,*1 :',1*/,*1 <',:*+', -.1 0/8/2H*8.1 -) 2'-H(. .* -',* (.1 .=08.11+',1
1. *8')&.,* -/,1 5,,.=. M" P')1 ,'1 CG0'*CH1.16 :.* .1*+2/*.)8 .1* (. 2.+((.)8 0/82+ *')1 (.1
.1*+2/*.)81 (+,9/+8.1" >( .1* :(/+8.2.,* /)*'8978.11+< -;'8-8. #"
J'*.Q B+., ?);/) (+.) -. :C.8:C.8 (. 2.+((.)8 .1*+2/*.)8 (+,9/+8.6 ?).(?);), 0.)* 0/8*+8 -;),
.1*+2/*.)8 -. <'82. 0/8*+:)(+H8.6 0/8 .=.20(. (. D(*8. -. R/(2/, :'22. ,')1 (;/&',1 </+* -/,1 STU6
.* .11/G.8 -. *8')&.8 (.1 2.+((.)81 0/8/2H*8.1 -. (;.1*+2/*.)8 ?)+ 890',-.,* @ :.8*/+, 'BV.:*+<
:'22. 0/8 .=.20(. (/ 2+,+2+1/*+', -. (;.88.)8 ?)/-8/*+?). 2'G.,,."

!"! #$%&'$(&)* +( $,,%&-$(&)*

60

12 Sep 01

13 Sep 01

14 Sep 01

(c) on-times

1s 1m

1d

1s 1m 1h

(b) inter-join times

14 Sep 01

12 Sep 01
1

13 Sep 01

14 Sep 01

Exponential fit
Weibull fit

0.5

0.5

Lognormal fit

13 Sep 01

Weibull linear fit for on-times, shape 0.345, scale 3700

Exponential fit

tel-00408745, version 1 - 3 Aug 2009

12 Sep 01

Lognormal linear fit for inter-join times, = 3.38, d = 1.49

!"#

0.25

0.5

0.75

0.25

0.5

0.75

"# $ %&'()'*+,-& .& /0&123*+2&4' 5/41 12+2312364&1 14' /&1 -),,&73),1 .&1 *&*8'&1

!" #$%&'$(&)* +( $,,%&-$(&)*

94' ,)1 64+2'& 2'+-&1 -32:&1 -3;.&1141< ,)41 +=),1 =:'3>: /+ 5':-313), .& ,)2'& &123*+2&4'"
?&1 ':14/2+21 1& 2')4=&,2 .+,1 /0+,,&7&< ,)41 ,)41 -),2&,2),1 3-3 .& 5':1&,2&' /&1 ':14/2+21 5)4'
/+ 5'&*3@'& 2'+-&" A3B4'& "# *),2'& -)**&,2 ,)2'& &123*+2&4' &12 -+5+8/& .& 83&, 2'+64&' /+
2+3//& .& /+ 5)54/+23), *+/B': /& (+32 64& /&1 12+2312364&1 14' /&1 5+'23-35+,21 ,& -)''&15),.&,2 5+1
='+3*&,2 +47 CD5)2C@1&1 (+32&1 .+,1 /& -+/-4/ E/&1 .&47 5&232&1 >B4'&1 &, 8+1F" G)41 ,)2),1 &,
5+'23-4/3&' /& (+38/& =)/4*& .&1 *&11+B&1 HI??J &,=)D:1 K -0&12 4,& 5)54/+23), 643 (+32 &,=3'), &2
&, *)D&,,& 4,& 2+3//& .& LMM &2 p &12 -C)313 .& 2&//& (+N), 6404, *&11+B& HI??J &12 &,=)D: 2)42&1
/&1 S 1&-),.1 ES O #" 1&-),.&1 .+,1 /+ >B4'&F" ?&1 +42'&1 >B4'&1 *),2'&,2 .&1 -)*5)'2&*&,21
13*3/+3'&1 &2 4,& &''&4' '&/+23=& *)D&,,& +42)4' .& PQ"
R+,1 /+ 5'+2364& &2 +>, .0423/31&' ,)2'& &123*+2&4'< 3/ (+42 /0+--)*5+B,&' 5+' 4,& +42'&
&123*+23), .4 2+47 .0+''3=:& .&1 *&*8'&1 &2 .& /+ .4':& *)D&,,& .& /&4' 3,1-'3523)," S&-3 5&42
1& (+3'& &, .&*+,.+,2 T -C+64& *&*8'& .0&,=)D&' .+,1 1), *&11+B& HI??J .&1 12+2312364&1 14'
1), 2&*51 .& 5+'23-35+23)," U5'&1< -&1 12+2312364&1 5&4=&,2 V2'& -)*83,:&1 &,1&*8/& 5)4' ()'*&'
/& 2&*51 .& 5+'23-35+23), *)D&," ?& 2+47 .0+''3=:& ,0&12 +42'& 64& /+ 2+3//& *)D&,,& .& /+
5)54/+23), &123*:& .3=31:& 5+' /& 2&*51 *)D&, .& 5+'23-35+23)," W,& +42'& 1)/423), 5)4''+32 V2'&
/0&,=)3 5+' -C+64& ,)4=&+4 5+'23-35+,2 T 1), +''3=:& &2 T 1), .:5+'2 4, *&11+B& .0+8),,&*&,2
&2 .& .:1+8),,&*&,2 +=&- 4,& 5')8+83/32: >7& 83&, .:>,3&" R& -&22& *+,3@'& /&1 5+'+*@2'&1

!"#$%&' () *" +,&&-."%$,/ %'0#,&'..' "1 2'&3$+' 4' .5-+!"/%$..,//"6'

"#$%#&' (')# #*'+,-* . "/)'+) 0# *#* *'/'+*'+1$#*2 3/+* 1$# 4# *5+' "5$) 4#''# *56$'+5& 5$ 67/$')#8
+6 *7/9+' #&45)# 0#* "+*'#* 5$%#)'#* . #:"65)#)2 ;# "6$* 0$) *#)/ 0# 45,<+&#) 67#*'+,/'+5& 0# 4#*
"/)/,=')#* /%#4 67#*'+,/'+5& 0# 6/ '/+66# 0# 6/ "5"$6/'+5& #66# ,(,# "5$) >5),#) $& ")5<6=,#
?5+&' 4/64$6/&' 6# '5$' . "/)'+) 0#* ,#**/9#* @A;;B *#$6#,#&'2

!" #$%&'()*$%)
C/&* 4# 4D/"+')# &5$* /%5&* ,5&')- *$) $& #:#,"6# 45&4)#' 67+&'-)(' 0# 6/ 45))-6/'+5& 0/&*
67#*'+,/'+5& #' 6/ ")-0+4'+5&2 ;/ 0+E4$6'- )-*+0# 0/&* 6# >/+' 1$# 6/ 45&*+0-)/'+5& 0# 6/ 45))-6/'+5&
0#,/&0# 0#* *'/'+*'+1$#* *$) 6# ")54#**$* . #*'+,#) 6$+ ,(,#8 0#* *'/'+*'+1$#* "6$* 0+E4+6#* .
5<'#&+) 1$# 0/&* 6# 4/* */&* 45))-6/'+5&2 F7#*' '5$'# 6/ 0+E4$6'- 0$ ")5<6=,#2 ;# ,+#$: #*' 0#
"/**#) "/) $& ,50=6# 0$ *G*'=,# 1$+ "#),#' 0# )-0$+)# 6# &5,<)# 0# "/)/,=')#* H45,,# 0/&*

tel-00408745, version 1 - 3 Aug 2009

&5')# 4/* 6/ )-0$4'+5& #& 0#$: "/)/,=')#* #' I8 #&*$+'# 67#*'+,/'+5& 0# 4#* "/)/,=')#* "/)
$&# ")54-0$)# /$:+6+/+)#2 ;7$'+6+*/'+5& 07$&# #*'+,/'+5& ?5+&'# 0# '5$* 6#* "/)/,=')#* #*' 0/&*
4#)'/+&* 4/* "5**+<6#* ,/+* 0#,/&0# <#/$45$" 07#J5)' . ,#'')# #& 5#$%)#2

!"#$%&' (
! "#$%&'&()* !+$),*
tel-00408745, version 1 - 3 Aug 2009

!"#$%%! #&' (&&)*) +

! "#$ %&'()*+#$ ,-#. )/,%%&'.0# ,.12-# "# ), +31&')'42# #$1 ), .0,&4# 56/#))# .,6$# $6& )#
&3$#,6 562 ",!$ .#&1,2!$ .,$ %#61 7,6$$#& )#$ &3$6)1,1$ "#$ +#$6&#$8 9,& #:#+%)#; ",!$ )# .,$
"/6! .'))#.1#6& 562 #!-'2# "#$ 1&,.#&'61#$ %'6& "3.'6-&2& ), 1'%')'42# "6 &3$#,6; )# !'+(&# "#$
1&,.#&'61#$ 43!3&3$ #1 )#$ +#$$,4#$ &#<6$ #! &#1'6& %#6-#!1 =1&# > )/'&242!# "/6!# .'!4#$12'!
,61'6& "6 .'))#.1#6& '6 ",!$ )# .'#6& "6 &3$#,6 #1 "/6!# %#&1# "#$ %,56#1$ "# +#$6&#8 ?,!$ ),
%&,1256#; )#$ 1&,.#&'61#$ $'!1 #!-'@3$ > 6! "3(21 .'!$1,!1 #1 7,2()# %'6& !# %,$ .'!4#$12'!!#&
)# &3$#,6 ABCDE8 F, 56#$12'! #$1 1'6G'6&$ .'++#!1 &34)#& .# "3(21 "#$ 1&,.#&'61#$ #1 .'++#!1
$/,$$6&#& 56/2) !/#$1 %,$ 1&'% 7,2()# %'6& &,)#!12& ), +#$6&#8 F# (61 #$1 $'6-#!1 "/,-'2& 6!#
+#$6&# "6 &3$#,6 )# %)6$ &,%2"#+#!1 %'$$2()# ,-,!1 56# $'! 31,1 .0,!4#8 F# +=+# %&'()*+#
#:2$1# ",!$ "/,61&#$ ,%%)2.,12'!$ .'++# %,& #:#+%)# )# !"# "/6! 4&,!" !'+(&# "# +,.02!#$
%'6& )/2!73&#!.# "# )#6& 31,1 '6 (2#! ), .'))#.1# "#$ $1,12$1256#$ "/6! 4&,!" !'+(&# "# +,.02!#$
&3%,&12#$ > 1&,-#&$ )# &3$#,6 %,& 6!# ,%%&'.0# ,.12-#8 H6,!1 > ), .'!4#$12'! 56# .#&1,2!$ '612)$ "#
+#$6&# .,6$#!1 $6& 6! .0#+2! $%3.2I56# "# )/J!1#&!#1 .'++# %,& #:#+%)# )'&$ "# ), +#$6&# "# ),
(,!"# %,$$,!1# "2$%'!2()# '6 )# "3),2; .#11# .'!4#$12'! %#61 $# &34)#& ,-#. "#$ %&'1'.')#$ $2+%)#$
> ), KL9 '6 ), KMNL ADOE 56,!" ./#$1 !3.#$$,2&#8 ?,!$ .# "#&!2#& .,$ 2) 7,61 7,2&# ,11#!12'!
> .# 56# )# .'!1&P)# "# .'!4#$12'! 2!1&'"621 !# %#&16&(# %,$ ), +#$6&#8 L#&1,2!$ '612)$ .'++#
$%&'&!( ABCCE 43!*&#!1 )#$ %,56#1$ ,-#. 6! %&'I)# (2#! "3I!2 56# )/2!1&'"6.12'! "/6!# ('6.)#
"# .'!1&P)# "# ('61Q#!Q('61 %#61 ,)13&#& )#6& 7'!.12'!!#+#!1 #1 %,& .'!$356#!.# 7,6$$#& )#6&$
&3$6)1,1$8
?,!$ 6! 1&,-,2) &3.#!1 ABCRE !'6$ !'6$ $'++#$ %#!.03$ $6& )# %&'()*+# ,%&*$ ,-'2& &#+,&563
)/,($#!.# "/6! %&'1'.')# "# .'!1&P)# "# .'!4#$12'! #1 "# S6: 562 %#&+#1 "# &34)#& )# "3(21 "#$
$'!"#$ ,.12-#$ #!-'@3#$ ",!$ )# &3$#,6 %,& 6! .'))#.1#6& "'!!3 > 6! 4&,!" !'+(&# "# +,.02!#$8 F#
+'1 ')**+'%+,( ",!$ ), $621# #$1 612)2$3 %'6& "#$24!#& 6!# +,.02!# $'!",!1 )# &3$#,6 #1 .'))#.1,!1
)#$ &3$6)1,1$ %'6& 6! 6$,4# 6)13&2#6&8 F/612)2$,12'! "# KL9 ",!$ .# .'!1#:1# !/#$1 %,$ #T.,.#
31,!1 "'!!3 )# %#121 -')6+# "# %,56#1$ 3.0,!43$ #!1&# )# .'))#.1#6& #1 .0,.6!# "#$ +,.02!#$8 U!
%)6$; )# %&'()*+# "6 .0'2: "6 !'+(&# "#$ .'!!#:2'!$ KL9 > 31,()2& #1 "# ), 7,<'! ,-#. ),56#))#
#))#$ $'!1 ),!.3#$ $# %'$#8 J) @ , ,)'&$ (#$'2! "# 1&'6-#& 6! !'6-#,6 %&'1'.')# 562 %#&+#1 "# &34)#&
OO

!"#$%&' () *" +,%&-.-/$' "0%$1'

"# $%&'( $# )#* *+,$#* -)('.#* #, /+,)('+, $# "0%(-( $1 2%*#-13 #( $# *0-**12#2 41# "#* *+,$#* +,(
&'#, -((#',( "#12* $#*(',-('+,* #( 41# "#* 2%5+,*#* *+,( &'#, 2#61#* 5-2 "# )+""#)(#127 8, )-* $# ,+,
2%5+,*#3 '" /-1( *0-**12#2 4101,# *+,$# #*( 2#,.+9%# :1*410; "- 2%)#5('+, $# "- 2%5+,*# +1 :1*410;
)# 41# "- $%)'*'+, 41# "- <-)=',# $#*(',-(2')# ,0#*( 5-* $'*5+,'&"# *+'( 52'*#7 >+1* .+1"+,* %.'(#2
$#* *+"1('+,* 41' '<5"'41#,( "# 2%*#-1 $-,* "- 52+)%$12#3 #7?73 @ABC3 +1 &'#, )#""#* 41' 5#1.#,( *#
)+,(#,(#2 $01, 5#('( #,*#<&"# $# 2%5+,*#* D#7?73 )-* $#* *+"1('+,* 41' )=#2)=#,( ; *-.+'2 *' 1,#
<-)=',# ,0- 5-* 2#61 1,# )#2(-',# ',/+2<-('+, *-,* )+,,-'(2# "0'$#,('(% $# "- <-)=',# @EBCF7 >+1*
.'*+,* 1,# *+"1('+, 41' G
H *+'( &-*%# *12 "# 52',)'5# $# &+1(I#,I&+1( *1'.-,( "#41#" "# 2%*#-1 ,0#*( 5-* '<5"'41%7
H )+,(2J"# "- )+,?#*('+, $1 2%*#-1 #( 2#(2-,*<#( "#* *+,$#* #, )-* $# ,+, 2%)#5('+, $# "0',I
/+2<-('+,7
H -<')-" -.#) "# (2-K) L,(#2,#( #M'*(-,(3 )# $#2,'#2 5-**-,( 5+12 "- 5"15-2( $# (#<5* 5-2

tel-00408745, version 1 - 3 Aug 2009

"0',(#2<%$'-'2# $# NOP7
H *+'( ?%,%2-"# ',$%5#,$-,(# $# "- ,-(12# $# "0',/+2<-('+, ; )+""#)(#2 $#* <-)=',#* *+,$%#*7
N+1( )+<<# NOP - %(% )+,61 5+12 2%)15%2#2 (+1(# *+2(# $# $+,,%#*3 ,+1* .+1"+,* 41#
,+(2# *+"1('+, *+'( )-5-&"# $0#Q#)(1#2 (+1(# *+2(# $# *+,$-?# #( $# 2%)15%2#2 (+1(# *+2(#
$0',/+2<-('+,* $# <-)=',#* $'*(2'&1%#* ; (2-.#2* "0L,(#2,#( *-,* "# $#.+'2 $0%(-&"'2 1,#
)+,,#M'+, NOP -.#) )=-)1,# $# )#* <-)=',#*7
H *+'( K-&"# $-,* "# *#,* 41# (+1(#* "#* *+,$#* 2%1**'**#,(7 O#)' ,# $+'( 5-* #<5R)=#2 "#
)=-,?#<#,( $1 ,'.#-1 $# K-&'"'(% *' &#*+',7
8, .1# $# (+1(#* *#* )+,$'('+,*3 ,+1* -.+,* )+,61 ,+(2# *+"1('+, #, *0',*5'2-,( "-2?#<#,( $#*
<%)-,'*<#* $# NOP <-'* #, "#* %"-2?'**-,( 5+12 410'"* 51'**#,( /+,)('+,,#2 $-,* )# )-* 5-2(')1"'#2
$1 5+',(I;I<1"('5+',(7 S- *+"1('+, #*( -55#"%# NLOP 5+12 NOPI/2'#,$"9 L,/+2<-('+, O+""#)('+,
P2+(+)+"7 NLOP - %(% )+,61 )+<<# 1, 52+(+)+"# (2-,*5+2( -.#) 1,# .#2*'+, )+""#)(#12 ; ',*(-""#2
$-,* "- <-)=',# 41' *+,$# #( 1,# -1(2# .#2*'+, (+1(# *'<5"# 41' *0-55#""#

2-3&0' #( 41' *0',*(-""#

$-,* (+1(# <-)=',# $%*'2-,( 5-2(')'5#2 ; "- *#**'+,7 T-,* NLOP3 (+1(# "0',(#""'?#,)# *# (2+1.#
-1 ,'.#-1 )+""#)(#12 41' "1' -**12# "# )+,(2J"# $# )+,?#*('+, #( "# )+,(2J"# $0#22#12*7 S- *+12)#
*# )+,(#,(# *'<5"#<#,( $# 2%5+,$2# -1M *+,$#* $1 )+""#)(#12 #, #,.+9-,( '<<%$'-(#<#,( *#*
$+,,%#*7 T-,* )#2(-',* )-* "- 52%*#,)# $# NLOP )J(% *+12)# ,0#*( 5-* ,%)#**-'2# *' "- *+12)# #*(
$%:; $+(% $01, <%)-,'*<# 41' "1' 5#2<#( $# 2%5+,$2# -1M *+,$#*7 S0#M#<5"# $01, )+""#)(#12
NLOP #,.+9-,( $#* 5',?* ; $#* <-)=',#* 2%5+,$-,( 5-2 LOUP #, #*( 1, &+, #M#<5"#7 T-,* )#
)=-5'(2# ,+1* 52%*#,(+,* 1,# &2V.# $#*)2'5('+, $# NLOP #( $# )#* <%)-,'*<#*7 S#* $%(-'"* *#
(2+1.#,( $-,* @WXAC -',*' 41# $-,* Y,,#M# T7

!" #$ %&'()*+$ ,$ %&'-$.(/&' $( +$ %&'()*+$ ,0$))$1). ,2'. 3456

S+2* $# "- )+,)#5('+, $# NLOP ,+1* ,+1* *+<<#* "-2?#<#,( ',*5'2%* $1 52+(+)+"# NOP @ !C7
N+1( )+<<# $-,* NOP3 "# )+,(2J"# $# )+,?#*('+, $-,* NLOP *# /-'( ; &-*# $# /#,R(2# ?"'**-,(#3

!"! #$ %&'()*+$ ,$ %&'-$.(/&' $( +$ %&'()*+$ ,0$))$1). ,2'. 3456

" #$ %&'()*+,* -.* #$ /*+01)* %* 2345 *61 7*6.)(* *+ +879)* %* 68+%*6 *+:8;(6 %$+6 #* )(6*$.
*1 <8.) #*6-.*##*6 .+* )(<8+6* +=*61 <$6 *+,8)* <$):*+.*> ?$ /*+01)* $.@7*+1* 1$+1 -.* #*6
)(<8+6*6 <$):&*++*+1 ,8))*,1*7*+1 *1 *61 %&:&6(* <$) %*.A ,877* %$+6 245 -.$+% %*6 )(<8+6*6
+* <$):&*++*+1 <$6> ?$ /*+01)* $.@7*+1* *A<8+*+1&*##*7*+1 $. %(9.1 %* #$ 6*66&8+ 6.&:$+1 .+*
<B$6* 7+&8 7(2)(C *1 <$66* " #$ <B$6* 5&'-$.(/&' 9:&/,2'%$ $<)D6 #* <)*7&*) 6&@+$# &+%&-.$+1 #$
,8+@*61&8+ %. )(6*$.>
?* ,8+1)E#* %* ,8+@*61&8+ *1 #* ,8+1)E#* %=*))*.)6 %$+6 2345 )*<86*+1 6.) #*6 <)&+,&<*6 6.&:$+16 F

!"#$%"#&' ()*'+ ,&-'+ !*./-#+*-+ !*- 0+ "+$.! %00+-1-+"&*- 23456

28.1 ,877* %$+6 245C #* <)818,8#* 2345 7*6.)* #* %(#$& $##*)G)*18.) )(6*$. *1 ,$#,.#* $&+6&
.+* *61&7$1&8+ %. %(#$& 78;*+ *1 %* #$ %(:&$1&8+ 78;*++* %. %(#$&> ?$ 98)+* 6.<()&*.)* 6.) #*
1*7<6 $##*)G)*18.) *61 *61&7(* ,877* #$ :$#*.) 78;*++* <#.6 -.$1)* /8&6 #$ %(:&$1&8+ 78;*++*>

tel-00408745, version 1 - 3 Aug 2009

4*11* 98)+* 6.<()&*.)* H2I *61 .+* *61&7$1&8+ ,8+6*):$1)&,* 6$+6 *A,D6 %. 1*7<6 -.* <)*+% .+*
68+%* <8.) $11*&+%)* 6$ %*61&+$1&8+ *1 <8.) -.* #$ )(<8+6* ,8))*6<8+%$+1* $11*&@+* #* ,8##*,1*.)>
J+* )(<8+6* -.& +=$))&:* <$6 <*+%$+1 ,* 1*7<6 H2I *61 .+* 98++* &+%&,$1&8+ %* #$ <*)1* %* #$
68+%* 8. 9&*+ %* #$ )(<8+6* 6.) #* ,B*7&+ %* )*18.)>

7&'8+!"#&' (* -/!+%*
?* ,8##*,1*.) %$+6 2345C 68.),* %* 68+%*6C *+:8&* $.1$+1 %* 68+%*6 -.* #$ /*+01)* #* <*)7*1>
?*6 )(<8+6*6 ,8))*6<8+%$+1*6 " .+* /*+01)* %* 68+%*6 68+1 6.<<86(*6 $))&:*) <*+%$+1 .+ 1*7<6
(@$# " H2I> K* ,* /$&1C #* ,8##*,1*.) 2345 7*6.)* #* +879)* %* )(<8+6*6 )*L.*6 <*+%$+1 #*
<)8,B$&+ H2I 6.&1* " .+* /*+01)* %* 68+%*6 *+:8;(*6 %$+6 #* <)(6*+1 H2I *1 ,$#,.#* $&+6& #*
1$.A %* <*)1*6 %* 68+%*6M)(<8+6*6> N& ,* 1$.A *61 <#.6 (#*:( -.=.+ ,*)1$&+ 6*.&#C #* )(6*$. *61
,8+6&%()( ,877* ,8+@*61&8++( *1 2345 %&:&6* $#8)6 6$ /*+01)* <$) %*.A> N& ,* )$<<8)1 *61 <#.6
(#*:( -.=.+ $.1)* 6*.&# <#.6 &7<8)1$+1C #* )(6*$. *61 ,8+6&%()( ,877* 6(:D)*7*+1 ,8+@*61&8++(
*1 2345 <$66* *+ 78%* N#8O N1$)1>

3+"-%'!$#!!#&' ()+--+*-!
2345 *61 6.<<86(* ,8++$P1)* #$ #&61* %*6 7$,B&+*6 " 68+%*)> 4*11* #&61* *61 <$),8.).* %=.+*
/$L8+ ,;,#&-.*> K$+6 #* <)*7&*) ,;,#* #*6 7$,B&+*6 68+1 68+%(*6 6.,,*66&:*7*+1> ?=8)%)* 6.&:$+1
#*-.*# *##*6 68+1 68+%(*6 *61 %&6,.1( %$+6 #$ 6*,1&8+ 6.&:$+1*> K$+6 #* %*.A&D7* ,;,#* #*6 7$,B&+*6
-.& +=8+1 <$6 )(<8+%. #8)6 %. <)*7&*) ,;,#* 68+1 68+%(*6> ?* 68+%$@* ,8+1&+.* Q.6-.=" ,* -.*
18.1*6 #*6 7$,B&+*6 )(<8+%*+1> 4*,& @$)$+1&1 #$ R$9&#&1( *1 #$&66* 6.S6$77*+1 %* 1*7<6 $.A
)(<8+6*6 *+ )*1$)% <8.) $))&:*) $. ,8##*,1*.)>

9*#"/ %:+; 47<

2345 %&6<86* %=.+ <$)$7D1)* -.& &+%&-.* #$ @)$+.#$)&1( %* 68+ ,8+1)E#* %* ,8+@*61&8+> 4*
<$)$7D1)*C $<<*#( ! <8.) H*-.*61 N&T*C *61 #=(-.&:$#*+1 %. 1$&##* %* <$-.*1 %$+6 245> K$+6

!"#$%&' () *" +,%&-.-/$' "0%$1'

tel-00408745, version 1 - 3 Aug 2009

!"# "#$ % &'(')'*+, -. /01,2. ('./ )2 32)+-24+'5 -, &678

!"# "#9 % 7':(2/2+1'5 ,54/, &678 ,4 &78

&678; )2 */25.)2/+40 ,14 ,5 5':</, -, 1'5-,1# 8,5-254 )2 (=21, -, 7'5*,14+'5 >3'+-25?,; )2
@,5A4/, -, &678 2.*:,54, -, ! 1'5-,1 4'.1 ),1 B&&1# C251 )2 (=21, D)'E D42/4; )2 @,5A4/, -'.<),
?=2F., B&&# C. ?'.(; ?, (2/2:G4/, ! -0?+-, -, )H0F.+40 ,54/, &678 ,4 &78# I5 /2+1'55254
,5 4,/:,1 -, (2F.,41J1; ), -0<+4 -H.5, 1,11+'5 &678 (,.4 A4/, 3. ?'::, 04254 ! @'+1 ), -0<+4
-H.5, ?'55,K+'5 &78 F.+ 4'./5, -251 ),1 :A:,1 ?'5-+4+'51#
L,1 -042+)1 -, ?,1 -,.K ?'54/M),1 -251 &678 1'54 -0?/+41 -251 N$O9P 2+51+ F., -251 )H>55,K, C#
L, (/'4'?'), 2 040 +:()0:,540 -251 ), 1+:.)24,./ -,1 /01,2.K 51Q9 N!!P 2+51+ F.H,5 7RR ('./
4'./5,/ 1./ )+5.K ,5 -,11.1 -, SC8# L,1 (,/@'/:25?,1 -, &678 '54 040 32)+-0,1 -251 -+T0/,541
1?,52/+'1 U )2 @'+1 (2/ 1+:.)24+'51 ,4 (2/ ,K(0/+:,5424+'51 8)25,4L2<# L, /01.)424 ), ().1 +540/,1Q
1254 042+4 U (/'('1 -, )2 ?':(2/2+1'5 ,54/, &78 ,4 &678# I5 1+:.)254 &678 -251 51Q9 1./ )2
4'(')'*+, -0?/+4, -251 V+*./, "#$ ,4 ('./ OO 1'./?,1 -, -'550,1 -+14/+<.0,1 2)024'+/,:,54 -251

!"! #$%&'(%) *+) ,$&-./+) *$/) 012#

"# $%&#'( ')*&) +(# "# ,-""#,.#($/ *-(& '0-*& 1( .$',#$ "#& 1#$2-$3'*,#& )""(&.$%#& 4'*& 5)6($# 7898
:#..# 4#$*);$# <6($# 3-*.$# "' 4($%# 4# "' &#&&)-* 4# &-*4'6# #* 2-*,.)-* 4( 0-"(3# 4=)*2-$3'>
.)-*& ? ,-""#,.#$ 4# ,@'+(# &-($,# &-*4%#8 A' <6($# 3-*.$# '(&&) "#& 1#$2-$3'*,#& -B.#*(#& #*
(.)")&'*. 4#& ,-**#C)-*& D:E #* 1'$'"";"# F #. GH8 E-($ "# ,'& 4#& ,-**#C)-*& D:E/ 4;& +(=(*#
,-**#C)-* &# .#$3)*# (*# '(.$# #&. %.'B")# '0#, (*# '(.$# &-($,# 4# 4-**%#& '<* 4# $%,(1%$#$ &#&
)*2-$3'.)-*&8 I" #&. ? *-.#$ 4'*& ,#..# <6($# "#& #C,#""#*.#& 1#$2-$3'*,#& 4# DI:E F(* -$4$# 4#
6$'*4#($ 1"(& +(# D:EH +('*4 )" &='6). 4# &-($,#& '0#, 1#( 4# 4-**%#& ? $#3-*.#$ '( ,-""#,.#($8
:#,) &=#C1")+(# 1'$ "# 2'). +(# DI:E '6$;6# "#& 4-**%#& #* 1$-0#*'*. 4#& 4)J%$#*.#& &-($,#& #.
%0).# 1'$ ,-*&%+(#*,# "#& 1@'&#& K"-L K.'$. 4#& ,-**#C)-*& D:E8 :#1#*4'*./ +('*4 "# 0-"(3#
4=)*2-$3'.)-*& 1'$ &-($,# 4#0)#*. )31-$.'*./ DI:E &# ,-31-$.# )4#*.)+(#3#*. '(C ,-**#C)-*&

tel-00408745, version 1 - 3 Aug 2009

D:E/ ,#& 4#$*);$#& 1-(0'*. 3')*.#*'*. '..#)*4$# "#($& 1@'&#& ,-*6#&.)-* '0-)4'*,#8

!" #$%&'(%) *+) ,$&-./+) *$/) 012#

A#& 3',@)*#& ? &-*4#$ 1#(0#*. &# .$-(0#$ $%1'$.)#& 4'*& "# $%&#'( #. &).(%#& 4#$$);$# 4#& 6-(>
"-.& 4=%.$'*6"#3#*. 4)J%$#*.&8 M#3'*4#$ ? DI:E 4# "#& &-*4#$ )*4%1#*4'33#*. 4# "#($ "-,'").%
,=#&. ,-31")+(#$ "' .N,@# 1-($ "# 1$-.-,-"# 1()&+(=)" '($' ? 6%$#$ 1"(&)#($& 6-("-.& 4=%.$'*6"#3#*.
#* 1'$'"";"# #. '0#, (*# &#("# 2#*O.$# 4# ,-*6#&.)-*8 DI:E ' '"-$& 4( 3'" ? 4)J%$#*,)#$ "#& 6-("-.&
,-*6#&.)-**%& 4# ,#(C *-* ,-*6#&.)-**%& ,# +() $%&(".# #* (* ,-*.$P"# 4# ,-*6#&.)-* &-(&>#Q,',#
#. 4#& $#&&-($,#& $%&#'(C 3'" (.)")&%#&8 E-($ ,#"' *-(& '0-*& #*$),@) DI:E '0#, (*# &-"(.)-*
1-($ "# $#6$-(1#3#*. 4#& 3',@)*#& ? &-*4#$8 A#& 3',@)*#& &-*. 6$-(1%#& '0#, (* 3%,'*)&3#
4# .R1# ,--$4-**%#& 0)$.(#""#& -( *-3& MSK -( '4$#&&#& IE/ #. &-*. '1$;& &-*4%#& #* 2-*,.)-*
4# "#($ "-,'").% #* '""'*. 4#& 3',@)*#& 1$-,@#& 4# "' &-($,# 0#$& "#& 3',@)*#& 4)&.'*.#&8 A# 1"(&
)31-$.'*. #&. +(# "#& 3',@)*#& 1$-,@#& "#& (*#& 4#& '(.$#& &-)#*. &-*4%#& #*&#3B"# '0'*. 4#
1'&&#$ ? (* '(.$# 6$-(1# 4# 3',@)*#& ? ,P.% #. ,#,) T(&+(=? 1'$,-($)$ .-(.# "' ")&.# 4#& 3',@)*#&8
:#,) ' 1"(&)#($& '0'*.'6#& U
V W* 1'$.'*. 4( 1"(& 1$;& '( 1"(& "-)*/ -* '&&($# (* 3'C)3(3 4# ,-""#,.# '( 4%B(. 4# "'
&#&&)-*8
V W* -$4-**'*. "#& 3',@)*#& 4'*& "' ")&.# ? &-*4#$ #* 2-*,.)-* 4# "#($ "-,'").%/ -* 1#$3#. '(
,-*.$P"# 4# ,-*6#&.)-* 4# .$').#$ &)3(".'*%3#*. '0#, "#& 3',@)*#& 4#$$);$# "# 3O3# 6-("-.
4=%.$'*6"#3#*. ,# +() ,-*4(). ? (*# (.)")&'.)-* #Q,',# 4#& $#&&-($,#&8
V 5)*'"#3#*./ #* ,"'&&'*. "#& 3',@)*#& 4# 1"(& 1$;& '( 1"(& "-)*/ -* '&&($# (* ,@'*6#3#*.
#* 4-(,#($ 4#& ,-*4).)-*& $%&#'( ,# +() 1#$3#. '(C 1'$'3;.$#& 4# DI:E 4# &='T(&.#$
$'1)4#3#*. ? ,# ,@'*6#3#*. &'*& 1%*'")&#$ "#& 1#$2-$3'*,#&8
S-(& '0-*& .#&.% "=)4%# 4( $#6$-(1#3#*. 4#& 3',@)*#& #* (.)")&'*. 4#& ,--$4-**%#& 0)$.(#""#&8
:@'+(# 3',@)*# #&. &(11-&%# 4-.%# 4# ,--$4-**%#& 4'*& (* #&1',# ? 4#(C 4)3#*&)-*& $#1$%&#*>
.'*. "' .-1-"-6)# 4# "=I*.#$*#.8 X1$;& "#& &-($,#& &-*. 6$-(1%#& #* 4#& ,#""("#& 4# 2-$3# ,'$$%# #.
"#& ,#""("#& &-*. 1'$,-($(#& 1'$ DI:E 4'*& "# &#*& -11-&% 4#& ')6()""#& 4=(*# 3-*.$# #* '""'*. 4#

!"#$%&' () *" +,%&-.-/$' "0%$1'

"# $ %& '()*+,-.& '& /0&1,-)& &2 /& 3&.3*+,&4&52 '&1 4-)675&1 '-51 89:;

Constant congestion window protocol

TICP with clustering
TICP without clustering

55
50
Collect Session duration (in s)

tel-00408745, version 1 - 3 Aug 2009

!"#

45
40
35
30
25
20
15
50

!"#

100
150
200
cwnd (Congestion window size)

250

300

"< $ 89:; -=&) 3&.3*+,&4&52 '&1 4-)675&1 )*4,-3( -+ )-1 1-51 3&.3*+,&4&52

,/+1 ,3>1 -+ ,/+1 /*75? =*73 @7.+3& "#" %- '+3(& 4*A&55& '0+5& 1&117*5 89:; -=&) 3&.3*+,&4&52
'&1 4-)675&1 &12 )*4,-3(& '-51 @7.+3& "< B /- '+3(& '& /- 1&117*5 1-51 3&.3*+,&4&52" C5 =*72
D7&5 +5& 3('+)27*5 '& /- '+3(& '& ,3&1E+& F!G" %- H.+3& 4*523& -+117 +5& )*4,-3-71*5 B +5
,3*2*)*/& '& 1*5'-.& B I&5J23& )*512-52& &2 E+7 )6-5.& 1+3 /0-K& '&1 L '-51 /- H.+3&" ;*+3 )&
'&357&3 ,3*2*)*/& &2 &5 -+.4&52-52 /- I&5J23&? /- '+3(& '& /- 1&117*5 D-711& -+ '(D+2 M+1E+0B
-22&75'3& /- I&5J23& *,274-/& B +27/71&3? &51+72& 1& '(.3-'& ,+71E+& /- I&5J23& '&=7&52 23*, .3-5'&
,*+3 E+& /& 3(1&-+ ,+711& /- 1+,,*32&3" 89:; -=&) 3&.3*+,&4&52 &12 )-,-D/& '& ,3('73& /- =-/&+3
*,274-/& '& /- I&5J23& &2 '& '*55&3 /7&+ -+K 4J4&1 ,&3I*34-5)&1 ,3-27E+&4&52"
N-51 OP!QR 5*+1 -=*51 -+117 3&.-3'( /- 2-7//& *,274-/& '& /- )&//+/& B +27/71&3 &2 /074,-)2 '& )&22&
2-7//&" :/-73&4&52 )&22& 2-7//& - +5& 74,*32-5)&" S+-5' &//& &12 23*, .3-5'&? 89:; I*5)27*55&
)*44& '-51 /& )-1 1-51 3&.3*+,&4&52 '&1 4-)675&1" T2 &5 3('+71-52 /- 2-7//& '& )6-E+& )&//+/&?

tel-00408745, version 1 - 3 Aug 2009

!"! #$%&'()*$%)

!"#

"# $ %&'()*+,(-),./ .0),1-&* ('+ 2&-/*)3-4 5'* 0-+ 67%2

&*( 0*+8.+1-/9*( :* 67%2 (;-1<&,.+*/) ='(>';? '/ 9*+)-,/ 0.,/) .0),1-&@ -0+A( *&&*( (* :<B+-:*/)
0',(>';./ +*5,*/) -' 9-( (-/( +*B+.'0*1*/) 0.'+ :* )+A( 0*),)*( 9*&&'&*(" %*))* )-,&&* .0),1-&* *()
9&-,+*1*/) 8./9),./ :' /.14+* :* (.'+9*(@ 1-,( '/* :* /.( .4(*+5-),./( :-/( 9* )+-5-,& 9;*() >'*
? 0-+),+ :;'/ 9*+)-,/ /.14+* :* (.'+9*(@ &- )-,&&* .0),1-&* :*5,*/) 0+*(>'* ,/5-+,-4&* :<)*+1,/<*
0-+ &- :,()+,4'),./ :*( B.'&.)( :;<)+-/B&*1*/) 0&'( >'* 0-+ &- :,()+,4'),./ :*( (.'+9*(" C,B'+* "#
,&&'()+* 9* +<('&)-) ('+ 2&-/*)3-4 .D ./ 5.,) '/ :<9.'0-B* .0),1-& */ :*( 9-++<( :* !EE 1(
:;-++F)* -' :*&? :;'/ 9*+)-,/ /.14+* :* (.'+9*("

!" #$%&'()*$%)
G.)+* 0+.).9.&* 67%2 0+.0.(* '/* (.&'),./ ? '/ 0+.4&A1* >', <)-,) ='(>';? &.+( (.'(H*(),1<@
9*&', :' 9./)+I&* :* 9./B*(),./ *) :' 9./)+I&* :;*++*'+( 0.'+ &*( -00&,9-),./( (./:-/) '/ B+-/:
/.14+* :* 1-9J,/*( :-/( &;7/)*+/*)" %* )+-5-,& *() */ 9.'+( *) - *' '/* 0+*1,A+* 5-&,:-),./ -5*9
/(HK *) ('+ 2&-/*)3-4" 7& +*()* ? &* 5-&,:*+ -5*9 :*( 5+-,*( -00&,9-),./( 9.11* 0-+ *L*10&* &*
)+-M-B* :* &- ).0.&.B,* :* &;7/)*+/*) .' &- 5<+,N9-),./ :* &;<)-) :* 8./9),.//*1*/) :;'/* 0&-)*8.+1*
>'*&9./>'* 8.+1<* :;'/ B+-/: /.14+* :* 1-9J,/*(" 3;<)':* :;-')+*( 1<)J.:*( :* +*B+.'0*1*/)
:*( 1-9J,/*( *() -'((, ,10.+)-/)* 0.'+ &* 8')'+ :* 9*))* (.&'),./"

tel-00408745, version 1 - 3 Aug 2009

!"#$%&' () *" +,%&-.-/$' "0%$1'

!"#$%&' (

tel-00408745, version 1 - 3 Aug 2009

!"#$%&'!"&
!""! "#$%! &'(%!)"! *) !)%!+,-! .! %/-*"0/)% &/*' *)! +("'/-/10! !2343! .! -56)"!')!"7 5!%"
*) ./+40)! !) &-!0)! (8/-*"0/) ("4)" ./))(! -4 3'/0%%4)3! !9&/)!)"0!--! .! -56)"!')!" !) "!'+!% .!
+43#0)!% !" '(%!4*9 3/*&-(! : -54,%!)3! .! %/-*"0/)% %"4).4'. &/*84)" '(&/).'! 4*9 ,!%/0)% .!%
*"0-0%4"!*'% !" .!% /&('4"!*'%7 ;/"'! /,<!3"0= -! -/)1 .! 3!""! '!3#!'3#! 4 ("( -! .(8!-/&&!+!)"
.!% %/-*"0/)% >*0 &4%%!)" : -5(3#!--! !" >*0 '(.*0%!)" -4 3#4'1! %*' -! '(%!4* %4)% &/*' 4*"4)"
3/+&'/+!""'! -4 &'(30%0/) .!% '(%*-"4"% /,"!)*%7 ;/*% 48/)% 410 &/*' 3!-4 %*' .!*9 ='/)"%7 ?*' -!
='/)" .!% +!%*'!% &4%%08!%@ )/*% 48/)% &'/&/%( *)! 4'3#0"!3"*'! &/*' -4 34&"*'! .* "'4A3 : "'48!'%
*) '(%!4* .5/&('4"!*'7 B* -0!* .! 3/--!3"!' "/*" -! "'4A3 !" .! -54)4-C%!' *-"('0!*'!+!)"@ )/"'!
4'3#0"!3"*'! &!'+!" 4*9 /&('4"!*'% .! .(A)0' *)! "D3#! .! +!%*'! !" .! &'/1'4++!' -!% '/*"!*'%
.! "!--! %/'"! >*! 3#43*) .5!*9 3/--!3"! -4 >*4)"0"( .! "'4A3 )(3!%%40'! %4)% .(&4%%!' *) 3!'"40)
%!*0- %*' -! 8/-*+! .* "'4A3 1-/,4- : '!+/)"!' 4*9 *)0"(% 3!)"'4-!%7 !""! '(.*3"0/) .* 8/-*+! .*
"'4A3 3/--!3"( %5!E!3"*! &4' -! ,040% .! -5(3#4)"0--/))41! %&4"04-! !" "!+&/'!--!7 F5(3#4)"0--/))41! 4
*) 0+&43" )(14"0= %*' -4 &'(30%0/) .! -4 +!%*'!@ *) 0+&43" >*0 &!*" G"'! '(.*0" &4' .!% "!3#)0>*!%
.! A-"'41! !" .!% %/-*"0/)% &'/"/3/-40'!%7 ;/*% 48/)% 0--*%"'( 3!30 %*' .!*9 !9!+&-!% 3/)3'!"%@
*) 3!-*0 .! -4 .("!3"0/) .!% 1'/% H/"% *"0-0%4"!*'% .! "'4A3 !" *) 4*"'! 3!-*0 .! -4 +("'/-/10! .!
1'4).!% &/&*-4"0/)% 84'04)" 4* 3/*'% .* "!+&%7 I/*' -!% .!*9 !9!+&-!% )/*% 48/)% +!)( *)!
("*.! &/*' (84-*!' -50+&43" .! -5(3#4)"0--/))41! %*080 &4' .!% %/-*"0/)% '(.*0%4)" -5!''!*' -/'% .!
-50)8!'%0/) .!% 0)=/'+4"0/)% (3#4)"0--/))(!%7 J)%*0"!@ )/*% 48/)% 4,/'.( -54&&'/3#! 43"08! .4)%
-4 +("'/-/10! !" 48/)% &'(%!)"( *)! %/-*"0/) 4&&!-(! K6 I >*0 &!'+!" .! 3/)"'L-!' -! .(,0" .!%
%/).!% !" .54%%*'!' -4 A4,0-0"( >*4). 0- %5410" .5*)! 4&&-034"0/) %/).4)" *) 1'4). )/+,'! .!
+43#0)!% .0%"'0,*(!% : "'48!'% -! '(%!4*7 K6 I &!*" G"'! 8* 3/++! *) +*-"0&-!941! .! &-*%0!*'%
3/))!90/)% K I &4'"4)" .5*)! +G+! +43#0)! !" .!%"0)(!% : .!% +43#0)!% .0E('!)"!% !) *)!
%!*-! 3/))!90/) &/0)"M:M+*-"0&/0)"7 K/*"!% )/% %/-*"0/)% /)" ("( 84-0.(!% &4' .!% %0+*-4"0/)% !"
.!% "'43!% '(!--!%7
F!% "'484*9 .4)% 3!""! "#$%! )! %/)" >*5*) !9!+&-! .! )/% 43"080"(% .! '!3#!'3#! .4)% -!
./+40)! .! -4 +("'/-/10! !) &4'"03*-0!' !" .!% '(%!4*9 !) 1()('4-7 ;/*% 48/)% &4'4--$-!+!)"
3/)"'0,*( : -4 +("'/-/10! .!% '(%!4*9 %4)% A- NOPQ !" 4* .(8!-/&&!+!)" .!% %/-*"0/)% &'/"/3/-40'!%
RS

!"#$%&' ()

*+,-./$*+/

"#$ %&'(&))&*) +,-./0*1& 2,$*34'(0)$4*5 &*)'& 2&5 0%%0'&$+5 50*5 6+5 "#& .& 54$) 2&5 4'2$*0)&#'5
%4')07+&5 4# 2&5 4'2$*0)&#'5 2& %4./& 8 9: ;<=>? @4#5 0A4*5 0#55$ .4*)'$7#- B +,-)#2& 2&5 '-5&0#C
D0$'EBED0$' &?1? 8;<=> &) B +0 5-.#'$)- &) 0# 2$(&*5$4**&(&*) 2&5 '-5&0#C $*34'(0)$"#&5 8FG>? H&
+4*1 2& )4#)& .&))& '&./&'./& *4#5 0A4*5 024%)- #*& 0%%'4./& 2& 2-A&+4%%&(&*) 2&5 54+#)$4*5
4%)$($5-&5 "#$ %&'(&))&*) #*& (&$++&#'& #)$+$50)$4* 2&5 '-5&0#C &) "#$ $*)'42#$5&*) 2& *4#A&0#C
5&'A$.&5? H,05%&.) %&'34'(0*.&5 2&5 54+#)$4*5 0 )4#I4#'5 -)0$) %'-5&*) 20*5 *45 )'0A0#C: .4#%+- B
#*& A0+$20)$4* 2&5 54+#)$4*5 %0' 2&5 )'0.&5 '-&++&5 &) 2&5 &C%-'$(&*)0)$4*5?
H&5 )'0A0#C &* .4#'5 &) 3#)#'5 .4*)$*#&*) B 34.0+$5&' 5#' +0 (-)'4+41$& %#$5"#& *4#5 .'4J4*5 "#&
.,&5) #* 24(0$*& )'K5 $(%4')0*) %4#' +,0A&*$' 2& +,L*)&'*&)? M6* 2& %4#A4$' %'4%45&' 2& *4#A&++&5
0'./$)&.)#'&5 $+ 30#) 0A4$' #*& 74**& .4(%'-/&*5$4* 2&5 +$($)0)$4*5 2& +,0'./$)&.)#'& 0.)#&++&? N*&
1'0*2& %0')$& 2&5 5&'A$.&5 %'4%45-5 0.)#&++&(&*) %4#''0 5& A4$' $*)-1'-& 20*5 +,L*)&'*&) 2& 3#)#'
.4((& %0' &C&(%+& +&5 5&'A$.&5 2& .0+.#+ 2& .44'24**-&5 A$')#&++&5 %4#' +&5 (0./$*&5 4# +&5

tel-00408745, version 1 - 3 Aug 2009

0'./$)&.)#'&5 D0$'EBED0$' &) OPQ? R* %+#5 2& .&5 5&'A$.&5: +,L*)&'*&) 2& 3#)#' 5&'0 .&')0$*&(&*)
24)- 2& *4#A&0#C 5&'A$.&5 "#$ %&'(&))&*) +0 2-)&.)$4* 2&5 %'47+K(&5 &) +&#' 2$01*45)$"#&? H&5
'4#)&#'5 2& +,L*)&'*&) 2& 3#)#' 5& A&''4*) 0#55$ 24)-5 2& (-.0*$5(&5 $*)&++$1&*)5 +&#' %&'(&))0*)
2& 2$0+41#&' &*)'& &#C &) 2& %'&*2'& +&5 74**&5 2-.$5$4*5 %4#' #*& (-)'4+41$& &S.0.& 2# '-5&0#
0# +$&# 2& 5& .4*)&*)&' .4((& .,&5) 30$) 0.)#&++&(&*) 2& )4#) &*A4J&' B #*& #*$)- .&*)'0+& "#$ 30$)
+& )'0$)&(&*)? @4#5 .'4J4*5 "#& .& 54*) 2&#C 2$'&.)$4*5 )'K5 $*)-'&550*)&5 0#C"#&++&5 *4#5 0++4*5
.4*.&*)'- +0 %+#%0') 2& *45 &T4')5? D4#' #* (&$++&#' L*)&'*&) .&5 &T4')5 2&A'4*) .&')0$*&(&*)
5,0..4(%01*&' 2& *4#A&++&5 0%%+$.0)$4*5 "#$ %055&*) B +,-./&++& &) "#$ .4((#*$"#&*) 0A&. +&5
5&'A$.&5 2& (-)'4+41$& &) 2& 5-.#'$)- 2# '-5&0# %4#' #* (&$++&#' 34*.)$4**&(&*) &) #*& (&$++&#'&
"#0+$)- %4#' )4#5 +&5 0.)&#'5?

!"#!$%&'()!*
!" #$%&'(' ) #*+,(-'* ('./012%(3 401 &',*%(35'*3' 1'6',1-7 ,(* '*8-,.%0 (9 7..: );;,$%5
&'('9%(.'1('.<9'*89
<" #9 #*,=6> ?9 @8> A9 B,-'1'6> C9 D8E'&*> ?9F1%'*=,(> G9 H010/%.I> F9 J0 K1'6.%> L9@9
M00(> N9 K,O60(> D9 ?0/6&'P> Q?7' R6' 04 S(*5.05'(* M8&.%-,6. M',681'='(.6 401 B7,5

tel-00408745, version 1 - 3 Aug 2009

1,-.'1%I%(3 T(.'1(,& C'./012 @'7,+%01Q> TSSS B0==8(%-,.%0(6 M,3,I%('> M,P <UUU9

V" W9 #%'&&0> F9 B78(3> J9 J8> A,(*0= '+0&8.%0( %( =,66%+' 31,:76> H,(*$002 0( M,66%+'
D,., L'.6> XS*69 G,='6 #$'&&0 '. ,&9Y> Z&8/'1 #-,*'=%- K8$&%67'16> X<UU<Y [\5!<<9
]" Z9

B9

#&='10.7

,(*

M9

H9

#==,19

MJ%6.'(>

![[^9

7..: );;///9--93,.'-79'*8;-0=:8.%(3;?'&'-0==;=$0(';9
^" L9 #&084> S9 #&.=,(> B9 @,1,2,.> K9 C,%(> QS6.%=,.%(3 M'=$'167%: %( , M8&.%-,6. L'66%0(Q>
%( :10-''*%(36 04 #BM LT_MS?ATBL> G8(9 <UUV9 #+,%&,$&' %( #(('O B9
`" L9 #&084> S9 #&.=,(> B9 @,1,2,.> K9 C,%(> Qa:.%=,& S6.%=,.%0( 04 M8&.%-,6. M'=$'167%:Q>
TSSS ?1,(6,-.%0(6 0( L%3(,& K10-'66%(3 5 L:'-%,& T668' 0( L%3(,& K10-'66%(3 %( C'./012%(3>
+0&9 ^!> (09 b> ::9 <!`^5<!\`> #8386. <UUV9
\" S9 #&.=,(> Z9 #+1,-7'(20+> B9 @,1,2,.> Q# L.0-7,6.%- M0*'& 04 ?BK;TK /%.7 L.,.%0(,1P
A,(*0= J066'6Q> TSSS;#BM ?1,(6,-.%0(6 0( C'./012%(3> +0&9 !V> (09 <> ::9 V^`5 V`[>
#:1%& <UU^9
b" S9 #&.=,(> Z9 #+1,-7'(20+> B9 @,1,2,.> K9 D8$'> QK'1401=,(-' #(,&P6%6 04 #TMD M'5
-7,(%6=6 0+'1 , M8&.%5L.,.' M,120+%,( K,.7Q> B0=:8.'1 C'./0126 G081(,&> +0&9 ]\> (09
V> ::9 VU\5V<`> F'$18,1P <UU^9
[" S9 #&.=,(> B9 @,1,2,.> N9 A,=06> cd8'8'%(3 #(,&P6%6 04 L%=:&' FSB L-7'='6 401 TK
?'&':70(Pe> TSSS TCFaBaM> #:19 <UU!9
!U" #CDSALSC> D9 _9> @#J#ZATLHC#C> H9> Z##LHaSZ> M9 F9> #CD MaAATL> A9>
QA'6%&%'(. a+'1&,P C'./0126Q9 T( K10-9 !b.7 #BM LaLK X@,(f> B,(,*,> a-.9 <UU!Y> ::9
!V!5!]^9
!!" D9 _9 #(*'16'(> C9 F',=6.'1> L9 @,8'1> H9 @,&,21%67(,(> Q?0:0&03P T(4'1'(-' 410= @_K
A08.%(3 DP(,=%-6Q> #BM LT_BaMM T(.'1('. M',681'='(. W012670:> M,16'%&&'> C0+9
<UU<9

`^

!"#!$%&'()!*
!"#$ %& '()(*+,-. /& 0(1233,24. %& 54)6)676. 8& 9,)24. :& 84,2;<=-. >& ?=+=@A. 0& >=)-,2-.
=-; B& :2,C2,4= D E'16,;,-) +4=F246(+2 =-6<=3,2* G,+H I=4,* +4=F246(+2E. ,- @46F22;,-)* 6J
'0> K-+24-2+ >2=*(42<2-+ 06-J242-F2. 5F+6L24 #MM N&
!"O$ '& ?& %=4=L=*, =-; B& '3L24+. EP<24)2-F2 6J QF=3,-) ,- B=-;6< R2+G64S*E. QF,2-F2. 163&
#T . @@& UMVDU"#. "VVV&
!"W$ 0& %=4=S=+. '& '3 8=G=3. E '-=3A*,* 6J 3,-SD32123 HAL4,; 8P0X'BYDQB J64 G,4232** 3,-S*
=-; 36-)D3,12; :0I +4=ZFE. I24J64<=-F2 P1=3(=+,6- [6(4-=3. 163& U\. -6& W. @@& W#ODUMM.
'()(*+ #MMW&
!"U$ 0& %=4=S=+. ]& K=--=FF6-2. =-; 0& ^,6+. EB=-S,-) _6G* J46< *=<@32; +4=ZFE. K- I46F22D
;,-)* 6J 06R2C+. :6(36(*2. 5F+6L24 #MMU& '1=,3=L32 ,- '--2C %&
!" $ 0& %=4=S=+. I& :H,4=-. ]& K=--=FF6-2. 0& ^,6+. I& 5G27=4*S,. E>6;23,-) K-+24-2+ L=FSL6-2

tel-00408745, version 1 - 3 Aug 2009

+4=ZF =+ +H2 _6G 32123E. KPPP :4=-*=F+,6-* 6- Q,)-=3 I46F2**,-) D Q@2F,=3 K**(2 6- Q,)-=3
I46F2**,-) ,- R2+G64S,-). 163& U". -6& T. @@& #"""D#"#W. '()(*+ #MMO&
!"\$ I& %=4J64;. '& %2*+=146*. [& %A24*. =-; >& 0461233=. E5- +H2 <=4),-=3 (+,3,+A 6J -2+G64S
+6@636)A <2=*(42<2-+*.E ,- '0> K-+24-2+ >2=*(42<2-+* `64S*H6@. aQ=- 84=-F,*F6. 0'.
bQ'N. R612<L24 #MM"&
!"T$ c& %2d24=-6 =-; B& B=*+6),& EB6L(*+ <6-,+64,-) 6J 3,-S ;23=A* =-; J=(3+* ,- KI -2+G64S*E&
K- I46F22;,-)* 6J KPPP K-J6F6<. '@4& #MMO&
!"V$ Q& %2- 842;d. :& %6-=3;. '& I46(+,242. ]& B2)-,2. [& B6L24+*. EQ+=+,*+,F=3 %=-;G,;+H
QH=4,-) e ' Q+(;A 6J 06-)2*+,6- =+ 836G ?2123E.
(%&) *$ +,-./011,

2$ 3,04,0-/

@D E@F<G3
(%') K:E$ +@1@9
A.9,-.,9;

(%X) +1@@8

/.5 6$ 7/1/8/90/. : ;</-1= !>>10?/90@. A5,.90B?/90@.;

0. C-@?$

;<4>,-0,.?, L09M -/9, ?@.9-@1 8,?M/.0N8N D@- >/?O,9 P05,@ 0. 9M,

E@8>"9,- E@88".0?/90@. 2,P0,L

3,1,>M@.=U

!"#$ %&&'$

H,?,8I-, %&&J$

3$ 3"-1,990

(%%) K:E$ +@1@9

!" #$%!&""

7$ S@NN,:C/-0N0N

H$ 3@LN1,=

A<<< AFSVEVW

+"-9@. Y$ Z'R[&\

P@1$ %Q

.@$ '

K/."/-= 'RRQ$

T!5/>90P, S<E:+/N,5 ,--@- ?@.9-@1 D@- A.9,-.,9

W/-$ 'RRR$

;7>/?,]908, 9-/5,:@^N 0. M/NM ?@50.# L09M /11@L/I1, ,--@-N;

E@88".0?/90@.N @D 9M, !EW 'X Z[\ _ `%%:`%J

(%`) H$ +-/"?OM@^ +$ 3,11,.I/?M !$ a/#.,- !$ */OM0./ W$ W/= ;A8>/?9 @D C/?O,9 7/8>10.#
@. !.@8/1= H,9,?90@. W,9-0?N;

!EW A.9,-.,9 W,/N"-,8,.9 E@.D,-,.?, %&&J

20@ 5, K/.,0-@

%&&J$

(%c) +-"91/# K$H$

+-/b01

V?9@I,-

AWE %&&J

;!I,--/.9 +,M/P0@- H,9,?90@. 0. 308, 7,-0,N D@- F,9L@-O W@.09@-0.#;

0.

C-@?$ '`9M 7=N9,8N !580.0N9-/90@. E@.D,-,.?, Z*A7! %&&&\ F,L V-1,/.N *! d7! %&&&$
(%J) 3$ +" /.5 H$ 3@LN1,=
L@-O;

;S04,5 C@0.9 !>>-@408/90@. D@- 3EC I,M/P0@- 0. /. !eW F,9:

C-@?,,50.#N @D !EW 7AfW<32AE7 %&&'$

!
"#!$ %&'()*+ ,-. /.01 2*345)6 ,7.01 89''36: ,;.01 <)=>&' ,?.01 @)6 ,<.01 A%56B&:+ C D):3>6
&6E FGH')G)6+&+356 5I &6 ?B+3() J)+K5*L <)*(3B)A. FMMM 759*6&' 56 <)')B+)E ?*)& 36
%5GG963B&+356: ,7<?%01 (5'. NO1 65. P1 8&*B= #QQN.
"#R$ 2. S. %&6+3)631 2. F&66&BB56)1 %. ;&*&L&+1 %. D35+1 T. U=3*&61 AS)I5*G9'&+36> +=) G563+5*
H'&B)G)6+ H*5V')G C WH+3G&' J)+K5*LXK3E) <&GH'36>A1 & :=5*+ ()*:356 3: 36(3+)E +5 +=)
%56I)*)6B) 56 F6I5*G&+356 <B3)6B) ,%F<<01 T*36B)+56 J71 8&*B= #QQ . ? '56> ()*:356 36
H*5B))E36>: 5I %5J)Y+1 /3:V5&1 T5*+9>&'1 D)B)GV)* #QQ . ?(&3'&V') 36 ?66)Y ?.
"#O$ 2. S. %&6+3)631 Z. J31 %. ;&*&L&+1 U. U9*')++31 AT)*I5*G&6B) ?6&'[:3: 5I \363+) /5&E
<59*B): 36 RQ#.NNV 89'+3*&+) M6(3*56G)6+:A1 %5GH9+)* %5GG963B&+356: 759*6&'1 :H)B3&'
3::9) 56 T)*I5*G&6B) F::9): 5I @3*)'):: /?J:1 T?J:1 &6E ?E ]5B J)+K5*L:1 (5'. #R1 65.
NQ1 HH. NQO^XNNQO1 796) #QQ^.

tel-00408745, version 1 - 3 Aug 2009

"PQ$ J. %&*EK)''1 <. <&(&>)1 U. ?6E)*:561 A85E)'36> U%T /&+)6B[A1

!!! "#$%$&1

8&*.

#QQQ.
"PN$ %55H)*&+3() ?::5B3&+356 I5* F6+)*6)+ D&+& ?6&'[:3: ,%?FD?01 =++H C__KKK.B&3E&.5*>_
"P#$ %&*+)*1 S. /. &6E %*5()''&1 8. M. NOO . A8)&:9*36> V5++')6)BL '36L :H))E 36 H&BL)+X:K3+B=)E
6)+K5*L:A. T)*I5*G. M(&'. #!X#R ,WB+. NOO 01 #O!XPNR.
"PP$ 8. %*5()''&1 ?. ;):+&(*5:1 A<)'IX<3G3'&*3+[ 36 @5*'E @3E) @)V U*&4B C M(3E)6B) &6E
T5::3V') %&9:):A1

!!!'(%& )*+,-+./01,- 1, "2/31*40,51

(5'. ^1 65. 1 HH. RP^XR` 1 D)B.

NOO!.
"P`$ \. D&V)L1 S. %5Y1 \. -&&:=5)L &6E S. 85**3:1 Aa3(&'E3 C ? E)B)6+*&'3b)E 6)+K5*L B55*E3X
6&+) :[:+)GA1 36 T*5B))E36>: 5I +=) ?%8 <F2%W881 T5*+'&6E1 W*)>561 ?9>9:+ #QQ`.
"P^$ ;. D566)+1 T. S&59'+1 U. \*3)EG&61 8. %*5()''&1 AD)H'5[G)6+ 5I &6 ?'>5*3+=G I5* /&*>)X
<B&') U5H5'5>[ D3:B5()*[A1 36 FMMM 759*6&' 56 <)')B+)E ?*)&: 36 %5GG963B&+356: ,7<?%01
D)B. #QQ 1 (5'. #`1 65. N#.
"P $ J.2. D94)'E1 \. /5 T*):+31 a. T&Y:561 D. U5K:')[1 AF6I)**36> '36L '5:: 9:36> :+*3H)E 963B&:+
H*5V):A1 T*5B. FMMM F6I5B5G #QQN1 ?6B=5*&>)1 ?'&:L&1 ?H*3' ##X# 1 #QQN.
"P!$ J. 2. D94)'E1 %. /96E1 &6E 8. U=5*9H. AM:+3G&+36> c5K E3:+*3V9+356: I*5G :&GH')E c5K
:+&+3:+3B:A. F6 T*5B))E36>: 5I ?%8 <3>B5GG1 ?9>. #QQP.
"PR$ MJD?%M.

D?2

J)+K5*L

8563+5*36>

F6+)*I&B)

%&*E.

=++H

C__)6X

E&B).B5G_6)+K5*L8%&*E:.=+G
"PO$ %. M:+&6 &6E 2. a&*>=):). AJ)K E3*)B+356: 36 +*&4B G)&:9*)G)6+ &6E &BB596+36>A. F6
T*5B))E36>: 5I ?%8 <3>B5GG1 ?9>. #QQ#.
"`Q$ U. <. M9>)6) J>1 ]93 d=&6>1 AT*)E3B+36> F6+)*6)+ J)+K5*L D3:+&6B) K3+= %55*E36&+):X
;&:)E ?HH*5&B=):A1 FJ\W%W8 #QQ#.
"`N$ 8. \&'59+:5:1 T. \&'59+:5: &6E %. \&'59+:5:1 AW6 T5K)* /&K S)'&+356:=3H: 5I +=) F6+)*6)+
U5H5'5>[A1 <F2%W88 NOOO.

!"#!$%&'()!*

"#$% &' ()*+,-../ &' 0' 12*3)45/ 6' 78-.9/ -.+ :' :2**2.9)4' ;<=.-,2>? @A B6 54-C> D &
?58+= @A 5E) 4@*) @A F-42-32*25= -.+ 5E) 2,G->5 @A >@.54@*;' B. 64@>' &0H IB10JHHKLL/
GG' MNOPMOM/ OLLL'
"#M% &' ()*+,-../ &' 14)).3)49/ 0' Q8.+/ R' S)2.9@*+/ T' S)UA@4+/ -.+ (' V48)' ;<)42F2.9
54-C> +),-.+? A@4 @G)4-52@.-* 2G .)5W@4X? D H)5E@+@*@9= -.+ )UG)42).>);' B. 64@>))+2.9?
@A &0H I29>@,,/ I)G5' $NNN'
"##% I' (*@=+/ Y' (-**/ ;64@,@52.9 5E) Z?) @A [.+\V@\[.+ 0@.9)?52@. 0@.54@* 2. 5E) B.5)4.)5;/
B[[[]&0H V4-.?->52@.? 2. R)5W@4X2.9/ F@*' ^/ .@' #/ GG' #_!\#^$/ &89' OLLL'
"#_% I' (*@=+/ H' 7-.+*)=/ T' 6-+E=)/ `[a8-52@.\3-?)+ >@.9)?52@. >@.54@* A@4 8.2>-?5 -GG*2>-\
52@.?b/ 2. G4@>))+2.9? @A &0H IB10JHH/ &89' $NNN'
"# % (4-*)29E/ 0'/ <2@5/ 0'/ Q=*)?/ c'/ H@@./ I'/ JW)d-4?X2/ 6'/ 6-G-92-..-X2/ <'/ V@3-92/ (' D

tel-00408745, version 1 - 3 Aug 2009

;<)?29. -.+ +)G*@=,).5 @A - G-??2F) ,@.25@42.9 2.A4-?548>584);' Q)>584) R@5)? 2. 0@,\

G85)4 I>2).>) $O^N e$NNOf'
"#^% H' (8**,)4 -.+ I' S@,29' ;VE) JIZ g@W\5@@*? G->X-9) -.+ 02?>@ R)5(*@W *@9?;' B.
64@>))+2.9? @A Z?).2U QBI&/ <)>' $NNN'
"#!% 1[&RV' E55G D]]WWW'+-.5)'.)5'
"#L% Y' 7-4A@8?E/ &' c)?5-F4@? -.+ T' c=)4?' ;S@38?5 B+).52h>-52@. @A IE-4)+ Q@??)? Z?2.9
[.+\5@\[.+ Z.2>-?5 64@3)?;/ 64@>' @A 5E) !5E B[[[ B.5)4.-52@.-* 0@.A)4).>) @. R)5W@4X
64@5@>@*? eB0R6f/ J?-X-/ T-G-./ R@F),3)4 $NNN/ GG' $$\MM'
"_N% I' 7-=X2.' H@+)4. (2*5)4?' H->,2**-./ R)W i@4X/ OL!L'
"_O% R' 7@E. -.+ <' j)25>E' ;B.F)452.9 ?-,G*)+ 54-C>;' B. 64@>))+2.9? @A &0H B.5)4.)5 H)-\
?84),).5 0@.A)4).>)/ J>5' $NNM'
"_$% T' 7@45@. -.+ &' Q@G)d\J452d' ;J. 5E) .8,3)4 @A +2?542385)+ ,)-?84),).5 G@2.5? A@4
.)5W@4X 5@,@94-GE=;' B. 64@>))+2.9? @A &0H B.5)4.)5 H)-?84),).5 0@.A)4).>)/ R@F'
$NNM'
"_M% R' 78 -.+ 6' I5)).X2?5)' ;[F-*8-52@. -.+ 0E-4->5)42d-52@. @A &F-2*-3*) c-.+W2+5E V)>E\
.2a8)?;' B[[[ TI&0 IG)>2-* B??8) 2. B.5)4.)5 -.+ ::: H)-?84),).5/ H-GG2.9/ -.+
H@+)*2.9/ $NNM'
"_#% R' 78/ Q' [' Q2/ k' H' H-@/ 6' I5)).X2?5)/ -.+ T' :-.9/ ;Q@>-52.9 2.5)4.)5 3@55*).)>X? D
-*9@425E,?/ ,)-?84),).5?/ -.+ 2,G*2>-52@.?;/ B. IB10JHH $NN# D 64@>))+2.9? @A 5E)
$NN# >@.A)4).>) @. &GG*2>-52@.?/ 5)>E.@*@92)?/ -4>E25)>584)?/ -.+ G4@5@>@*? A@4 >@,G85)4
>@,,8.2>-52@.?/ GG' #O\_#/ R)W i@4X/ Ri/ ZI&/ $NN#/ &0H 64)??'
"__% B6 (*@W B.A@4,-52@. )lG@45 :@4X2.9 14@8G' B.5)4.)5 [.92.))42.9 V-?X (@4>)'
E55G D]]WWW'2)5A'@49]E5,*'>E-45)4?]2GhU\>E-45)4'E5,*'
"_ % j' T->@3?@./ `0@.9)?52@. -F@2+-.>) -.+ >@.54@*b/ &0H IB10JHH/ &89' OL!!'

!
"#$% &'( )'*+,-+(. /0'12*2'34' 1++5 1+ 6(783 *2'3'*1836-16*- +7 9(183(81 0'12-/. :38-8(4
18; '1 128 <'128='16*'5 >*68(*8- ?8-8'3*2 9(-161@18A<>?BC D -56;8- 'E'65',58 73+=
710 FGG710.88.5,5.H+EG0'12*2'3GI J0365 B!!$.
"#K% )'6(I <. '(; L+E3+56-I M. NOON. /P(;41+48(; 'E'65',58 ,'(;Q6;12 F =8'-@38=8(1 =812+;+4
5+HRI ;R('=6*-I '(; 385'16+( Q612 SM: 123+@H20@1/. >9TMU<< M+=0@1. M+==@(. ?8E.
VNI W AU*1. NOONCI N!#4VOK.
"#!% >. )'6-Q'5I T. 9'(('**+(8I M. L6+1I ). X@3+-8I L. S+Q-58RI /9(78336(H SM: M+((8*16+(
M2'3'*1836-16*- S23+@H2 :'--6E8 <8'-@38=8(1-/I 9PPP 9(7+*+=I Y+(H X+(HI <'3*2 NOOW.
" O% >. )'=6(I M. )6(I Z. )6(I L. ?'[I Z. >2'E611I '(; \. ]2'(H. U( 128 05'*8=8(1 +7 6(183(81
6(-13@=8(1'16+(. 9( :3+*88;6(H- +7 9PPP 9(7+*+=I J03. NOOO.
" B% ^6(H )6(I _.4:. X8( Z6@I >.4Y. T'3R M2'( '(; Z'`@( _'(HI /a81Q+3b S+0+5+HR 9(7838(*8

tel-00408745, version 1 - 3 Aug 2009

c'-8; +( P(;41+4P(; <8'-@38=8(1-/I 9PPP )>JM -08*6'5 6--@8 +( -'=056(H 128 9(183(81I
&+5. NWI a+. BNI 00. NBKN4NB!#I L8*. NOO .
" N% )@(6083 S3'd* >'=056(H. 2110 FGGQQQ.`@(6083.(81. a81e+Q4*+=0'16,58 6=058=8(1'16+(.
" V% <. J. X''7'3I \. <'12RI M. c'3'b'1I X. >'5'='16'(I S. S@358116I _. L',,+@-I />8*@36(H
9(183(81 M++3;6('18 P=,8;;6(H >R-18=-/I 6( 03+*88;6(H- +7 JM< >9TMU<<I XR+1+I
)'0'(I J@H@-1 NOO$.
" W% ?.P. X'5='(I J a8Q J003+'*2 1+ \6(8'3 f651836(H '(; :38;6*16+( :3+,58=-I 6( S3'(-'*4
16+(- +7 128 J><P 4 )+@3('5 +7 c'-6* P(H6(8836(H &+5. KN F 00. V#4W#I B! O.
" #% >. X'116I L. X'1',6I M. c5'b8I P. X+2583I '(; ). >13'@--I /<@516g F J@1+='18; L818*16+(
+7 <@516058 c+1158(8*b- J5+(H ' :'12/I JM< 9<M NOOW.
" % >. X8-2'EI hJ *+(13+54128+3816* '003+'*2 1+ e+Q *+(13+5I JM< >9TMU<<I >80. B!!B.
" $% J. X367'I M. c'3'b'1I S. >0R3+0+@5+-I /U016='5 c@i83 <'('H8=8(1 :+56*68- 7+3 L85'R
S+583'(1 a81Q+3b-/I 6( 03+*88;6(H- +7 128 #12 9PPP M+(7838(*8 +( >8(-+3I <8-2 '(; J;
Y+* M+==@(6*'16+(- '(; a81Q+3b- A>PMUa NOOKCI >'( f3'(*6-*+I )@(8 NOOK.
" K% \JXY9aJI J.I M?U&P\\JI <.I JaL L9USI M. /L6'H(+-6(H (81Q+3b4Q6;8 13'd* '(+4
='568-/. 9( :3+*88;6(H- +7 JM< >9TMU<< ANOOWC.
" !% J. \'b26('I <. M3+E855' '(; M. L6+1. /<6(6(H J(+='568- j-6(H S3'd* f8'1@38 L6-136,@4
16+(-/. J008'3- 6( JM< >9TMU<<I :265';85026'I :JI J@H@-1 NOO#.
"$O% _. \85'(;I <. S'kkI _. _6556(H83I L. _65-+(I /U( 128 -8574-6=65'3 ('1@38 +7 P1283(81
13'd*I/ JM< >9TMU<<I>80. B!!V.
"$B% \6,0*'0. 2110 FGGQQQ.1*0;@=0.+3H.
"$N% <'*263'`@I >.I &861*2I L.I c'**8556I f.I '(; c+5+1I ). M. /J;;6(H ;8l(616+( 1+ '*16E8 03+4
,6(H/. >9TMU<< M+=0@1. M+==@(. ?8E. V$I N A<'3. NOO$CI B$4NK.
"$V% \61Q6(I _. AB!KOC. /\6(8'3 2'-26(H F J (8Q 1++5 7+3 l58 '(; 1',58 ';;38--6(H/. :3+*. 12
M+(7838(*8 +( &83R \'3H8 L'1','-8- F 0'H8- NBN4NNV.

!"#!$%&'()!*

" #$ %& '()& *+, -./0102(3 '4563 47 899: ;6<=41> 91(?2*@ :1426650,AB 47 C;DEFE' GH @
I4J6@ K(/(,@ +/103 LHH &
" M$ K& '(0@ +& N105)(1(,@ FO;& F)P()@ 9& Q6@ (,5 8& R(,A@ *C./(2< 47 :(2>6< N(./30,A
4, :41<B2(, +,4.(3S T6<62<04,@* C--- K4P1,(3 4, N6362<65 +16(B 47 F4..P,02(<04,B O
N/620(3 CBBP6 4, N(./30,A <)6 C,<61,6<@ U43& V#@ ,4& LV@ //& VVWMOVVHW@ T626.J61 V!!X&
" X$ '& '(330@ F& %(1(>(<@ Y& T(JJ4PB@ *+//302(<04,O36U63 U61BPB ;6<=41>O36U63 :14Z0.0<S*@ 0,
/1426650,AB 47 <)6 +C;9-F 24,7616,26@ %(,A>4>@ T626.J61 V!!M&
" $ '& '(330@ F& %(1(>(<@ Y& T(JJ4PB@ *F8-NN [ +, (//302(<04,O(=(16 B/(26 741 6,)(,265
B2(3(J36 B61U026B 0, 4U613(S ,6<=41>B*@ F4./P<61 F4..P,02(<04,B K4P1,(3@ U43& \L@ ,4& X@
//& LV\HOLVM\@ +/103 V!!W&
" W$ ;& '43361@ F& %(1(>(<@ I& +U1(2)6,>4U@ (,5 -& +3<.(,@ *C,<61O/14<4243 7(01,6BB J6<=66,

tel-00408745, version 1 - 3 Aug 2009

9F: ;6= ]6,4 (,5 9F: Y6B<=445^*@ 0, /1426650,AB 47 ;_C V!! `F4,7616,26 4, ;6Z<
_6,61(<04, C,<61,6< ;6<=41>Ba@ 914,5)60.@ ;41=(S@ '(S V!! &
" H$ R& '& '(4@ K& ]6Z7415@ K& Y(,A@ (,5 ]& 8& I(<b& *94=(15B (, (22P1(<6 +NO36U63 <1(2614P<6
<443*& C, NC_FE'' V!!\&
"W!$ N& '2F(,,6 (,5 c& K(24JB4,& *9)6 %NT /(2>6< d3<61 [ + ,6= (12)0<62<P16 741 PB61O36U63
/(2>6< 2(/<P16*& C, :1426650,AB 47 <)6 eN-;Cf Y0,<61 F4,7616,26@ K(,P(1S LHH\&
"WL$ +& '650,(@ ;& 9(7<@ I& N(3(.(<0(,@ N& %)(<<(2)(1SS(@ (,5 F& T04<& *91(?2 .(<10Z 6BO
<0.(<04, [ -Z0B<0,A <62),0gP6B (,5 ,6= 50162<04,B*& C, :1426650,AB 47 +F' N0A24..@
:0<<BJP1A)@ +PA& V!!V&
"WV$ %& '63(,561@ '& %h41>.(,@ (,5 :& _P,,0,AJ61A& *+ ,6= 6,5O<4O6,5 /14J0,A (,5 (,(3SO
B0B .6<)45 741 6B<0.(<0,A J(,5=05<) J4<<36,62>B*& C, C--- _34J624. O _34J(3 C,<61,6<
NS./4B0P.@ N(, D1(,20B24@ ;4U6.J61 V!!!&
"W\$ '4AP3@ K& (,5 N& T6610,A@ * :(<) '9e 50B24U61S*@ ]DF LLHL@ ;4U6.J61 LHH!&
"W#$ N& '44,@ :& N>633S@ (,5 T& 94=B36S@ *-B<0.(<04, (,5 ]6.4U(3 47 F342> N>6= 714. ;6<=41>
T63(S '6(BP16.6,<B@* :1426650,AB C--- C,7424.@ ;Q@ '(1& LHHH&
"WM$ F0B24 ;6<i4=@ )<</ [jj===&20B24&24.j
6,jeNj/145P2<Bj/BXX!Lj/145P2<Bk04Bk/14<4243kA14P/k)4.6&)<.3
"WX$ 8& ;APS6, (,5 :& 9)01(,& *+2<0U6 .6(BP16.6,< 741 .P3<0/36 30,> 7(03P16B 50(A,4B0B 0, C:
,6<=41>B*& C, :1426650,AB 47 :+' Y41>B)4/@ +/1& V!!#&
"W $ K& ;4,,6,.(2)61 (,5 -& %061B(2>& *N2(3(J36 7665J(2> 741 3(1A6 A14P/B*& C---j+F' 91(,B&
4, ;6<=41>0,A@ `\a [\ MO\WX@ KP,6 LHHH&
"WW$ 9)6 l%;l ;6<=41> N0.P3(<41@

+,-.@ )<</ [jj===&0B0&65Pj,B,(.j,Bj

"WH$ +& ;P220@ ;& 9(7<@ F& %(1(>(<@ :& 9)01(,@ *F4,<143365 eB6 47 -Z26BB %(2>J4,6 %(,5=05<)
741 :14U050,A ;6= N61U026B 0, C:O4U61OYT' ;6<=41>B*@ C--- K4P1,(3 4, N6362<65 +16(B
0, F4..P,02(<04,B@ U43& VV@ ,4& H@ //& LXHVOL ! @ ;4U6.J61 V!!#&

!
"#$% &' &())*+ ,' &)-.(/-+ )-0 1' 2.3//' 45-26- .67867)9:6-;< <=*6> ? @ A/9(60 B67 A6-:96C
7:-D 97)E. :- <>:9.(/0 )-0 76F9/0 -/9>67G<4' H=I J! K+ ,/89' L$$!'
"#!% @- 68/-+ 0:<97:MF9/0 8*)9B67A B67 0/N/*68:-D+ 0/8*6O:-D )-0 )../<<:-D 8*)-/9)7OC <.)*/
-/9>67G </7N:./<+ <// (998 ?PP>>>'8*)-/9C*)M'67DP'
"#L% Q' &*6-G)' 4=*6><.)- ? @ -/9>67G 97)E. R6> 7/8679:-D )-0 N:<F)*:S)9:6- 966*4' 5- &76.//C
0:-D< 6B T</-:U V5,@+ Q/.' L$$$'
"#J% &6<9/*+ W'+ 45-9/7-/9 I6-976* 2/<<)D/ &7696.6*4+ H=I

+ T,IP5-B67A)9:6- ,.:/-./< 5-<C

9:9F9/+ @87:* !#X!'

"#Y% &).G/9

,)A8*:-D

Z67G:-D

[76F8'

5-9/7-/9

\-D:-//7:-D

])<G

=67./'

(998 ?PP>>>':/9B'67DP(9A*'.()79/7<P8<)A8C.()79/7'(9A*'

tel-00408745, version 1 - 3 Aug 2009

"#^% _' H)A6<+ I' `)7)G)9+ \' @*9A)-+ 4@ 26N:-D @N/7)D/ &7/0:.967 B67 &*)O6F9 Q/*)O I6-976*
:- _65&4+ :- 876.//0:-D< 6B 9(/ \*/N/-9( 5-9/7-)9:6-)* Z67G<(68 6- aF)*:9O 6B ,/7N:./
b5Za6, L$$Jc+ 26-9/7/O+ I@+ WF-/ L$$J'
"#K% ,' H)9-)<)AO+ 2' d)-0*/O+ H' 3)78+ ,' ,(/-G/7+ 4]686*6D:.)**OC@>)7/ eN/7*)O I6-<97F.C
9:6- )-0 ,/7N/7 ,/*/.9:6-4+ 5\\\ 5-B6.6A L$$L'
"# % f' H/G(9/7 /9 ]' V:+ 4@ `670/7 [)9/>)O &7696.6* b`[&CYc4' H=I ! !+ 2)7< !##^'
"#X% H/-)9/7' (998 ?PP>>>'7/-)9/7'B7'
"##% `' =' H:M/:76+ Q' =' ]6><*/O+ ]' f/+ W' `6*69+ 4=:<(/7 :-B67A)9:6- 6B <)A8*/0 8).G/9< ? ))88*:.)9:6- 96 R6> <:S/ /<9:A)9:6-4' 5-9/7-/9 2/)<F7/A/-9 I6-B/7/-./ L$$K ? !^CLK
"!$$% _'W' H:M/:76+ H'd' H:/0:+ H'[' `)7)-:FG+ W' 1)N7)9:*+ V' I6997/**+ 4&)9(I(:78 ? \E.:/-9
@N):*)M*/ `)-0>:09( \<9:A)9:6- B67 1/9>67G &)9(<4+ &)<<:N/ )-0 @.9:N/ 2/)<F7/A/-9
Z67G<(68+ L$$J'
"!$!% T-:N/7<:9O 6B e7/D6- H6F9/ _:/>< &76g/.9' (998 ?PP>>>'76F9/N:/><'67DP
"!$L% 3' ,M):+ I' `)7)G)9+ 4\U8/7:/-./< 6- /-()-.:-D 0)9) .6**/.9:6- :- *)7D/ -/9>67G<4+ 51H5@
]/.( H/8679 :-7:)C$$!Y$#J + 96 )88/)7 :- I6A8F9/7 1/9>67G<' \U9/-0/0 N/7<:6- 6B 9>6 8)C
8/7< ? 6-/ :- @--)*< 6B ]/*/.6AAF-:.)9:6-<+ N6*' K!+ -6' !CL+ 88' !K C!#L+ W)-F)7OC=/M7F)7O
L$$K+ )-0 9(/ 69(/7 :- 876.//0:-D< 6B 1]2, L$$ bI6-B/7/-./ 6- 1/> ]/.(-6*6D:/<+ 26C
M:*:9O )-0 ,/.F7:9Oc+ &)7:<+ 2)O L$$ ' @N):*)M*/ :- @--/U Q'
"!$J% 2' 3' ,M):+ I' `)7)G)9+ W' I(6:+ @' @* d)A7)+ ]' ]F7*/99:+ 4@0)89:-D `:9]677/-9 96 >:7/*/<<
)0 (6. -/9>67G<4 :- 876.//0:-D< 6B 9( 5-9/7-)9:6-)* .6-B/7/-./ 6- )0 (6. -/9>67G< )-0
>:7/*/<< L$$X b@QCdeI 1eZc+ ,68(:) @-9:86*:<+ =7)-./+ ,/89/AM/7 L$$X'
"!$Y% ,G:99/7 :-B7)<97F.9F7/ MO I@5Q@+ (998 ?PP>>>'.):0)'67DP966*<PA/)<F7/A/-9P<G:99/7P
"!$^% 1' ,87:-D+ H' 2)()g)-+ Q' Z/9(/7)**+ 42/)<F7:-D 5,& ]686*6D:/< >:9( H6.G/9BF/*4+ @I2
,5[Ie22 L$$L'

!"#!$%&'()!*

"#$%& '( )*+,-. '( /012340. 5( 6078. 9( :0;0<3044013. 9( )0,0=08304. '( 5+>>3. ?( @A*B-,,0 04C
@( D3*8( E6A0F> ?08A3>-G H I0,04>34< ?-0G+A-=-48G. J47-A-4>- 04C ?*C-,34<E( J4 '@?
)3<=-8A3>G( K+4- !$$L(
"#$ & )8A0+GG. K(. 9080M3. D(. 04C 900G2*-1. E' =-0G+A-=-48 G8+CN *7 0B03,0M,- M04CO3C82
-G83=083*4 8**,GE( J4 :A*>--C34<G *7 82- PAC '@? )JQ@R?? @*47-A-4>- *4 348-A4-8
?-0G+A-=-48 S?30=3 I-0>2. T/. U)'. R>8*M-A !$$PV(
"#$W& 9( )+2. X( Q+*. K( 9+A*G-. 04C D( 6*OG,-N( E/*>0834< 4-8O*A1 =*438*AG H @*=;,-Y38N.
2-+A3G83>G 04C >*B-A0<-E( J4 :A*>--C34<G *7 JZZZ J47*>*=. ?0A( !$$L(
"#$[& ?( \04<A3,,3 04C I( I( /*O-10=;. E';;,N34< :A34>3;,-G *7 '>83B- 'B03,0M,- I04CO3C82
',<*A382=G 8* :0GG3B- 6@: 6A0>-G.E %82 ]*A1G2*; *4 :0GG3B- 04C '>83B- ?-0G+A-=-48G
S:'?!$$LV. /5@) P^P#. ?0A>2 !$$L(

tel-00408745, version 1 - 3 Aug 2009

"##$& _( \2-4<. Z( 9( /+0. ?( :30G. 6( Q( QA3F4( EJ48-A4-8 `*+834< :*,3>3-G 04C `*+4Ca6A3;a
63=-GE. :0GG3B- 04C '>83B- ?-0G+A-=-48 ]*A1G2*; S:'?V !$$L(
"###& ]'QR5 H ' ]-M )-AB-A I-4>2=0A134< 6**,. J5`J'. ?3G8A0, A-G-0A>2 <A*+;(
"##!& `(]( ]*,b( )8*>20G83> ?*C-,34< 04C 82- 62-*AN *7 c+-+-G( :A-483>-a _0,,. Z4<,-O**C
@,3bG. 5K. #[W[(

tel-00408745, version 1 - 3 Aug 2009

! "#$%&''#
!"#
$%&'(
)!"%*+ ,'!'-.
/%'%0#1.(%2
344&#..
5(61# 789:#;
5(61# 7,#:&#'!&%!';
)!< 789:#;
="!%*
>#? 5!@#

!"#$% &'(')'*
+#%$# ,-./#01023 456" 780. 459:
;#<<%.$ 61 718=#0# ;##>>#?%
@<.0A"
BC(B' D EFF:3 <186. $.G -8A%1>.G D FH5FE +1I"%# '06%I1>%G D @<#0A.
JKK : 5E KL 9M 5H
JKK : 5E KL 9L EM
JKK : 5E KL 95 9L
!"#$%N&#<#O#6PG1I"%#N%0<%#NQ<
"66I

RRI>#0.6.N%0<%#NQ<RA"#$%

!"#$%%&"' (
S FKRFE D

(.G.#<A" GA%.06%G63 T>#0U6. I<1 V.A6D6.#=3 BC(B' +1I"%# '06%I1>%G D ;W$%D

6.<<#0W.3 @<#0A.N
S FKRF: D FLRF:

B0X%6.$ @#A8>6Y ;.=/.<3 B06.> (.G.#<A" !#=/<%$Z.3 [)N

S FMRF4 D FKRFE

T1G6$1A61<#> Q.>>1?3 -!' D @B! D \T@-3 -#8G#00.3 +?%6].<>#0$N

S F5R5L D FMRF4

T"^ G68$.063 ;%G6<#> I<1V.A6D6.#=3 BC(B'3 @<#0A.N

S FKR5L D F9R5L

B06.<03 0.6?1<O%0Z $.I#<6=.063 \C+*DT#<%G3 @<#0A.N

)*+,-.&"' (
S 'I<%> EFF4
S
S
S
S

T"^ $.Z<.. %0 !1=I86.< +A%.0A.G

A1%B#&.%'+ C
D!?6&!'6&+ C
34B%.6& C
,-?E#:' C

S 780. 455L

C%A. D +1I"%# '06%I1>%G3 @<#0A.N

;%G6<#> I<1V.A6D6.#=3 BC(B'3 @<#0A.N

\%6#0 '>6=#0N

T.<Q1<=#0A. .X#>8#6%10 1Q *!T A10Z.G6%10 A106<1>N

^\' ,^%I>_=. $`a68$.G 'II<1Q10$%.G3 6". Q<.0A" .b8%X#>.06 1Q 6". ;#G6.<

$.Z<..2 %0 C.6?1<O%0Z #0$ ^%G6<%/86.$ +YG6.=G3 [0%X.<G%6Y 1Q C%A. D +1I"%#D'06%I1>%G3

@<#0A.N
S 78>Y 4559

\>.A6<%A#> #0$ \>.A6<10%AG \0Z%0..<%0Z $.Z<..3 -./#0.G. [0%X.<G%6Y3 &.%<863

-./#010N

9K

! "#$%&''#

!"#$% &
" #$%& '('$) (*()+ ,-) ./ '('$) 01'&2.(3 #45$) 6(7(8$.$7& 9-32:2$% ,-) ;$3(/ <-3$)(7&
=$&*-)>%0? @AAA BAC1= :-7,$)$7:$? B(7 D)(7:2%:-? E47$ FGGHI
" #$%& 9J; %&4+$7& '('$) (*()+ ,-) ./ '('$) 017 KCL M3&$)278 -7 ( %3-* )$N$)%$ :J(77$30?
M)%& 27&$)7(&2-7(3 *-)>%J-' -7 O4(32&/ -, ,4&4)$ @7&$)7$& B$)N2:$%? #$)327? B$'&$.P$) FGGGI
" Q()2)2 ,-47+(&2-7 (*()+? &-' RGG ,47+$+ P/ &J$ ,-47+(&2-7 27 S$P(7-7 (7+ (P)-(+? MN$
&2.$% ,)-. RTTU &- RTT I

'(%("#)* +,-(#(%-% &

@7&$)7$& .$(%4)$.$7& (7+ &)(V: (7(3/%2%? :-78$%&2-7 (7+ $))-) :-7&)-3

27 :-.'4&$) 7$&*-)>%? '$),-).(7:$ $N(34(&2-7 -, *2)$3$%% 7$&*-)>%? N-2:$ -N$) @9? 9$$)W&-W9$$)
():J2&$:&4)$%? 7$&*-)> :-+278? ;$3(/ <-3$)(7& =$&*-)>278I

tel-00408745, version 1 - 3 Aug 2009

'(%("#)* .#/0()-% &

" =$&*-)> +2(87-%2% P/ :-33(P-)(&2N$ .-72&-)278I B4''-)&$+ P/ &J$ C61= K=X ')-Y$:&I
" C-7&$7& %J()278 27 *2)$3$%% 7$&*-)>%I B4''-)&$+ P/ &J$ @<AK AZ'$%J()$ ')-Y$:&I
" K+('&2N$ &)(V: %(.'3278I B4''-)&$+ P/ &J$ AC1;A D9 ')-Y$:&I
" [7+$)%&(7+278 &J$ 9)-'$)&2$% -, &J$ #2&<-))$7& 1N$)3(/I @72&2(33/ %4''-)&$+ P/ &J$ 1BCKX
K=X ')-Y$:&I
" B:J$+43278 (7+ )$%%-4):$ :-7&)-3 27 ;$3(/ <-3$)(7& =$&*-)>%I
" B:(3(P3$ %-34&2-7% ,-) +(&( :-33$:&2-7I @72&2(33/ %4''-)&$+ P/ Q2&(:J2 A4)-'$I
" 9$),-).(7:$ $N(34(&2-7 -, :-78$%&2-7 :-7&)-3 27 &J$ @7&$)7$&I

.#/1(%%2/,"3 ")-242-2(% &

" C-W-)8(72\$) -, &J$ ,)(7:-'J-7$ %4..$) %:J--3 X$%C-. FGG I
" ]4$%& $+2&-) ,-) ( EBKC %'$:2(3 -7 B(.'3278 &J$ @7&$)7$& ^ <$:J72_4$% (7+ K''32:(&2-7%I
" K)$( $+2&-) ,-) KC6 C-.'4&$) C-..472:(&2-7 X$N2$* `CCXa? FGGb W ')$%$7&I
" B&$$)278 :-..2&&$$ .$.P$) ,-) 9K6 *-)>%J-' `9(%%2N$ (7+ K:&2N$ 6$(%4)$.$7&aI
" 9)-8)(. C-..2&&$$ :-W:J(2) ,-) &J$ B(.'3278 FGGb *-)>%J-' `9()2%aI
" ]$7$)(3 :J(2) ,-) c21'& FGGb *-)>%J-'% `X2N( ;$3 ]()+(? @&(3/aI
" ]$7$)(3 :J(2) ,-) 9K6 FGG! `E4(7W3$%W927%? D)(7:$aI
" <$:J72:(3 9)-8)(. C-..2&&$$ .$.P$) ,-) K38-&$3 FGGT `D)(7:$a? @<C FGGT `9()2%a? @7,-W
:-. FGGT `X2-a? #)-(+7$&% FGGH `S-7+-7a? c=BF FGGH `K&J$7%a? c276$$ FGGH `#$)327a?
9DS;7$& FGGH `6(7:J$%&$)a? =B<--3% FGG `=(7&$%a? BAC1= FGG `B(7 ;2$8-a? cc@C
FGG `C-2.P)(a? @cO-B FGGd `=$* Q(N$7a? cc@C FGGd `#$)7a? c1=B FGGd `S/-7a?
c2=6$$ FGGd `#-%&-7a? @6C FGGb `#$)>$3$/a? @C=9 FGGb `#-%&-7a? 9K6 FGGb `#-%&-7a?
c1=B FGGb `B&I 6-)2&\a? @=D1C16 FGGb `62(.2a? KBc= FGG! `#-%&-7a? ]@e=]= %/.W
'-%24. (& ]3-P$:-. FGG! `;(33(%a? 9K6 FGG! `E4(7W3$%W927%a? @=D1C16 FGG! `Q-78
L-78a? KB@K= FGGF `Q(7-2a? @C=9 FGGF `9()2%aI
" B$%%2-7 :J(2) (& @6C FGGb `#$)>$3$/a? c21'& FGGU `B-'J2( K7&2'-32%a? @C=9 FGGF `9()2%aI

!
" #$%&'() *+(,-(. ,' /.0$(1 2$&%(.*&'3 4256 7 899:; <=>? 8:! @$,1 *3A+B*&0A 4>0..(3; 256
7 899 ; C*&,$ *DEBB1 4/,$F-B-6 7 899!; G7$(H' *DEBB1 4IB0%,&$6 7 899!; 5?J 4>'BD-EB1A6
7 899!; GK># 4?0$&*6 7 899!; L#? 4/B*'B$6 7 899M; 2LC>> 4CAE(.*'6 7 899M; /B*'B$
2$&%(.*&'3 4/B*'B$6 7 899M; #$'(1 N(*(,.DE <,AO.&)F( 7 899M; GK># 4?0$&*6 7 899M; C>#CK
*DEBB1 4/,$F-B-6 7 899P; GDB1( KB.A,1( >0+Q.&(0.( 4R,.&*6 7 899P; GDB1( )SQ'Q #$'(.$('
KB0%(11( TQ$Q.,'&B$ 4RB.U0(.B11(*6 7 899P; GK># 4?0$&*6 7 899P; GRVI >0AA(. N(*(,.DE
#$*'&'0'( 4I,0*,$$(6 7 8998; #/L 4W0.&DE6 7 899XY
" >+(,-(. &$ A,$3 &$'(.$,'&B$,1 DB$Z(.($D(* [ >G<=K 899: 4>,$ V.,$D&*DB6; <BK(H' 899\
4I&*OB,6; <BK(H' 899! 4?B01B0*(6; ]&=+' 899P 4R,.&*6; RVI^$(' 899P 4T($(%,6; #L]
8998 4L,.*(&11(6; #<K 899X 4<B1A,.6; #KV=<=L 899X 4C1,*-,6; _BZ#> 8999 4/(.1&$6; >#T7
<=LL 8999 4>'BD-EB1A6; >#TLG?N#<> 8999 4>,$', <1,.,6; KG?]=N5#KT 8999 4R,7
.&*6; TI=/G<=L X``` 4N&B6Y

tel-00408745, version 1 - 3 Aug 2009

" N(Z(.(( ZB. A,$3 &$'(.$,'&B$,1 DB$Z(.($D(* 4>#T<=LL; #KV=<=L; <B7$(H'; >#TLG7
?N#<>; #<KR; TI=/G<=L; KG?]=N5#KT; ]&=+'; #><<; C>#CK; K('7<B$;YYY6 ,$)
aB0.$,1* 4?.,$*,D'&B$* B$ K('bB.-&$F; ?.,$*,D'&B$* B$ L01'&A()&,; ?.,$*,D'&B$* B$ ]&7
.(1(** <BAA0$&D,'&B$*; c>C<; R(.ZB.A,$D( G%,10,'&B$; #GGG K('bB.-*; <BAA0$&D,'&B$
L,F,d&$(; C<L <<N; <BA+0'(. K('bB.-*; <BA+0'(. <BAA0$&D,'&B$*;YYY6Y
" L(AO(. BZ 'E( )&.(D'B.3 OB,.) ZB. 'E( K('bB.-&$F L,*'(. ,' 'E( 2$&%(.*&'3 BZ K&D(7>B+E&,
C$'&+B1&*; .(*+B$*&O1( BZ 'E( &$'(.$*E&+* ZB. 'E( *,A( L,*'(.; ,$) (H'(.$,1 A(AO(. &$ 'E(
E&.&$F DBAA&''(( ,' 'E( DBA+0'(. *D&($D(* )(+,.'A($' BZ 'E( 2$&%(.*&'3 BZ K&D(7>B+E&,
C$'&+B1&*Y
" L(AO(. BZ 'E( G7$(H' G0.B+(,$ K('bB.- BZ GHD(11($D(Y 899P7899\Y
" 4<B6=.F,$&d(. BZ L&*'.,1 *(A&$,.* 4X```7899X6; BZ R1,$e'( *(A&$,.* 4899P 7 6; BZ f*(A&$,.
D.B&*Q NQ*(,0f 4=D'BO(. 899P6 ,$) BZ #KN#C <BA/ (%,10,'&B$ *(A&$,. 4KB%(AO(. 899 6Y

!"#$%&'() *+))(,+&(%'+!$ -

" R,.'$(. BZ 'E( <L=K CKN V.($DE +.Ba(D' 1() O3 ?EBA*B$Y

" R,.'$(. BZ 'E( G<=^G >?NGR G0.B+(,$ +.Ba(D' 1() O3 C1D,'(1 /(1F&0AY
" R,.'$(. BZ 'E( GH+(*E,.( #?GC G0.B+(,$ +.Ba(D' &$%B1%&$F B%(. M9 +,.'$(.* 4KB-&,; RE&7
1&+*; g??; >&(A($*; ?(1(ZB$&D,; ('D6Y
" R,.'$(. BZ 'E( =><CN V.($DE NKN? 1() O3 V.,$D(?(1(DBA ,$) &$%B1%&$F L&'*0O&*E&Y
" C %&*&' BZ *&H AB$'E* 'B #$'(1 N(*(,.DE <,AO.&)F(Y
" C DB11,OB.,'&B$ BZ B$( 3(,. b&'E C1D,'(1 L,.DB0**&* B$ #$'(.$(' ?B+B1BF3 #$Z(.($D(Y
" ?NCK>C? [ G>C R.Ba(D' 1() O3 C1D,'(1 *+,D( B$ &A+.B%&$F *,'(11&'( ,DD(** $('bB.-*Y
" ?#<R [ R.Ba(D' b&'E J&',DE& B$ DB11(D'&$F &$ZB.A,'&B$ &$ b&)( ,.(, $('bB.-*Y
" <B11,OB.,'&B$ b&'E >+.&$' C?I ,$) GRVI B$ '.,hD AB)(1&$F ,$) +.B'(D'&B$Y
" GH+(.'&*( BZ 'bB DBA+,$&(* ,' >B+E&, C$'&+B1&* ZB. 'E(&. <.Q)&' ^S#A+i' N(DE(.DE(Y

! "#$%&''#

!"#$%&' (
" #$%&'( ) *+,-./0 1./ 02,3$+4 '5/3 6$3/./00 ,7 2'( +/%6'3809
" :;<= ) *+,-./0 7,%, ('../(%$'+ $+ .,34/ +/%6'3809
" > 0/% '? 0(3$@%0 ?'3 :<= ,+,.A0$0 $+ %2/ +0BC 0$DE.,%'39

')'*#'+ ,%)-. (
" F+ D/00,4/ 73'@ ,+7 0(2/7E.$+4 $+ G/.,A :'./3,+% H/%6'380I #3E+/. J+$5I K'+7'+I >@3$.
CLLM9
" ;00E/0 3/.,%/7 %' %2/ D'7/..$+4 '? %2/ :<= @3'%'('.I *B+/N% 0(2''.I K'E5,$+I H'5/D-/3
CLLO9
" P(,.,-./ 0'.E%$'+0 ?'3 $+?'3D,%$'+ ('../(%$'+I Q:&I P%'(82'.DI P/@%/D-/3 CLLO9

tel-00408745, version 1 - 3 Aug 2009

" F+ :<= @/3?'3D,+(/ $+ 2/%/3'4/+/'E0 +/%6'380I PEDD/3 R/0/,3(2 ;+0%$%E%/I K,E0,++/I

SE.A CLLC9

,'%*/012 %*#030#0'. (
" <'E30/ '+ J+7/30%,+7$+4 H/%6'380I TL 2'E30I U,0%/3 ;V;I J+$5/30$%A '? H$(/ P'@2$, >+B
%$@'.$0I CLLMB@3/0/+%9
" <'E30/ '+ ;+%/3+/% U/,0E3/D/+% ,+7 :3,W( >+,.A0$0I XO 2'E30I Y$Z H/%6'38$+4 ,+7 G$0B
%3$-E%/7 PA0%/D0 U,0%/3 ,% %2/ J+$5/30$%A '? H$(/ P'@2$, >+%$@'.$0I CLL[BCLL I ,+7 Y$$Z
U,0%/3 R;UI *HP;I :E+$0I CLLTB@3/0/+%9
" \/+/3,. ('E30/ '+ :<=];=I CL 2'E30I K$(/+(/ =3'?/00$'++/../ K=P;KI ;J: '? %2/ JHP>I
CLL!B@3/0/+%9
" <'E30/ '+ .'(,. ,3/, +/%6'380I XO 2'E30 ^ 9O 2'E30 @3,(%$(,. 6'38I 130% A/,3I ;J: '? %2/
J+$5/30$%A '? H$(/ P'@2$, >+%$@'.$0I CLL BCLLM9
" <'E30/ '+ _'$(/ '5/3 ;=I

2'E30I Y$Z *('./ 7`;+4a+$/E3 ;+%/3+/%I U,30/$../I CLLTI Y$$Z U,0%/3

:;UI JHP>I CLL BCLLMI Y$$$Z U,0%/3 R:UI ;J= >5$4+'+I CLLM9
" <'E30/ '+ ;+?/3/+(/ '? ;+%/3+/% :'@'.'4AI

2'E30I U,0%/3 R:UI ;J= >5$4+'+I CLL!B

@3/0/+%9
" *N/3($0/0 '+ =/3?'3D,+(/ *5,.E,%$'+ $+ H/%6'38$+4I ! 2'E30I ;+0%$%E% *E3/('DI P'@2$,
>+%$@'.$0I CLLT9
" <'E30/ '+ :<=];=I ! 2'E30I U,b%3$0/ ;+?'3D,%$cE/I J+$5/30$%A '? H$(/ B P'@2$, >+%$@'.$0I
V3,+(/I CLLT9
" *N/3($0/0 '+ H/%6'38$+4I G'(%'3,. P(2''.I GP<B*=VKI K,E0,++/I P6$%d/3.,+7I CLLX9
" <'E30/ 6$%2 /N/3($0/0 '+ !"#I

2'E30I Y$Z H/%6'38$+4 ,+7 G$0%3$-E%/7 PA0%/D0 U,0%/3I

J+$5/30$%A '? H$(/ B P'@2$, >+%$@'.$0I V3,+(/I CLLXBCLLTI Y$$Z U,0%/3 R:UI ;J= >5$4+'+I
CLLM9
" <'E30/ 6$%2 3/0/,3(2 @3'e/(%0 '+ ;+%/3+/% '5/3 P,%/..$%/0I ! 2'E30I H/%6'38$+4 ,+7 G$0%3$B
-E%/7 PA0%/D0 G*>I J+$5/30$%A '? H$(/ B P'@2$, >+%$@'.$0I V3,+(/I CLLL9

! "#$%&'& () *+,&(-& () &#.' /'01)'&' +(2+ &-+##3&4 566 7

!" #$%&'($# )
!

!"# $%&&'("#

8 9:1%:'; () 9'*:'.0'% <==>7 ?+'&(& :(:3' 8 @;1*:(A' B#)(:#%()2 1); B1)1C

2'.'): #D E):'%)': ?%1F-7

!

)'*%!"# +%,-! ./%-

8 "#C&$*'%A(&'; G(:+ H13(; I100#$&7 9:1%:'; () J$)' <== 7 ?+'&(&

:(:3' 8 @%-+(:'-:$%' D#% ;1:1 &+1%()2 () .#0(3' G(%'3'&& )':G#%K&7

!

)'*%!%# 0%/",

8 "#C&$*'%A(&'; G(:+ L+(3(**' M1()7 9:1%:'; () N-:#0%' <== 7 ?+'&(& :(:3' 8

I':'-:(#) 1); :%#$03'&+##:()2 #D E):'%)': 1)#.13('&7

!

8 O%1;$1:'; () 9'*:'.0%' <==P7 ?+'&(& :(:3' 8 E)D'%%()2 E):'%)': :#*#3#2,

)'*%!%# )%11-

D%#. 1**3(-1:(#) *#(): #D A('G7

tel-00408745, version 1 - 3 Aug 2009

2-34', 5%!'& 6

"#C&$*'%A(&'; G(:+ Q(:1) @3:.1)7 O%1;$1:'; () I'-'.0'% <==R7 ?+'&(&

:(:3' 8 S'3(103' 1); %#0$&: :%1)&.(&&(#) #D .$3:(.';(1 -#):'):& #A'% :+' E):'%)':7

*+,-(''# )
!

7!8% .%1*- 6

"#):'): 3##K$* () G(%'3'&& )':G#%K&4 <==>4 I(*3T.' ;UE)2V)('$% ') E)D#%.1C

:(W$'4 QM9E4 ?$)(&7

!

9!-, +,-:% 6

S'13(&:(- 9(.$31:(#) #D I'31, ?#3'%1): M':G#%K&4 <== 4 I(*3T.' ;UE)2V)('$%

') E)D#%.1:(W$'4 QM9E4 ?$)(&7

!

)'(8% 9/#"1!'(!"8 6

9$**#%: #D :%1)&*#%: D$)-:(#)13(:('& () I'31, ?#3'%1): M':G#%K&4

<==PC= 4 Q)2()''% () ?'3'-#..$)(-1:(#)&4 9$*"#.4 ?$)(&7

!

+%,-! ./%- 6

"#33'-:()2 E)D#%.1:(#) D%#. 1 31%2' )$.0'% #D )':G#%K '):(:('&4 <==P4 I(*3T.'

;UE)2V)('$% ') E)D#%.1:(W$'4 QM9E4 ?$)(&7

!

;-"1& )'11", 6

@)13,&(& #D ?"L H'&:G##;X4 <==P4 L+I &:$;'): 1: Y?Z4 9G';')7

!%8 =%8>%,48",

5'<

8 I'&(2) 1); 1)13,&(& #D *%#:#-#3& D#% .$3:(.';(1 1**3(-1:(#)& #A'% L<L

)':G#%K&4 <==R4 ')2()''%()2 1); .1&:'% ;'2%'' #D QL[/4 /1$&1))'7

!

8 @)13,&(& #D :+' B@" 31,'% () >=<755 H/@M4 ')2()''%()2 ;'2%'' #D

?-'8<5"4' @%84-"8-

QL[/4 /1$&1))'7
!

5%A% 9/#"1!'(!"8

8 9:$;, #D 1 3()KC3'A'3 [Q"\@S] &-+'.' D#% G(%'3'&& 3()K& 1); &+#%:C

3(A'; ?"L :%1F-4 <==^4 I(*3T.' ;UE)2V)('$% ') E)D#%.1:(W$'4 QM9E4 ?$)(&7
!

91%"##-8" 91 B%C%1 6

9:$;, #D 1 3()KC3'A'3 [Q"\@S] &-+'.' D#% 3#)2C3(A'; ?"L :%1DC

_-4 <==<4 M':G#%K()2 1); ?'3'-#..$)(-1:(#)& IQ@4 /'01)'&' `)(A'%&(:, C 91(): J#&'*+
`)(A'%&(:,4 /'01)#)7
!

."D'( .-#-/" 6

@)13,&(& #D ?"L\EL #A'% 1&,..':%(- &1:'33(:' 3()K&4 56664 B1&:'% #D :+'

[%')-+ M1:(#)13 E)&:(:$:' D#% ?'3'-#..$)(-1:(#)& aEM?b4 QA%,4 [%1)-'7

!

7!!%8("1 $%/',#" 6

cE4 [%1)-'7

B1%K#A(1) 1)13,&(& #D ?"L\EL4 56664 IQ@ #D :+' `)(A'%&(:, #D L1%(&

! "#$%&''#

!
"

!&($)* +%,)- . #$%$&'( )')*+,&, -. ,&/0*% 123 ,45%/%, .-6 78 9%*%05-'+: ;<<<: =%>?-6@

tel-00408745, version 1 - 3 Aug 2009

A&'( )'B C&,>6&D$>%B E+,>%/, C2F: G'&H%6,&>+ -. =&4% E-05&)@F'>&0-*&,: 16)'4%I

!"#$ %$ &'()!*+#!,-"
!"#$! %&'()*+, *)! -&&.,
! "#$%& '($)%&* +,&-$%&$#& './%012&3,.* 41%5# 6%/%3%$* +%/$#327% 41%&5/%7%&%* 8' 9:(;

tel-00408745, version 1 - 3 Aug 2009

$#(2.2( '<</,%01 =,/ $12 9,52((#&> ,= ?%/>2 @4ABCA D2$E,/3-8* 41%<$2/ F #& 8G202&$
'5.%&02- #& 9,52(#&> %&5 H#):(%$#,& @,,(- =,/ 4,)):&#0%$#,& D2$E,/3- %&5 H2/.#02-8
<:I(#-125 I7 H</#&>2/* H2<$2)I2/ JKKL!
J! 41%5# 6%/%3%$* M#%&(:0% C%&&%00,&2* N#) +:/,-2* %&5 O%//7( P2#$01* 8H%)<(#&> $12 C&;
$2/&2$ Q $201&#R:2- %&5 %<<(#0%$#,&-8* C""" N,:/&%( ,= H2(20$25 '/2%- #& 4,)):&#0%$#,&-*
.,(! JF* &,! J* O202)I2/ JKKS!
T! 41%5# 6%/%3%$* C%& A/%$$ Q A%--#.2 %&5 '0$#.2 D2$E,/3 92%-:/2)2&$* U$1 C&$2/&%$#,&%(
V,/3-1,<* A'9 JKKF* '&$#I2- N:%&;(2-;A#&-* W/%&02* '</#( X;JK* JKKF* A/,0225#&>-! ?20;
$:/2 D,$2- #& 4,)<:$2/ H0#2&02 TK U H</#&>2/ JKKF* CH6D T;UFK;J FXJ;U!

/*0$(, ") 1$2$($$! %&'()*+,

! +%/#) HI%# %&5 41%5# 6%/%3%$* 8"Y<2/#2&02- ,& 2&1%&0#&> 5%$% 0,((20$#,& #& (%/>2 &2$;
E,/3-Z* $, %<<2%/ #& 4,)<:$2/ D2$E,/3-!
J! 9,1%))%5 9%((#* 41%5# 6%/%3%$* V%(#5 O%II,:-* 84["HH Q '& %<<(#0%$#,&;%E%/2 -<%02
=,/ 2&1%&025 -0%(%I(2 -2/.#02- #& ,.2/(%7 &2$E,/3-8* 4,)<:$2/ 4,)):&#0%$#,&- N,:/&%(*
.,(! T * &,! S* <<! JTX; JUT* '</#( JKK\!
T! @! @%&-:<%-#/#* +! +%&01%&%-:$* 41%5# 6%/%3%$* A1#(#<<2 N%0R:2$* 8]-#&> '0$#.2 D2$E,/3@201&,(,>7 =,/ O7&%)#0 ^,H8* 4,)<:$2/ D2$E,/3- N,:/&%(* .,(! UK* &,!

* <<! SXJ; LKX*

':>:-$ JKKS!
F! 41%5# 6%/%3%$* 9,1%))%5 9%((#* D%,)#01# D,&%3%* 8@C4A Q @/%&-<,/$ C&=,/)%$#,& 4,(;
(20$#,& A/,$,0,(8* '&&%(- ,= @2(20,)):&#0%$#,&-* .,(! S * &,! ;J* <<! SL; XJ* JKKS!
U! "#$%& '($)%&* +,-$7% './%012&3,.* 41%5# 6%/%3%$* '/_%5;'(%) +12/%&#* 6%(%3/#-1&%
A/%I1:* 8'&%(7-#- ,= 9C9O 4,&>2-$#,& 4,&$/,( '(>,/#$1) =,/ [#>1 H<225 D2$E,/3-8*
4,)<:$2/ D2$E,/3- N,:/&%(* .,(! F\* &,! S* <<! XLJ;X\X* ':>:-$ JKKU!
LX

!"#$ %$ &'()!*+#!,-"

"# $%&'( )*&+'(, -.'/% 0'1'2'&, 3%4&51 6# 7'+58 7#, 9)('*:8%8 5; )<6= >15&545*8 5?@1
>'&.8 A%&. ?'1%'B*@ /@*':9, -5+>C&@1 D@&A5128 E5C1('*, ?5*# F , (5# ", >># G"!HGIJ, )CKC8&
L!!M#
I# <+'/ )'/, N%'(K D%, -.'/% 0'1'2'&, O.%@11: OC1*@&&%, 9$(.'(4%(K <$$$ !L#JJ 6)- %(
-5(K@8&@/ $(?%15(+@(&89, -5+>C&@1 -5++C(%4'&%5(8 E5C1('*, ?5*# L , (5# JF, >># J"!MH
J"JI, P@>&@+B@1 L!!M#
# Q%5( 7@&5 -'(&%@(%, N%'(K D%, -.'/% 0'1'2'&, O.%@11: OC1*@&&%, 9R@1;51+'(4@ )('*:8%8
C(/@1 S%(%&@ T5'/ '(/ <+>15?@+@(&8 ;51 6C*&%1'&@ !L#JJ9, -5+>C&@1 -5++C(%4'&%5(8
E5C1('*, 8>@4%'* %88C@ 5( R@1;51+'(4@ 5; U%1@*@88 T)D8, R)D8, '(/ )/ V54 D@&A5128, ?5*#
L , (5# J!, >># J!GMHJJ!G, EC(@ L!!M#
G# $%&'( )*&+'(, W58&:' )?1'4.@(25?, -.'/% 0'1'2'&, 9) P&54.'8&%4 65/@* 5; O-RX<R A%&.

tel-00408745, version 1 - 3 Aug 2009

P&'&%5('1: 7'(/5+ T588@89, <$$$X)-6 O1'(8'4&%5(8 5( D@&A512%(K, ?5*# JY, (5# L, >>#
YM"H Y"G, )>1%* L!!M#
J!# $%&'( )*&+'(, W58&:' )?1'4.@(25?, -.'/% 0'1'2'&, R'1%Z'& =CB@, 9R@1;51+'(4@ )('*:8%8
5; )<6= 6@4.'(%8+8 5?@1 ' 6C*&%H8&'&@ 6'125?%'( R'&.9, -5+>C&@1 D@&A5128 E5C1('*,
?5*# FI, (5# Y, >># Y!IHYL", S@B1C'1: L!!M#
JJ# )(&5(%5 DC44%, D%(' O';&, -.'/% 0'1'2'&, R'&1%42 O.%1'(, 9-5(&15**@/ [8@ 5; $\4@88 0'42H
B5(@ 0'(/A%/&. ;51 R15?%/%(K D@A P@1?%4@8 %( <RH5?@1HU=6 D@&A51289, <$$$ E5C1('*
5( P@*@4&@/ )1@'8 %( -5++C(%4'&%5(8, ?5*# LL, (5# G, >># J"GLHJI!I, D5?@+B@1 L!!F#
JL# -.'/% 0'1'2'&, )*'@//%(@ )* S'A'*, 9)('*:8%8 5; *%(2H*@?@* .:B1%/ S$-X)7NHP7 ;51 A%1@H
*@88 *%(28 '(/ *5(KH*%?@/ O-R &1']49, R@1;51+'(4@ $?'*C'&%5( E5C1('*, ?5*# MI, (5# F, >>#
FLYHM!!, )CKC8& L!!F#
JY# -.'/% 0'1'2'&, $%&'( )*&+'(, 9) 6'125?%'( 65/@* ;51 O-R )('*:8%8 %( ' =%^@1@(&%'&@/
P@1?%4@8 D@&A5129, O@*@45++C(%4'&%5( P:8&@+8, ?5*# LM, (5# J,L, >># JLGHJMM, E'(C'1: L!!F#
JF# P'1' )*5C;, $%&'( )*&+'(, -.'/% 0'1'2'&, R.%*%>>@ D'%(, 9_>&%+'* $8&%+'&%5( 5; 6C*&%4'8&
6@+B@18.%>9, <$$$ O1'(8'4&%5(8 5( P%K('* R154@88%(K H P>@4%'* <88C@ 5( P%K('* R154@88%(K
%( D@&A512%(K, ?5*# MJ, (5# , )CKC8& L!!Y#
JM# -.'/% 0'1'2'&, R'&1%42 O.%1'(, Q%'(*C4' <'(('445(@, -.1%8&5>.@ =%5&, R.%*%>>@ _A@`'182%,
965/@*%(K <(&@1(@& B'42B5(@ &1']4 '& &.@ a5A *@?@*9, <$$$ O1'(8'4&%5(8 5( P%K('* R15H
4@88%(K H P>@4%'* <88C@ 5( P%K('* R154@88%(K %( D@&A512%(K, ?5*# MJ, (5# , )CKC8& L!!Y#
J"# -.'/% 0'1'2'&, $%&'( )*&+'(, 9_( )-W b*&@1%(K 5( ' 8*5A 1@?@18@ 4.'((@*9, <(&@1('&%5('*
E5C1('* 5; P'&@**%&@ -5++C(%4'&%5(8, ?5*# LJ, >># LFJHLM , L!!Y#
JI# $%&'( )*&+'(, W58&:' )?1'4.@(25?, -.'/% 0'1'2'&, 7C/@8%(/5 Dcd@`HNC@%Z', 9P&'&@H
/@>@(/@(& 6XQXJ &:>@ eC@C@%(K '('*:8%8 ;51 45(K@8&%5( 45(&15* %( /'&' (@&A51289, -5+H
>C&@1 D@&A5128 E5C1('*, ?5*# YG, (5# ", >># I GH ! , )CKC8& L!!L#
J # $%&'( )*&+'(, -.'/% 0'1'2'&, 3%4&51 7'+58, 9NC@C%(K '('*:8%8 5; 8%+>*@ S$- 84.@+@8
;51 ?5%4@ 5?@1 <R9, -5+>C&@1 D@&A5128, ?5*# YG, (5# L, >># J MHL!", EC(@ L!!L#

!
!"# $%&'( )&*&+&,- .(,&/ 01,2&/- 3)&/'4(',% ,*&'567 85,455/ 9$: &/' 1(/+;15<51 =.$3$62>?,5* @5,46*+A B6?*/&1- <61# C"- /6# D- >># !CC;!EF- B?/5 DFFD#
DF# $%&'( )&*&+&,- 39$: 26'51(/G &/' <&1('&,(6/3- H... @5,46*+- <61# !E- /6# C- >># C ;IJK&L DFF!#
D!# .(,&/ 01,2&/- M6A,L& 0<*&N%5/+6<- $%&'( )&*&+&,- 39$: (/ >*5A5/N5 6O 8?*A,L 16AA5A3:5*O6*2&/N5 .<&1?&,(6/- <61# ID- /6# D;C- >># !D";!IJ- PN,685* DFFF#
DD# $%&'( )&*&+&,- .(,&/ 01,2&/- Q&1(' R&886?A- 3P/ 9$: :5*O6*2&/N5 (/ & S5,5*6G5/56?A
@5,46*+ T 0 U?*<5L3- H... $622?/(N&,(6/ K&G&V(/5- <61# C - /6# !- >># IF;IW- B&/?&*L
DFFF# 0 26*5 '5,&(15' <5*A(6/ T H@XH0 X5A5&*N% X5>6*, XX;CJCJ- B?1L !"""#

tel-00408745, version 1 - 3 Aug 2009

!"#$% &' $()##*&'+% (, -#,#$##* .(',#$#')#%

!# 02(* MXH=0- K6%&25' M&*(2 U)0H- $%&'( )0X0M09- 9%(5**L 9YXZ.99H- 3)(,S6N T
0 N6/,5/, A%&*(/G &>>1(N&,(6/ O6* Q(*515AA 0' %6N @5,46*+A3- '526 '5AN*(>,(6/ ,6 &>>5&*
(/ >*6N55'(/GA 6O ,%5 H... :5*N62 N6/O5*5/N5- [&1<5A,6/- 95\&A- K&*N% DFF"#
D# K6%&25' M&*(2 U8&(- $%&'( )&*&+&,- B&5L6?/G $%6(- 0/4&* 01 S&2*&- &/' 9%(5**L
9?*15,,(- 30'&>,(/G )(,96**5/, ,6 4(*515AA &' %6N /5,46*+A3- (/ >*6N55'(/GA 6O ,%5 0'S6N;
@64 @5,46*+A &/' Q(*515AA N6/O5*5/N5- U6>%(& 0/,(>61(A- U5>,5285* DFF #
C# 02(* M*(O&- $%&'( )&*&+&,- 9%*&AL<6?16A U>L*6>6?16A- 30/ P>,(2&1 B6(/, UN%5'?1(/G &/'
R*6> :61(NL O6* R51&L 9615*&/, @5,46*+A3- (/ >*6N55'(/GA 6O ,%5 Q6QK6K Q6*+A%6> 6/
0?,6/62(N &/' P>>6*,?/(A,(N $622?/(N&,(6/A- @54>6*, )5&N% ]$0^- B?/5 DFF #
I# 02(* M*(O&- $%&'( )&*&+&,- 9%*&AL<6?16A U>L*6>6?16A- 3P>,(2&1 )?75* K&/&G525/, :6;
1(N(5A O6* R51&L 9615*&/, @5,46*+A3- (/ >*6N55'(/GA 6O ,%5 E,% H... $6/O5*5/N5 6/ U5/A6*K5A% &/' 0' S6N $622?/(N&,(6/A &/' @5,46*+A ]U.$P@ DFF ^- U&/ =*&/N(AN6- B?/5
DFF #

!"# $%&!'(

E# K6%&25' 01( M&&O&*- Z&?*5/, K&,%L- $%&'( )&*&+&,- M&<5 U&1&2&,(&/- 9%(5**L 9?*15,,(Q&1(' R&886?A- 3U5N?*(/G H/,5*/5, $66*'(/&,5 .285''(/G ULA,52A3- (/ >*6N55'(/GA 6O
0$K UH[$PKK- ML6,6- B&>&/- 0?G?A, DFFJ#
W# @(51A K6115*- $%&'( )&*&+&,- M6/A,&/,(/ 0<*&N%5/+6<- .(,&/ 01,2&/- 3H/,5*;>*6,6N61 O&(*;
/5AA 85,455/ 9$: @54 X5/6 &/' 9$: Q5A,466'_3- (/ >*6N55'(/GA 6O @[H DFFJ ]$6/O5;
*5/N5 6/ @5\, [5/5*&,(6/ H/,5*/5, @5,46*+A^- 9*6/'%5(2- @6*4&L- K&L DFFJ#
J# M&*(2 U)0H- $%&'( )0X0M09- 39*&/A>6*, H/O6*2&,(6/ $6115N,(6/ :*6,6N61 4(,% N1?A,5;
*(/G 6O (/O6*2&,(6/ A6?*N5A3- (/ >*6N55'(/GA 6O @9KU DFFJ ]$6/O5*5/N5 6/ @54 95N%/616;
G(5A- K68(1(,L &/' U5N?*(,L^- :&*(A- K&L DFFJ#
# [# )6GG(&- :# $&2&*'&- 0# R`01N6/V6- Z#0# [*(5N6- U# K&AN616- .# 01,2&/- $# )&*&+&,3K6'51(/G ,%5 0H0RR :&*&'(G2 (/ @5,46*+A 4(,% a&*(&815 R51&LA3- (/ >*6N55'(/GA 6O
$6@5\,- Z(A86&- :6*,?G&1- R5N5285* DFFW#

!"#$ %$ &'()!*+#!,-"

"# $%&' ()*& +,'*%)'%- $%,'./0, 1,'',00&')- +2,3% 4,5,6,*- +25%7*&82) 9%&*- :,*5%06 ;2%5,'<()=&5>/.,*%'? *2) >&'%*&5 8.,0)>)'* 85&@.)> A B8*%>,. C)*D&56ED%3) F,>8.%'?<- %'
85&0))3%'?7 &= +&C)G*- H%7@&,- :&5*/?,.- 9)0)>@)5 !IIJ#
KI# L,50 H%@)5,*&5)- 45%,' H)M%')- +2,3% 4,5,6,*- <L,G%>%N%'? ;5,'7=)5 B88&5*/'%*%)7 %'
9;C<- %' 85&0))3%'?7 &= +&C)G*- H%7@&,- :&5*/?,.- 9)0)>@)5 !IIJ#
KK# O'D,5 O. P,>5,- +2,3% 4,5,6,*- ;2%)55Q ;/5.)**%- <C)*D&56 +&3%'? =&5 R%5).)77 L)72
C)*D&567 A O +,7) F*/3Q<- %' 85&0))3%'?7 &= R&RL&L- 4/S,.&ECT- U/') !IIJ#
K!# L&2,>>,3 L,..%- +2,3% 4,5,6,*- R,.%3 9,@@&/7- <O88.%0,*%&'E.)M). M)57/7 C)*D&56E.)M).
:5&G%>%*Q<- %' 85&0))3%'?7 &= *2) O1C;V+ 0&'=)5)'0)- 4,'?6&6- 9)0)>@)5 !IIW#
KX# +2,3% 4,5,6,*- $%,'./0, 1,'',00&'')- +25%7*&82) 9%&*- <(,'6%'? Y&D7 =5&> 7,>8.)3

tel-00408745, version 1 - 3 Aug 2009

*5,Z0<- %' 85&0))3%'?7 &= +&CV[;- ;&/.&/7)- B0*&@)5 !IIW#

K\# P&77)%' L,'72,)%- $%&' ()*& +,'*%)'%- +2,3% 4,5,6,*- ;2%)55Q ;/5.)**%- <:)5=&5>,'0)
O',.Q7%7 &= *2) 1VVV I!#KK LO+ ,'3 :2Q7%0,. H,Q)5 :5&*&0&.<- %' 85&0))3%'?7 &= R&RE
L&L- ;,&5>%',- U/') !IIW#
KW# 1>,3 O,3- ]%,'? C%- +2,3% 4,5,6,*- ;2%)55Q ;/5.)**%- <V'2,'0%'? 1VVV I!#KK LO+
%' +&'?)7*)3 V'M%5&'>)'*7<- %' 85&0))3%'?7 &= *2) O88.%0,*%&'7 ,'3 F)5M%0)7 %' R%5).)77
C)*D&567 D&5672&8 ^ORFC_- 4&7*&'- O/?/7* !II\#
KJ# F,5, O.&/=- V%*,' O.*>,'- +2,3% 4,5,6,*- :2%.%88) C,%'- <B' *2) 3Q',>%0 )7*%>,*%&' &=
>/.*%0,7* ?5&/8 7%N)7<- %' 85&0))3%'?7 &= *2) 1'*)5',*%&',. FQ>8&7%/> &' L,*2)>,*%0,.
;2)&5Q &= C)*D&567 ,'3 FQ7*)>7 ^L;CF_- H)/M)'- 4).?%/>- U/.Q !II\#
K`# a,*>, H&/,*%- +2,3% 4,5,6,*- R,.%3 9,@@&/7- <O3,8*%M) +.,77E@,7)3 ]/)/%'? =&5 2,'3E
.%'? ;D&ER,Q *5,Z0 %' O7Q>>)*5%0 C)*D&567<- %' 85&0))3%'?7 &= *2) 0&'=)5)'0) &' P%?2
F8))3 C)*D&567 ,'3 L/.*%>)3%, +&>>/'%0,*%&'7 ^PFCL+_- ;&/.&/7)- a5,'0)- U/.Q !II\#
K # V%*,' O.*>,'- b&7*Q, OM5,02)'6&M- +2,3% 4,5,6,*- O5N,3EO.,> b2)5,'%- 4,.,65%72',
:5,@2/- <O',.Q7%7 &= F0,.,@.) ;+:<- %' 85&0))3%'?7 &= *2) 0&'=)5)'0) &' P%?2 F8))3
C)*D&567 ,'3 L/.*%>)3%, +&>>/'%0,*%&'7 ^PFCL+_- ;&/.&/7)- a5,'0)- U/.Q !II\#
K"# L&2,>>,3 L,..%- ]%,'? C%- ;2%)55Q ;/5.)**%- +2,3% 4,5,6,*- <O3,8*%M) a,%5 +2,'').
O..&0,*%&' =&5 ]&F V'2,'0)>)'* %' 1VVV I!#KK R%5).)77 HOC7<- %' 85&0))3%'?7 &= 1VVV
1++- :,5%7- U/') !II\#
!I# (,c, O@3).>&/>)'- +2,3% 4,5,6,*- <O',.Q7%7 &= ;+: .,*)'0Q &M)5 D%5).)77 .%'67 7/88&5E
*%'? aV+dO(]EF( =&5 )55&5 5)0&M)5Q<- %' 85&0))3%'?7 &= 1VVV 1'*)5',*%&',. +&'=)5)'0)
&' +&>>/'%0,*%&'7 ^1++_- :,5%7- U/') !II\#
!K# P/'? C?/Q)'- :,*5%06 ;2%5,'- +2,3% 4,5,6,*- <B' *2) 0&55).,*%&' &= ;+: *5,Z0 %' @,06E
@&') ')*D&567<- %' 85&0))3%'?7 &= *2) 1VVV 1'*)5',*%&',. FQ>8&7%/> &' +%50/%*7 ,'3
FQ7*)>7 ^1F+OF_- e,'0&/M)5- +,',3,- L,Q !II\#

!
""# $%&'( )&*&+&,- .(/0 12,/ $&0,(20(- 34 '(5,&0627&8&*2 9/'2: /; <"#== >4$ :&?2*3- (0
@*/622'(0A5 /; B(C@, 8/*+5%/@ D >/'2:(0A &0' C@,(9(E&,(/0 (0 >/F(:2- 4' G/6 &0'
B(*2:255 H2,8/*+5 I2J,20'2' &F5,*&6,K- $&9F*('A2 ILMK- >&*6% "<<N#
"!# O(,&0 4:,9&0- $%&'( )&*&+&,- P(6,/* ># 1&9/5 1#- 340&:?5(5 /; 4Q>R @*/,/6/:5 /S2*
@&,%5 8(,% S&*(&F:2 '2:&?3- (0 @*/622'(0A5 /; QOOO QHTC$C>- G/0A M/0A- >&*6% "<<N#
"N# U(&0A H(- Q9&' 4&'- $%&'( )&*&+&,- V%(2**? VW*:2,,(- 3>/'2:(0A &0' 40&:?5(5 /; X:/8
$B R26*2&52 ;/* QOOO <"#== BY4H3- (0 @*/622'(0A5 /; ,%2 Q0,2*0&,(/0&: X?9@/5(W9 /0
Z2*5/002:- Q0'//* &0' >/F(:2 1&'(/ $/99W0(6&,(/05 IZQ>1$K- )2([(0A- $%(0&- X2@,29F2*
"<<!#
"\# L*,E( 4?25,&- M/5,?& 4S*&06%20+/S- O(,&0 4:,9&0- $%&'( )&*&+&,- Z&*([&, RWF2- 3>W:7
,(:2S2: 4@@*/&6% ;/* >/'2:(0A X%/*, V$Z X255(/053- (0 @*/622'(0A5 /; QV$7= - )2*:(0-

tel-00408745, version 1 - 3 Aug 2009

X2@,29F2* "<<!#
"]# P(6,/* 1&9/5- $%&'( )&*&+&,- O(,&0 4:,9&0- 34 >/S(0A 4S2*&A2 Z*2'(6,/* ;/* Z:&?/W,
R2:&? $/0,*/: (0 P/QZ3- (0 @*/622'(0A5 /; ,%2 O:2S20,% Q0,2*0&,(/0&: B/*+5%/@ /0 UW&:(,?
/; X2*S(62 IQBU/X "<<!K- >/0,2*2?- $4- ^W02 "<<!#
"_# X&*& 4:/W;- O(,&0 4:,9&0- $%&'( )&*&+&,- Z%(:(@@2 H&(0- 3O5,(9&,(0A >29F2*5%(@ (0 &
>W:,(6&5, X255(/03- (0 @*/622'(0A5 /; 4$> XQ.>OV1Q$X- X&0 R(2A/- $4- ^W02 "<<!#
" # 4:&2''(02 4: T&8&:- $%&'( )&*&+&,- 3X(9W:&,(/07F&52' 5,W'? /; :(0+7:2S2: %?F*(' TO$`41U7
X1 ;/* 8(*2:255 :(0+5 &0' :/0A7:(S2' V$Z ,*&a63- (0 @*/622'(0A5 /; B(C@,b<! D >/'2:(0A
&0' C@,(9(E&,(/0 (0 >/F(:2- 4' G/6 &0' B(*2:255 H2,8/*+5- X/@%(& 40,(@/:(5- T*&062>&*6% "<<!#
"c# 1&*25 X2*F&0- $%&'( )&*&+&,- B&:(' R&FF/W5- 3R?0&9(6 125/W*62 4::/6&,(/0 (0 $/*2
1/W,2*5 /; & R(d52*S H2,8/*+3- (0 @*/622'(0A5 /; ,%2 45(&0 $/9@W,(0A X6(2062 $/0;2*2062
I4XQ4HK- G&0/(- P(2,0&9- R2629F2* "<<"#
!<# $%&'( )&*&+&,- Z&,*(6+ V%(*&0- .(&0:W6& Q&00&66/02- $%*(5,/@%2 R(/,- Z%(:(@@2 C82E&*5+(34 e/87F&52' 9/'2: ;/* Q0,2*02, F&6+F/02 ,*&a63- (0 @*/622'(0A5 /; 4$> XQ.$C>>
Q0,2*02, >2&5W*2920, B/*+5%/@- >&*52(::2- T*&062- H/S29F2* "<<"#
!=# $%&'( )&*&+&,- Z&,*(6+ V%(*&0- .(&0:W6& Q&00&66/02- $%*(5,/@%2 R(/,- fC0 Q0,2*02, F&6+7
F/02 ,*&a6 9/'2:(0Ag- 2J,20'2' &F5,*&6, (0 @*/622'(0A5 /; 4$> XQ.>OV1Q$X- >&*(0&
R2: 12?- $4- ^W02 "<<"#
!"# O(,&0 4:,9&0- M/5,?& 4S*&6%20+/S- $%&'( )&*&+&,- fV$Z 02,8/*+ 6&:6W:W5 D V%2 6&52 /;
:&*A2 F&0'8(',%7'2:&? @*/'W6,g- (0 @*/622'(0A5 /; QOOO QHTC$C>- Hh- ^W02 "<<"#
!!# O(,&0 4:,9&0- $%&'( )&*&+&,- P(6,/* 1&9/5- fC0 ,%2 W,(:(,? /; TO$ 926%&0(595 ;/* &W'(/
&@@:(6&,(/05g- (0 @*/622'(0A5 /; ,%2 (0,2*0&,(/0&: 8/*+5%/@ /0 UW&:(,? /; ;W,W*2 Q0,2*02,
X2*S(625 IU/;QXK- $/(9F*&- Z/*,WA&:- X2@,29F2* "<<=#

!"#$ %$ &'()!*+#!,-"

"!# $%&'( )*&+'(, -./&0' )12'345(6.1, 74'8% 9'2'6'&, :;85/%(8. <;(5= >;5%?', @A7B +.C
85*%(D %( &45 E25/5(35 .F (.(*%(5'2 G%(8.G D2.G&4@, %( E2.3558%(D/ .F HA7CIJ, K'*1'8.2 8'
9'4%', 92'=%*, K5E&5+L52 MNNI#
"O# 74'8% 9'2'6'&, $%&'( )*&+'(, @9'(8G%8&4 A2'85.P L5&G55( A7B '(8 Q%(6CQ515* R$7@,
%( E2.3558%(D/ .F H$$$ H(&52('&%.('* 7.(F525(35 .( <5&G.26%(D, 7.*+'2, R2'(35, S;*0
MNNI#
"T# $%&'( )*&+'(, -./&0' )12'345(6.1, 74'8% 9'2'6'&, :;85/%(8. <;(5= >;5%?', @K&'&5C
85E5(85(& UVWVI A0E5 >;5;5%(D )('*0/%/ F.2 7.(D5/&%.( 7.(&2.* %( X'&' <5&G.26/@,
%( E2.3558%(D/ .F H$$$ H<RY7YU, )(34.2'D5, )*'/6', )E2%* MNNI# ) +.25 85&'%*58 152C
/%.( Z 7[H 25E.2& B<)C:NNNO, S;*0 MNNN#
"J# $%&'( )*&+'(, 74'8% 9'2'6'&, \%3&.2 :'+./, ]>;5;5%(D )('*0/%/ .F K%+E*5 R$7 K345+5/

tel-00408745, version 1 - 3 Aug 2009

F.2 HB A5*5E4.(0^, %( E2.3558%(D/ .F H$$$ H<RY7YU, )(34.2'D5, )*'/6', )E2%* MNNI#

" # $%&'( )*&+'(, 74'8% 9'2'6'&, $+'(;5* Q'L.285, B'&2%36 92.G(, X5(%/ 7.**'(D5, ]R'%2(5//
)('*0/%/ .F A7BVHB^, %( E2.3558%(D/ .F H$$$ 7.(F525(35 .( X53%/%.( '(8 7.(&2.*, K08(50,
);/&2'*%', X535+L52 MNNN#
"_# $%&'( )*&+'(, -./&0' )12'345(6.1, 74'8% 9'2'6'&, B'2%?'& X;L5, @A7B .152 ' U;*&%C
K&'&5 U'26.1%'( B'&4@, %( E2.3558%(D/ .F &45 %(&52('&%.('* 3.(F525(35 .( &45 B52F.2+'(35
'(8 >.K .F <5`& W5(52'&%.( <5&G.26%(D aBb><5&c, <'D.0', S'E'(, <.15+L52 MNNN#
!N# 74'8% 9'2'6'&, $%&'( )*&+'(, @Y( )7- R%*&52%(D .( ' K*.G :5152/5 74'((5*@, %( E2.C
3558%(D/ .F &45 d2/& %(&52('&%.('* G.26/4.E .( >;'*%&0 .F F;&;25 H(&52(5& /521%35/, 952*%(,
W52+'(0, K5E&5+L52 MNNN#

!"# $%& '#()!*# $+,!-.

!I# 74'8% 9'2'6'&, $%&'( )*&+'(, @) U'26.1%'( U.85* F.2 A7B )('*0/%/ %( ' X%P525(&%'&58
K521%35/ <5&G.26@, %( E2.3558%(D/ .F &45 d2/& %(&52('&%.('* G.26/4.E .( >;'*%&0 .F F;&;25
H(&52(5& /521%35/, 952*%(, W52+'(0, K5E&5+L52 MNNN#
!M# $%&'( )*&+'(, -./&0' )12'345(6.1, 74'8% 9'2'6'&, @) K&.34'/&%3 U.85* .F A7BVHB G%&4
K&'&%.('20 :'(8.+ Q.//5/@, %( E2.3558%(D/ .F )7U KHW7YUU, K&.364.*+, KG585(, );C
D;/& MNNN#
!"# $%&'( )*&+'(, -./&0' )12'345(6.1, 74'8% 9'2'6'&, @ A7B %( E25/5(35 .F L;2/&0 *.//5/@,
%( E2.3558%(D/ .F )7U KHWU$A:H7K, K'(&' 7*'2', 7'*%F.2(%', S;(5 MNNN#
!!# 74'8% 9'2'6'&, $%&'( )*&+'(, @B52F.2+'(35 .F K4.2& A7B A2'(/F52/@, %( E2.3558%(D/ .F
<5&G.26%(D MNNN aB52F.2+'(35 .F 7.++;(%3'&%.( <5&G.26/c, B'2%/, U'0 MNNN#
!O# 74'8% 9'2'6'&, $%&'( )*&+'(, @)('*0/%/ .F A7B G%&4 K5152'* 9.&&*5(536 <.85/@, %( E2.C
3558%(D/ .F H$$$ WQY9$7YU aW*.L'* H(&52(5& K0+E./%;+c, :%., 92'=%*, X535+L52 I___#
) +.25 85&'%*58 152/%.( Z H<:H) :5/5'234 :5E.2& ::C"TMN, R5L2;'20 I___#
!T# 74'8% 9'2'6'&, $%&'( )*&+'(, @)('*0/%/ .F &45 B45(.+5(.( .F K5152'* K*.G K&'2& B4'/5/
%( A7B@, 5`&5(858 'L/&2'3& %( E2.3558%(D/ .F )7U KHWU$A:H7K, K'(&' 7*'2', 7'*%F.2(%',
S;(5 MNNN# ) +.25 85&'%*58 152/%.( Z H<:H) :5/5'234 :5E.2& ::C"OJ!, X535+L52 I__ #

!
"#$ %&'() *'+','-. /01+)20 %&'&0+. 3'4)( 5'66781. 9)-'2 :4-;'2. <=;>+7?)2@ A%BC=B 7?0+
D071-'-)72'+E F'-044)-0 G)2,1<. )2 >+7H00()2@1 7I =999 DGJ*9%JK LD020+'4 %72I0+02H0M.
N)7. *+'O)4. 50H0;60+ PQQQ$ : ;7+0 (0-')40( ?0+1)72 R =/N=: N010'+H& N0>7+- NNST!#T.
50H0;60+ PQQ $

!"#$% &' $()##*&'+% (, -('./#,#$##* 0(',#$#')#% !'* 1!+!2&'#%

P$ 9)-'2 :4-;'2. U721-'2-)2 :?+'H&02,7?. %&'() *'+','-. U'+-),0E' %&'2(+'E'2'. <K84-)S
40?04 '>>+7'H& I7+ -&0 ;7(04)2@ 7I 4'+@0 A%BC=B 20-V7+,1<. )2?)-0( -7 -&0 FE;>71)8; 72
N0H02- :(?'2H01 I7+ K7(04)2@ '2( F);84'-)72 A7741 I7+ %7;;82)H'-)721 /0-V7+,1 '2(
F0+?)H01. 7+@'2)O0( 6E %JFA W ! )2 F8++0E. XU$

tel-00408745, version 1 - 3 Aug 2009

W$ %&'() *'+','-. <9?'48'-)72 (01 >0+I7+;'2H01 (01 +Y10'8Z 3=[= \W$PP<. A0H&2)]801 (0
4^=2@Y2)08+. G0--+0 +Y10'8Z 1'21 _4. /8;Y+7 /7?0;6+0S5YH0;6+0 W\\`$
T$ D)72 N0-7 %'2-)02). D)'248H' ='22'HH720. %&'() *'+','-. %&+)1-7>&0 5)7-. B'-+)H, A&)+'2.
<N0I7+;84'-)2@ -&0 ;72)-7+ >4'H0;02- >+7640; R J>-);'4 /0-V7+,SV)(0 F';>4)2@<. )2?)-0(
-7 -&0 %72I0+02H0 72 =2I7+;'-)72 FH)02H0 L%=FFM. B+)2H0-72 /a. K'+H& W\\`$
"$ 9$ :4-;'2. %$ *'+','-. F$ K'1H747. /$ K7440+. '2( a$ F82. <:2'4E1)1 7I A%B 301-V77(b
)2 &)@& 1>00( 20-V7+,1<. )2 >+7H00()2@1 7I B[G520-^\` R V7+,1&7> 72 B+7-7H741 I7+ ['1G72@S5)1-'2H0 20-V7+,1. /'+'. a'>'2. [06+8'+E W\\`$
!$ N'+01 F0+6'2. %&'() *'+','-. 3'4)( 5'66781. <92&'2H)2@ %*c V)-& ' 5E2';)H *'2(S
V)(-& :447H'-)72 K7(840<. )2 >+7H00()2@1 7I F0-)- W\\T. A82)1)'. K'+H& W\\T$
`$ X+-O) :E01-'. U71-E' :?+'2H&02,7?. 9)-'2 :4-;'2. %&'() *'+','-. B'+)d'- 5860. <K84S
-)40?04 :>>+7'H& I7+ K7(04)2@ A%BC=B<. )2 >+7H00()2@1 7I B[G520-^\T R V7+,1&7> 72
B+7-7H741 I7+ ['1- G72@S5)1-'2H0 20-V7+,1. %9N/SD020?'. FV)-O0+4'2(. [06+8'+E W\\T$
#$ U71-E' :?+'-H&02,7?. X+-O) :E01-' . 9)-'2 :4-;'2. B&)4)>>0 /')2. %&'() *'+','-. <A&0
0e0H- 7I +78-0+ 68e0+ 1)O0 72 -&0 A%B >0+I7+;'2H0<. )2 >+7H00()2@1 7I GJ/==F V7+,1&7>
72 A040H7;;82)H'-)72 /0-V7+,1 '2( A040-+'fH A&07+E. F-$ B0-0+168+@. N811)'. a'28'+E
W\\W$
$ %&'() *'+','-. 9)-'2 :4-;'2. <=;>'H- 7I /0-V7+, *8e0+1 72 A%B F-'+-SX><. )2 >+7H00S
()2@1 7I =999 K)((40 9'1- 37+,1&7> 72 /0-V7+,)2@. *0)+8-. G06'272. /7?0;60+ PQQQ$
Q$ %&'() *'+','-. a0'2SG78)1 N78@)0+. <J>-);)O'-)72 7I g)0+'+H&)H'4 K84-)H'1- A+001 )2 :AK
/0-V7+,1<. )2 >+7H00()2@1 7I =[=B :AK^Q . =4,40E. XU. a84E PQQ $

/#%#!$)3 /#"($4%5 6789&44#* !"#$%

P$ K7&';'( a'60+. %&'() *'+','-. <=-0+'-)?0 B'H,0-S*'10( :>>+7'H& [7+ 9'+4E :>>4)H'-)72
=(02-)_H'-)72<. 186;)--0( I7+ >864)H'-)72. 50H0;60+ W\\ $

!"#$ %$ &'()!*+#!,-"

"# $%&'( $) *'+(', $(%'-. /012-3, 45'.6 7'('8'3, 9:%.0(;3'%.6%1 350 <(2<0(360; 2= 350
2>0()'? 10%0('30. @? 763A2((0%39, ;-@+6330. =2( <-@)6B'362%# CDEC$ (0<2(3 '>'6)'@)0 '3
533< FGG5')#6%(6'#=(G6%(6'HIIJKILMN#
M# $+6( OECP$, Q25'+0. O'(6+ R7$C, 45'.6 7$E$O$A, A560((? A:E/SAAC, 9763*2B F
$ B2%30%3 ;5'(6%1 '<<)6B'362% =2( T6(0)0;; $. 52B D03&2(8;9, ;-@+6330. =2( <-@)6B'362%,
D2>0+@0( "II #
K# Q25'+0. O'(6+ R@'6, S+%' R')56, 45'.6 7'('8'3, 9$.'<36>0 2>0()'? =2( U"U +0+@0(;56<
+'%'10+0%3 6% Q$DSA9, CDEC$ 30B5%6B') (0<2(3 6%(6'HIIMK"!LJ, D2>0+@0( "II #
V# Q25'+0. O'(6+ R7$C, 45'.6 7$E$O$A, 9E0>6;636%1 B2%30%3 ;5'(6%1 6% &6(0)0;; '. 52B
%03&2(8;9, ;-@+6330., W0B0+@0( "II #

tel-00408745, version 1 - 3 Aug 2009

!"#$#
J# 45'.6 7'('8'3, XU0(=2(+'%B0 S>')-'362% 2= A4U 42%10;362% 42%3(2)Y, U5W 350;6;, $<(6)
"IIJ#
"# 45'.6 7'('8'3, 9/0 E2-3'10 Q-)36B';3 *6Z('(B56[-09, WS$ A50;6;, SRRC, :%6>0(;63? 2=
D6B0 R2<56'H$%36<2)6;, P('%B0, \-%0 JLL #
M# 45'.6 7'('8'3, E'%6 Q'886, *'?'% Q'5=2-], 9W0;61% '%. C+<)0+0%3'362% 2= ' ^6.02 A('H
B80(9, S)0B3(6B') '%. S)0B3(2%6B; S%16%00(6%1 W01(00, /0@'%0;0 :%6>0(;63?, 706(-3, /0@'%2%,
\-)? JLLN#

!!"#"

tel-00408745, version 1 - 3 Aug 2009

!"#$%&'()*+, )-! .#+*)#$ /'(0!%!+)

/$#1'!% 2 34)*%(' 5!)6#$789*:!
;(%4'*+,
!"# $%&" '(#&!%#!) !(#*+,( -(##(,,"#%) '.(/! 0(1(2(&)
'.1!3&"4.% 5!"&) 6(&1!,2 7.!1(#
!"#$% &'$!(#) *+! ()&(%', %# %"' -#).'$')/' #) 0).#$1+%(#) 2/(')/'
3-02245 6$()/'%#) 785 9+$/" :;;<= >#)? &'$!(#) +@@'+$', () @$#/'',()?!
#. -#7'A%5 B(!C#+5 6#$%D?+>5 E'/'1C'$ :;;<=

tel-00408745, version 1 - 3 Aug 2009

Reformulating the Monitor Placement Problem:

Optimal Network-Wide Sampling
Gion Reto Cantieni, Gianluca Iannaccone, Chadi Barakat, Christophe Diot, Patrick Thiran

tel-00408745, version 1 - 3 Aug 2009

ABSTRACT
Confronted with the generalization of monitoring in operational networks, researchers have proposed placement algorithms that can help ISPs deploy their monitoring infrastructure in a cost effective way, while maximizing the benefits of their infrastructure. However, a static placement of
monitors cannot be optimal given the short-term and longterm variations in traffic due to re-routing events, anomalies
and the normal network evolution. In addition, most ISPs
already deploy router embedded monitoring functionalities.
Despite some limitations (inherent to being part of a router),
these monitoring tools give greater visibility on the network
traffic but raise the question on how to configure a networkwide monitoring infrastructure that may contain hundreds
of monitoring points.
We reformulate the placement problem as follows. Given
a network where all links can be monitored, which monitors
should be activated and which sampling rate should be set
on these monitors in order to achieve a given measurement
task with high accuracy and low resource consumption? We
provide a formulation of the problem, an optimal algorithm
to solve it, and we study its performance on a real backbone
network.

1.

INTRODUCTION

Network operators perform traffic measurements as part

of their day by day network management activities that include traffic engineering, anomaly detection, accounting and
capacity planning. There exist several ways to perform traffic measurements. Some involve router support (e.g., SNMP
counters, Netflow [4]), while others require additional equipment to be installed in the network to perform passive or
active measurements.
The various solutions present a trade-off between the ac
G.Cantieni and P.Thiran are with EPFL, Switzerland.
G.Iannaccone is with Intel Research, Cambridge UK.
C.Barakat is with INRIA Sophia Antipolis, France. C. Diot
is with Thomson Research, Paris, France.

Permission to make digital or hard copies of all or part of this work for
personal or classroom use is granted without fee provided that copies are
not made or distributed for profit or commercial advantage and that copies
bear this notice and the full citation on the first page. To copy otherwise, to
republish, to post on servers or to redistribute to lists, requires prior specific
permission and/or a fee.
Copyright 2006 ACM 1-59593-456-1/06/0012 ...$5.00.

curacy of the measurement and the amount of computing

resources they require. SNMP counters, for example, represent a very low cost solution (in terms of router processing resources) but give the operator only a rough idea on
the traffic that is traversing the network. The aggregate
counters are of little use to operators interested in users
perceived performance [16] or in estimating network traffic demands [20]. At the other extreme, passive monitoring
equipment, that captures every packet on a link, allows extremely accurate measurements, but scales very poorly for
large networks, given the high unit cost for deployment and
maintenance. In addition, space and power constraints may
prevent operators from deploying this equipment where it
would be the most useful, while, once deployed, routing or
network changes may render it useless.
The solution that many network operators have adopted is
then to use Netflow [4] or similar solutions [23, 19]. NetFlow
allows a ubiquitous deployment of traffic monitors and provides a detailed enough view of the traffic streams. Netflow
is today widely used by Internet Service Providers (ISPs)
and several tools exist to process, analyze and visualize NetFlow data [9, 24]. However, enabling NetFlow can have an
impact on packet forwarding performance. To address this
problem, router vendors have introduced versions of Netflow that sample the incoming packets and update the flow
information only with sampled packets.
Network operators then face two options: (i) enable Netflow on all routers but using very low sampling rates to
minimize potential network impact, or, (ii), enable Netflow
on a chosen set of routers where the sampling rates are set
depending on the measurement task and the target accuracy.
Currently the first option is the one followed by ISPs because no automated method exists for the second. This work
aims at filling this gap.
The contributions of the paper are threefold. First, we
define a general framework to approach the problem of sampling traffic data in large IP networks. Our framework allows to combine and solve in one step the selection of traffic monitors and the setting of the sampling rates for each
monitor. We show how this framework can be applied to
a general class of measurement tasks. Second, we provide
an optimal algorithm to solve the sampling and placement
problem. We validate the algorithm using network data collected from the GEANTs backbone network [10]. Finally,
we discuss how to deploy our solution in a real backbone network and introduce additional methods that allow to adapt
the sampling rates to changes in the traffic due to time-ofday effects, failures or anomalies.

The paper is organized as follows. In Section 2, we describe the objectives and challenges of this work. Section 3
presents related work. Sections 4 and 5 formally define the
problem and our approach to solve it. Section 6 evaluates
the performance of our algorithm in the GEANT backbone
network. Section 7 describes how to deploy our solution in a
real backbone and how to deal with traffic changes when the
measurement task runs in a continuous fashion. Section 8
concludes the paper.

tel-00408745, version 1 - 3 Aug 2009

2.

OBJECTIVES AND CHALLENGES

The nature of the measurement task has a clear impact

on the choice of traffic monitors to be used. For example,
in order to estimate the traffic demands, a network operator
would ideally monitor all ingress links in the network or
at least all peering links [8]. However, when the network
operators want to focus on an individual network prefix or
Autonomous System (AS), they may require a very different
layout of monitors in the network.
Very often, network operators do not have prior knowledge of the measurement tasks the monitoring infrastructure
will have to perform. This is particularly true with security
applications. For example, a specific network prefix that is
below the radars for traffic engineering purposes may play
an important role in the early detection of anomalies.
Furthermore, network traffic demands are subject to short
term variations due to failures and other anomalous events
as well as longer term variations due to the addition of new
customers, peering links, Points of Presence, etc. These
changes quickly make a static placement of traffic monitors
perform sub-optimally.
For these reasons, Internet Service Providers (ISP) prefer
widespread monitoring infrastructures that provide visibility over the entire network. Netflow [4] is a perfect example
of such monitoring system. Embedded into the routers, it
maintains a list of flow records that describe the traffic forwarded by the router. The flow records are then exported
to a collector for analysis and storage. Netflow has become
today the de-facto standard of flow monitoring solutions: it
is the most widely deployed and various router manufactures support compatible monitoring applications [19]. A
standardization effort is also in progress at the IETF [15].
ISPs configure Netflow on all routers to the same safe
sampling rate router vendors usually recommend a value
of 1/1000 packets while little attention is paid to the accuracy of the results. Low sampling rates reduce the stress on
the routers but introduce large errors in the measurement.
Clearly, this approach leads to an inefficient use of the
resources both in the routers and in the collector, and to
a less than desirable measurement accuracy (we will quantify this in Section 6.3). Our objective is to find a method
that, given a measurement task and a target measurement
accuracy, has the following properties:
It selects the monitors that need to participate in the
measurement. This reduces the processing overhead
on the collector side where it is desirable that a large
proportion of the records received, processed and stored
are actually relevant to the measurement task.
It provides a set of sampling rates for the active monitors to guarantee an optimal use of the resources while
making sure that the sampling rates are low enough to
not be a concern for the operator.

The resource consumption, as measured by the total

number of sampled and processed packets, is minimal
and adapts to the changes in the traffic. The objective
is to maintain the overall resource consumption stable
while keeping the accuracy close to target.
The method requires a minimal amount of configuration from the network operator and is robust to incorrect input parameters.
In addition, our method should support a general class of
monitoring applications and be easily adapted to new measurement tasks. We choose one example of such tasks to
illustrate the contributions of this paper. We find the sampling rates to estimate the amount of traffic flowing among
a set of origin-destination pairs selected by the network operator. In our terminology, origin and destination could refer to any combinations of end-hosts, network prefixes, Autonomous Systems, etc. We have chosen this task because it
helps in illustrating our contribution and it is, at the same
time, a canonical measurement task for classical traffic engineering, security and accounting applications.

3.

RELATED WORK

Identifying the strategic locations for traffic monitors is a

hard problem that has attracted significant interest in the
literature. Several solutions have been proposed for different contexts. For example, in [17], the authors focus on
the placement of measurement devices for active monitoring
(more specifically for the construction of distance maps).
Others have addressed the placement problem in an active
monitoring infrastructure to measure delays and detect link
failures [1, 13, 21].
In the passive monitoring domain, Suh et al. [28] address
the problem of placing monitors and set their sampling rates
in order to maximize the fraction of IP flows being sampled.
They propose a two phase approach where they first find
the links that should be monitored and then run a second
optimization algorithm to set the sampling rates. Their approach bears some similarities with our work, but the analysis is limited to a generic monitoring goal (maximize the
overall sampled traffic) and only considers a static placement
of monitors. Their formulation leads to a set of heuristics
that find near-optimal solutions. Our approach, instead, allows to indicate whether a solution corresponds to the global
optimum.
Many researchers have also explored ways to improve NetFlow to automate some of the features and help network
operators configure the routers. Estan et al. [7] propose a
set of techniques to let a router running NetFlow adapt the
sampling rate in order to keep a fixed resource consumption.
The adaptation is a local and independent decision of the
router, and is not tied to any measurement objective. Our
work is complementary to [7] in the sense that it provides a
global sampling strategy for a specific monitoring goal. The
individual routers could then apply local decisions in order
to minimize their memory usage.
There is a large body of literature that addresses the problem of inversion of traffic properties from sampled traffic [6,
5, 12]. Duffield et al. [6] show that periodic and random sampling provide roughly the same result on high speed links.
Random sampling can thus be used in the mathematical
analysis for its appealing features.

i i L}. As the utility function M is strictly concave,

the optimization problem, given by (2), (4) and (5), has a
unique maximizer, that we denote p (see e.g., [3, Chapter
2]).
Note that an alternative objective could be to maximize
the minimum of the utilities, i.e., minkF M (k (p)). The
two formulations have their advantages and limitations. Maximizing the sum of utilities gives us more flexibility in setting
4. PROBLEM FORMULATION
the sampling rates. Indeed, we can compensate the poor acIn this section we formalize the placement problem. Our
curacy of one OD pair increasing the accuracy of another.
approach to solve it will be described in the next section.
We will discuss the implications of this in greater detail in
We represent the network by a directed graph G(V, E)
Section 6. On the other hand, maximizing the minimum
where V corresponds to the set of nodes and E is the set of
utility may lead to increase significantly the sampling rate
edges. The traffic load on edge e E is denoted by Ue . The
on those links that carry small OD pairs. Furthermore, the
routing of each origin-destination pair (OD pair) is specified
minimum of the utilities is not a differentiable function over
by the routing matrix R, whose entries ri,j = 1 if the OD
the whole parameter space and this may impact the converpair i traverses edge j and 0 otherwise.
The measurement tasks are defined over a set F = {1, 2, . . . , F } gence of the algorithm (see Section 5.4). We leave the study
of alternative objective functions for future work.
of OD pairs. We indicate the subset of links traversed by

tel-00408745, version 1 - 3 Aug 2009

Traffic matrix estimation techniques (e.g., [20, 29, 26])

address a measurement task similar to the one we are considering as example. However, the focus in those works is
the inference of the traffic matrix from partial information
(e.g., link loads). Indeed, they often use sampled Netflow
data to validate their methods.

the OD pairs in F by L E. The optimization framework

we propose in this paper is general and can be applied to
any definition of node (e.g., end-host, network prefix, Autonomous System, etc.).
The quality of a measurement for an OD pair k F is
computed via a utility function M : R+ R+ , whose argument k is the effective sampling rate of OD pair k F ,
defined as the probability that a packet of the kth OD pair
is sampled at least once by at least one monitor deployed in
the network. Note that this definition assumes that we have
means to discern whether the same packet is sampled at multiple locations in the network. In Section 6, we will address
this aspect of the monitoring infrastructure in more detail.
Assuming that the packets are sampled in an i.i.d. (independent and identically distributed) fashion at each monitor, with pi denoting the packet sampling probability of the
monitor deployed on link i, and that the sampling processes
of different monitors are statistically independent, we have
that
Y
(1)
(1 pi )rk,i .
k = 1

5.

METHOD

In this section we first reformulate the optimization problem using Lagrange multipliers. Second, we introduce and
comment on the assumptions that simplify the computation
of the optimal solution. Third, we discuss the choice of a
suitable utility function M (). Last, we briefly present the
algorithm used to solve the optimization problem and discuss its advantages and pitfalls.

5.1

Our approach

The most common approach to solve a constrained optimization problem, such as the one in (2), (3), (4) and (5),
is to define the corresponding Lagrangian:

L(p, , , )

kF

under the constraints

pi

for all i L

(3)

pi

for all i L

(4)

pi Ui

(5)

iL

where is the capacity of the system, defined as the maximum total number of packets that can be sampled in the
entire network, and i represents the maximum sampling
rate that can be applied to the individual link i.
All constraints are linear and P
therefore define a convex
solution space defined by {p | iL pi Ui , 0 pi

X
iL

pi Ui

iL

kF

iL

Clearly, the larger the effective sampling rate, the more information it brings. However, the marginal rate of information is usually smaller for large values of k than for small
values. These two observations lead us to reasonably assume
that M (k ) is an increasing and strictly concave function of
k .
Our objective is to choose the vector of sampling rates
p = (pi )iL that maximizes
X
M (k (p)) ,
(2)

M (k (p))

i (pi i ) +

i pi ,

!
(6)

iL

where (, , ) = (, i , i )iL is the set of Lagrange multipliers. Each Lagrange multiplier enforces the satisfaction
of one of the constraints (3), (4) and (5) with an equality
sign. A constraint met with an equality sign is called an
active constraint. For example, if pi = i , then the ith
constraint (4) is active, whereas if pi < i , it is inactive.
In order to find the unique maximizer p , we can solve the
system of equations provided by the Karush-Kuhn-Tucker
(KKT) conditions [11, Chapter 5]. As the solution space is
a convex hull and the objective function is concave, the KKT
conditions are sufficient for optimality (see e.g., [3, Chapter
5.5]). The difficulty in using the KKT conditions to solve the
problem given by (2), (3), (4) and (5) is that it requires to
know the set of active and inactive constraints in advance,
which is not possible. Therefore, we need to rely on an
iterative procedure, such as a gradient projection method
to explore the solution space. Our approach to find the
optimal solution p is given in more details further below.

5.2

Assumptions

Let us recall the assumptions made so far: the utility

function M is strictly increasing and concave; each monitor
sampling process is i.i.d. and is statistically independent
from the sampling process of the other monitors.

From a practical perspective, we expect to obtain sampling rates that are in the order of 0.01 and lower. Moreover,
we rarely expect to have more than one or two monitors observing the traffic of the same OD pair. This allows us to
approximate the effective sampling rate (1) by
X
rk,i pi .
(7)
k =

1
0.9
0.8

Utility Function M()

0.7

iL

This set of assumptions allows to make the optimization

problem more tractable and, as we show in Section 6.2, have
a minor impact on the performance of our method.
Next, we force the constraint (5) to be met with an equality sign. This assumption is pretty straightforward to make,
as there is no practical interest not to use all the provided
resources. Hence (5) becomes
X
pi Ui = .
(8)
iL

tel-00408745, version 1 - 3 Aug 2009

5.3

Choosing the utility function

The choice of the utility function M (k ) is dictated by the

following conditions. First, the utility function has to quantify the information provided by the measurement for each
OD-flow in F . Second, the function has to comply with the
requirements set by the optimization framework: as mentioned earlier, it is strictly increasing and concave. Without
loss of generality, we assume that M (0) = 0, i.e. that the
utility is zero if no packet is sampled at all. Third, the function M must be easy to use in our optimization algorithm.
This requires the function to be twice continuously differentiable. We next derive a possible function which combines
the above specified properties.
A first straightforward choice for the utility function M is
to use the relative error between the actual metric of interest
that we want to measure (in the present case, the flow size)
and its value estimated from the sampled packets. Let Sk be
the actual size of the kth OD pair (number of packets of the
kth OD pair) in a given time interval, and Xk be the number
of sampled packets from this OD pair in the same time interval. Because of the assumptions that the sampling processes
at different monitors are independent and that packets are
sampled at most once (Section 5.2), the distribution of Xk
conditionally to Sk is binomial with parameters (Sk , k ). In
this paper, we consider the squared relative error (SRE) between the estimated size Xk /k and the actual size of the
flow Sk , i.e.

SRE =

Xk /k Sk
Sk

(9)

whose expected value is

E[SRE](k ) =

1 k
1
1
dP(Sk s) = E
1 .
k s
Sk
k

Define A(k ) = 1 E[SRE](k ), that we call the mean

squared relative accuracy. A possible candidate for M (k )
would be to take it equal to A(k ), since it is a strictly increasing, concave function of k . However, this function is
not yet adequate to be used as utility function M . The
problem is that the function A(k ) is not defined at the
origin. This is required, as we expect to have zero utility for zero sampling. To fix this problem, we divide the

x0
(0.000599, 0.666)

0.6

x0
(0.00598, 0.668)

0.5
0.4
0.3
0.2
average size S = 500
average size S = 5000

0.1
0

0.002

0.004
0.006
Effective sampling rate

0.008

0.01

Figure 1: The utility function M with respect to the

effective sampling rate k . The utility function is
defined over the interval [0,1], such as given in (10).
interval [0, 1] in two intervals [0, x0 ] and [x0 , 1]. We take
M (k ) = 1E[SRE](k ) if x0 k 1 and M (k ) = A" (k )
for 0 k < x0 , where A" is defined in such a way that it is
strictly concave and increasing and twice differentiable over
[0, x0 ], with A" (0) = 0, A" (x0 ) = A(x0 ), A" (x0 )/k =
A(x0 )/k and 2 A" (x0 )/2k = 2 A(x0 )/2k . A function
that complies with these requirements is the quadratic expansion of A(k ) at x0 , that reads
A" (k ) = A(x0 ) + (k x0 )

(k x0 )2 2 A
A
(x0 ),
(x0 ) +
x
2
x2

where x0 is given by the relation A" (0) = 0, that is,

A(x0 ) x0

x2 2 A
A
(x0 ) = 0.
(x0 ) + 0
x
2 x2

To summarize, we define the utility function M as

"
A (k ) if 0 k x0
M (k ) =
A(k ) if x0 k 1.

(10)

We plot the function M (k ) for E[1/Sk ] = 0.002 and for

E[1/Sk ] = 0.0002 in Figure 1. We emphasize that this manipulation on the function has the sole purpose of making
it suitable for our optimization framework. The interval
0 k < x0 , on which we use the quadratic expansion
A" (k ), is of little practical interest because it corresponds
to low values of utility. Hence, we expect that this necessary
manipulation does not affect much the optimization results.
We will validate this observation in Section 6.

5.4

The algorithm

We solve our optimization problem by use of the gradient

projection method for constrained optimization (Refer to
[11, Chapter 5]). At each iteration step n, this method consists in projecting first the gradient of the objective function
onto the subspace spanned by the active constraints (that is,
the set of points p that satisfy the active constraints). This
projected gradient gives the search direction s(n), along
which the current feasible solution p(n) is moved until either the objective function is maximized along this line or
an inactive constraint is hit. In the latter case, this inactive constraint has to be activated and incorporated into the
next computation of the gradient projection. In the former
case, the maximization of the objective function along s(n)

tel-00408745, version 1 - 3 Aug 2009

reduces to a one-dimensional search. We choose Newtons

method (refer to [3], chapter 9.5) to perform this search.
Newtons method shows fast convergence but requires the
objective function to be twice continuously differentiable.
Once the new solution p(n + 1) that maximizes the objective along s(n) is found, a new search direction s(n + 1) has
to be computed. The new search direction will be orthogonal to the previous one. The successive search directions
s(n), s(n + 1), . . . form therefore a zigzag path in the subspace spanned by the active constraints, which may result in
a poor convergence depending on the shape of this subspace.
A better approach is to add, with some weighting factor, the
previous search direction to the new one. This weighting factor is usually chosen according the Polak-Ribiere rule. More
details can be found e.g. in [18].
We start our search with a feasible solution p(0) arbitrarily chosen on the plane defined by the active constraints (5).
We then follow the search direction s(0), that coincides with
the projection of the gradient of the objective function (2)
on the subspace spanned by the active constraints, until we
hit the constraint p1 0. This forces us to activate this constraint, i.e. to set p1 = 0, and to recompute next the search
direction s(1). With Newtons method we finally find p(2).
We mentioned earlier that in order to find the optimal
solution, we eventually have to know the set of active and
inactive constraints of our optimization problem. That is,
we have to know which monitors i are going to be used
(whose corresponding constraint pi > 0 is inactive) and
which ones not (whose corresponding constraint pi = 0 is
active). This combinatorial explosion is unavoidable as the
monitor placement problem is NP-hard. Probing each possible combination is not feasible in practice, hence we use
the above introduced gradient method to find iteratively the
optimum. However, we do not have any guarantee that the
gradient projection method always converges to the optimum. It might well happen that the projection of the gradient s converges to a zero vector at some point, say p" ,
which forces us to stop the search, even though p" is not
optimal. This means that the algorithm has not generated
the optimal set of active and inactive constraints, and thus
that some monitors are turned on/off on suboptimal links.
Given the feasible solution p" we have obtained, we apply
the Karush-Kuhn-Tucker (KKT) conditions [11, Chapter 5]
that determine whether this solution is optimal or not. To
do so, we have to compute the Lagrange multipliers defined
in (6). Having some of them negative indicates that the algorithm cannot converge to the optimum with the current
set of active constraints. It may be possible to continue the
search by removing (making inactive) some of the currently
active constraints.
A strategy is to remove (make inactive) the subset of active constraints associated with negative Lagrange multipliers (a similar strategy is given in [11, Chapter 5]). After
this operation, we recompute the gradient projection on the
subspace spanned by the new set of active constraints and
proceed with the search. We continue until we either reach a
point that satisfies the KKT conditions and therefore is the
optimal solution by KKTs theorem, or we abort the search
because the number of iterations (a new iteration starts each
time a new search direction has to be computed) exceeds a
threshold we set on the maximum number of iterations. In
our experiments described in Section 6, this threshold is set
to 2000 to keep the execution time of the algorithm in the

order of few seconds. We observe that in 98.6 percent of

cases the optimum is found in less than 2000 iterations.
The performance of the gradient projection method can
also be measured by the number of times we end up in the
situation where we have to remove (make inactive) the active constraints associated with negative Lagrange multipliers. To provide a number, we conducted 200 independent
executions of the algorithm, each time with a different set
of input parameters (different OD pair sizes, different link
loads, different capacity ). The input parameters are given
from the data set discussed in Section 6. Even though we
cannot have a formal upper bound on the number of times
we have to remove some active constraints, we observe an
average of 1.64 situations per execution of the algorithm,
where we have to remove (make inactive) the constraints
associated with negative Lagrange multipliers, with a standard deviation of 1.17. These low numbers, that are specific
to the monitoring problem we solve, justify our choice of the
gradient projection method to solve efficiently our optimization problem.

6.

EVALUATION

We study the performance of our method by defining and

simulating a measurement task on GEANT, the European
Research network [10]. Our goal is to verify the following
properties of the solution we obtain with the method described in Section 5: (i) the sampling rates are low and
validate the approximation in (7); (ii) the solution results
in a fair allocation of resources to each OD pair; (iii) the
solution is clearly superior to other solutions that could be
derived without running any optimization algorithm.

6.1

Data

We use sampled Netflow data collected on all the 23 routers

of the GEANT network. The links have varying speeds from
OC-3 (155 Mbps) to OC-48 (2.5 Gbps). We also collect in
a continuous fashion BGP and ISIS updates and use them
to build GEANT routing matrix.
Every GEANT router has NetFlow-compatible monitoring capabilities [19] enabled with a sampling rate of 1/1000.
The packets are classified by the 5-tuple (source and destination IP address, source and destination port number and
protocol number) and the flow records are exported every
minute by the routers. Each record contains the following
information (in addition to the 5-tuple) that is relevant to
this study: (i) Flow start and end time. The start time is
the timestamp of the first sampled packet of the flow. The
end time is the timestamp of the last packet of the flow.
Flow termination is triggered either by a FIN packet or by
an idle timeout (set to 30 seconds). (ii) Sampled packets
and bytes. The total number of sampled packets in the flow
and their cumulative size in bytes. (iii) Source and Destination Autonomous System (AS). The AS numbers to which
the source and destination IP addresses belong. (iv) Input
and output interface. The index of the router interface on
which the flow was received and sent.
Before using NetFlow data, we need to perform some additional post-processing of the records. First, we aggregate
all flow records (exported by the routers every minute for all
active flows) in 5 minutes bins according to their start time.
We choose 5 minutes as our measurement interval to reduce
the impact of synchronization issues that could have arisen
when collecting flow data from different routers. Then, we

adjust the sampled packet and byte count by multiplying

them by the inverse of the sampling rate (1000 in our case).
Henceforth, we consider the post-processed NetFlow data
to represent the actual traffic traversing the GEANT network. Our experiments then consist in simulating a sampling process with the rates set by the optimal algorithm
and comparing the results with the post-processed data. Although our validation does not depend on a perfect reconstruction of the traffic dynamics, the sampled Netflow data
present a potential bias against small flows that can affect
the relative contribution of each OD pair of interest. Unfortunately, this is the only type of data available today for
research and we are not aware of any other public dataset
that contains full unsampled information for a network of
the size of GEANT.

tel-00408745, version 1 - 3 Aug 2009

6.2

Results

For the evaluation we choose to estimate the traffic sent

by JANET (UK Research Network, AS number 786) to each
individual GEANT PoP through the UK PoP.
This task gives us a set F of 20 OD pairs. The OD pairs
traverse 22 of the 72 unidirectional links of GEANT. We
use flow data from November 22nd, 2004 where we associate to each flow record the egress PoP, computed from the
destination IP address using the technique presented in [8].
The first two columns of Table 1 present a summary of the
OD pairs and their sizes in packets/sec. The results in this
section are computed over a single measurement interval.
This task shows that the method operates on OD pairs
where origins (i.e., the JANET AS) and destinations (i.e,
all the GEANT PoPs) are of a different nature. In addition,
the OD pair sizes cover the entire spectrum: JANET to
Netherlands (NL) consist of more than 30,000 packets/sec
while JANET to Luxembourg is made of a mere 20 packets/sec. As it will become clear later, this is one of the main
strengths of the method, i.e., the ability to track indifferently small and large OD pairs.
The remaining columns in Table 1 represent the sampling
rates that the optimal solution provides. For each link i, we
indicate the sampling rate pi and the OD pairs that traverse
(and are sampled on) that link. All links that are not present
in the table have pi = 0, i.e., those monitors do not need to
participate in the measurement task. Finally, the last two
rows of the table show the load in packets/sec on the links
and the relative contribution to the total capacity. In this
experiment we have chosen a value of = 100, 000, that is,
at most 100, 000 packets can be sampled in each 5 minutes
measurement interval in the entire network. We also set
i = 1 for all links, i.e., we do not define an upper limit for
the sampling rates. Thus, we assume the operator has no
prior knowledge of the network traffic.
We can immediately observe that the sampling rates are
extremely low on most links even if no upper limit was set.
Only two links (FR-LU and CZ-SK) need a sampling rates
somewhat higher (around 0.9%), but they are lightly loaded
links needed to accurately estimate the two smallest OD
pairs (JANET-SK and JANET-LU). Furthermore, the low
sampling rates we obtain and the fact that each OD pair is
sampled in at most two links validate the assumptions on
the effective sampling rate made in Section 5.2.
The last two columns in Table 1 show the value of the
utility function (10) and the accuracy of the measurement.
We define the accuracy of an OD pair size as 1 minus the

absolute relative error: 1 | x/ s | /s, where s is the

actual size of the OD pair, x is the sampled size and is the
effective sampling rate as in (7).
We use the accuracy instead of the utility M (10) to validate the impact of two assumptions we made in the definition of M : (i) the quadratic expansion to force M to zero
when the sampling rate is zero (see Section 5.3); (ii) the approximation (7) for the effective sampling rate , that may
result in overestimating the size of certain OD pairs.
We run 20 sampling experiments on the flow records and
compute the average accuracy over the 20 runs. Each sampling experiment consists in simulating a random sampling
process on the flow records observed on link i using the sampling rate pi in Table 1. The values in those two columns
demonstrate that the method achieves good fairness among
OD pairs. Although the algorithm maximizes the sum of the
utilities, the results indicate that the individual utilities are
well balanced. Moreover, the accuracy of the measurement
is extremely good being on average above 0.89 for any OD
pair.

6.3

Comparison with other solutions

The last aspect we want to address is how the optimal

solution compares to nave solutions. Clearly, if any solution
performs well enough, then there is no reason to add the
complexity of the optimization algorithm we propose.
The first nave solution would consist in monitoring only
the JANET access link to GEANT. This solution has the
advantage that every sampled packet would belong to one
of the OD pairs of interest. However, in order to track a
small OD pair (e.g., JANET to Luxembourg) with a similar
accuracy to the one we obtain in Table 1, the network operator would be forced to sample the link at a rate of about
1%, i.e., the effective sampling rate for JANET-LU. Given
the high load on that link, this would require the capacity
to be 70% higher than the one needed by our method to
give the same measurement accuracy1 .
Furthermore, apart from being a suboptimal solution, monitoring the access link may not be feasible in all scenarios given that, in corporate networks for example, the edge
routers (i.e., the router that is connected to the ISP) is often
directly owned and managed by the ISP that provides network connectivity. These routers, usually called CPE (Customer Premise Equipment), can only be accessed by the ISP
network operators. Therefore, in order to keep our method
general and applicable to a wide range of network scenarios,
we do not include the access links in the set of possible links
to monitor. Even with this additional constraint, Table 1
shows that our method performs extremely well.
An alternative to the monitoring of the access link is to
monitor all links that connect the UK PoP to the other PoPs
in GEANT. This solution allows to balance the sampling
rates over six links, instead of just one, and gives more freedom to reduce the resource consumption.
In order to compare this solution with the optimum we
run our algorithm and restrict the choice of available monitors to just the six UK links. Figure 2 shows the comparison
in terms of the accuracy over a wide range of values of the
capacity . With respect to the optimum, this simple solu1
Adding up the values in the second column of Table 1 we
obtain 57,933 packets per second. At a sampling rate of 1%,
this results in 173,798 sampled packets on average over a 5
minutes interval.

OD pair

pkt/s

p5
UK-FR
0.0013
0.0013
0.0013
0.0013
0.0013
0.0013
0.0013
0.0013
63603
24.5%

JANET-NL
30123
9387
JANET-NY
JANET-DE
4300
4080
JANET-SE
JANET-CH
4033
JANET-FR
1723
1400
JANET-PL
JANET-GR
1080
1003
JANET-ES
JANET-SI
913
873
JANET-IT
JANET-AT
790
JANET-CZ
590
490
JANET-BE
JANET-PT
463
JANET-HU
377
237
JANET-HR
JANET-IL
87
JANET-SK
43
20
JANET-LU
Link Loads (pkt/s)
Contribution to

p7
UK-SE
0.0003
0.0003
51833
5.1%

p8
UK-NL
0.0016
0.0016
0.0016
0.0016
0.0016
0.0016
0.0016
0.0016
57756
26.9%

p9
UK-NY
0.0002
37286
2.1%

p17
SE-PL
0.0003
23680
2.1%

p30
UK-PT
0.0011
19950
6.8%

p31
IT-IL
0.0018
15213
8.3%

p33
FR-BE
0.0002
11173
0.7%

p2
FR-LU
0.0090
6133
16.5%

p28
CZ-SK
0.0092
2600
7.1%

Utility
0.9999
0.9982
0.9995
0.9973
0.9994
0.9985
0.9960
0.9981
0.9974
0.9977
0.9971
0.9968
0.9965
0.9955
0.9935
0.9945
0.9912
0.9877
0.9929
0.9840

Avg
Accuracy
0.993
0.965
0.982
0.960
0.979
0.969
0.950
0.964
0.959
0.961
0.956
0.954
0.952
0.946
0.937
0.940
0.924
0.910
0.932
0.897

0.9

0.8

0.7

Average accuracy

tel-00408745, version 1 - 3 Aug 2009

Table 1: Optimal sampling rates pi for each link i. For each OD pair, it indicates the size (pkts/sec), the
link(s) where it is monitored, their respective sampling rates and the average accuracy. All other links have
zero sampling rate.

0.6

0.5

0.4

0.3
Average over all OD pairs
Worst OD pair
Best OD pair
Average over all OD pairs (UK links only)
Worst OD pair (UK links only)
Best OD pair (UK links only

0.2

0.1

10

10

Resource constraint

Figure 2: Accuracy of the measurement for two different solutions as a function of .

tion has poor performance with respect to small OD pairs.
This is expected given that the UK links are heavily loaded
and the high sampling rate required to accurately estimate
a small OD pair results in a high resource consumption.
In summary, we have seen that both in terms of resource
usage and measurement accuracy there is a clear advantage
in using our method for setting the sampling rates when
compared to other nave solutions. This advantage comes
from the network-wide approach to packet sampling. Indeed, the optimization method finds those links across the
entire network, where the small OD pairs manifest themselves with a small amount of cross traffic from other large
OD pairs. Several studies [2, 8] have shown that this is a
general property of current network design.

7.

DEPLOYMENT ON A BACKBONE NETWORK

Before deploying our method in a real backbone network,

we need to address two additional issues: (i) how to set the

input parameters and (ii) how to deal with traffic variations

over time.
As described in Section 5, the input parameters consist of
the quantities , E[1/Sk ] and Ui , for all k F and i L. In
order to obtain these quantities, the network operator needs
to have some prior knowledge about the network. In particular, the operator needs to have a rough idea of the sizes,
Sk , of the OD pairs that are under study. This is a classical bootstrapping problem and we will present a solution in
Section 7.1.
The second deployment issue has to do with the temporal
fluctuations of the OD pair sizes and link loads. In the flow
data, we observe that these fluctuations may be very large
and abrupt. This phenomenon is also more pronounced for
small OD pairs and confirms previous measurement studies [27].
It is out of the scope of this paper to investigate the reasons behind these fluctuations. They may be due to normal time-of-day effects, failures in the network, flash crowd
events or other anomalies. What is important to observe
is that these fluctuations may lead our method to overshoot the capacity when the link loads increase abruptly,
or to miss the target accuracy when an OD pair size decreases. Section 7.2 shows the impact of the fluctuations on
the performance of our method, while Section 7.3 introduces
a simple heuristic to effectively cope with them.
In the remainder of this section we will refer to flow data
collected over a 12 hours period from 8AM to 8PM on November 22nd, 2004.

7.1

Bootstrapping phase

The bootstrapping method we propose is very simple.

Given the presence of traffic monitors on every link in the
network, an operator can get a rough estimate of the OD
pair sizes and link loads by activating all monitors at once.
The monitors should be activated at a very low sampling
rate, that we call background sampling rate (
p), to guarantee a minimal impact on the performance of the routers and
a limited overall resource usage.

70

Actual
Estimated

60

Packets/sec

0.95

0.9

Accuracy

40
30
20
10

0.85

10

12

14

16

18

20

0.8
12000

0.75

Packets/sec

Average over all OD pairs (one iteration)

Worst OD pair (one iteration)
Average over all OD pairs (two iterations)
Worst OD pair (two iterations)

0.65

0.6
7
10

Actual
Estimated

10000

0.7

8000
6000
4000
2000
0

10

10

10

10

10

10

10

Figure 3: Performance of the bootstrapping algorithm as a function of the background sampling rate.

We can then use these estimates as input parameters to

our method, derive a new set of sampling rates for all the
routers (many of which will be set to zero), re-configure
all routers, wait one measurement interval and get a new
estimate of the input parameters. We iterate this procedure
until the set of sampling rates converges and no router needs
to change its settings.
This approach, however, raises three questions: (i) how do
we estimate the size of the OD pairs? (ii) what is an appropriate background sampling rate? (iii) how many iterations
are needed to perform the bootstrapping?
The optimization algorithm needs the values of E[1/Sk ]
for all OD pairs of interest. After the first measurement
interval, we obtain Xk , the sampled size of each OD pair k.
If Xk > 0, we replace E[1/Sk ] in (10) with k /Xk note that
pi = p, for all i L. In the case Xk = 0, the OD pair has
not been sampled at all. This means that the original OD
pair size is likely smaller than 1/
p. For our purposes, here,
any value below 1/
p is a good enough estimate of its size.
Hence, we replace E[1/Sk ] with c
p, where c is a constant
arbitrarily set to 10 in all our experiments.
Note that finding an accurate estimator for E[1/Sk ] is outside the scope of this paper. In fact, we deliberately choose
to use one single sample of the size as the estimator because it is simple and clearly imperfect. Our aim is to show
that our bootstrapping method is robust to incorrect input
parameters and requires no configuration effort.
To answer the last two questions, we run an experiment
over the flow data where we vary the value of p from 107
to 1. We estimate the size of the OD pairs from JANET to
all PoPs and the link loads across the entire network. Then,
we run the optimization algorithm using this set of input
parameters and a value of = 100, 000 (as in Section 6) to
derive a set of sampling rates.
Figure 3 shows the average and worst accuracy over all
OD pairs as a function of the background sampling rates.
The figure compares the performance after one or two iterations. It is clear that two iterations are sufficient to complete
the bootstrapping. Indeed, after two iterations the accuracy
of the measurement does not depend anymore on the background sampling rate and thus on the quality of the input
parameters. We also observe that a background sampling
rates of 107 is enough if the algorithm is allowed at least
two iterations. This demonstrates that the bootstrapping

10

12

14

16

18

20

Hour (UTC)

Background sampling rate

tel-00408745, version 1 - 3 Aug 2009

50

Figure 4: Estimated size of JANET to Luxembourg (top) and JANET to Switzerland (bottom)
OD pairs. Sampling rates are fixed over the entire
duration of the experiment
algorithm is very robust against incorrect input parameters.

7.2

Performance over time

We evaluate the performance of our method by estimating the size of all OD pairs originating from JANET over
the entire period 8AM to 8PM. As in Section 6, we set a
capacity = 100, 000. We perform the bootstrapping algorithm as described in the previous section at 7.55AM with
a sampling rate of 107 . The sampling rates obtained from
the optimization algorithm are then set on all the routers
and unchanged for the entire 12 hour period.
Figure 4 compares the estimated and actual size for two
OD pairs (JANET to Luxembourg and JANET to Switzerland). The figure shows the result for one run of the sampling process over the 12 hour period. The estimate is extremely accurate and closely follows all fluctuations over
time for the JANET to Switzerland OD pair (the two lines
overlap completely in the figure). The estimate is somewhat less accurate for JANET to Luxembourg (our worst
case) but this is well expected given the small size of this
OD pair (it peaks at 70 packets/sec).
In Figure 5 we plot the accuracy of the measurement over
20 experiments. Each experiment consists in one run of the
sampling process across the entire 12 hour period. We show
the average accuracy over all OD pairs and the accuracy
for the worst OD pair over time. The average accuracy
is always above 0.95 indicating that we are able to closely
track the fluctuations in all OD pairs. Even, the worst OD
pair is tracked with an accuracy around 0.90, but in certain
measurement intervals, the accuracy drops below 0.85.
A second metric that plays a crucial role in our method
is the evolution over time of the measured resource consumption defined as the total number of sampled (and thus
= P pi Ui (t). We are interested in
processed) packets, (t)
i
comparing this value with , the overall capacity. Figure 6
)/ .
shows the relative difference computed as ((t)
As we can see, the actual resource consumption is always
above the constraint and it may exceed the constraint by
as much as 120%. This is a consequence of having set the
sampling rate early in the morning without taking into account time-of-day effects that lead to an increase in traffic
in the later hours of the day. Therefore, the sampling rates
are too high for the resource constraint, but result in good

0.95

0.95

0.9

0.9

0.85

0.85

Accuracy

Accuracy

0.8

0.75

0.8

0.75

0.7

0.7
Average over all OD pairs
Worst OD pair

0.65

10

12

14

16

18

Average over all OD pairs

Worst OD pair
20

Hour (UTC)

0.65

10

12

14

16

18

20

Hour (UTC)

Figure 5: Accuracy over all OD pairs averaged over

20 experiments. Sampling rates are fixed over the
entire duration of the experiments

Figure 7: Average accuracy over all OD pairs and

worst accuracy. Monitors are reconfigured as soon
is 10% above or below the
as the resource usage (t)
capacity .

Relative difference between actual and target *

tel-00408745, version 1 - 3 Aug 2009

1.4

1.2

0.8

0.6

0.4

0.2

10

12

14

16

18

20

Hour (UTC)

Figure 6: Relative difference between the actual resource consumption (t) and the target constraint .
Sampling rates are fixed over the entire duration of
the experiment
accuracy.

7.3

Adapting to traffic fluctuations

Ideally, the network operator would like to define a measurement task and a target accuracy (or a limit on the
resource consumption) and then have the monitoring infrastructure that provides the best possible results independently of when the measurement is started or which network
events occur during the measurement.
There are three specific events that make a given set of
sampling rates suboptimal: (1) The monitored links start to
carry a larger amount traffic that is not of interest for the
measurement task. This leads to an increase in the resource
consumption that does not correspond to an increase in the
accuracy (as described in Section 7.2). (2) An OD pair of
interest decrease in size requiring a higher sampling rate
in order to keep the target accuracy. (3) The OD pairs
of interest are routed differently in the network and some
monitors are only capturing traffic that is not of interest.
In the following, we present solutions that allow us to cope
with each one of these events.
Fluctuations in the link loads. The link loads vary over
time for a variety of reasons, including time-of-day effects,

anomalies, flash crowds, etc. At peak times the traffic on

a link may be orders of magnitude larger than during quiet
times. If the sampling rates are fixed and set during quiet
times, this may lead to exceed the resource constraint as
shown in Figure 6.
A possible solution is to define two thresholds L and H
around the capacity and then predict the number of sam + 1). When the prepled packets in the next interval, (t
diction exceeds those thresholds, the collector can trigger
a recomputation of the optimal sampling rates and (potentially) reconfigure the monitors.
At the end of each measurement interval, the collector
+ 1) from the number of sampled packets
can estimate (t
derived from the received flow records. Identifying an
(t)
accurate predictor of the resource usage is outside the scope
of this paper. Our goal is to provide an analysis of the
performance of this method with the simplest predictor: we
as the predictor of the resource usage in the next
use (t)
+ 1).
interval, (t
Figures 7 and 8 show the performance of this method
when L = 0.9 and H = 1.1. The results are averaged
over 20 experiments. Each experiment is started at 8AM
with the bootstrapping taking place at 7.55AM (UTC). As
we can see this method succeeds in keeping the resource
consumption within the desired bounds even if the bootstrapping is performed during quiet times (early morning).
With this adaptive method, the network operator can start
the measurement at any time and for any duration without
incurring the risk of overloading the monitoring infrastructure.
OD pair fluctuations. The case of fluctuations in an OD
pair traffic volume is the most difficult to address. Indeed,
if the size of an OD pair decreases, in order to preserve
the same measurement accuracy, there exists no alternative
but to increase the sampling rates. If the OD pair accounts
only for a small portion of the traffic on the monitored link,
increasing the sampling rate causes a large amount of traffic
not relevant for the measurement task to be collected.
We can see this phenomenon in Figure 7, where the accuracy of the worst OD pair (JANET to Luxembourg in our
network scenario) drops significantly around 1PM and 7PM.

0.4

Resource Consumption

0.3

0.2

0.1

x 10

1
Capacity
Resource usage
0

10

12

14

16

10

12

14

16

18

20

0.1

0.2

0.3

0.4

10

12

14

16

18

20

Resource Consumption

Relative difference between actual and target

Hour (UTC)

x 10

1
Average Capacity
Capacity with perfect knowledge
0

18

20

Hour (UTC)

Figure 8: Relative difference between the resource

usage (t)
and the capacity . Monitors are recon is 10% above or below .
figured as soon as (t)
1

Figure 10: Adapting to preserve the target accuracy. Top: actual resource consumption over one
experiment. Bottom: average over 20 experiments
compared with the case of complete knowledge of
the traffic.

0.95

0.95

0.85
0.9

0.8

Accuracy

Accuracy

tel-00408745, version 1 - 3 Aug 2009

0.9

0.75

0.85

0.8

0.7
0.75
Actual
Estimate
0.65

10

12

14

16

18

20

Hour (UTC)

0.7
Average over all OD pairs
Worst OD pair

Figure 9: Comparison between estimated and actual

accuracy for JANET to Luxembourg OD flow

0.65

10

12

14

16

18

20

Hour (UTC)

Figure 11: Average accuracy over all OD pairs when

is adapted to meet target accuracy (0.85)
These drops are consistent with the decrease in size of the
JANET to Luxembourg OD pair (see Figure 4), but they
do not correspond to equivalent decreases in the link loads.
Therefore, in the cases where the network operator is interested in preserving a target measurement accuracy, the
capacity (i.e., the total number of sampled packets) needs
to adapt to network conditions as well.
We now face a second challenge. Given a target accuracy
set by the network operator, how can we estimate if, given
the sampled size Xk (t) of each OD pair k, the sampling rate
is good enough to meet our target in the next intervals?
To this end, we approximate the accuracy as 1 minus the
square root of (10). As we did in Section 7.1, we replace
E[1/Sk ] by k /Xk (t), where k is the effective sampling rate
for OD pair k.
Figure 9 compares the estimated accuracy with the actual
accuracy from the data for the JANET to Luxembourg OD
pair (the smallest OD pair). As we can see, our estimate
closely follows the actual accuracy.
We can then use this approximation of the measurement
accuracy to trigger a new reconfiguration of the traffic monitors as soon as the worst accuracy drops below the defined
target.
We first run the optimization algorithm with the same
value of used in the previous interval. We then compute

again the estimate accuracy using the same size estimate

Xk (t) but with the new sampling rates. If the estimated
accuracy is still below the target, we increase by 10% until
we reach the target accuracy for all OD pairs.
Furthermore, in order to avoid keeping a unnecessary
large for long periods, we trigger a new reconfiguration if
the estimated worst accuracy is above our target for one
consecutive hour.
Figure 10 shows how the capacity varies over time when
the target accuracy is set to 0.85. The top graph shows, for
one experiment, the evolution of the actual resource usage
and the capacity. The bottom graph shows the average over
20 experiments and compares it with the capacity that could
be set knowing the exact size of each OD pair. Finally,
Figure 11 illustrates the performance in terms of accuracy
over all OD pairs and worst OD pair.
Around 1PM we observe a significant increase of the capacity . This corresponds to a significant drop in accuracy for the worst OD pair. Note that in the previous nonadaptive schemes, the poor accuracy would persist for more
than one hour (see Figure 7). Increasing allows instead to
quickly bring the accuracy above the target. Moreover, is
decreased as soon as there is no more need of the additional

Packet sampling rate pi

0.1
link 9 (FR-BE)
link 17 (SE-PL)
link 30 (IT-IL)
Link 33 (FR-LU)
Link 28 (CZ-SK)

0.01

0.001

0.0001
09h

10h

12h

13h

15h

16h

18h

19h

Packet sampling rate pi

0.1
Link 5 (UK-FR)
link 7 (UK-SE)
link 8 (UK-NL)
link 31 (UK-NY)
Link 2 (UK-PT)

0.01

0.001

0.0001
09h

10h

12h

13h

15h

16h

18h

19h

tel-00408745, version 1 - 3 Aug 2009

Figure 12: Evolution of the sampling rates over

time. Top: lightly loaded links. Bottom: heavily
loaded links.

resources.
A last aspect we want to investigate is the evolution of
the sampling rates over time. We are interested in having
sampling rates that are always low. Moreover, we want to
verify that to track effectively OD pairs, a static placement
of monitors is not sufficient.
Figure 12 shows the sampling rates for the 10 monitors
that are activated during the day. The top graph shows the
sampling rates for the lightly loaded links, while the bottom
graph shows the sampling rates on the heavily load links.
The number of monitors used to track the 20 OD pairs of
interest varies between eight and ten. Often the algorithm
includes one or more of the lightly loaded links in order to
improve the accuracy or reduce the resource consumption.
The sampling rates exceed 1/100 only for the lightly loaded
France to Luxembourg link that is instrumental to accurately track the fluctuations of the JANET to Luxembourg
OD pair. The heavily loaded links (all links from the UK, a
major PoP in GEANTs network) are all set to low sampling
rates and experience much less variability over time.
Routing matrix changes. Network failures or BGP updates may modify the path taken by an OD pair under study
across the network. When this occurs, some monitors may
be recording traffic that is not of interest for the operator.
Detecting routing changes is very simple given that routing messages need to propagate to all routers. ISPs often
deploy systems to record and process all intra-domain and
inter-domain routing messages [14, 25]. The monitoring infrastructure can therefore be aware of routing changes that
affect the OD pairs of interest and trigger a new bootstrapping phase. The bootstrap would allow to find where the
OD pairs are now entering the network (e.g., for multihomed
customers) and get a new estimate of their size. Therefore,
after two measurement intervals the monitoring infrastructure will have reconfigured itself with a new set of optimal
monitors and sampling rates (see Section 7.1).

7.4 Implementation feasibility

We have shown that it is possible to design a self-configuring,
self-adapting monitoring infrastructure in todays backbone
networks. Our proposed method can be deployed today
in any network running NetFlow (or any other widespread
monitoring infrastructure). The network operator has just
to select a measurement task (e.g., track any set of OD pairs)

and a measurement strategy (i.e., focus on the resource usage or on the accuracy). The proposed mechanisms will then
choose the monitors that allow to track those OD pairs with
an optimal use of the resources and adapt to changes in the
network conditions.
However, we did not look at the time it takes to reconfigure NetFlow on todays routers. All our results assume
zero time to run the optimization algorithm (that is instead
in the order of a few seconds), and to set the new sampling
rates on the routers. Unfortunately, we do not have any figures about the time needed to remotely configure NetFlow,
as we did not have any access to the routers in the network
but only to the flow records stored at the collector. We leave
this task as future work.
Our approach uses only the information stored (or that
can be derived) from NetFlow records. This is one of the
strengths of our method given that it is generic and can
be applied to new NetFlow versions without modifications.
Newer versions of NetFlow are constantly increasing the
amount of information that is exported in each record. For
example, NetFlow v9 records contain BGP next-hop information that would significantly simplify the task of finding
the egress PoP for a given network prefix [22]. Moreover,
router vendors are reportedly working on ways to support
variable sampling rates as in [7]. In this case, our method
can be used to define lower bounds on the sampling rates
across the entire network, while the individual routers adapt
to the local traffic conditions given their available resources.
The method gives the best performance when compared to
alternative solutions (Section 6.3) if different network links
have different load and carry a diverse set of OD pairs. Indeed, lightly utilized links are candidate for higher sampling
rates that allow to compensate for the low sampling rates
that we set on heavily loaded links. This is true for the
GEANT network and other measurement studies have indicated that this is a common case for other ISP networks as
well [2, 8].
Finally, there are a number of ways in which the adaptive
methods described in this section can be extended. The
choice to increase by steps of 10% and to decrease it after
an hour is in fact somewhat arbitrary (see Section 7.3). It
showed to perform well in practice but one can find smoother
ways to adapt to the traffic conditions. It is also possible
to use different predictors for the resource usage, as well
as estimators for the OD pair size. All the results show
the baseline performance when the prediction is done using
just the previous sample. Even with this simple techniques,
our method has demonstrated to work well and achieve the
target accuracy. However, other predictors may perform
better for different datasets.

8.

CONCLUSION

We reformulated the monitor placement problem to adapt

it to the reality of network operations and management in
the Internet. We have proposed an optimization method
to select and configure passive monitors in a backbone network. The method receives as input the network topology,
the routing matrix and the set of OD pairs of interest. It
returns a set of monitors (and their sampling rates) that is
optimal with respect to the measurement task to perform.
We have described the performance of our method considering a canonical measurement task, i.e., the estimation of
the size of a set of OD pairs.

Although the evaluation in this paper is in terms of estimating OD pair sizes, the optimization method is not specific to them. The method can be applied to a wide range
of measurement tasks for which a utility function can be
sought. Our ongoing work is centered on defining new expressions for the utility function for applications such as
anomaly detection and network performance analysis. We
are also studying alternative formulations for the objective
function as well as trying to evaluate the performance of our
method in larger networks and different datasets.

ACKNOWLEDGMENTS
We wish to thank DANTE for granting us access to the
NetFlow, BGP and ISIS dataset of the GEANT network to
perform this study.

tel-00408745, version 1 - 3 Aug 2009

9.

REFERENCES

[1] Y. Bejerano and R. Rastogi. Robust monitoring of

link delays and faults in IP networks. In Proceedings
of IEEE Infocom, Apr. 2003.
[2] S. Bhattacharyya, C. Diot, J. Jetcheva, and N. Taft.
Geographical and temporal characteristics of
inter-PoP flows: View from a single PoP. European
Transactions on Telecommunications, 13(1):522, Feb.
2002.
[3] S. Boyd and L. Vandenberghe. Convex Optimization.
Cambridge University Press, 2004.
[4] Cisco Systems. NetFlow services and applications.
White Paper, 2000.
[5] N. G. Duffield, C. Lund, and M. Thorup. Properties
and prediction of flow statistics from sampled packet
streams. In Proceedings of ACM Internet
Measurement Workshop, Nov. 2002.
[6] N. G. Duffield, C. Lund, and M. Thorup. Estimating
flow distributions from sampled flow statistics. In
Proceedings of ACM Sigcomm, Aug. 2003.
[7] C. Estan, K. Keys, D. Moore, and G. Varghese.
Building a better NetFlow. In Proceedings of ACM
Sigcomm, Aug. 2004.
[8] A. Feldmann, A. Greenberg, C. Lund, N. Reingold,
J. Rexford, and F. True. Deriving traffic demands for
operational ip networks: Methodology and experience.
In Proceedings of ACM Sigcomm, Sept. 2000.
[9] M. Fullmer and S. Romig. The OSU flow-tools
package and Cisco NetFlow logs. In Proceedings of
Usenix LISA, Dec. 2000.
[10] GEANT. http://www.dante.net.
[11] R. T. Haftka and Z. Guerdal. Elements of Structural
Optimization. Kluwer Academic Publishers, third
edition.
[12] N. Hohn and D. Veitch. Inverting sampled traffic. In
Proceedings of ACM Internet Measurement
Conference, Oct. 2003.
[13] J. Horton and A. Lopez-Ortiz. On the number of
distributed measurement points for network
tomography. In Proceedings of ACM Internet
Measurement Conference, Nov. 2003.
[14] G. Iannaccone, C.-N. Chuah, S. Bhattacharyya, and
C. Diot. Feasibility of IP restoration in a tier-1
backbone. IEEE Network, 18(2):1319, Mar. 2004.

[15] IP Flow Information eXport Working Group. Internet

Engineering Task Force.
http://www.ietf.org/html.charters/ipfix-charter.html.
[16] S. Jaiswal, G. Iannaccone, C. Diot, J. Kurose, and
D. Towsley. Inferring TCP connection characteristics
through passive measurements. In Proceedings of
IEEE Infocom, Mar. 2004.
[17] S. Jamin, C. Jin, Y. Jin, D. Raz, Y. Shavitt, and
L. Zhang. On the placement of internet
instrumentation. In Proceedings of IEEE Infocom,
Apr. 2000.
[18] A. K. John Hertz and R. G. Palmer. Introduction to
the Theory of Neural Computation. Addison-Wesley.
[19] Juniper Traffic Sampling. http://www.juniper.net.
Netflow-compatible implementation.
[20] A. Medina, N. Taft, K. Salamatian, S. Bhattacharyya,
and C. Diot. Traffic matrix estimation: Existing
techniques and new directions. In Proceedings of ACM
Sigcomm, Pittsburgh, Aug. 2002.
[21] H. Nguyen and P. Thiran. Active measurement for
multiple link failures diagnosis in IP networks. In
Proceedings of PAM Workshop, Apr. 2004.
[22] K. Papagiannaki, N. Taft, and A. Lakhina. A
distributed approach to measure IP traffic matrices. In
Proceedings of ACM Internet Measurement
Conference, Oct. 2004.
[23] P. Phaal, S. Panchen, and N. McKee. InMon
corporations sFlow: A method for monitoring traffic
in switched and routed networks. RFC 3176, Sept.
2001.
[24] D. Plonka. Flowscan: A network traffic flow reporting
and visualization tool. In Proceedings of Usenix LISA,
Dec. 2000.
[25] A. Shaikh, M. Goyal, A. Greenberg, R. Rajan, and
K. Ramakrishnan. An OSPF topology server: Design
and evaluation. IEEE Journal on Selected Areas in
Communications, 20(4), May 2002.
[26] A. Soule, A. Lakhina, N. Taft, K. Papagiannaki,
K. Salamatian, A. Nucci, M. Crovella, and C. Diot.
Traffic matrices: Balancing measurements, inference
and modeling. In Proceedings of ACM Sigmetrics,
June 2005.
[27] A. Soule, A. Nucci, E. Leonardi, R. Cruz, and N. Taft.
How to identify and estimate top largest OD flows in
a dynamic environment. In Proceedings of ACM
Sigmetrics, June 2004.
[28] K. Suh, Y. Guo, J. Kurose, and D. Towsley. Locating
network monitors: Complexity, heuristics and
coverage. In Proceedings of IEEE Infocom, Mar. 2005.
[29] Y. Zhang, M. Roughan, N. Duffield, and
A. Greenberg. Fast accurate computation of
large-scale IP traffic matrices from link loads. In
Proceedings of ACM Sigmetrics, June 2003.

!!"#" $

!"#$"% &'()* +,(- .!-/'01 2,!34

tel-00408745, version 1 - 3 Aug 2009

!"#$ %"&"'"() *$"+,-." /"++"..0+1) !&$2(03!1 4$0(

!!"#$"% &' !$()""%&'*+ (, -(."/01 2(34(3+"1 5)0(6"$ 7889:

tel-00408745, version 1 - 3 Aug 2009

tel-00408745, version 1 - 3 Aug 2009

Ranking Flows from Sampled Traffic

Chadi Barakat

Gianluca Iannaccone, Christophe Diot

INRIA - Planete
` group
Sophia Antipolis, France

Intel Research
Cambridge, UK

Chadi.Barakat@inria.fr

Gianluca.Iannaccone@intel.com,
Christophe.Diot@intel.com

ABSTRACT

1. INTRODUCTION

Most of the theoretical work on sampling has addressed the

inversion of general traffic properties such as flow size distribution, average flow size, or total number of flows. In
this paper, we make a step towards understanding the impact of packet sampling on individual flow properties. We
study how to detect and rank the largest flows on a link. To
this end, we develop an analytical model that we validate
on real traces from two networks. First we study a blind
ranking method where only the number of sampled packets
from each flow is known. Then, we propose a new method,
protocol-aware ranking, where we make use of the packet
sequence number (when available in transport header) to
infer the number of non-sampled packets from a flow, and
hence to improve the ranking. Surprisingly, our analytical
and experimental results indicate that a high sampling rate
(10% and even more depending on the number of top flows
to be ranked) is required for a correct blind ranking of the
largest flows. The sampling rate can be reduced by an order
of magnitude if one just aims at detecting these flows or by
using the protocol-aware method.

The list of the top users or applications is one of the most

useful statistics to be extracted from network traffic.
Network operators use the knowledge of the most popular
destinations to identify emerging markets and applications
or to locate where to setup new Points of Presence. Content delivery networks use the popularity of sites to define
caching and replication strategies. In traffic engineering, the
identification of heavy hitters in the network can be used to
treat and route them differently across the network [20, 17,
10]. Keeping track of the network prefixes that generate
most traffic is also of great importance for anomaly detection. A variation in the pattern of the most common applications may be used as a warning sign and trigger careful
inspection of the packet streams.
However, the ability to identify the top users in a packet
stream is limited by the network monitoring technology.
Capturing and processing all packets on high speed links still
remains a challenge for todays network equipment [16, 9].
In this context, a common solution is to sample the packet
stream to reduce the load on the monitoring system and to
simplify the task of sorting the list of items. The underlying
assumption in this approach is that the sampling process
does not alter the properties of the data distribution.
Sampled traffic data is then used to infer properties of the
original data (this operation is called inversion). The inversion of sampled traffic is, however, an error-prone procedure
that often requires a deep study of the data distribution to
evaluate how the sampling rate impacts the accuracy of the
metric of interest. Although the inversion may be simple
for aggregate link statistics (e.g., to estimate the number
of packets transmitted on a link, it is usually sufficient to
multiply the number of sampled packets by the inverse of
the sampling rate), it is much harder for the properties of
individual connections or flows [9, 11, 8].
For these reasons, in this paper, we address this simple,
and so far unanswered, question: which sampling rate is
needed to correctly detect and rank the flows that carry the
most packets?
We define the problem as follows. Consider a traffic monitor that samples packets independently of each other with
probability p (random sampling) and classifies them into
sampled flows. At the end of the measurement period, the
monitor processes the list of sampled flows, ranks them
based on their size in packets, and returns an ordered list of
the t largest flows.
We are interested in knowing (i) whether the ordered list
contains all the actual largest flows in the original packet

Categories and Subject Descriptors

C.4 [Computer Systems Organization]: PERFORMANCE
OF SYSTEMSMeasurement techniques

General Terms
Measurement, Performance, Experimentation

Keywords
Packet sampling, largest flow detection and ranking, performance evaluation, validation with real traces

This work was done while the author was visiting Intel
Research Cambridge.

Permission to make digital or hard copies of all or part of this work for
personal or classroom use is granted without fee provided that copies are
not made or distributed for profit or commercial advantage and that copies
bear this notice and the full citation on the first page. To copy otherwise, to
republish, to post on servers or to redistribute to lists, requires prior specific
permission and/or a fee.
CoNEXT05, October 2427, 2005, Toulouse, France.
Copyright 2005 ACM 1-59593-097-X/05/0010 ...$5.00.

tel-00408745, version 1 - 3 Aug 2009

stream (detection), and (ii) if the items in the list appear in

the correct order (ranking).
We build an analytical model and define a performance
metric that evaluates the accuracy of identification and ranking of the largest flows. We consider a flow to consist of a
single TCP connection. However, our results are general
and can be applied to alternative definitions of flow, as well.
We evaluate two approaches to sort the list of flows:
(i) Blind, where the sampled flows are ranked just based
on their sampled size. This method can be applied to any
definition of flow.
(ii) Protocol-aware, where we make use of additional information in the packet header (e.g., the sequence number
in TCP packets) to infer the number of non-sampled packets
between sampled ones. This method can only be applied to
flow definitions that preserve the protocol level details.
The contributions of this work are the following: (1) We
perform an analytical study of the problem of ranking two
sampled flows and compute the probability that they are
misranked. We propose a Gaussian approximation to make
the problem numerically tractable. (2) We introduce the
protocol-aware ranking method that uses protocol level information to complement the flow statistics and render the
detection and ranking of the largest flows more accurate. (3)
Based on the model for the ranking of two flows, we propose
a general model to study the detection and ranking problem,
given a generic flow size distribution. We define a performance metric and evaluate the impact of several metrics
parameter on the accuracy of the ranking. (4) We validate
our findings on measurement data using publicly-available
packet-level traces. Our results indicate that a surprisingly
high sampling rate is required to obtain a good accuracy
with the blind approach (10% and even more depending on
the number of flows of interest). As for the protocol-aware
approach, it allows to reduce the required sampling rate by
an order of magnitude compared to the blind approach.
The paper is structured as follows. Next, we discuss the
related literature. In Section 3 and 4, we present our model.
Section 5 analyzes the model numerically and Section 6 validates it on real packet-level traces. Section 7 concludes the
paper and provides perspectives for our future research.

2.

RELATED WORK

The inversion of sampled traffic has been extensively studied in the literature. The main focus has been on the inversion of aggregate flow properties such as flow size distribution [9, 11], average flow size or total number of flows [8] on
a given network link. Duffield et al. [8] study the problem of
flow splitting and propose estimators for the total number
of flows and for the average flow size in the original traffic
stream. [9, 11] study the inversion of the flow size distribution with two different methods. They both show that the
major difficulty comes from the number of flows that are not
sampled at all and that need to be estimated with an auxiliary method. As an auxiliary method, [8, 9] propose the use
of the SYN flag in the TCP header to mark the beginning of
a flow. [9] shows that periodic and random sampling provide
roughly the same result on high speed links, and so random
sampling can be used for mathematical analysis due to its
appealing features. [4] finds the sampling rate that assures
a bounded error on the estimation of the size of flows contributing to more than some predefined percentage of the
traffic volume. [14] studies whether the number of sampled

packets is a good estimator for the detection of large flows

without considering its impact on the flow ranking.
Given the potential applications of finding the list of top
users, it does not come as a surprise that there has been a
significant effort in the research community to find ways to
track frequent items in a data stream [5, 7, 3, 10]. However,
this problem has usually been addressed from a memory requirement standpoint. All the works in the literature assume
that if the algorithm and the memory size is well chosen, the
largest flows can be detected and ranked with a high precision. However, in the presence of packet sampling, even if
the methods rank correctly the set of sampled flows, there
is no guarantee that the sampled rank corresponds to the
original rank. The problem we address in this paper complements these works as it focuses on the impact of sampling
on the flow ranking.

3. BASIC MODEL: RANKING TWO FLOWS

In this section, we study the probability to misrank two
flows of original sizes S1 and S2 in packets. This probability
is the basis for the general model for detecting and ranking
the largest flows that we will present later. Indeed, the
detection and ranking of the largest flows can be transformed
into a problem of ranking over a set of flow pairs.
Without loss of generality, we assume S1 < S2 . We consider a random sampling of rate p. Let s1 and s2 denote
the sizes in packets of both flows after sampling. The two
sampled flows are misranked if (i) s1 is larger than s2 , or
(ii) both flows are not sampled, i.e., their sampled sizes
equal to zero. By combining (i) and (ii), one can see that
the necessary condition for a good ranking is to sample at
least one packet from the larger flow (i.e., the smaller of
the two flows can disappear after sampling). The probability to misrank the two flows can then be written as
Pm (S1 , S2 ) = P {s1 s2 }. For the case S1 = S2 , we consider the two flows as misranked if s1 6= s2 , or if both flows
are not sampled at all, i.e. s1 = s2 = 0.
We compute and study the misranking probability of two
flows of given sizes in the rest of this section. First, we consider the blind ranking method where only the number of
sampled packets from a flow is known. For this method,
we express the misranking probability as a double sum of
binomials, then we present a Gaussian approximation to
make the problem tractable numerically. Second, we consider the protocol-aware ranking method for which we calculate a numerical-tractable closed-form expression of the
misraking probability. Note that the misranking probability
is a symmetric function, i.e., Pm (S1 , S2 ) = Pm (S2 , S1 ).

3.1

Blind ranking

With this method, s1 and s2 represent the number of

sampled packets from flows S1 and S2 . Under our assumptions, these two variables are distributed according to a binomial distribution of probability p. Hence, we can write
for S1 < S2 ,
Pm (S1 , S2 ) = P {s1 s2 } =

S1
X
i=0

bp (i, S1 )

i
X

bp (j, S2 ). (1)

j=0

bp (i, S) is the probability density function of a binomial distribution of probability p, i.e., the probability of
obtaining i
successes out of S trials. We have bp (i, S) = Si pi (1 p)Si
for i = 0, 1, ..., S, and bp (i, S) = 0 for i < 0 and i > S. The

probability to misrank two flows of equal sizes is given by

P {s1 6= s2 or s1 = s2 = 0} = 1 P {s1 = s2 6= 0}
P 1 2
=1 S
i=1 bp (i, S1 ).
Unfortunately, the above expression for the misranking
probability is numerically untractable since it involves two
sums of binomials. For large flows of order S packets, the
number of operations required to compute such a probability
is on the order of O(S 3 ), assuming that the complexity of
the binomial computation is on the order of O(S). The
problem becomes much more complex if one has to sum
over all possible flow sizes (i.e., O(S 5 )). For this reason, we
propose next a Gaussian approximation to the problem of
blind ranking that is accurate and easy to compute. We use
this approximation to study the ranking performance as a
function of the sampling rate and the flow sizes.

tel-00408745, version 1 - 3 Aug 2009

3.1.1 Gaussian approximation to blind ranking

Consider a flow made of S packets and sampled at rate
p. The sampled size follows a binomial distribution. However, it is well known that the binomial distribution can be
approximated by a Normal (or Gaussian) distribution when
p is small and when the product pS is on the order of one
(flows for which, on average, at least few packets are sampled) [21, pages 108109]. We assume that this is the case
for the largest flows, and we consider the sampled size of
a flow as distributed according to a Normal distribution of
average pS and of variance p(1 p)S. Using this approximation, one can express the misranking probability for the
blind ranking problem in the following simple form.
Proposition 1. For any two flows of sizes S1 and S2
packets (S1 6= S2 ), the Gaussian approximation gives,

!
|S2 S1 |
1
Pm (S1 , S2 ) ' erf c p
,
(2)
2
2(1/p 1)(S1 + S2 )
R
2
where erfc(x) = ( 2 ) x eu du is the complementary error cumulative function.

Proof: Consider two flows of sizes S1 and S2 in packets

such that S1 < S2 . Their sampled versions s1 and s2 both
follow Normal distributions of averages pS1 and pS2 , and
of variances p(1 p)S1 and p(1 p)S2 . We know that the
sum of two Normal variables is a Normal variable. So the
difference s1 s2 follows a Normal distribution of average
p(S1 S2 ) and of variance p(1 p)(S1 + S2 ). We have then
this approximation for the misranking probability:
Pm (S1 , S2 ) =
'
=

P {s1 s2 0}
)
(
p(S2 S1 )
P V > p
p(1 p)(S1 + S2 )
!

1
S2 S1
. (3)
erfc p
2
2(1/p 1)(S1 + S2 )

V is a standard Normal random variable. Given the symmetry of the misranking probability, one can take the absolute
value of S2 S1 in (3) and get the expression stated in the
proposition, which is valid for all S1 and S2 .

For S1 = S2 , one can safely approximate the misranking

probability to be equal to 1. This approximation is however
of little importance given the very low probability of having
two flows of equal sizes, especially when they are large.

3.2

Protocol-aware ranking

Packets can carry in their transport header an increasing

sequence number. A typical example is the byte sequence
number in the TCP header. Another example could be the
sequence number in the header of the Real Time Protocol
(RTP) [19]. One can use this sequence number, when available, to infer the number of non-sampled packets (or bytes
in the case of TCP) between sampled ones, and hence to improve the accuracy of ranking. The size of the sampled flow
in this case is no longer the number of packets collected, but
rather the number of packets that exist between the first and
last sampled packets from the flow. Although this solution
is limited to flows whose packet carry a sequence number, we
believe that the study of this ranking method is important
given the widespread use of the TCP protocol. Our objective
is to understand how the use of protocol-level information
can supplement the simple, and more general, blind method
and if it is worth the additional overhead it introduces (i.e.,
storing two sequence numbers per flow record).
In the following, we calculate the misranking probability
of two flows of given sizes when using the protocol-aware
method. This probability will be used later in the general
ranking problem. The main contribution of this section is a
closed-form expression for the misranking probability that
is numerically tractable, without the need for any approximation.
Let S be the size of a flow in packets. Let sb , sb =
1, 2, ..., S, denote the (packet) sequence number carried by
the first sampled packet, and let se , se = S, S 1, ..., sb ,
denote the sequence number carried by the last sampled
packet. Given sb and se , one can estimate the size of the
sampled flow in packets to s = se sb + 1. The error in
this estimation comes from the non-sampled packets that
are transmitted before sb and after se . We give next the
distribution of s, which is needed for the computation of
the misranking probability, then we state our main result.
Before presenting the analysis, note that this new flow size
estimator only counts the packets that are transmitted with
distinct sequence numbers. In the case of TCP, this corresponds to the number of bytes received at the application
layer, rather then the number of bytes carried over the network. It is equivalent to assuming that the probability of
sampling a retransmitted (or duplicated) packet is negligible. This is a reasonable assumption if the loss rate is low.
We will address this aspect in more detail in Section 6.
Consider a flow of size S 2 in packets. Using the above
definition for s, the sampled flow has a size of i packets,
i 2, with probability:
P {s = i} =

Si+1
X
k=1

P {sb = k} P {se = k + i 1} .

We have P {sb = k} = (1 p)k1 p, and P {se = k + i 1} =

(1 p)Ski+1 p. This gives
P {s = i}

Si+1
X
k=1

(1 p)k1 p(1 p)Ski+1 p

p2 (1 p)Si (S i + 1).

(4)

As for i = 0, we have P {s = 0} = (1 p)S for S 1. And

for i = 1, we have P {s = 1} = p(1 p)S1 S for S 1. It
is easy to prove that the cumulative distribution of s is the

following for all values of S:

P {s i 6= 0} = p(1 p)Si (S i + 1) + (1 p)Si+1 . (5)
We come now to the misranking probability, which we recall
is a symmetric function. For S1 < S2 , we have
Pm (S1 , S2 ) = P {s2 s1 } =

S1
X
i=0

P {s1 = i}

i
X

j=0

P {s2 = j} .
(6)

And for S1 = S2 , we have

Pm (S1 , S2 ) = 1

S1
X
i=1

P {s1 = i}2 .

(7)

Our main result is the following.

Proposition 2. For S1 < S2 , the misranking probability
is equal to
Pm (S1 , S2 )

=
+
+

(1 p)S1 (1 p)S2

p(1 p)S1 1 S1 [p(1 p)S2 1 S2 + (1 p)S2 ]

p3

2 F (1 p, 1 p)
F (1 p, 1 p)
+ p2
,
xy
x

tel-00408745, version 1 - 3 Aug 2009

where
F (x, y)

xy S2 S1 +1 + ... + xS1 1 y S2 1

xy S2 S1 +1 (1 (xy)S1 1 )/(1 xy).

For S1 = S2 = S, the misranking probability is equal to

Pm (S, S) = 1 p2 (1 p)2(S1) S 2 p4

2 G(1 p, 1 p)
,
xy

where
G(x, y) = xy + x2 y 2 + xS1 y S1 = (xy (xy)S )/(1 xy).
Proof: One can validate the results by plugging (4) and (5)
into (6) and (7).

Note that the main gain of writing the misraking probability in such a condensed form is a complexity that drops
from O(S 3 ) in (6) to O(S) in our final result. This gain
comes from the closed-form expression for the cumulative
distribution in (5), and from introducing the two functions
F (x, y) and G(x, y). These two latter functions transform
two series whose complexity is O(S 2 ) into a closed-form expression whose complexity is O(S).
We solve the derivatives in the above equations using the
symbolic toolbox of matlab, which gives explicit expressions
for the misranking probability. These expressions are simple
to compute, but span on multiple lines, so we omit them for
lack of space.

3.3

Analysis of the misranking probability

3.3.1 The blind case

We use the Gaussian approximation to study how the misranking probability varies with the sampling rate and with
the sizes of both flows, in particular their difference. The
study of the impact of the flow sizes is important to understand the relation between flow size distribution and ranking
of the largest flows.
The misranking probability is a decreasing function of the
sampling rate. It moves to zero when p moves to 1 and to 0.5

when p approaches zero1 . Therefore, there exists one sampling rate that leads to some desired misranking probability,
and any lower sampling rate results in larger error.
We study now how the misranking probability varies with
the sizes of both flows. Take S1 = S2 k, k a positive
integer. From (2) and for fixed k, the misranking probability
increases with S1 and S2 (erfc(x) is an increasing function in
x). This indicates that it is more difficult to rank correctly
two flows different by k packets as their sizes increase in
absolute terms. The result is different if we take the size of
one flow equal to < 1 times
the size of the second,
i.e.,
S1 =
S2 . Here, (S1 S2 )/ S1 + S2 is equal to S1 (1
)/ 1 + , which increases with S1 . Hence, the misranking
probability given in (2) decreases when S1 increases. We
conclude that, when the two flow sizes maintain the same
proportion, it is easier to obtain a correct ranking when they
are large in absolute terms.
We can now generalize the result above. One may think
that the larger the flows, the better the ranking of their
sampled versions. Our last two examples indicate that this
is not always the case. The ranking accuracy depends on
the relative difference of the flow sizes. In general, to have
a better ranking, the difference between the two flow sizes
must increase with the flow sizes and the increase must be
larger than a certain threshold. This threshold is given by
(2): the difference must increase at least as the square root
of the flow sizes. This is an interesting finding. In the context of the general ranking problem, it can be interpreted
as follows. Suppose that the flow size has a cumulative distribution function y = F (x). As we move to the tail of the
distribution2 , the size of the flows to be ranked increases.
The ranking performance improves
if the difference between
flow sizes increases faster than x. This is equivalentto
saying that dx/dy should increase with x faster than x.
All common distributions satisfy this condition, at least at
their tails. For example, with the exponential distribution
we have dx/dy ex (1/ is the average), while for the
Pareto distribution we have dx/dy x+1 ( is the shape).

3.3.2 The protocol-aware case

The first difference with the blind case is in the estimation
error (S s = sb 1 + S se ), which can be safely assumed
to be independent of the flow size for large flows (only dependent on p). This means that if two large flows keep the
same distance between them while their sizes increase, their
ranking maintains the same accuracy. Their ranking improves if the difference between their sizes increases as well,
and it deteriorates if the difference between their sizes decreases. So in contrast to the blind case, the threshold for
the ranking here to improve is that the larger flow should
have its size increasing a little faster than the smaller one. In
the context of the general ranking problem where flow sizes
are distributed according to a cumulative distribution function y = F (x), and when the top flows become larger, the
protocol-aware ranking improves if the derivative dx/dy increases with x. This is equivalent to saying that the function
F (x) should be concave, which is satisfied by most common
distributions at their tail. For blind ranking, concavity was
1
The Gaussian approximation does not account for the case
p = 0 where the misranking probability should be equal to
1 based on our definition.
2
Because we are more and more focusing on large flows or
because the number of available flows for ranking increases.

not enough to obtain a better

ranking; the derivative dx/dy
had to increase faster than x. So in conclusion, the condition to have a better ranking when we move to the tail of the
flow size distribution is less strict with the protocol-aware
method, which is an indication of its good performance.
The second difference with the blind case is in the relation between the ranking accuracy and the sampling rate.
Consider two large flows of sizes S1 and S2 in packets, and
let s1 and s2 denote their sampled sizes. The coefficient of
variation of the difference s2 s1 is an indication on how
well the ranking performs (a small coefficient of variation
results in better ranking3 ). It is easy to prove that this coefficient of variation scales as 1/p for protocol-aware ranking

and as 1/ p for blind ranking. This is again an important

finding. It tells that when the sampling rate is very small,
blind ranking could (asymptotically) perform better than
protocol-aware ranking. Our numerical and experimental
results will confirm this finding.

tel-00408745, version 1 - 3 Aug 2009

4.

GENERAL MODEL: DETECTING AND

RANKING THE LARGEST FLOWS

We generalize the previous model from the ranking of

two flows to the detection and ranking of the top t flows,
t = 1, 2, . . . , N . The misranking probability Pm (S1 , S2 ) previously calculated is the basis for this generalization. Let
N t denote the total number of flows available in the measurement period before sampling. We want the sampled list
of top t flows to match the list of top t flows in the original
traffic. Two criteria are considered to decide whether this
match is accurate. First, we require the two lists to be identical. This corresponds to the ranking problem. The second,
less constrained, criterion requires the two lists to contain
the same flows regardless of their relative order within the
list. This corresponds to the detection problem. For both
problems, the quality of the result is expressed as a function
of the sampling rate p, the flow size distribution, the number
of flows to rank t, and the total number of flows N .

4.1

Performance metric

In order to evaluate the accuracy of detection and ranking, we need to define a performance metric that is easy
to compute and that focuses on the largest flows. A flow
at the top of the list can be misranked with a neighboring
large flow or a distant small flow. We want our metric to
differentiate between these two cases and to penalize more
the latter one; a top-10 flow replaced by the 100-th flow in
the sampled top list is worse than the top-10 flow being replaced by the 11-th flow. We also want our metric to be zero
when the detection and ranking of the top flows are correct.
We introduce our performance metric using the ranking
problem. The performance metric for the detection problem is a straightforward extension. Lets form all flow pairs
where the first element of a pair is a flow in the top t and
the second element is anywhere in the sorted list of the
N original flows. The number of these pairs is equal to
N 1 + N 2 + + N t = (2N t 1)t/2. We then count
the pairs in this set that are misranked after sampling and
we take the sum as our metric for ranking accuracy. This
3
For S1 < S2 , we are interested in P {s1 s2 }. According
to Tchebychev inequality, this probability can be supposed
to behave like VAR[s1 s2 ]/E [s1 s2 ]2 , which is the square
of the coefficient of variation.

sum indicates how good the ranking is at the top of the list.
It is equal to zero when the ranking is correct. When the
ranking is not correct, it takes a value proportional to the
original rank of the flows that have taken a slot in the top-t
list. For example, if the top flow is replaced by its immediate
successor in the list, the metric will return a ranking error
of 1. Instead, if the same flow is replaced by a distant flow,
say the 100-th, the metric will return an error of 99. Also,
note that our metric does not account for any misranking of
flows outside the list of top t flows. For any two flows n and
m, such that n > m > t, the fact that n takes the position
of m does not add anything to our performance metric since
our metric requires at least one element of a flow pair to be
in the original list of top t flows.
In the detection problem, we are no longer interested in
comparing flow pairs whose both elements are in the top t
list. We are only interested in the ranking between flows
in the top t list and those outside the list. Therefore, our
detection metric is defined as the number of misranked flow
pairs, where the first element of a pair is in the list of top t
flows and the second element is outside this list (non top t).
The above metrics return one value for each realization
of flow sizes and of sampled packets. Given that we want
to account for all realizations, we define the performance
metrics as the number of misranked flow pairs averaged over
all possible values of flow sizes in the original list of N flows
and over all sampling runs. We deem the ranking/detection
as acceptable when our metric takes a value below one (i.e.,
on average less than one flow pair is misranked).
In addition to the above, our metrics have the advantage
of being easily and exactly calculable. Performance metrics
based on probabilities (e.g.,[12]) require lot of assumptions
that make them only suitable for computing bounds, but
not exact values.

4.2

Computation of the performance metric

for the ranking problem

Consider a flow of i packets belonging to the list of top t

flows in the original traffic (before sampling). First, we compute the probability that this flow is misranked with another
flow of general size and general position. Denote this probability by Pmt (i), where m stands for misranking and t for
top. Then, we average over all values of i to get Pmt 4 . This
latter function gives us the probability that, on average, the
top t-th flow is misranked with another flow. Thus, our performance metric, which is defined as the average number of
misranked flow pairs where at least one element of a pair
is in the top t, is equal to (2N t 1)tPmt /2. Next, we
compute the value of Pmt .
Let pi denote the probability that the size of a general
flow is equal to i packets, and Pi denote the flow
P size complementary cumulative distribution, i.e., Pi =
j=i pj . For
a large number of flows N and a high degree of multiplexing,
we consider safe to assume that flow sizes are independent
of each other (see [2] for a study of the flow size correlation
on a OC-12 IP backbone link). A flow of size i belongs to
the list of top t flows if the number of flows in the original
total list, with a size larger than i, is less or equal than t 1.
Since each flow can be larger than i with probability Pi independently of the other flows, we can write the probability
that a flow of size i belongs to the list of the top t flows
4
Note that the distribution of the size of a flow at the top
of the list is different from that of a generic flow.

Pt1
as Pt (i, t, N ) =
k=0 bPi (k, N 1), where bPi (k, N 1)
is the probability to obtain k successes out of N 1 trials, Pi being the probability of a success. The probability
that the t-th largest flow has a size of i packets is equal
to Pt (i) = pi Pt (i, t, N )/Pt (t, N ). Pt (t, N ) is the probability
that a flow of general size is among the top t in the original
total list, which is simply equal to t/N .
Using the above notation, one can write the misranking
probability between a top t flow of original size i packets
and any other flow as follows
i1
X
1
pj Pt (i, t, N 1)Pm (j, i)+
Pmt (i) =
Pt (i, t, N ) j=1
!

X
pj Pt (i, t 1, N 1)Pm (i, j) . (8)

tel-00408745, version 1 - 3 Aug 2009

j=i

Trace
Link speed
Duration
TCP connections
Packets

Jussieu
GigE (1 Gbps)
2 hours
11M
112M

Abilene
OC-48 (2.5 Gbps)
30 minutes
15M
125M

Table 1: Summary of the traces

P
i is PiP
=
k=i pk . The probability that it is larger than j is

Pj = k=j pk . The probability that a flow size is between

j and i given that it is smaller than i is (Pj Pi )/(1 Pi ).
We call it Pj,i . It follows that:
Pt? (j, i, t, N ) =

t1
X

k=0

bPi (k, N 2)

NX
k2

l=tk1

bPj,i (l, N k 2).

In this expression, we sum over all possible original sizes

of the other flow (the variable j) and we separate the case
when this other flow is smaller than i from the case when it
is larger than i 5 . Pm (i, j) is the misranking probability of
two flows of sizes i and j packets, which we calculated in the
previous P
section for the two ranking methods. Pmt is then
equal to
i=1 Pt (i)Pmt (i).
For protocol-aware ranking, Pm (i, j) is given explicitly
in Proposition 2 and can be easily computed. For blind
ranking, we use the Gaussian approximation summarized in
Proposition 2, which we recall holds when at least one of the
two flows to be compared is large.

The first sum accounts for the probability to see less than t
flows above i packets. The second sum accounts for the
probability to see more than t flows above j given that
k flows (k < t) were already seen above i. For t = 1,
?
Pt? (j, i, t, N ) is no other than Pt (i, t, N 1), and both Pmt
and Pmt are equal (i.e., the ranking and the detection problems are the same).
?
Once Pmt
is computed, we multiply it by the total number
of flow pairs whose one element is in the top t and the other
one is not. This total number is equal to t(N t). Our metric
for the detection problem is the result of this multiplication.
As for the ranking problem, we want this metric to be less
than one for the detection of the top t flows to be accurate.

4.3

5. NUMERICAL RESULTS

Computation of the performance metric

for the detection problem

Consider the probability that a flow among the top t is

swapped with a flow that does belong to the top t. Let
?
Pmt
denote this probability. Following the same approach
described in Section 4, we can write
i1
1 XX
?
= ?
pi pj Pt? (j, i, t, N )Pm (j, i).
Pmt
Pt i=1 j=1
?
To get this expression for Pmt
, we sum over all possible
values for the size of the flow in the top t (index i) and
all possible values for the size of the other flow not among
the top t (index j). In this expression, pi and pj represent
the probability that the size of a flow is equal to i or j
packets, respectively. Pm (j, i) is the probability that two
flows of sizes i and j are misranked it is given by the
Gaussian approximation described in Proposition 1 for the
blind method and the result stated in Proposition 2 for the
protocol-aware method. Pt? (j, i, t, N ) is the joint probability
that a flow of size i belongs to the list of the top t flows while
another flow of size j does not belong to it (i.e., it is in the
bottom N t flows). Pt? is the joint probability that a flow
of any size belongs to the list of the top t flows while another
flow of any size does not belong to this list. It is equal to
t(N t)/(N (N 1)).
We now compute Pt? (j, i, t, N ) for j < i, i.e., the probability that flow i belongs to the top list while flow j does
not. The number of flows larger than i should be smaller
than t, while the number of flows larger than j should be
larger than t. The probability that a flow size is larger than
5
In the case j i, at most t 2 flows can be larger than i
packets if we want the flow of size i to be in the top t.

We analyze now the accuracy of identifying and ranking

the largest flows in a packet stream for both the blind and
protocol-aware methods. Our metrics require the following
input: pi , the flow size distribution and N , the total number
of flows observed on the link during the measurement period.
To derive realistic values for these two quantities, we consider two publicly available packet-level traces. The first
trace is Abilene-I collected by NLANR [15] on an OC-48
(2.5 Gbps) link on the Abilene Network [1]. The second
trace has been collected by the Metropolis project [13] on
a Gigabit Ethernet access link from the Jussieu University
campus in Paris to the Renater Network [18]. Table 1 summarizes the characteristics of the two traces.
We model the flow size distribution in the traces with
Pareto. We opted for Pareto since it is known to be appropriate to model flow sizes in the Internet due to its heavy
tailed feature [6]. Note that it is not our goal to find an accurate approximation of the distribution of flow sizes in our
traces, but rather to find a general, well-known, distribution
that approaches the actual flow size. In this section we analyze a wide range of parameters while Section 6 focuses on
the performance we observe in the two packet-level traces.
The Pareto distribution is continuous with a complementary cumulative distribution function given by P {S > x} =
(x/a) . > 0 is a parameter describing the shape of the
distribution and a > 0 is a parameter describing its scale.
The Pareto random variable takes values larger than a, and
has an average value equal to a/( 1). The tail of the
Pareto distribution becomes heavier as decreases.
We use our traces to derive an indicative value of the
shape parameter . To this end, we compute the empirical
complementary cumulative distribution of flow sizes and we

10

10

Abilene trace
x2

10

Abilene trace, N=487K, beta=2, blind ranking

Jussieu trace
1.5
x

10

t=25
t=10
t=5
t=2
t=1

10
2
2

10

10

10

10

10

10

10

10

10

10

10
0

10

10

10
10
Flow size in Kbytes

10

10

10

10

10
10
Flow size in Kbytes

10

10

Average number of misranked flow pairs

10
Complementary CDF

Complementary CDF

10

10

10

10

10

Figure 1: Empirical flow size distribution

tel-00408745, version 1 - 3 Aug 2009

10

5.1

Blind ranking

5.1.1 Impact of the number of flows of interest

The first parameter we study is t, the number of largest
flows to rank. The purpose is to show how many flows can
be detected and ranked correctly for a given sampling rate.
We set , N , and the average flow size to the values described before. The performance of blind ranking the top t
flows is shown in Figure 2 for both traces. We observe that
the larger the number of top flows of interest, the more difficult it is to detect and rank them correctly. In particular,
with a sampling rate on the order of 1%, it is possible to
rank at most the top one or two flows. As we focus at larger
values of t, the required sampling rate to get a correct ranking increases well above 10%. Note that with a sampling
rate on the order of 0.1%, it is almost impossible to detect
even the largest flow. We also observe that the ranking on
the Jussieu trace behaves slightly better than that on the
Abilene trace. The Jussieu trace has a heavier tail for its
flow size distribution, and so the probability to get larger
flows at the top of the list is higher, which makes the ranking more accurate. This will be made clear next as we will
study the impact of the shape parameter .

5.1.2 Impact of the flow size distribution

We consider the blind ranking of the top 10 flows varying

10

Jussieu trace, N=103K, beta=1.5, blind ranking

t=25
t=10
t=5
t=2
t=1

10
Average number of misranked flow pairs

plot it on a log-log scale. A heavy-tailed distribution of

shape parameter decays linearly on a log-log scale at rate
. The empirical distributions are shown in Figure 1. The
plots show that equal to 2 suits the Abilene trace and
equal to 1.5 suits the Jussieu one. This means that the flow
size distribution has a heavier tail in the Jussieu trace.
Then, we compute the average flow size in packets to get
the starting point a for the Pareto distribution. As an average flow size we measure 5.76 Kbytes and 7.35 packets on
the Abilene trace, and 9.22 Kbytes and 9.9 packets on the
Jussieu trace. The total number of flows N is set by taking a
measurement interval equal to one minute, then multiplying
this interval by the average arrival rate of flows per second
on each trace. This gives N = 487 Kflows for the Abilene
trace and N = 103 Kflows for the Jussieu one.
In the rest of this section, all figures plot the ranking metric versus the packet sampling rate p on a log-log scale. We
vary p from 0.1% to 50%. Each figure shows different lines
that correspond to different combinations of t, , and N .
We are interested in the regions where the value of the metric is below one, indicating that the ranking is accurate on
average. To ease the interpretation of results in the figures,
we plot the horizontal line of ordinate 1.

10
Packet sampling rate (%)

10

10

10

10

10

10
Packet sampling rate (%)

10

Figure 2: Performance of blind ranking varying the

number t of top flows of interest
the shape parameter for the Pareto distribution among five
distinct values: 3, 2.5, 2, 1.5 and 1.2. Note that for 2
the Pareto distribution is known to be heavy tailed (infinite
variance). The other parameters of the model (N and the
average flow size) are set as before. The values taken by our
metric are shown in Figure 3 for both traces. We can make
the following observations from the figure:
Given a sampling rate, the ranking accuracy improves
as becomes smaller, i.e., the tail of the flow size
distribution becomes heavier. Indeed, when the distribution tail becomes heavier, the probability to obtain
larger flows at the top of the list increases, and since it
is simpler to blindly rank larger flows (for distributions
satisfying the square root condition, see Section 3.1.1),
the ranking becomes more accurate.
The ranking is never correct unless the sampling rate is
very high. In our setting, one needs to sample at more
than 50% to obtain an average number of misranked
flow pairs below one for a value of equal to 1.5 (i.e,
heavy tailed distribution), and at more than 10% for
a value of equal to 1.2 (i.e., pronounced heavy tailed
distribution). For larger values of (i.e., lighter tail),
the sampling rate needs to be as high as 100%.

5.1.3 Impact of the total number of flows

Another important parameter in the ranking problem is
N , the total number of flows available during the measurement period. When N increases, the flows at the top of
the list should become larger, and therefore as we saw in
Section 3.1.1, the blind ranking accuracy should improve

Abilene trace, N = 487K, t = 10 flows, blind ranking

10
beta=3
beta=2.5
beta=2
beta=1.5
beta=1.2

Average number of misranked flow pairs

10

10

10

10

10

10

10

10

10
Packet sampling rate (%)

10

10

Jussieu trace, N = 103K, t = 10 flows, blind ranking

10

10
Packet sampling rate (%)

10

Jussieu trace, beta = 2, t = 10, blind ranking

10

10
beta=3
beta=2.5
beta=2
beta=1.5
beta=1.2

N=52K
N=103K
N=258K
N=515K

10
Average number of misranked flow pairs

10
Average number of misranked flow pairs

10

10

10

10

10

10

10

10

10

10

10

10

tel-00408745, version 1 - 3 Aug 2009

10

10

10

10

N=244K
N=487K
N=1.2M
N=2.4M

10
Average number of misranked flow pairs

10

10

Abilene trace, beta=2, t = 10, blind ranking

10

10

10
Packet sampling rate (%)

10

10

10

10
Packet sampling rate (%)

10

Figure 3: Performance of blind ranking varying the

shape parameter of the flow size distribution

Figure 4: Performance of blind ranking varying the

total number of flows

for flow size distributions satisfying the square root condition (in particular the Pareto distribution we are considering
here). N varies with the utilization of the monitored link
the higher the utilization, the larger the number of flows. N
can also vary with the duration of the measurement period
the longer we wait before ranking and reporting results,
the larger the number of flows.
We study the impact of N on the blind ranking accuracy. We take the same value of N used in the previous
sections and computed over one minute measurement period (487 Kflows for the Abilene trace and 103 Kflows for
the Jussieu trace), then we multiply it by some constant
factor ranging from 0.5 (2 times fewer flows) to 5 (5 times
more flows). Results are shown in Figure 4. The lines in
the figures correspond to a factor value equal to: 0.5, 1,
2.5, and 5. In these figures, we consider the ranking of the
top 10 flows with the values of and average flow size set
from the traces. Clearly, the ranking accuracy improves as
N increases. However, in our setting, this improvement is
still not enough to allow a perfect ranking. One can always
imagine increasing N (e.g., by increasing the measurement
period) until the top t flows are extremely large and hence,
perfectly detected and ranked.

sary to remove the discrepancy between carried data volume

(throughput) and application data volume (goodput).
Using the previous values for N , and average flow size,
we reproduce Figure 2, but this time for the protocol-aware
case. This leads to Figure 5, which illustrates the impact of
the number of largest flows to rank. For lack of space, we
omit the other figures.
We compare this new figure to its counterpart in the blind
case. We make the following two observations:
(i) The protocol-aware method improves the accuracy of
the largest flows ranking by an order of magnitude for high
sampling rates (above 1%). For example, for the Abilene
trace, a sampling rate on the order of 50% was necessary to
detect and rank the largest 5 flows with the blind method.
Now, with the protocol-aware method, a sampling rate on
the order of 5% is sufficient. The same conclusion applies
to the Jussieu trace. A sampling rate on the order of 10%
is needed. With the protocol-aware method, it becomes on
the order of 1%.
(ii) The protocol-aware method does not improve the performance when applied at low sampling rates (above 1%).
This can be clearly seen if we compare the plots between
both figures for sampling rates below 1%. This results confirms our observations in Section 3.3.2.

5.2

Protocol-aware ranking

Protocol-aware ranking takes advantage of the information carried in the transport header of the sampled packets
to infer the number of non-sampled packets of a flow. We
use our model to check whether this improvement exists and
to evaluate it. Remember that we are always in the context
of low retransmission and duplication rates, which is neces-

5.3

Largest flows detection

To illustrate the difference between ranking and detection,

we consider the same scenario as in Section 5.1.1. We plot
the detection metric as a function of the sampling rate for
different values of t (the number of top flows of interest) and
for both Abilene and Jussieu traces. This gives Figure 6 for

Abilene trace, N=487K, beta=2, protocolaware ranking

t=25
t=10
t=5
t=2
t=1

10
Average number of misranked flow pairs

10
Average number of misranked flow pairs

Abilene trace, N=487K, beta=2, blind detection

t=25
t=10
t=5
t=2
t=1

10

10

10

10

10

10

10

10

10

10
Packet sampling rate (%)

10

10

Jussieu trace, N=103K, beta=1.5, protocolaware ranking

Average number of misranked flow pairs

10

10

10

tel-00408745, version 1 - 3 Aug 2009

10

10

10

10

10

10

10
Packet sampling rate (%)

10

Figure 5: Performance of protocol-aware ranking

varying the number t of top flows of interest
blind ranking and Figure 7 for protocol-aware ranking. A
comparison between these results and their counterparts in
Figure 2 and 5, respectively, shows a significant improvement in the detection case for both ranking methods. All
plots are shifted down by an order of magnitude. For example, in the case of blind ranking, the required sampling
rate to correctly rank the top 5 flows was around 50% for
the Abilene trace and 10% for the Jussieu trace. Now, with
blind detection, it is around 10% and 3%, respectively. Another example is with the protocol-aware method where a
sampling rate around 10% was required to rank the largest
10 flows (Figure 5), whereas now, a sampling rate around
1% is sufficient to only detect them. The same gain can be
observed if we reconsider the other scenarios in Section 5.1
(not presented here for lack of space). Also, note how in
the detection case the protocol aware method allows a better accuracy for high sampling rates when compared to the
blind method. For low sampling rates (e.g., below 1%), the
accuracy does not improve.

6.

t=25
t=10
t=5
t=2
t=1

10
Average number of misranked flow pairs

10

Jussieu trace, N=103K, beta=1.5, blind detection

t=25
t=10
t=5
t=2
t=1

10

10
Packet sampling rate (%)

EXPERIMENTAL RESULTS

In this section we present the results of running random

sampling experiments directly on the packet traces. We use
the traces described in Section 5 and compute the performance metrics defined in Section 4.1.
In our traces we consider only TCP packets. Since TCP
sequence numbers count bytes, we express the flow sizes in
bytes instead of packets throughout this section.
Our experiments are meant to address four major issues
that arise when we move from the analytical study to a real
network setting: (i) how to deal with invalid TCP sequence

10

10
Packet sampling rate (%)

10

Figure 6: Only detecting the largest flows: Performance of blind ranking varying the number t of top
flows of interest
numbers in the packet stream; (ii) the importance of flow
size distributions and duration of the measurement interval;
(iii) the impact of packet loss rates on individual flows
lost packets trigger retransmissions by the TCP senders; (iv)
the variability of the detection/ranking performance across
multiple bins and packet sampling patterns.

6.1

Implementation of protocol-aware
ranking

The protocol-aware method depends on TCP sequence

numbers to perform the ranking. For a given flow, it keeps
track of the lowest and highest sequence number observed
(taking care of packets that wrap around the sequence number space), sb and se respectively.
Note that an actual implementation of this method would
just require two 32 bit fields per flow to store the two sequence numbers.
At the end of the measurement period, we compute the
difference between the highest and lowest sequence numbers
for each sampled flow, and we use the obtained values to
rank flows. We then compare this ranking with the one
obtained by counting all the bytes each flow transmits in
the original non sampled traffic.
In order to discard invalid packets carrying incorrect sequence numbers that would corrupt the ranking, we implement a simple heuristic to update se and sb . A sampled packet with sequence number S > se causes an update
se S if S se < ( MTU)/p. The same rule applies to
the updates of sb . This way we set a limit on the maximum
distance in the sequence space between two sampled pack-

Abilene trace, N=487K, beta=2, protocolaware detection

Jussieu trace, blind ranking

10
6

Average number of misranked flow pairs

10

10

10

10

top 1
top 2
top 5
top 10
top 25

10

average number of misranked flow pairs

t=25
t=10
t=5
t=2
t=1

10

10

10

10

10

10

10

10
Packet sampling rate (%)

10

10

10

10

10

packet sampling rate %

Jussieu trace, N=103K, beta=1.5, protocolaware detection

Jussieu trace, protocolaware ranking

10
6

Average number of misranked flow pairs

10

10

10

10

top 1
top 2
top 5
top 10
top 25

10

average number of misranked flow pairs

t=25
t=10
t=5
t=2
t=1

10

10

10

10

10
2

tel-00408745, version 1 - 3 Aug 2009

10

10

10
Packet sampling rate (%)

10

10

10

10

10

packet sampling rate %

Figure 7: Only detecting the largest flows: Performance of protocol-aware ranking varying the number t of top flows of interest

Figure 8: Performance of blind and protocol-aware

ranking on Jussieu trace (60s measurement interval).

ets. This distance is inversely proportional to the sampling

rate and depends on the Maximum Transmission Unit.
Furthermore, we use the parameter that allows to make
this threshold more or less permissive in order to account
for the randomness of the sampling process and for other
transport-layer events (e.g., packet retransmissions when the
TCP window is large). We have run several experiments
with different values of and the results have shown little
sensitivity to values of > 10. All the results in this Section
are derived with = 100.

dom sampling on the packet traces and classify the sampled

packets into flows. At the end of each measurement interval
(set to 1 or 5 minutes), we collect the flows and rank them
by the number of bytes sampled for each flow. We compare
the ranking before and after sampling using our performance
metric (Section 4.1). For each sampling rate we conduct 15
runs and we calculate averages.
The results of the experiments confirm the numerical results of the previous section. In the interest of space, we
plot the results of two representative experiments on which
we make several observations. The difference between numerical and experimental results, especially at low sampling
rates, is caused by the non perfect match of the empirical
flow size distribution with Pareto (Figure 1).
Figure 8 shows the performance of ranking flows on the
Jussieu trace when the measurement bin is 60s. We consider
a wide range of sampling rates from 0.1% to 50% and study
the performance when ranking the top 1, 2, 5, 10 and 25
flows in the packet stream. The top graph in Figure 8 is derived using the blind method while the bottom graph shows
the performance of the protocol-aware methods. These results are very similar to the numerical results. For sampling
rates above 1%, protocol-aware ranking gives approximately
an order of magnitude gain on the performance when compared to blind ranking. When the sampling rate is lower
than 1%, however, the performance of the two methods is
similar. Overall, the blind method requires a sampling rate
of 10% to correctly identify the largest flow in the packet
stream. The same sampling rate allows to correctly rank
the largest 5 flows when using the protocol-aware method.

6.2

Flow size distribution and measurement

interval

As shown in Figure 1, flow size distributions do not follow

a perfect Pareto. Furthermore, the measurement interval
itself plays a major role in shaping the distribution: it caps
the size of the largest flows, that is not unbounded but now
depends on the link speed. Indeed, network operators often
run measurements using a binning method, where packets
are sampled for a time interval, classified into flows, ranked,
and then reported. At the end of the interval, the memory is
cleared and the operation is repeated for the next measurement interval. With this binning method, all flows active at
the end of the measurement interval are truncated, so that
not all sampled packets of the truncated flow are considered at the same time for the ranking. The truncation may,
therefore, penalize large flows and alter the tail of the flow
size distribution (where flows are of large size and probably
last longer than the measurement interval).
Each experiment consists of the following. We run ran-

tel-00408745, version 1 - 3 Aug 2009

Impact of loss rate

The protocol-aware method keeps performing better

than the blind method when the sampling rate is above
1%. At lower sampling rates, the blind method performs better although it presents very large errors.
For sampling rates above 2%, the curve relative to
the detection of the top-25 flows in the protocol-aware
method flattens to a value around 70. This is due
to the presence of a few flows that experience a high
loss rate when compared to other flows. Increasing
the sampling rate does not help the protocol-aware
method in detecting the largest flows when the volume of bytes sent is used to define the flow size. However, the protocol-aware method can correctly detect
the top-25 flows when their size is defined in terms of
transport data (see Figure 10).
In summary, the network operator has to choose the metric of interest that depends on the application. For example, for anomaly detection or traffic engineering, a metric
that counts the number of bytes sent may be more appropriate. Instead, for dimensioning caches and proxies, the
metric that considers the size of the objects transferred may
be preferred. This latter metric suits more the protocolaware method.

6.4

Variability of the results

A last important aspect that we need to address is the

variability of the results across multiple measurement intervals and different realizations of the sampling process.
Indeed, moving from one measurement interval to another,

top 1
top 2
top 5
top 10
top 25

10

average number of misranked flow pairs

In the analysis of the protocol-aware method in Section 3.2,

we made the assumption of negligible number of retransmissions for all the flows in the packet stream.
A retransmitted packet may cause inconsistency between
the blind and protocol-aware method depending on the location of the monitoring point. Indeed, the blind method
counts the total number of bytes sent by the flow while the
protocol-aware method considers only the data sent by the
transport layer. Therefore, if the packet is lost before the
monitoring point, the blind and protocol-aware method will
have a consistent view of the number of bytes sent. Instead,
if the packet is lost after the monitoring point, the blind
method may count this packet twice.
The impact of packet losses on the detection and ranking
of the largest flows depends on the metric used to estimate
the size of the flows. If flow sizes are estimated according to
the total number of bytes sent (i.e., the throughput), then
the protocol-aware method may incur in an underestimation
error that is independent of the sampling rate (it will occur
even if all packets are sampled!). On the other hand, if the
flow sizes are estimated according to the transport data sent
(i.e., the goodput), then the blind method may incur in an
overestimation error independently of the sampling rate.
To illustrate the effect of packet loss rates, we plot in
Figure 9 the performance of detecting the largest flows in the
Abilene trace when the measurement bin is 5 minutes and
the flow sizes are measured using the total number of bytes
sent over the link. The top graph shows the performance
of the blind method, while the bottom graph presents the
results for the protocol-aware method.
We can make the following observations:

Abilene trace, blind detection

10

10

10

10

10

10

10

10

10

10

packet sampling rate %

Abilene trace, protocolaware detection

10

top 1
top 2
top 5
top 10
top 25

10

average number of misranked flow pairs

6.3

10

10

10

10

10

10

10

10

10

packet sampling rate %

Figure 9: Performance of blind (top) and protocolaware (bottom) detection on Abilene trace (300s
measurement interval).
the composition of flows varies and with it the flow size distribution. Moreover, the sampling process may get lucky
in certain cases and provide good results. The opposite is
also possible.
Figure 11 shows the average performance over 15 sampling experiments of the detection of the top-10 flows in
the Abilene trace over the 5-minute measurement intervals.
The error bars indicate the standard deviation across the
15 experiments. As usual, the top graph refers to the blind
method, while the bottom graph presents the protocol-aware
method results.
As we can see the average performance shows limited
variability. A sampling rate of 0.1% gives poor results for
all bins, while increasing the sampling rates consistently
helps. With a sampling rate of 10% the performance metric (i.e., average number of misranked flow pairs) for the
blind method is always below 100 while the protocol-aware
method is always below 1.
Looking at the standard deviation, we observe large values for the blind method and much smaller values for the
protocol-aware method. This indicates that the blind method
is more sensitive to the sampling process than the protocolaware method. The explanation is given in Section 3.3.2
where we showed that that the blind method presents a
larger error for large flow sizes (expect when the sampling
rate is very low).

7. CONCLUSIONS
We study the problem of detection and ranking the largest
flows from a traffic sampled at the packet level. The study is

sampling 0.1%
sampling 1%
sampling 10%

Abilene trace, protocolaware detect

10

average number of misranked flow pairs

10

10

10

10

10

Abilene trace, detection

10

average number of swapped flow pairs

top 1
top 2
top 5
top 10
top 25

10

10

10

10

10

200

400

600

800

1000

1200

1400

1600

1800

time (sec)
0

sampling 0.1%
sampling 1%
sampling 10%

10

Abilene trace, detection

10

10

10

packet sampling rate %

Figure 10: Performance of protocol-aware detection

on Abilene trace (300s measurement interval) when
using actual amount of data sent by the transport
layer application.

10

average number of swapped flow pairs

10

10

10

10

10

tel-00408745, version 1 - 3 Aug 2009

10

done with stochastic tools and real packet-level traces. We

find that the ranking accuracy is strongly dependent on the
sampling rate, the flow size distribution, the total number
of flows and the number of largest flows to be detected and
ranked. By changing all these parameters, we conclude that
ranking the largest flows requires a high sampling rate (10%
and even more). One can reduce the required sampling rate
by only detecting the largest flows without considering their
relative order.
We also introduce a new method for flow ranking that
exploits the information carried in transport header. By
analysis and experimentation, we demonstrate that this new
technique allows to reduce the required sampling rate by an
order of magnitude.
We are currently exploring two possible future directions
for this work. First, we want to study the accuracy of the
ranking when the sampled traffic is fed into one of the mechanisms proposed in [10, 12] for sorting flows with reduced
memory requirements. Second, we are exploring the use of
adaptive schemes that set the sampling rate based on the
characteristics of the observed traffic.

Acknowledgements
We wish to thank NLANR [15], Abilene/Internet2 [1] and
the Metropolis project [13] for making available the packet
traces used in this work.

8.

REFERENCES

[1] Abilene: Advanced networking for leading-edge research and

education. http://abilene.internet2.edu.
[2] C. Barakat, P. Thiran, G. Iannaccone, C. Diot, and
P. Owezarski. Modeling Internet backbone traffic at the flow
level. IEEE Transactions on Signal Processing (Special Issue
on Signal Processing in Networking), 51(8):21112124, Aug.
2003.
[3] M. Charikar, K. Chen, and M. Farach-Colton. Finding frequent
items in data streams. In Proceedings of ICALP, 2002.
[4] B. Y. Choi, J. Park, and Z. Zhang. Adaptive packet sampling
for flow volume measurement. Technical Report TR-02-040,
University of Minnesota, 2002.
[5] G. Cormode and S. Muthukrishnan. Whats hot and whats
not: Tracking most frequent items dynamically. In Proceedings
of ACM PODS, June 2003.
[6] M. Crovella and A. Bestravos. Self-similarity in the World
Wide Web traffic: Evidence and possible causes. IEEE/ACM
Transactions on Networking, 5(6):835846, Dec. 1997.

200

400

600

800

1000

1200

1400

1600

1800

time (sec)

Figure 11: Performance of blind (top) and protocolaware (bottom) detection over multiple 300s intervals (Abilene trace). Vertical bars show the standard deviation over multiple experiments.
[7] E. Demaine, A. Lopez-Ortiz, and I. Munro. Frequency
estimation of internet packet streams with limited space. In
Proceedings of 10th Annual European Symposium on
Algorithms, 2002.
[8] N. G. Duffield, C. Lund, and M. Thorup. Properties and
prediction of flow statistics from sampled packet streams. In
Proceedings of ACM Sigcomm Internet Measurement
Workshop, Nov. 2002.
[9] N. G. Duffield, C. Lund, and M. Thorup. Estimating flow
distributions from sampled flow statistics. In Proceedings of
ACM Sigcomm, Aug. 2003.
[10] C. Estan and G. Varghese. New directions in traffic
measurement and accounting. In Proceedings of ACM
Sigcomm, Aug. 2002.
[11] N. Hohn and D. Veitch. Inverting sampled traffic. In
Proceedings of ACM Sigcomm Internet Measurement
Conference, Oct. 2003.
[12] J. Jedwab, P. Phaal, and B. Pinna. Traffic estimation for the
largest sources on a network, using packet sampling with
limited storage. Technical Report HPL-92-35, HP
Laboratories, Mar. 1992.
[13] Metropolis: METROlogie Pour lInternet et ses services.
http://www.laas.fr/ owe/METROPOLIS/metropolis eng.html.
[14] T. Mori, M. Uchida, R. Kawahara, J. Pan, and S. Goto.
Identifying elephant flows through periodically sampled
packets. In Proceedings of ACM Sigcomm Internet
Measurement Conference, Oct. 2004.
[15] NLANR: National Laboratory for Applied Network Research.
http://www.nlanr.net.
[16] Packet Sampling Working Group. Internet Engineering Task
Force. http://www.ietf.org/html.charters/psamp-charter.html.
[17] K. Papagiannaki, N. Taft, and C. Diot. Impact of flow
dynamics on traffic engineering design principles. In
Proceedings of IEEE Infocom, Hong Kong, China, Mar. 2004.
[18] Renater. http://www.renater.fr.
[19] H. Schulzrinne, S. Casner, R. Frederick, and V. Jacobson.
RTP: A transport protocol for real-time applications. RFC
1889, Jan. 1996.
[20] A. Shaikh, J. Rexford, and K. G. Shin. Load-sensitive routing
of long-lived IP flows. In Proceedings of ACM Sigcomm, Sept.
1999.
[21] M. Spiegel. Theory and Problems of Probability and
Statistics. McGraw-Hill, 1992.

!!"#" $

tel-00408745, version 1 - 3 Aug 2009

!"#$%"#&' ()$*)+!,#- #& % (./"#0%!"

1)!!#2&
!"! #$%&'( )*+!, #$+-!,( ./!0* 1!"!2!+( 3/*$*445 6!*,

!!"#$"% &' !$()""%&'*+ (, -. /01.2340-/5 /#' 6&"*(5 - 5 78'"

9::;< =('* >"$+&(' ?#+ !8@=&+A"% &' 0222 3$#'+#)B&('+ (' /&*'#=
C$()"++&'* D /!")&#= 0++8" (' /&*'#= C$()"++&'* &' E"B?($F&'*5 >(=< GH5 '(<
I5 !!< 9HJGD9HKJ5 8*8+B 9::;<

tel-00408745, version 1 - 3 Aug 2009

Estimating Membership in a Multicast Session

tel-00408745, version 1 - 3 Aug 2009

Sara Alouf

Eitan Altman

Vrije Universiteit
Department of Mathematics
De Boelelaan 1081a
1081 HV Amsterdam, The Netherlands

INRIA
2004 Route des Lucioles
B.P. 93
06902 Sophia Antipolis, France

alouf@cs.vu.nl

altman@sophia.inria.fr

Chadi Barakat

Philippe Nain

INRIA
2004 Route des Lucioles
B.P. 93
06902 Sophia Antipolis, France

INRIA
2004 Route des Lucioles
B.P. 93
06902 Sophia Antipolis, France

cbarakat@sophia.inria.fr

nain@sophia.inria.fr

ABSTRACT

Keywords

We propose two novel on-line estimation algorithms to determine the size of a dynamic multicast group. We first use a
Wiener filter to derive an optimal estimator for the membership size of the session in case the join process is Poisson and
the lifetime of participants is distributed exponentially. We
next develop the best first-order linear filter from which we
derive an estimator that holds for any lifetime distribution.
We apply this approach to the case where the lifetime distribution is hyperexponential. Both estimators hold under
any traffic regime. Applying both estimators on real traces
corresponding to video sessions, we find that both schemes
behave well, one of which performs slightly better than the
other in some cases. We further provide guidelines on how
to tune the parameters involved in both schemes in order to
achieve high quality estimation while simultaneously avoiding feedback implosion.

On-line estimation, multicast applications, membership size,

M/G/ queue, Wiener filter, first-order linear filter

Categories and Subject Descriptors

G.3 [Probability and Statistics]: Queueing theory, Stochastic processes; H.4.3 [Information Systems Applications]:
Communications ApplicationsComputer conferencing, teleconferencing, and videoconferencing

General Terms
Algorithms, Performance, Measurement
This work was done when the author was at Inria Sophia
Antipolis, France.

Permission to make digital or hard copies of all or part of this work for
personal or classroom use is granted without fee provided that copies are
not made or distributed for profit or commercial advantage and that copies
bear this notice and the full citation on the first page. To copy otherwise, to
republish, to post on servers or to redistribute to lists, requires prior specific
permission and/or a fee.
SIGMETRICS03, June 1014, 2003, San Diego, California, USA.
Copyright 2003 ACM 1-58113-664-1/03/0006 ...$5.00.

1.

INTRODUCTION

Since its introduction, IP multicast has seen slow deployment in the Internet. As stated in [6], the service model
and architecture do not efficiently provide or address many
features required for a robust implementation of multicast.
However, the fact remains that IP multicast is very appealing in offering scalable point-to-multipoint delivery specially
in satellite communications. Current research efforts tend to
propose alternatives to IP multicast like the so-called application layer multicast [4, 9, 14, 19], the idea being to
deploy multicast at the application layer. Also, new models to support multicast communications in a more effective
way have been proposed, such as the EXPRESS multicast
[13]. The latter is an extension to IP multicast that provides explicit support for large-scale multicast applications
such as real-time stock quote dissemination, live sports video
feeds or Internet radio and TV. EXPRESS provides as well
a best-effort count of the number of subscribers.
This paper is motivated by the conviction that large-scale
multicast applications will be widely deployed in the future
as soon as the capability becomes available. We believe
that membership estimates will be an essential component
of this widespread deployment as they can be very useful
for scalable multicast. The membership of a session can
be used for feedback suppression as it is the case in current
protocols such as RTP [21] and SRM [8]. In order to regulate
the amount of session/control messages sent by receivers
the idea being not to exceed 5% of overall session bandwidth
these protocols use delay timers that are tuned based on
the membership estimates.
The membership of a multicast session can be used for
charging the sources in large-scale applications. ISPs traditionally charge their customers on an input-rate basis. An
alternative pricing scheme would be to charge sources based

tel-00408745, version 1 - 3 Aug 2009

on their audience size which is more profitable in the case

of millions of subscribers.
Estimating the size of a multicast session can be quite useful to many applications. Bolot, Turletti and Wakeman [3]
use membership estimation to further estimate the proportion of congested receivers as needed in their videoconference system IVS. Future Internet radios and TVs will need
to characterize their audience preferences and to follow the
fluctuations of the audience size. Dutta, Schulzrinne and
Yemini proposed an architecture for Internet radio and TV
called MarconiNet [7] that relies on RTCP [21, 20]. Even
though RTCP provides an easy mechanism for collecting
statistics on the size of the audience, it does not scale well
to large multicast session [7]. In such applications, samplingbased techniques are more appropriate.
There has been a significant research effort in devising
sampling-based schemes for the estimation of the membership in multicast sessions [3, 17, 18, 10, 16, 2]. The feedback
algorithms presented in [3, 17, 18, 10, 16] are all at-least-one
scenarios in the sense that the membership estimation is
based on at least one acknowledgement (ACK) coming from
the receivers. In these probabilistic schemes, the receivers
send ACKs to the source as a reply to a specific request,
either with a certain probability as in [3] or after some random time like in [17, 18, 16]. But what is common to these
schemes (except for [17]) is that they all assume that the
size of the group does not change during the estimation process. In a recent work [2], we propose a dynamic scheme that
tracks the variations of the membership in an optimal way.
The estimation algorithm used is quite simple: the source
requests from its receivers to send ACKs with probability p
every S seconds; it collects the amount of ACKs received at
each observation step and filters out these measurements to
estimate the membership. To derive the optimal estimator
we rely on a diffusion approximation for the heavy-traffic
regime. Under the assumptions of Poisson join times and
exponentially distributed connection times, the diffusion approximation yields linear dynamics which enables the design
of the optimal filter using Kalman filter theory.
In this paper we propose two novel algorithms for estimating the membership based on the polling scheme presented
in [2]. Our purpose is to develop an estimator under more
general assumptions than the ones used in [2]. Our first
approach is based on a Wiener filter, which provides the optimal dynamic estimator among all linear estimators. The
dynamics is not required to be linear as in the case of the
Kalman filter, which allows us to remove the heavy-traffic
assumption made in [2]. Yet, in order to obtain explicit expressions for the parameters of the Wiener filter, we still
have to assume that participants join the session according to a Poisson process, and that the time during which
they stay in the multicast session, hereafter referred to as
on-time, has an exponential distribution. Under these assumptions we design the optimal linear estimation scheme
that turns out to require a filter of order one. Motivated by
this structure, we then design an efficient estimation scheme
for generally distributed on-times. To that end, we identify
the optimal filter among all linear filters of order one. We illustrate this approach in the case of hyperexponentially distributed on-times. Alike the estimator developed with the
Wiener filter, the latter estimator is valid under any traffic regime. Both estimators are then tested on real traces.
Despite the fact that these traces violate the assumptions

under which the estimators have been derived, very good

performance are observed.
Remark 1.1. The material presented in this paper does
not require the specific use of IP multicast, nor any other
multicast protocol. The solutions proposed hereafter are
meant to be deployed at the application layer, and only require a multicast delivery of the requests for ACKs. This
delivery can be achieved either by IP multicast, or by an
application-layer multicast, or even by a new multicast technology. It is however assumed that the IP address of the
source is available to all receivers in the session.
The paper is organized as follows: the mathematical model
of the membership is introduced in Section 2. The theory
of Wiener filters is briefly presented in Section 3 and its application to the M/M/ model comes in Section 3.1. The
optimal first-order linear filter is developed in Section 4. Section 5 proposes some guidelines on how to choose parameters p and S. The robustness of both estimators is addressed
through validations on real traces in Section 6. Finally, open
issues are discussed in Section 7 and concluding remarks are
given in Section 8.

2.

MULTICAST GROUP MODELED AS AN

M/G/ QUEUE

We consider a multicast group that participants join and

leave at random times. Let Ti and Ti + Di be the join time
and leave time, respectively, of the i-th participant. In the
following, Di denotes the on-time of the i-th participant. Let
N (t) be the number of participants in the multicast group
at time t or, equivalently, the size of the multicast session
at time t. Clearly
N (t) =

1(Ti t < Ti + Di )

(1)

i1

where 1(E) equals 1 if the event E occurs and 0 otherwise.

We shall assume that the join times form a homogeneous
Poisson process (with constant intensity 0 < = 1/E[Ti+1
Ti ]) and that the on-times form a renewal sequence of random variables (rvs) with common probability distribution
(x) = P (Di < x), 0 < E[Di ] < , further independent of
the join times. In the following D will denote a generic rv
with probability distribution (x).
In the queueing terminology, {N (t), t 0} represents the
occupation process (number of busy servers) in an M/G/
queue [15].
At times t = nS, n = 0, 1, . . ., with S > 0 a constant,
each participant to the multicast session sends an ACK to
the source with probability 0 < p < 1 and does not send
any feedback information to the source with probability 1
p. We assume that ACKs cannot be lost. However, this
assumption can be relaxed if one knows the loss probability
as it is possible to incorporate it in our feedback mechanism.
The ACK interval S between two consecutive polling instants has to be larger than the largest round-trip time between a receiver and the source, so that all of the ACKs
produced in a round reach the source before the (automatic)
start of the next round. Note that in practice the source will
have to regularly multicast the pair (p, S) to ensure that each
participant will know these values. Throughout the paper, p
and S are held fixed (see Section 5 for possible extensions).

Let Yn be the number of ACKs received by the source at

time nS. Based on the knowledge of Y1 , . . . , Yn , our objective is to find an optimal estimator (in a sense to be defined
n for Nn := N (nS), the size of the multicast group
below) N
at time nS. In filtering parlance, Yn is an input signal and
n that is as close as
we want to generate another signal N
possible to an unknown signal Nn (e.g. by minimizing the
mean square error).
For later use we briefly review some results on the M/G/
queue. In steady-state, the number N of busy servers is
a Poisson random variable with parameter := E[D],
namely, P [N = j] = j exp()/j!. In particular, both
the mean and the variance of the number of busy servers
are equal to . The autocovariance function of the stationary version of the process {N (t), t 0}, also denoted by
{N (t), t 0}, is given by [5, Eqn (5.39)]

|z| = 1. Introduce
Sy (z) =

P (D > u) du.

Covy (k)z k

k=

the z-transform of the autocovariance function (also called

the power spectrum) of {yn }n .
Let Covy (k) = E[nk yn ] be the cross-correlation function of processes {n }n and {yn }n . We also introduce1
Sy (z) =

Covy (k)z k

k=

the z-transform of Covy (k). We can express Covy (k) and

Covy (k) in terms of Cov (k) as follows

Cov(N (t), N (t + h)) =

(2)

Covy (k)

p2 Cov (k) + 1(k = 0)p(1 p)

(6)

Covy (k)

p Cov (k)

(7)

tel-00408745, version 1 - 3 Aug 2009

|h|

In the following we will denote by CovX () the autocovariance function of any second-order discrete-time stationary
process {Xn , n = 0, 1, . . .}. With this notation and the definition of the process {Nn }n , we see from (2) that
CovN (k) = |k| ,

k = 0, 1, . . . ,

(3)

with := exp(S), when the on-times {Di }i are exponentially distributed with mean 1/.
Throughout the paper, we will assume that

CovN (k) < .

(4)

k0

In other words, we will exclude the situation where the ontimes are heavy-tailed (e.g. Pareto distribution).

3.

WIENER FILTER

Our objective is to transform a signal Yn (noisy observan (estimator) that is the closest to
tion) into another signal N
an unknown signal Nn . By closest we mean that the mean
n ] = E[Nn ]) and that the mean square
error is zero (i.e. E[N
error is minimized.
Such a transformation can be achieved by the Wiener filter
that identifies the optimal linear filter [11]. This approach
gives the transfer function of the linear filter, which can be
transformed back to the time domain to obtain the impulse
response of the filter. From the impulse response of the filter,
n as a function of Yn and, possibly, of
the expression of N
n1 , N
n2 , . . ., can be found. We will detail this procedure
N
below.
Since a filter that minimizes the mean square error when
the underlying processes are centered also minimizes the
mean square error when the same processes are non-centered,
we will derive the Wiener filter for the centered (stationary)
n }n and {Yn }n , denoted by
versions of processes {Nn }n , {N
{n }n , {
n }n and {yn }n , respectively. We have observed in
the previous section that E[Nn ] = . On the other hand

where we have used the identity Cov (k) = CovN (k). We

are now in position to derive the Wiener filter. First, we
write Sy (z) as
Sy (z) = G(z)G(z 1 )

(8)

where is a constant. This operation is called the canonical

factorization of the power spectrum of {yn }n . The function
G(z) is the part of Sy (z) that has all its zeros and poles
inside the unit circle. The function 1/G(z) is the transfer
function of the whitening filter: it transforms {yn }n into a
white noise process with variance .
Next, we form the ratio Sy (z)/G(z 1 ). This ratio is interpreted as the transfer function of a linear filter. The
impulse response of this filter has values at the left and the
right of the time origin (non-causal filter). We look for the
transfer function of the part of the impulse response at the
right of the time origin. This can simply be done by expanding the transfer function into fractions and then taking
only the fractions with zeros and poles in the unit circle. In
other words, we transfer the filter from a non-causal filter
into a causal one. We denote the transfer function of the
causal version of the filter by

"

H(z) =

Sy (z)
G(z 1 )

.
+

The transfer function of the optimal filter is given by [11]

Ho (z) =

H(z)
.
G(z)

It remains to invert this transfer function back into the time

domain to find the desired recurrence between n and yn
n
and, subsequently, between the non-centered variables N
and Yn . This procedure is illustrated in Section 3.1 for the
case that the underlying model is the M/M/ queue.

(5)

Application to the M/M/ model

In light of the results reported in Section 3, all what
we have to do is to find expressions for Sy (z) and Sy (z).
This can easily be done when the underlying model is the
M/M/ queueing model, as shown below.

n and yn = Yn p.
Therefore n = Nn , n = N
Throughout the paper, z is a complex number such that

1
Observe from (6) and (7) that both Sy (z) and Sy (z) are
well-defined for |z| = 1 under the assumption (4).

E[Yn ] = E[E[Yn | Nn ]] = E[p Nn ] = p.

3.1

Let us first determine Sy (z). By using (6) and (3) together

with the property that CovN (k) = Cov (k), we find
p2 |k| ,
p,

Covy (k) =

for k 6= 0
for k = 0.

Since < 1 and |z| = 1, the z-transform of Covy (k) is

Sy (z) =

"

p (p 1)z 2 + [1 + 2 (1 2p)]z + (p 1)
.
z(1 z)(1 z 1 )

The second-order polynomial in the variable z in the numerator has two positive real roots given by r and 1/r, with
r=

1 + 2 (1 2p)

min =

' ("
1!
S (z) Ho (z)Sy z 1
z
where z1 , . . . , zM are the poles (if any) of the function F (z)
inside the unit circle. The notation Res [F (z), zk ] stands for
the residue of F (z) at point z = zk , namely, the coefficient
of 1/(z zk ) in the Laurent series expansion of F (z) in the
vicinity of zk .
Specializing F (z) to the values of S (z), Sy (z) and Ho (z)
found earlier, yields
F (z) :=

G(z) G(z 1 )

tel-00408745, version 1 - 3 Aug 2009

p(1 p)
r

and

G(z) :=

1 rz 1
.
1 z 1

We now compute Sy (z). From (7) and (3) we find

Covy (k) = p |k|

F (z) =

so that
Sy (z) =

p(1 2 )
.
(1 z)(1 z 1 )

$S

y (z)
G(z 1 )

=
+

p(1 2 )
(1 r)(1 z 1 )

min = 1

and the transfer function Ho (z) of the optimal filter takes

here the simple form
p(1 )
B
=
(1 r)(1 rz 1 )
1 Az 1

min =

where
A = r,

(1 2 ) +

(1 2 )(1 2 (1 2p)2 )
.
2 2 p

(10)

This expression for min can be used to tune the parameters

p and or equivalently S (see Section 5).

r(1 2 )
p(1 2 )
=
.
B=
(1 r)
(1 p)(1 r)

The impulse response of this linear filter is given by the

first-order recurrence relation [11]
n = A
n1 + Byn
with n the estimator of n . We now return to the original
processes {Nn }n and {Yn }n , to finally obtain the optimal
linear filter:
n = AN
n1 + BYn + (1 A pB).
N

Bp
1 A

By using the expressions of A and B, we finally obtain

Ho (z) =

(1 2 )((1 Bp)z A)
.
(1 z)(z )(z A)

This function has two poles inside the unit circle which are
located at z = A and z = ; the residues of F (z) at these
poles are given by pAB(1 2 )/((1 A)(A )) and
(1+pB/(A)), respectively. Summing up these residues
gives

The transfer function H(z) is given by

H(z) =

Res [F (z), zk ]

with

with
:=

M
&
k=1

p(1 p) (1 rz)(1 rz 1 )
r
(1 z)(1 z 1 )

We conclude this section by computing the mean square

n )2 ] of our estimator. It is known
error min := E[(Nn N
that [11]

(1 2 )[1 2 (1 2p)2 ]
.
2(1 p)

Note that r < 1. Hence

Sy (z) =

and the Wiener filter are optimal (among the class of linear
filters) in the sense that they minimize the mean square
error. The key point is that the Kalman filter used in [2] was
derived under a heavy traffic assumption, while the Wiener
filter computed in the present paper holds for any value of
the model parameters and . This partly explains why the
estimator in [2] behaves well under light or moderate traffic
as experimentally observed in that paper.

(9)

It is interesting to compare this filter with the Kalman

filter derived in [2]2 . They appear to be the same! This
result is somehow expected, since both the Kalman filter
2
Recall that a Kalman filter is the optimal filter under the
condition of linear dynamics and observation, which does
not hold in our case. However, the dynamics does converge
to a linear diffusion as the traffic load tends to infinity, allowing us in [2] to obtain a Kalman filter which is optimal
for the asymptotic heavy traffic regime.

4.

OPTIMAL FIRST-ORDER LINEAR FILTER

The theory reported in Section 3 applies to any on-time

distribution (x) (with the exception of heavy-tailed distributions). However, it is not easy to identify the function
G(z) that appears in the canonical factorization of the spectrum Sy (z) (see (8)) and thereby the optimal filter, except
when the on-times are exponentially distributed rvs (see Section 3.1).
In this section we will determine the first-order linear filter
that minimizes the mean square error. Observe that, unlike
the Wiener filter, the proposed approach will not return the
optimal filter among all linear filters but simply the optimal
linear filter among all first-order linear filters. We will illustrate this approach at the end of this section in the case
where (x) is an hyperexponential distribution.
Recall the definition of the centered processes {n }n , {
n }n
and {yn }n made at the beginning of Section 3.

The methodology is simple: we want to find constants

A (0, 1) and B such that := E[(n n )2 ] is minimized
when the process {
n }n satisfies the following first-order recurrence relation
n = A
n1 + Byn .

(11)

In steady-state this implies that

Ak ynk .

n = B

(12)

k=0

!" !"

The mean square error is equal to

= E n2 + E n2 2E [
n n ] .

If this equation has a unique solution A (0, 1), then substituting this value of A into (15) will give the optimal pair
(A, B).
It is shown in Appendix A that (16) has always a unique
solution in [0, 1) (in particular) if g 0 (x) > 0 in [0, 1). This
condition will hold as long as P (D > S) > 0. In practice,
one can always select S such that this condition is true.
The reader can check that the filter defined in (11) with
the optimal pair (A, B) is the same as the Wiener filter
found in Section 3.1 when the on-times are exponentially
distributed.
We now illustrate the approach developed in this section
by considering the situation where on-times have a hyperexponential distribution. More precisely, we assume that

We have E[n2 ] = E[(Nn )2 ] = (see Section 2). From

(12) and (7) we find

pl el x

(x) = 1

l=1

Ak Cov (k) = pBg(A)

E [
n n ] = pB
k=0

where

tel-00408745, version 1 - 3 Aug 2009

z k Cov (k).

g(z) :=

(13)

k=0

The power series g(z) converges for |z| < 1 (Hint: k

Cov (k) is nonincreasing) and is therefore differentiable for
|z| < 1. We will denote by g 0 (z) its derivative.
It remains to express E[
n2 ] in terms of the parameters
A and B. Squaring both sides of (11) and then taking the
expectation, yields

!" #

E n2 =

B
1 A2

$%

! "&

2AE[
n1 yn ] + BE yn2

!" #

pB 2
1 A2

=
A

2pB
1 A2

(2pg(A) + (1 2p)) .

pB 2
1 A2

+ ,
AB

L
l=1

|k|

l l

for k = 0
for k =
6 0

so that
L

g(A) =

l
.
1 Al

1/1
1/2
1/

=
3897s 1
= 480061s 2
=
18316s

= 19.5
= 75.1
= 94.7

1
2

= 0.999359
= 0.999995

n = 0.99879456 N
n1 + 0.10720289 Yn + 0.006540864.
N

2pg(A) + (1 2p)
1 A2

#,

Numerical example : L = 2, p = 0.0106 and S = 2.5s. Also

The optimal first-order filter is

(2pg(A) + (1 2p)) . (14)

+g (A)(pB (1 A ))

=0

For comparison the Wiener filter found in Section 3.1 (for

exponential on-times) for these values is
n = 0.99828589 N
n1 + 0.14885344 Yn + 0.012900081.
N

5.

2pg(A) + (1 2p)

g(A)
= 2p B
B
1 A2

= 0.

The 2nd equation gives

B=

0/

l=1

In order to minimize , A (0, 1) and B must be the solution

of the following system of equations:

'(
(((
)
(((
(*

Cov (k) =

Finally, the mean square error is

= 2pBg(A) +

L
with 0 < pl < 1, l = 1, 2, . . . , L, and
l=1 pl = 1. In
this setting the underlying queueing model can be seen as
L independent M/M/ queues in parallel. The arrival
rate to queue l is pl and the service rate is l . Define
L
l := exp(l S), l := pl /l so that =
l=1 l . The
autocovariance function of the process {n }n is equal to

With the identities E[yn2 ] = Covy (0) = p (see (6)) and

E[
n1 yn ] = Bp2 (g(A) )/A (Hint: use (12), (7) and
Cov (0) = ), we obtain
E n2 =

(17)

g(A)(1 A2 )
.
2pg(A) + (1 2p)

(15)

Substituting this value of B into the 1st equation shows that

A must satisfy
Ag(A)(2pg(A) + (1 2p))
g 0 (A)(1 A2 )(pg(A) + (1 2p)) = 0.

(16)

GUIDELINES ON CHOOSING ACKS PARAMETERS

A good pair (p, S) should (i) limit the feedback implosion while at the same time (ii) achieve a good quality of the
estimator. Of course (i) and (ii) are antinomic and therefore
a trade-off must be found. This trade-off will be formalized
as follows: we want to select a pair (p, S) so that the mean
number of ACKs generated every S seconds (see (5)) and
the relative error of the variance of the estimator (denoted
as ) are bounded from above by given constants, namely

')
*

E[Yn ] = p
n )
V ar(Nn ) V ar(N
=
.
V ar(Nn )

(18)

3
The values of the parameters come from the trace called
video1 investigated in Section 6.

tel-00408745, version 1 - 3 Aug 2009

n is optimal then V ar(Nn ) V ar(N

n ) = E[(Nn
When N
2

Nn ) ] and becomes the normalized mean square error

[12, p. 202]. Optimality was shown for the M/M/ queue,
therefore
min
=

with min given in (10).

For given constants and , it is easy to solve the constrained optimization problem defined in (18), provided that
is known.
For the M/M/ model, where min is given in (10), we
find that p = / and that S, or equivalently , is the unique
positive solution of the equation min = .
The problem now is to choose constants and so that
conditions (i) and (ii) are satisfied. We have found that
in the range [0.5, 1] and 0.15 give satisfactory results.
We conclude this section with general remarks on how to
adapt the parameters p and S to important variations in the
membership. The estimation schemes in Sections 3.1 and 4
have been obtained under the assumption that parameters p
and S are fixed. However, the filters constructed in Sections
3.1 and 4 can still be used if p and/or S change over time,
provided that these modifications do not prevent the system
to be most of the time in steady-state. In that setting, a new
filter will have to be recomputed after each modification.
Such a modification can be called for each time the number
of ACKs received during a given period of time significantly
deviates from the current expectation (i.e. p).

6.

Table 2: Mean and percentiles

Trace
Mean
25
50
nE
video1 N
6.82
1.09 2.42
nH2
N
6.12
1.08 2.55
nE
video2 N
4.19
1.41 3.08
nH2
N
4.12
0.98 2.14
nE
video3 N
4.20
1.55 3.26
nH2
3.98
1.07 2.36
N
nE
video4 N
3.79
1.23 2.57
nH2
N
4.06
1.02 2.21
nE
overall N
4.44
1.33 2.88
nH2
4.34
1.02 2.26
N

n |/Nn
of |Nn N
75
90
95
5.25
6.31
5.43
4.41
5.71
4.83
4.51
4.39
5.22
4.73

11.5
13.5
8.66
8.78
8.71
9.35
7.50
8.98
8.60
9.61

19.4
20.6
11.9
12.6
11.0
12.6
11.0
14.7
12.0
14.2

n
Table 3: Empirical mean and variance of Nn N
Trace
Mean Variance min ,

nE
video1 N
-0.112
12.664 13.942 0.147

NnH2 -0.047
12.851 12.120
nE
video2 N
0.006
0.495
1.407 0.099
nH2
N
0.019
0.785
0.396
nE
video3 N
0.037
0.207
0.737 0.091
nH2
N
0.019
0.229
0.208
nE
video4 N
0.052
0.911
1.566 0.087
nH2
N
0.065
1.423
0.676

VALIDATION WITH REAL TRACES

In this section we apply the estimators developed in Sections 3.1 and 4 to four traces corresponding to video muticast sessions. Two types of estimators will be used: the
nE found in (9) when the popestimator denoted as N
ulation is modeled as an M/M/ queue; the estimator
nH2 derived in Section 4 in case the join times
denoted as N
are Poisson and the on-times have a 2-stage hyperexponential distribution (M/H2 / model).
The objective is twofold: we want to investigate the quality of both estimators when compared to real life conditions,
and we want to identify the best one.
We have collected four MBone traces denoted videoi , i =
1, . . . , 4 between August 2001 and September 2001 using
the MListen tool [1]. Each trace corresponds to a long-lived
video session (see duration of each session in Table 1, where
the superscript d stands for days). We have run both
algorithms (estimators) on each trace.
For each trace we have identified the parameters of the
M/M/ model (parameters and , or equivalently parameters and ) and of the M/H2 / model (parameters
, 1 , 2 , p1 and p2 = 1 p1 see definitions in Section 4).
The values of these parameters are reported in columns 38
in Table 1. Details on how these values are retrieved come
in Appendix B.
Parameters p and S have been chosen by following the
guidelines presented in Section 5, namely {0.5, 1} and
{0.1, 0.15}. Values of these parameters are listed in
columns 910 in Table 1. The performance of estimators
nE and N
nH2 are reported in Tables 2 and 3.
N
Table 2 reports several order statistics (columns 37) and

the sample mean of the relative error |NnNnNn | (column 2),

n is either N
nE or N
nH2 . All results are expressed in
where N

percentages. The first observation is that both estimators

perform reasonably well. The sample mean of the relative
error is always less than 6.82% and is as low as 3.79%; when
averaging over all experiments, this sample mean is less than
nE and N
nH2 (see last two rows). The last
4.5% for both N
column gives the 95th percentile and reads as follows: the
nE (resp. N
nH2 ) is
relative error achieved on trace video3 by N
95% of the time less than 11.00% (resp. 12.56%). The second observation is that no scheme is uniformly better than
the other over an entire session but their sample means are
nE
very close to each other (see column 2). For instance, N
H
2
n regarding the 90th and the 95th
performs better than N
percentiles whereas the result is reversed regarding the 25th
nH2 is empirpercentile. It looks like the relative error on N
ically more dispersed around its mean than is the relative
nE , and has a longer tail. Across all sessions (see
error on N
nH2 performs better than
last two rows), 75% of the time N
E

Nn . This improvement does not come for free, since it requires the identification of 4 parameters (, 1 , 2 and p1 )
nE .
instead of 2 ( and ) for N
Table 3 reports the sample mean and the sample variance
n . In the 4th column, we list the theoof the error Nn N
nE (see (10)) and
retical variance. It is given by min for N
H2

n ]
by for Nn (see (14)). The expected average E[Nn N
E

is zero in both approaches. Both estimators Nn and NnH2

have almost no bias (see column 2), and their empirical variances closely match the theoretical ones given by min and
, respectively. It is of interest to point out that for the 4
traces studied, , the theoretical mean square error provided
nH2 , is smaller than min , the theoretical mean square
by N
nE (however, this result is reversed if we
error provided by N
nH2 is
consider the empirical mean square errors). Thus, N

Trace
video1
video2
video3
video4

Session lifetime
3d 13h 33m 20s
11d 1h 46m 8s
50d 22h 13m 20s
29d 16h 43m 13s

Table 1: Parameter identification

1/
1/1
1/2
p1
p2
94.7 18316 3897 480061 0.97 0.03
14.1 16476
1 226498 0.93 0.07
8.1 66823
1 900854 0.93 0.07
17.9 83390
1 473268 0.82 0.18

tel-00408745, version 1 - 3 Aug 2009

Table 4: Distributions that best fitted into the

Trace
Best fit for inter-arrivals sequence
video1 Lognormal with = 3.38, d = 1.49
video2 Lognormal with = 5.20, d = 1.68
video3 Weibull with shape 0.65, scale 3500
video4 Weibull with shape 0.55, scale 2700

nE (again, N
nE is empirically more efmore efficient4 than N
H2

ficient than Nn ). The last column provides the relative

nE ), called (= min /) in Section 5.
error on V ar(N
In Fig. 1 (resp. 2, 3 and 4) we plot the variations of
membership for session video1 (resp. video2 , video3 and
nE and
video4 ), together with the estimates returned by N
H
n 2 . Among all 4 sessions, session video1 presents the highN
est variations in Nn . Fig. 1(a) (resp. 2(a), 3(a) and 4(a))
displays three curves: the membership of the video session,
nE , labeled Exponential, and
the estimation returned by N
nH2 , labeled Hyperexponenthe estimation returned by N
tial. It is clearly visible, especially at the left-hand side of
nE tracks better the session dynamics than
graph 1(a), that N
nH2 . Both estimators N
nE and N
nH2 have been derived unN
der some specific and restrictive assumptions: Poisson join
times for both of them, exponential (resp. 2-stage hyperexponential) on-times for the first (resp. second) one. It is
interesting to know whether or not these assumptions were
violated in each session videoi , i = 1, . . . , 4. We have therefore carried out a statistical analysis of each trace in order
to determine the nature of their join time process and of
their on-time sequence.
As shown in Table 4 and Figs. 1, 2, 3 and 4, parts (b) and
(c), neither is the join time process Poisson nor are the ontimes exponentially distributed (or hyperexponentially distributed) for any of the traces. The inter-join times and
the on-times appear to follow subexponential distributions
(Weibull and Lognormal distributions), a situation quite
different from the assumptions under which the estimators
have been obtained. Despite these significant differences,
the estimators behave well and therefore show a good robustness to assumption violations.
In summary, both estimators perform very well when applied to real traces and are robust to significant deviations
nH2
from their (theoretical) domain of validity. Estimator N
returns the best global performance for the relative error crinE .
terion, but does not track high fluctuations as well as N
nE is a good estimator, both
Overall, we have found that N
in terms of its performance and its usability since it only
requires the knowledge of two parameters: and .

4
An estimator is said to be more efficient if it has a smaller
variance.

7.

p
0.011
0.034
0.062
0.028

S
2.5
3.2
20.0
10.0

1.0
0.5
0.5
0.5

0.15
0.1
0.1
0.1

inter-arrivals and on-times sequences.

Best fit for on-times sequence
Weibull with shape 0.35, scale 3700
Weibull with shape 0.26, scale 1400
Lognormal with = 5.08, d = 3.32
Weibull with shape 0.18, scale 4000

OPEN ISSUES

The main pending issue concerns the knowledge of parameters and (or equivalently any couple of parameters
among , and , since = / in steady-state). When
these parameters are not known, the source should estimate
them. Again, the source could estimate any two parameters
among , and and infer the third one.
One possible way of estimating is to let a newly arrived
receiver send a heartbeat to the source with a certain (constant) probability q (q should be small enough to avoid overwhelming the source with heartbeats). The source would
then use the arrival time tm of the mth heartbeat to esti = m/(qtm ).
mate . The maximum likelihood estimator is
This estimator is unbiased and consistent by the strong law
of large numbers (limm tm /m = 1/(q)).
In a similar way, the source can estimate if receivers
probabilistically send a goodbye message reporting their
lifetime when they leave the session. Let m0 be the lifetime indicated in the m0 th goodbye message received at the
source, then the maximum likelihood estimator of is sim0
0
ply
= m0 /( m
is unbiased and
i=1 m ). The estimator
consistent.
n ]. As long as there is
A natural estimator for is = E[N
no estimation of both and , it is not possible to compute
the filter coefficient A and B. Then only a naive estimator
for Nn can be used, defined as the ratio of the number of
ACKs received Yn over the ACK probability p. Notice that
E[Yn /p] = .
One might want to treat the estimation of Nn , and
as a joint parameter estimation problem. Unfortenately, the
problem becomes much more complicated, and we are not
able at the moment to say anything about its resolution.
n ]
Alternatively, one might want to replace with = E[N
in (9) and study the performance of the resulting equation.
It is possible to go even farther by replacing with Yn /p,
instead of using E[Yn /p]. This will add more burstiness to
n and (9) becomes
the dynamics of N
n = AN
n1 + (1 A) Yn .
N
p
The latter equation is nothing but an exponential weighted
moving average (EWMA) estimator for the membership process Nn . We expect this EWMA estimator to perform better
n , as given in (9), over highly dynamic sessions, as its
than N

tel-00408745, version 1 - 3 Aug 2009

auto-regressive equation has been derived by replacing the

constant term with the bursty term Yn /p. It is also expected that this EWMA estimator will react faster to important and sudden changes in the membership process. However, its performance over real sessions still needs to be investigated more carefully. Observe that the use of this EWMA
estimator relies on the prior knowledge of = exp(S)
solely, and therefore either goodbye or hello messages
are still required to compute
.
The first option is to directly estimate as indicated
above, but the first estimate will be delayed until the reception of the first goodbye message, which might necessitates
the departure of several customers. The second option is to
estimate both and as indicated earlier in the section and
, which is not expected to perform as
to estimate as /
well as the first option, but will return a first estimate much
faster. At the expense of a larger warm-up period, we believe that the first option is preferrable. Another advantage
of this option concerns the possible modification of q 0 by the
source, in order to control the volume of goodbye messages
sent, which clearly cannot be done with the hello messages.

8.

CONCLUSION

The major contribution of this work is the design of two

novel estimators for evaluating the membership in multicast
sessions. We have designed estimators capable of efficiently
tracking the dynamics of multicast sessions while simultaneously avoiding feedback implosion. In contrast to the estimator proposed in [2] which was designed under heavy traffic
assumptions, our schemes do not place such restrictions.
Relying on the appealing Wiener filter theory, we have
computed the optimal linear estimator for session membership when the underlying model is an M/M/ queue. The
optimality refers to the unbiasedness of the estimator and
to the fact that the mean square error is minimized. We
have also developed the optimal first-order linear filter in
case the on-time distribution is arbitrary and have derived
the associated estimator in case the on-times have a 2-stage
hyperexponential distribution.
Both obtained estimators have been validated on real traces.
Their performance have been shown to be excellent, one of
them showing a good ability to adapt to highly dynamic
multicast sessions.

9.

ACKNOWLEDGMENTS

The authors wish to thank Prof. Patrick Thiran for helpful suggestions.

Proof. Write f (x) as f (x) = f+ (x) f (x) where

f+ (x) := [2p(g(x) ) + ] xg(x)
f (x) := [p(g(x) ) + (1 p)] (1 x2 )g 0 (x).
0
The derivative of f (x) is given by f
(x) = x2 x +
using

!2

:= p g 0 (x)

:= 2[p(g(x) ) + (1 p)]g 0 (x).

Since g 0 (x) > 0 and g(x) > (see (13)), it is seen that > 0
00
and > 0 which implies that f
(x) = 2x < 0 for x
0
[0, 1). We therefore have that f
(x) is strictly decreasing in
0
0
[0, 1), with f
(0) = > 0 and f
(1) = < 0. Thus, the
0
function f (x) has only one zero in [0, 1).
Therefore, and under the assumptions of the lemma, it is
seen that:
(i) f+ (x) is continuous and strictly increasing in [0, 1)
with f+ (0) = 0;
(ii) f (x) is continuous in [0, 1) and f (1) = 0. There
exists x0 (0, 1) such that f (x) is strictly increasing
0
in [0, x0 ), strictly decreasing in (x0 , 1) and f
(x0 ) = 0.
We deduce from the above that f (x) has a unique zero in
[0, 1) if g 0 (x) > 0 in [0, 1). This condition will hold as long
as P (D > S) > 0. In practice, one can always select S such
that this condition is true.

B.

COMPUTING PARAMETERS FROM

TRACE

Each trace records (Ti , Di ), i 1. To use the M/M/

queue model, we identify 1/ = E[Ti+1 Ti ] and 1/ =
E[D], and deduce = /. To use the M/H2 / queue
model, is computed as before. Identifying 1 , 2 , p1 and
p2 requires the knowledge of the first three moments of D
(recall that p2 = 1 p1 ). For a 2-stage hyperexponential
distribution, the kth moment is given by
k

E[D ] =

2
"
pl k!
l=1

EXISTENCE AND UNIQUENESS OF THE

SOLUTION

Lemma A.1. Define

f (x)

:=

(2pg(x) + (1 2p))xg(x)
(pg(x) + (1 2p))(1 x2 )g 0 (x),

where g(x) is given in (13).

If g 0 (x) > 0 for x [0, 1), then f (x) has a unique zero in
[0, 1).

(l )k

= k!

p1
p2
+
(1 )k
(2 )k

, for k 1.

The parameters 1 , 2 , p1 and p2 are then solution to the

following system of four equations, where l stands for 1/l
with l = 1, 2.
p1 + p2 = 1

APPENDIX
A.

+ [p(g(x) ) + (1 p)]g 00 (x),

(19)

p1 1 + p2 2 = E[D]

(20)

p1 12
p1 13

(21)
(22)

+
+

p2 22
p2 23

= E[D ]/2
= E[D3 ]/6.

Equations (19) and (20) readily give

p1 =

2 E[D]
,
2 1

p2 =

E[D] 1
.
2 1

(23)

Substituting Equations (23) for p1 and p2 into (21) yields

E[D2 ]/2 = E[D](2 + 1 ) 1 2 .

(24)

Substituting them into (22) yields

E[D3 ]/6 = E[D]((2 + 1 )2 1 2 ) 1 2 (2 + 1 ). (25)

[12] S. Haykin. Adaptive Filter Theory. Prentice Hall, 3rd

edition, 1996.
[13] H. W. Holbrook and D. R. Cheriton. IP multicast
2
E[D ]/2 = E[D]S P
channels: EXPRESS support for large-scale
E[D3 ]/6 = E[D](S2 P ) P S
single-source applications. In Proc. of ACM
SIGCOMM 99, Cambridge, Massachusetts, pages
= (E[D2 ]/2) S E[D]P
6578, September 1999.
where the latter identity is obtained when using the first
[14] J. Jannotti, D. Gifford, K. Johnson, M. Kaashoek,
one. It then follows that
and J. OToole. Overcast: Reliable multicasting with
3E[D]E[D2 ] E[D3 ]
an overlay network. In Proc. of USENIX OSDI 00,
S =
,
(26)
San Diego, California, pages 197212, October 2000.
3(2E[D]2 E[D2 ])
[15] L. Kleinrock. Queueing Systems: Theory, volume 1.
3E[D2 ]2 2E[D]E[D3 ]
P =
,
(27)
John Wiley and Sons, 1975.
6(2E[D]2 E[D2 ])
[16] C. Liu and J. Nonnenmacher. Broadcast audience
and 1 = 1/1 and 2 = 1/2 are the (positive) solutions of
estimation. In Proc. of IEEE Infocom 00, Tel Aviv,
x2 S x + P = 0. Namely,
Israel, volume 2, pages 952960, March 2000.
"
!
[17]
J.
Nonnenmacher. Reliable multicast transport to large
1,2 = 1/2 S (S )2 4P .
groups. PhD thesis, Ecole Polytechnique Federale de
Lausanne, Switzerland, July 1998.
We now can compute p1 and p2 as given in Equations (23). It
[18]
J.
Nonnenmacher and E. Biersack. Scalable feedback
is then possible to calculate l = pl /l and l = exp(l S)
for large groups. IEEE/ACM Trans. on Networking,
for l = 1, 2. Last = 1 + 2 .
7(3):375386, June 1999.
[19] D. Pendarakis, S. Shi, D. Verma, and M. Waldvogel.
C. REFERENCES
ALMI: An application level multicast infrastructure.
[1] K. C. Almeroth and M. H. Ammar. MListen, 1995.
In Proc. of USENIX USITS 01, San Francisco,
http://www.cc.gatech.edu/computing/Telecomm/mbone/.
California, pages 4960, March 2001.
[2] S. Alouf, E. Altman, and P. Nain. Optimal on-line
[20] J. Rosenberg and H. Schulzrinne. Timer
estimation of the size of a dynamic multicast group. In
reconsideration for enhanced RTP scalability. In Proc.
Proc. of IEEE Infocom 02, New York, New York,
of IEEE Infocom 98, San Francisco, California,
volume 2, pages 11091118, June 2002.
volume 1, pages 233241, March/April 1998.
[3] J.-C. Bolot, T. Turletti, and I. Wakeman. Scalable
[21] H. Schulzrinne, S. Casner, R. Frederick, and
feedback control for multicast video distribution in the
V. Jacobson. RTP: a transport protocol for real-time
Internet. In Proc. of ACM SIGCOMM 94, London,
applications. RFC 1889, Network Working Group,
England, pages 5867, September 1994.
January 1996.
[4] Y. Chu, S. Rao, and H. Zhang. A case for end system
multicast. In Proc. of ACM SIGMETRICS 00, Santa
Clara, California, pages 112, June 2000.
[5] D. R. Cox and V. Isham. Point Processes. Chapman
and Hall, New York, 1980.
[6] C. Diot, B. N. Levine, B. Lyles, H. Kassem, and
D. Balensiefen. Deployment issues for the IP multicast
service and architecture. IEEE Network magazine,
special issue on Multicasting, 14(1):7888,
January/February 2000.
[7] A. Dutta, H. Schulzrinne, and Y. Yemini. MarconiNet
- an architecture for Internet radio and TV networks.
In Proc. of NOSSDAV 99, Basking Ridge, New
Jersey, June 1999.
[8] S. Floyd, V. Jacobson, S. McCanne, C. Liu, and
L. Zhang. A reliable multicast framework for
light-weight sessions and application level framing. In
Proc. of ACM SIGCOMM 95, Cambridge,
Massachusetts, pages 342356, August 1995.
[9] P. Francis. Yoid: Extending the internet multicast
architecture. Unrefereed report, April 2000. available
at http://www.icir.org/yoid/docs/index.html.
[10] T. Friedman and D. Towsley. Multicast session
membership size estimation. In Proc. of IEEE Infocom
99, New York, New York, volume 2, pages 965972,
March 1999.
[11] S. Haykin. Modern Filters. Macmillan, New York,
1989.

tel-00408745, version 1 - 3 Aug 2009

Introduce now S and P as the sum and the product of 1

and 2 , respectively. Equations (24) and (25) become

60
0

12 Sep 01

13 Sep 01

14 Sep 01

(c) on-times

1s 1m

1d

1s 1m 1h

(b) inter-join times

14 Sep 01

12 Sep 01
1

13 Sep 01

14 Sep 01

Exponential fit
Weibull fit

0.5

0.5

Lognormal fit

13 Sep 01

Weibull linear fit for on-times, shape 0.345, scale 3700

Exponential fit

0.25

0.5

0.75

0.25

0.5

0.75

Figure 1: Membership estimation of session video1 and corresponding probability plots

12 Sep 01

14 Sep 01

16 Sep 01

18 Sep 01

20 Sep 01

1s 1m

1s 1m 1h

17 Sep 01

21 Sep 01

13 Sep 01

Lognormal linear fit for inter-join times, = 5.20, d = 1.68

21 Sep 01

Exponential fit

0.5

0.5

Exponential fit

Weibull fit

Lognormal fit

17 Sep 01

Weibull linear fit for on-times, shape 0.26, scale 1400

13 Sep 01
1

22 Sep 01

(c) on-times
1d

(b) inter-join times

tel-00408745, version 1 - 3 Aug 2009

12 Sep 01

Lognormal linear fit for inter-join times, = 3.38, d = 1.49

0.25

0.5

0.75

0.25

0.5

0.75

Figure 2: Membership estimation of session video2 and corresponding probability plots

6
0

08 Aug 01

15 Aug 01

22 Aug 01

29 Aug 01

05 Sep 01

19 Sep 01

1s 1m

1s 1m 1h

29 Aug 01

12 Sep 01

15 Aug 01

Lognormal fit

Weibull fit

12 Sep 01

Exponential fit

0.5

Exponential fit

29 Aug 01

Lognormal linear fit for on-times, = 5.075, d = 3.323

Weibull linear fit for inter-join times, shape 0.65, scale 3500

0.5

15 Aug 01

0.25

0.5

0.75

0.25

0.5

0.75

Figure 3: Membership estimation of session video3 and corresponding probability plots

25 Aug 01

01 Sep 01

08 Sep 01

15 Sep 01

1s 1m

1s 1m 1h

08 Sep 01

22 Sep 01

Weibull linear fit for inter-join times, shape 0.55, scale 2700

22 Sep 01

Exponential fit
Weibull fit

Weibull fit

08 Sep 01

Weibull linear fit for on-times, shape 0.18, scale 4000

0.5

0.5

Exponential fit

25 Aug 01
1

25 Aug 01
1

22 Sep 01

(c) on-times
1d

(b) inter-join times

tel-00408745, version 1 - 3 Aug 2009

12 Sep 01

(c) on-times
1d

(b) inter-join times

0.25

0.5

0.75

0.25

0.5

0.75

Figure 4: Membership estimation of session video4 and corresponding probability plots

tel-00408745, version 1 - 3 Aug 2009

!!"#" $

tel-00408745, version 1 - 3 Aug 2009

!"#$%#&'#( )& #&*+&'%&, -+.+

')//#'.%)& %& /+$,# &#.0)$1(
!"#$ %&!#' ()!*# +!"!,!!" # $%&' "%()*+ ,-*,./001234245 +) .((%.* ,- 6)7(8+%* !%+9)*:;<
=>+%-?%? @%*;,)- )A +9) (.(%*; B )-% ,- #--.C; )A $%C%&)778-,&.+,)-;5 @)C<
D45 -)< 4/25 ((< 4DE/4F25 G.-8.*H/I%J*8.*H 200D5 .-? +'% )+'%* ,(*)&%%?,-K; )A !$LM 200E N6)-A%*%-&% )- !%9 $%&'-)C)K,%;5 L)J,C,+H .-?
M%&8*,+HO5 P.*,;5 L.H 200E<

tel-00408745, version 1 - 3 Aug 2009

Experiences on enhancing data collection in

large networks
Mohamed Karim Sbai, Chadi Barakat
Project-Team Plan`ete, INRIA Sophia-Antipolis mediterranee, France
Email: Mohamed Karim.Sbai@sophia.inria.fr, Chadi.Barakat@sophia.inria.fr

inria-00324121,
tel-00408745, version
version 12 -- 324Aug
Sep2009
2008

Abstract
We improve and validate TICP [3], our TCP-friendly reliable transport protocol to
collect information from a large number of Internet entities. A collector machine
sends probes to a set of information sources that reply by sending back their reports. TICP adapts the sending rate of probes in a way similar to TCP for the
purpose of avoiding network congestion and keeps requesting reports until they are
well received. In a first part of this work, we add to TICP a mechanism to cluster information sources in order to smooth the variation of network conditions during the
collection session and to ensure an efficient handling of congestion at network bottlenecks. Simulations in ns-2 and PlanetLab experiments prove the outperformance
of TICP over non adaptive solutions and the interest of the clustering mechanism in
shortening the duration of the collection session and in decreasing the ratio of lost
packets. In a second part, we adapt TICP to collect several packets of information
from each data source. By the means of simulations, we compare the performance
obtained by TICP to that obtained when using parallel short-lived TCP connections. Our main observation is that TICP yields shorter collection sessions due to
its inherent multiplexing capability and that it avoids parallel Slow Start phases.
Finally, in a last part, we study the impact of delegating collection to some proxy
collectors. We explain our method for delegation and we show by simulations that
for a judicious choice of proxy collectors, one can decrease considerably the collection
session duration.
Key words: Transport protocol, information collection, clustering, several packets,
short-lived TCP connections, delegation

? This paper summarizes our contributions on the TICP protocol. It relies on our
two previous publications [3] and [4] and goes beyond by presenting experimental
results over PlanetLab and two new and important components, that are the extension to the multiple packets per source case and the delegation of the collector
function to intermediate proxies.

Preprint submitted to Elsevier

24 September 2008

inria-00324121,
tel-00408745, version
version 12 -- 324Aug
Sep2009
2008

Introduction

TICP [3] stands for TCP-friendly Information Collection Protocol. It is a

TCP-friendly reliable transport protocol designed to collect information from a
large number of sources distributed throughout the Internet. TICP is a general
purpose end-to-end transport protocol that does not impose any constraint on
the type of the collected data and that does not require network collaboration.
This generality widens the spectrum of application of the protocol. TICP can
be used in all scenarios where entire data collection is needed. The data to be
collected can be the availability of network entities, statistics on hosts, routers
and network paths, quality of reception in a multicast session, numbering of
population, weather monitoring, vote results, etc. It is of particular interest
in the area of network monitoring and topology inference (e.g., [5]) where
the number of network entities involved is large (hundreds of thousands of
machines and routers) and where it is preferable to end up the measurement
in a relatively short time without incurring additional load on the network.
TICP is an appropriate congestion and error control protocol to be used in
this area for the purpose of adapting the rate of probe packets as the pings
and the traceroutes.
In TICP, a collector machine sends probes to information sources, which reply
by sending back report packets containing their information. Some difficulties
come into play when designing a data collection protocol like TICP:
There is a risk of network congestion due to bandwidth limitation and the
large number of sources. Furthermore, all sources are not behind the same
bottleneck which makes the congestion control more difficult.
The collection traffic can be aggressive towards traffic generated by other
applications. In particular, attention should be paid to prevent the collection
traffic from penalising the concurrent TCP traffic.
The loss of probes or reports lengthens the duration of the collection session,
which urges for an efficient retransmission scheme.
TICP does not only adapt the probing rate as a function of network conditions,
but also tries to minimize the collection session duration by implementing an
efficient retransmission strategy. Moreover, TICP shares network resources
fairly with concurrent traffic, namely TCP traffic, by adapting its probing
rate in a similar way to TCP.
TICP was first introduced in [3]. It answers a need for a reliable transport
solution to collect information in large IP networks. Current solutions used
in network monitoring and management as in Skitter [17] or in the SNMP
protocol [15] implement simple periodic probing at a low rate in order not
the overload the network. For example in SNMP and to collect average traffic

statistics over 5 minute intervals, routers are probed periodically by no more

than once every 5 minutes. For short intervals and large number of routers the
overhead on the network can become significant. Other solutions like Concast
[14] rely on routers to aggregate the collected information and so do not work
in the current Internet.

inria-00324121,
tel-00408745, version
version 12 -- 324Aug
Sep2009
2008

The present work enhances the protocol with three additional mechanisms
to help it achieve better performances and adapt to more situations. This
is in addition to a deep analysis of the protocol by the means of extensive
simulations and real experiments over the PlanetLab platform [9]. The first
mechanism has been introduced in [4]. The second and third mechanisms are
novel. The implementation of the code in C++ to run over UDP sockets in
the PlanetLab platform is also one of our new contributions in this area.
The collector in the former version of TICP [3] probes information sources in
a random order. We start by showing that this strategy causes many problems
when moving to large networks, which results in long collection sessions, high
loss ratios and out of control traffic. The reason is that only one control at the
TICP collector is used to limit the traffic at several bottlenecks simultaneously,
which is clearly suboptimal given that the congestion of one bottleneck router
can be hidden by the low utilization of another bottleneck router and vice
versa. To probe sources behind the same bottleneck together and separately
from other bottlenecks, we add to TICP a mechanism to gather information
sources into clusters. This mechanism is based on the modeling of the Internet
by a two-dimensional Euclidean space and its decomposition into clusters. We
use to this end the Global Network Positioning system (GNP) [6] that provides
Internet host coordinates. Our new mechanism makes it possible to probe
sources from the closest cluster to the collector in terms of RTT (Round Trip
Time) to the farthest one. This very likely results in sources behind the same
bottleneck probed together before the collector moves to neighboring sources
located behind another bottleneck. It is supposed that this behaviour improves
the efficiency of the congestion control and ensures a smooth variation of its
variables in TICP.
To evaluate the performances of the protocol thus obtained, we run simulations with the ns-2 simulator [7] over realistic and complex network topologies.
These simulations show that TICP with the new mechanism of clustering has
better performances than without clustering, and that it outperforms other
non adaptive data collection solutions. To generalize this result, we implement
the protocol in C++ in linux and run real experiments on the PlanetLab
testbed. This practical experience tells us that TICP can be easily deployed
in the Internet and that the outperformance of the protocol observed in simulations is real.
Another challenge in developing a protocol like TICP is how to extend it to

inria-00324121,
tel-00408745, version
version 12 -- 324Aug
Sep2009
2008

the case of several packets of information to be collected from each source.

This information can be present at the moment the collection is initiated
or can be generated later by sources. The old version of TICP [3] ensures
collecting only one packet from each information source. In this paper, we
explain how we have extended TICP to support collecting several packets per
source and we evaluate this extension. The hard issue in this extension is
how to handle acknowledgments and sequence numbers of packets and how to
enable parallel collection so that to fully utilize the available bandwidth. We
implement our extension to TICP into ns-2 and we compare its performance
to that of a collection application based upon parallel TCP connections. When
the number of packets to be collected from each data source is not important,
TCP connections will be short-lived. Simulation results show that TICP has
shorter collection session than parallel short-lived TCP connections and this
is for any number of parallel TCP connections used. One can explain this
result by the fact that TICP multiplexes better the packets coming from all
sources. Furthermore, three-way handshake is a problem in short- lived TCP
connections and having several Slow Start phases in parallel is sub-optimal.
TICP has only one slow start phase for all sources.
We believe that deploying TICP to collect very large amounts of data will
not bring much compared to the use of parallel TCP connections. TICP was
designed to collect one to several packets from each source. But one can note
that even in the case of long-lived TCP connections, TICP does not require
adjusting any number of connections since it has only one connection to all
sources. Our simulation results will confirm this observation.
Afterwards, we move into studying the impact of delegating collection to a
set of sources, called proxy collectors, on TICP performance. A proxy collector plays the role of a TICP collector from the viewpoint of sources on its
charge, and plays the role of a regular information source (with several packets
of data) when seen by the main collector. Thus, there is a global TICP session connecting the main collector to proxy collectors and local TICP sessions
connecting the proxy collectors to sources on their charges. The goal of this
delegation is to minimize the duration of the collection session by replicating
the collector probing functionality inside the network. This paper explains the
method used to choose the proxy collectors as well as the changes made to
TICP to support this mechanism of delegation. With the help of ns-2, we vary
the number of proxies in different scenarios and we measure the duration of the
collection session. As expected, results show that delegation improves performances. However, we observe the existence of an optimal number of delegated
proxies otherwise we get a negligible gain compared to the no delegation case.
The paper is organized as follows. In Section 2, we describe the main functionalities of TICP [3]. We explain in Section 3 our approach to cluster information
sources and we discuss therein simulation and experimental results. Section

4 shows how we extend TICP to support the collection of several packets of

information from each source together with the comparison by simulations
to the case of parallel TCP connections. In the fifth section, we present the
changes made to TICP to ensure collection delegation and we discuss the
choice of proxy collectors. Section 6 overviews the related work and Section 7
ends the paper with some conclusions and perspectives on our future research
in this area.

TCP-friendly Information Collection Protocol

inria-00324121,
tel-00408745, version
version 12 -- 324Aug
Sep2009
2008

TICP [3] is a reliable transport information collection protocol implementing diverse functionalities. We overview in this section those related to error
recovery and network congestion control.

2.1 Error recovery

The TICP collector has a list of all information sources. The way to obtain this
list is out of the scope of TICP. Every source is distinguished by an identifier
that can be for example its IP address. Sources whose reports are lost are
probed again and are requested to retransmit their reports until they are
correctly received by the collector. To make the retransmission of reports in
TICP efficient, the collection session is made as a succession of rounds. In the
first round, the collector sends request (probe) packets to all sources following
their ranking in the list. In a second round, the collector sends requests to
sources whose reports were not received in the previous round. The collector
continues in rounds until it receives all reports. This behavior in rounds is
meant to wait for transitory network congestion to disappear from one round
to another and to absorb the excessive delay that some reports may experience.

2.2 Congestion control

To control the rate of requests and reports across the network, TICP implements a report-clocked window based congestion control similar to the TCP
one [2]. The collector maintains one variable cwnd indicating the congestion
window size in number of requests or reports. cwnd is then the maximum
number of requests that the collector can send without receiving any report
packet. New requests are only transmitted when the number of expected reports pipe is less than cwnd. TICP adapts cwnd to the observed loss rate of

reports. It proposes two algorithms to achieve this objective: Slow start and
Congestion Avoidance.

2.2.1 Slow start

inria-00324121,
tel-00408745, version
version 12 -- 324Aug
Sep2009
2008

The collector starts a collection session by setting cwnd to RS (Request Size,

a protocol parameter) and sending RS request packets. 1 After some time,
reports start to arrive. Some of these reports come on time, others are delayed.
A timely report indicates that the network is not congested and that the
collector can continue increasing its congestion window: cwnd = cwnd + 1.
This yields a doubling of the probing rate for every window size of probes. The
window continues growing in this way until the network becomes congested.
At this point, the collector divides its congestion window by two and enters the
congestion avoidance phase. The protocol comes back to slow start whenever
a severe congestion appears (to be defined later).

2.2.2 Congestion avoidance

This represents the steady state phase of TICP. During this phase, the collector
increases slowly cwnd to probe the network for more capacity. We aim a linear
increase of the congestion window by RS probes every window size of probes.
Thus, upon each timely report, the congestion window is increased by: cwnd =
RS
cwnd + cwnd
. When congestion is detected, cwnd is divided by two and a new
congestion avoidance phase is started.

2.3 Congestion detection mechanism

TICP implements a congestion detection mechanism to compute report loss
rates and to decide whether a report is on time, delayed or lost. This mechanism is based upon a timer TO scheduled at the beginning of the session and
rescheduled every time it expires.

2.3.1 Round-trip time estimator

TICP sets the timer of the mechanism to a conservative estimate of RTT
(Round Trip Time), using the samples of RTT seen so far. The value of the
1

RS (Request Size) is the initial number of requests sent without receiving any
report packet. It is also the step by which the window size increases. Compared to
TCP, this can be seen as the packet size in bytes.

timer is computed using estimates of the average RTT and of its mean deviation. Let srtt and rttvar be the estimates of the average and the mean deviation
of the RTT. Let rtt be the measured round-trip time when a report arrives.
The collector updates the estimates and the timer (TO) in the following way
:
rttvar = 34 .rttvar + 14 .|srtt rtt|
srtt = 78 .srtt + 81 .rtt
T O = srtt + 4.rttvar

inria-00324121,
tel-00408745, version
version 12 -- 324Aug
Sep2009
2008

This dynamics and the coefficients it involves are inspired from TCP [8]. The
weights for the estimation of srtt and rttvar are known in TCP to be easy
to calculate and to provide good balance between convergence and precision.
Setting TO to srtt plus four times the rttvar has shown its efficiency in setting
an upper bound on the Round-Trip Time in TCP. Since our estimator has
almost the same target as in TCP, we adopt the same strategy.
2.3.2 Detecting network congestion
TICP computes the report loss ratio over a time window in the past equal to
TO. When the timer is scheduled, the collector saves in the variable torecv the
number of reports to be received before the expiration of the timer. Let recv
be the number of timely reports received between the scheduling of the timer
and its expiration. The collector considers then that torecv recv reports were
recv
lost in the network. Consequently, it estimates the loss ratio to 1 torecv
.
The network is considered congested if the loss ratio exceeds the Congestion Threshold (CT) and severely congested if the loss ratio exceeds a higher
threshold SCT > CT called the Severe Congestion Threshold. CT and SCT
are two parameters of our protocol. They can be set to some default values,
for example, to 10% for CT and to 90% for SCT. We set them as follows:
RS
CT = min(0.1, cwnd
)
cwndRS
SCT = max(0.9, cwnd )

The minimum and maximum functions in the expressions of CT and SCT

are necessary to ensure that these thresholds do not take unrealistic values
when the congestion window is of small size (close to RS). One can use other
RS
for large
default values than 0.1 and 0.9. Set as above, CT is equal to cwnd
congestion windows, which means that congestion is concluded when more
than RS reports are not received (resp. severe congestion is concluded when
less than RS reports are received in a window). We recall that a report is
not received if it is lost or delayed, or if the corresponding request itself is
lost or delayed. This way we are compliant with TCP, which considers that
the network is congested if at least one packet is lost, and severely congested

when all packets are lost or delayed (i.e., they arrive after the expiration
of the timer). Note that we are designing TICP so that a packet in TCP is
comparable to RS requests/probes in our context. A TICP session behaves like
this as RS TCP connections running in the same environnement, a parameter
that the administrator can tune to whatever desired value.
2.3.3 Delayed and timely reports

inria-00324121,
tel-00408745, version
version 12 -- 324Aug
Sep2009
2008

A timely report is a report received before its deadline. The deadline of a report
is given by the timer. A report not received before its deadline is assumed to
be lost. If it arrives later than its deadline, it is considered to be delayed.

Fig. 1. The two types of reports

Fig. 1 explains how the deadline of a report is set. Let startTO be the scheduling time of the timer. Let startprevTO be the previous scheduling time of
the timer. When a report is received, the collector extracts from its header
the timestamp reqtime indicating the time by which the corresponding probe
has been sent. The report is received on time if and only if startprevT O <
reqtime. The report is a delayed one in the opposite case.

Clustering of information sources

The congestion control mechanism of TICP described above relies on RTT

estimation for the calculation of the report loss ratio and for the setting of
deadlines to retransmit lost reports. In case sources are probed randomly
and independently of their locations, this estimation is errored because of
the high variability in the RTT process and its low auto-correlation. The
loss ratio and report deadlines are then badly calculated, which impairs the
performance of the collection. In this section, we study the impact of clustering

information sources on the performance of TICP. First, we detail the reasons

that motivated the addition of the clustering mechanism. Then, we describe
briefly the GNP Internet coordinate system[6] that provides the knowledge of
sources locations and we explain our approach to cluster information sources.
Finally, we validate TICP with clustering by the means of simulations and
real experiments.

Does a random ordering of sources yield a good estimation of RTT for the
next pairs of request/report? The accuracy of this estimation is essential for
the correct functioning of the protocol. An overestimation of RTT results in
a delay in the detection of network congestion; the collector waits more than
necessary for already lost reports. This delay means a waste of time and an
aggravation of network congestion since the probing rate will not be reduced
on time. On the other hand, an underestimation of RTT can cause errors in
the computation of the report loss ratio due to the premature expiration of
the timer. Thus, some reports can be declared lost while they are not. If it
is the case, TICP will reduce unnecessarily the size of the congestion window
(cwnd) which will lengthen unnecessarily the collection session.
rtt
srtt

200

150
values (in ms)

inria-00324121,
tel-00408745, version
version 12 -- 324Aug
Sep2009
2008

3.1 Need for clustering of information sources

100

50

10

15
20
order of the source

25

30

35

Fig. 2. Estimating RTT with random ordering of sources

To clarify this point, we plot in Fig. 2 a sequence of rtt and srtt for sources
picked randomly from the list of sources without any consideration of topology.
The x-axis shows the probing order. Results are obtained from the simulations
whose setting is to be described later. The figure illustrates how the real value
of the RTT oscillates; it can exceed the estimated average by large values and
go much below it from time to time. This certainly implicates a regression in

360

rtt
srtt

340
320

Values (in ms)

300
280
260
240
220
200
180
160

10

15
20
order of the source

25

30

35

inria-00324121,
tel-00408745, version
version 12 -- 324Aug
Sep2009
2008

Fig. 3. Estimating RTT with latency-based ordering of sources

the performances of TICP. Our idea is to cluster information sources so as to

probe them from the nearest to the collector to the farthest one. This way, the
RTT in the network varies smoothly and can be estimated more accurately.
Fig. 3 supports this claim. It plots rtt and srtt when sources are ordered based
on their latency to the collector. We can clearly observe how the RTT and the
srtt vary smoothly with a smooth shift that can be easily compensated by the
mean deviation rttvar.
There is another gain from clustering. If the collector probes the sources in a
random order, the request packets and the reports will circulate everywhere in
the network. Hence, the collection traffic will participate to the congestion of
several bottlenecks at the same time. But the collector cannot with one control
adapt the sending rate of probes to network conditions on several paths in
parallel. In case of random ordering, the Internet is viewed by TICP as a sole
bottleneck which is far from reality. TICP should treat the bottlenecks one by
one. The gathering of information sources into clusters and their probing from
the nearest to the collector to the farthest ensures this decoupling of network
bottlenecks and adapts correctly the probing rate to these bottlenecks one
after the other instead of having a joint adaptation.

3.2 Global Network Positioning System (GNP)

To support the clustering mechanism, we need the knowledge of hosts locations. Global Network Positioning (GNP) [6] System is a promising and accurate solution to this problem. It models the Internet as an Euclidean space S
of dimension n. This space has a well-defined coordinate base and a distance
function. An internet host is represented in this space by a point such that

the Round-Trip Time between two hosts can be estimated by computing the
distance between their coordinates. To assign coordinates to hosts, GNP first
computes the coordinates of a small set of well-known hosts called landmarks
based on the measurements of delay between them. The coordinates of landmarks are then disseminated to any ordinary host willing to compute its own
coordinates. An ordinary host derives its own coordinates based on landmarks
coordinates and the measurements of delay between itself and the landmarks.
We suppose these coordinates are available to the TICP collector in the way
the IP addresses of hosts and their names are. Finally, coordinates are meant
to form a part of the ID of a machine in the Internet helping to localize it.
The way to get this information is out of scope of TICP. One can do it using
an additional channel as the DNS or other at a frequency much lower than
the frequency of collecting information.

inria-00324121,
tel-00408745, version
version 12 -- 324Aug
Sep2009
2008

3.3 The clustering approach

A cluster is a set of hosts located in the same neighbourhood such that the
latency to reach a host in the same cluster is generally shorter than the latency
to reach a host not belonging to it. If the cluster is well chosen, its hosts
very likely meet the same network conditions when they communicate with
the collector and their traffic cross the same network bottleneck. This is the
basic idea behind our solution to remove (even partially) the bias caused by
random probing in TICP. To locate information sources in the Internet, we
use GNP as described above. We model the Internet as a two-dimensional
Euclidean space. The collector and sources participate to GNP as ordinary
hosts in addition to some well-defined landmarks. An information source S
has a couple of coordinates (xS , yS ). The collectors coordinates are (xco , yco ).
We define a cluster as being a set of information sources whose representing
GNP points are located in a square area. The side length of the square is
denoted by a, which is a parameter of the protocol. The central cluster is the
square whose centre is the point representing the collector.
A cluster is completely defined by a couple of coordinates (X,Y) being integer
values. These coordinates are those of the centre of the corresponding square
relatively to the collector coordinates and then normalized by a. An information source S whose coordinates are S(xS , yS ) belongs to the cluster (X,Y)
given by:
co
)
X = round( xS x
a
yS yco
Y = round( a )

TICP probes information sources from the nearest to the collector to the
farthest one. This is supposed to smooth traffic transitions between network

inria-00324121,
tel-00408745, version
version 12 -- 324Aug
Sep2009
2008

bottlenecks while maximizing the volume of collected information at the beginning of the session. The collector begins with the central cluster and traverses
the other clusters following a spiral trajectory. Fig. 4 illustrates this trajectory. One can with a simple procedure, find the coordinates of the next cluster
during the collection knowing the coordinates of the current cluster.

Fig. 4. Order of probing information sources

3.4 Performance evaluation

To assess the impact of clustering sources on the performance of TICP, we
run simulations in ns-2 [7] over realistic network topologies as well as real
experiments on the PlanetLab testbed using our C++ implementation of the
protocol. We are particularly interested in the evaluation of the collection session duration and in the understanding of the impact of the different protocol
parameters.

3.4.1 Simulations results

We discuss here the results of our simulations. These simulations are run in ns2 [7] after the implementation of both TICP and GNP. We generate realistic
network topologies for simulations using the GT-ITM tool (Georgia TechInternet Topology Modeling) [11]. We choose to work on transit-stub (TS)
topologies due to their ability to capture the complexity and the hierarchical
structure of the real Internet. TS topologies model the Internet using a 2-level
hierarchy of routing domains with transit domains interconnecting lower level
stub domains. We assign latencies of 35ms for intra-transit domain links, 10
ms for stub-transit links and 5ms for intra-stub domain links. Fig. 5 shows an
example of a TS topology. The size of each topology realization is depicted by

inria-00324121,
tel-00408745, version
version 12 -- 324Aug
Sep2009
2008

Fig. 5. Transit-stub network topology

the parameter values in Table 1. In each simulation we generate a TS topology

and we choose randomly 500 sources of information and a collector among the
nodes composing the topology. The parameters of TICP are set as in Table 2.
In particular, probe packets are of size 100 bytes and reports can fit in packets
of 1500 bytes. With these values, congestion appears on the return path from
sources to collector. As for the cluster size a, we set its value to 50 ms as
our simulations indicate that this value leads to the best results over our TS
topologies. Later in the paper, we change the value of a and we study how this
impacts the collection session. Routing tables are calculated at the beginning
of simulations to provide shortest paths in terms of number of hops. These
tables are used to carry probes and reports in unicast between the collector
and the information sources. In contrary to [3] and to make our evaluation of
TICP more realistic, multicast routing is disabled in routers.
Table 1
Transit-stub topology parameters
Parameter
Signification
T
Number of transit domains
Nt
Average Number of nodes / transit domain
K
Number of stub domains / transit node
Ns
Average number of nodes / stub domain

Scenario
5
7
8
7

We compare between the both versions of TICP with and without clustering of information sources. The comparison criterion is the collection session
duration. Fig. 6 shows this duration for several simulations of TICP without
clustering. In each simulation, the order of sources in the list of the collector is
different, that is why we obtain each time a different duration for the collection
session. For TICP with clustering, the result is the same since the topology
does not change, and so the ranking of the sources is the same defined by
their coordinates. TICP with clustering finds the good ranking of information
sources and guarantees the shortest collection session duration. We save on
average 30% of time by moving from TICP without clustering to TICP with

Table 2
TICP parameters
Parameter
RS
probe size
report size
a

Value
10
100 B
1500 B
50 ms

clustering.
26

TICP without Clustering

TICP with clustering

Collect session duration (in s)

inria-00324121,
tel-00408745, version
version 12 -- 324Aug
Sep2009
2008

24
22
20
18
16
14
12
10

10
12
simulations

14

16

18

20

Fig. 6. Collection session duration for different ordering of sources

To evaluate the optimality of TICP, we implement in ns-2 an information collection protocol having a constant window size. For each window size, we run
10 simulations and we record the minimum of the collection session duration
over them in order to identify the best combination. Figure 7 plots three curves
in the same space. The x-axis represents cwnd, the size of the congestion window for the constant congestion window protocol. The y-axis represents the
collection session duration (in seconds) for the constant window protocol. This
is the curve having a parabolic convex shape. For small congestion window
sizes, the collection session is long because we have a low probing rate. For
large window sizes, the network is congested which lengthens the collection
session. As for the two horizontal lines, they represent the session duration
obtained by TICP and are only drawn for comparison. TICP adapts dynamically cwnd to network conditions. The role of TICP is to find in a dichotomic
way the right congestion window size that minimizes the collection session
duration. This proves the out-performance of an adaptive solution compared
to a non adaptive one. We notice in the figure how TICP with clustering manages to reach the optimum unlike TICP without clustering which yields longer
durations.
Then, we vary the cluster size and we study its impact on the collection

Constant congestion window protocol

TICP with clustering
TICP without clustering

55

Collect Session duration (in s)

50
45
40
35
30
25
20
15
50

100
150
200
cwnd (Congestion window size)

250

300

session duration. Taking a very large a is equivalent to TICP without clustering

since sources will be probed independently of their locations within the large
cluster. Taking a very small results in clusters empty or with few number
of sources. This is not efficient since there will be no clustering of sources
behind common bottlenecks. So, there should be some average a that provides
the best performance. Fig. 8 validates this intuition where we can see that
over the network topologies we simulate in this work, a value of a around
50ms is optimal. Each point of the curve in the figure is the average over 5
simulation runs on different network topology realizations, all satisfying the
characteristics in Table 1. The number of sources is taken equal to 500.
simulations
Average on simulations

24
Collect Session Duration (in s)

inria-00324121,
tel-00408745, version
version 12 -- 324Aug
Sep2009
2008

Fig. 7. Optimality of the protocol

22

20

18

16
0

20

40

60
a (in ms)

80

100

120

Fig. 8. Impact of a on collection session duration

Fig. 9 studies how the number of sources impacts the choice of optimal a. We
can clearly see that the optimal cluster size decreases when the number of

120

Optimal a

100

Optimal a (in ms)

inria-00324121,
tel-00408745, version
version 12 -- 324Aug
Sep2009
2008

sources increases. Compared to the value used above, the optimal a is equal
to 85ms for 300 sources and to 45ms for 700 sources. Indeed, for small number
of sources, one needs to increase a to group more sources behind the same
bottleneck together. At the opposite, for more sources, one needs to decrease
a so that the collector can better probe them depending on their locations.
But, if we continue increasing the number of sources, the optimal a will stabilize and become equal to some minimum value dependent of the topology.
As mentioned above, a cluster is in somehow the set of sources located behind
the same bottleneck. Thus, increasing the number of sources (distributed uniformly in space) does not increase the number of bottlenecks in the network.
Once we reach the maximum number of bottlenecks, the optimal size of the
cluster will become constant. One can safely use this value for applications collecting data from a very large number of sources. We suggest that in practice,
one starts by calculating this value by running multiple collection sessions,
then adapts it as a function of the measured session duration to account for
any change in the underlying network topology.

80
60
40
20
0
200

300

400

500
600
700
Number of sources

800

900

1000

Fig. 9. Optimal a as a function of the number of sources

3.4.2 Experimental results

We implement TICP with clustering in C++ and run real experiments on
the PlanetLab testbed [9]. We use one node from those available in our PlanetLab slice as collector and the other nodes as information sources. These
experiments on PlanetLab aim to study the efficiency of deploying TICP in
the Internet and to validate the simulation results already obtained. We run
successfully all scenarios described in our simulations. For lack of space, we
include two samples of our results. Fig. 10 plots the collection session duration as a function of the number of sources for the optimal cluster size, a small
cluster size (100 ms) and a big cluster size (1000 ms); and Fig. 11 plots the

optimal cluster size as a function of the number of sources.

Collection session duration (in s)

120

Optimal cluster size

Small cluster size (100ms)
Big cluster size (1000ms)

100

80

60

40

20

100

150

200
250
300
350
Number of data sources

400

450

500

Fig. 10. Collection session duration as a function of number of sources in PlanetLab

experiments
1000

TICP

900
800
Optimal cluster size (in ms)

inria-00324121,
tel-00408745, version
version 12 -- 324Aug
Sep2009
2008

50

700
600
500
400
300
200
100

50

100

150

200

250
300
350
Number of sources

400

450

500

Fig. 11. Optimal cluster size as a function of the number of sources in PlanetLab
experiments

Both figures show that experimental results have the same shape compared to
simulation ones but clearly different values of the optimums. This is because
the real Internet topology is different from the topologies used in our simulations. The main observation we can make from Fig. 10 is that the collection
session duration grows almost linearly with the number of sources which means
that TICP scales well in practice. Since the large cluster curve corresponds almost to the no clustering case, one can notice that there is an important gain
when using TICP with clustering. Moreover, the small cluster curve shows

that having a few number of sources in each cluster is suboptimal. As for Fig.
11, it gives an idea on how to choose the cluster size when collecting information in the real Internet. It seems that clusters of 100ms side length lead to
the best performances for a large number of Internet sources.

inria-00324121,
tel-00408745, version
version 12 -- 324Aug
Sep2009
2008

Collecting several packets of data from each source

The first version of TICP [3] supposes that each information source has a
small amount of data to send to the collector that can fit in one packet. This
is a strong assumption that needs to be relaxed. Some sources might possess
more than one packet when the session is initiated, or generate data later
during the collection session. In this section, we present our extension to TICP
that allows collecting informational reports of several packets. We explain first
the new functionalities added to TICP. We then study the scalability of the
new version through simulations. Finally, we compare TICP to a collection
application based on parallel TCP connections while varying the number of
TCP connections.
This extension is far from being an easy task. First, it requires sequence numbers to detect the lost reports from each information source and to ask for
retransmissions. Second, it requires parallel collection from each information
source when possible to better utilize the network resources. Lastly, it should
adapt to the case when information is generated by the sources during the
collection session. As to be described later, our extension handles all these
issues in an efficient way.

4.1 Behaviour of information sources

An information source subdivides data into packets of equal sizes. Each packet
has a unique sequence number SEQ. When the collector probes an information
source for packet number N, the source sends back the data packet having SEQ
equal to N. In general, the role of sources is simple as before and consists in
answering the requests for packets sent by the collector.
An information source does not necessarily know the number of packets to
send to the collector at the beginning of the session. We want to account
for the scenario when data is continuously generated during the collection
session. To indicate to the collector that the current packet is the last one, an
information source sets a field MORE in this packet to 0. When MORE is set
to 1, the collector understands that the source has more packets to send and
that it should keep probing it for more packets.

4.2 Behaviour of the collector

inria-00324121,
tel-00408745, version
version 12 -- 324Aug
Sep2009
2008

The collector probes sources in the order given by the clustering mechanism.
The sending rate of probes respects the congestion window size as before.
In the first round, the collector probes all sources asking them to send their
first packet. To make this possible, we add to the request packet a field RSEQ
indicating the sequence number of the requested packet. In the second round,
the collector discovers that some packets requested during the first round have
arrived to the collector while others have been delayed or lost. The probing
scheme consists in asking a source which has correctly sent her first packet to
send its second one and to ask a source whose first packet has been declared lost
to retransmit it. As for sources whose first packet has not yet arrived (their
deadlines have not expired), an efficient solution is to ask for their second
packet, of course if the congestion window allows and if the field MORE has
not been set to 0 for these sources. By behaving in this way, TICP is able
to anticipate the arrival of packets and fill the congestion window which is
necessary for an efficient utilization of network resources. This behavior is
generalized to the following rounds.
To implement this behavior, we introduce the following structure at the collector that contains all required information about the sources during the
collection:
SOURCES LIST = {SOURCE}
SOURCE = [ORDER, ID SRC, MORE, REQUESTS, RECEIVED PACKETS]
REQUESTS = [LAST SENT REQUEST, REQUESTS LIST]
RECEIVED PACKETS = [LAST RECEIVED PACKET, PACKET NUMBERS LIST]
REQUESTS LIST = {REQUEST}
PACKET NUMBERS LIST = {SEQ}
REQUEST = [RSEQ, DEADLINE]
When it is the turn of an information source, the collector checks the MORE
field of the object SOURCE corresponding to this source. If MORE equals 0,
then the collector jumps over this source and treats the next source in the
list. If all sources have the MORE field equal to 0, the collector announces
the end of the collection session. In the case when the MORE field of a source
is still equal to 1, the collector first seeks in the list of sent requests related
to this source (REQUESTS LIST) a request packet whose deadline has been
reached. We recall that the deadline of a request/data packet is reached when
the corresponding request has been transmitted before startprevTO, the time
at which the previous timer has been scheduled (see Section 2). If such request is found, the collector considers that the corresponding report is lost,
cancels this request, retransmits it and jumps to the next source in the list. If

REQUESTS LIST is empty or the collector has not found any request whose
deadline has been reached, it sends a request packet having RSEQ equal to
LAST SENT REQUEST + 1. This means that it anticipates requesting the
next packet and gives more time to delayed packets to arrive at the collector.
At the beginning of the session, LAST SENT REQUEST is set to 0. Any newly
sent request must be added to REQUESTS LIST and LAST SENT REQUEST
must be incremented. In the case of a retransmitted request, a request packet
having the same RSEQ is sent and the transmission time is set equal to the
current time. The attribute LAST SENT REQUEST does not change in this
case. When the collector receives a data packet, it updates cwnd and pipe
as described earlier (see Section 2), then it deletes the corresponding request
from REQUESTS LIST.

We want to test the scalability and the efficiency of this new version of TICP
when the number of packets per source and the number of sources grow. For
this, we run simulations in ns-2 on the same network topology used to validate
the clustering approach. The sizes of a request packet and a data packet are
respectively taken equal to 100 bytes and 1000 bytes. We fix the number of
sources and study the impact of changing the number of packets per source on
the performance of TICP. Fig. 12 plots the evolution of the collection session
duration as a function of the number of packets per source for three values
of the number of sources N1=50, N2=100, N3=150. We observe that for a
160

N1
N2>N1
N3>N2

140
Collect session duration (in s)

inria-00324121,
tel-00408745, version
version 12 -- 324Aug
Sep2009
2008

4.3 Scalability of the new version of TICP

120
100
80
60
40
20
0

50

100
150
Number of packets/source

200

250

Fig. 12. Impact of the number of packets per source

given number of sources, the collection session duration varies linearly with
the number of sources. When the number of sources increases, the slope of the

graph remains constant. This evolution proves the scalability of TICP when
the number of packets per source increases. One can interpret this behavior
by the fact that the network is seen by TICP as a set of bottlenecks. The
collection session duration is the sum of the collection times over the different
bottlenecks. We know that the collection time of lets say S bytes of data over
a bottleneck having B bps as available bandwidth, is almost linear with S
and equal to BS . When adding packets in sources, it is as if we are increasing
linearly the number of packets to collect over each bottleneck, which results
in this linear behaviour.

inria-00324121,
tel-00408745, version
version 12 -- 324Aug
Sep2009
2008

4.4 Comparison with a collection application using parallel TCP connections

The intuitive question that one would ask at this level is how TICP compares
to a collection application using regular TCP to collect data from sources
(one TCP connection per source). As we are in the context of relatively small
number of packets from each data source, the TCP connections will be shortlived connections, and hence, TICP will provide better performances for the
simple reason that it avoids the three-way handshaking and the slow start
phases of the individual TCP connections. Note that for very large amounts
of data per source, TICP works as well but its gain compared to the use of
TCP is not guaranteed. It is still has the advantage of doing the collection in
one session for all sources without the need to establish and schedule a large
number of TCP connections.
We compare here TICP to a collection application using parallel TCP connections to gather packets from sources. One TCP connection is opened between
the collector and each source to retrieve its information. The collection application using TCP limits the number of parallel connections to lets say m
connections. When one connection ends, another connection is established
with another source and this continues until all data is retrieved. The main
challenge here is to well choose the number of parallel TCP connections that
minimizes the collection session duration. Several works have studied the problem of how many parallel TCP connections one needs to open to get the best
performances. All agree on that only few connections need to be established
in parallel. [1] shows through emulations that for a large number of connections, the network becomes congested and the throughput deteriorates. On
the other hand, for m very small, one does not use completely the available
bandwidth and so the throughput of reception is not the best. [1] observes that
the throughput reaches its maximum for a number of connections between 6
and 8.
Having this in mind, we design the collection application based on parallel
TCP as follows. The application begins by opening m connections from the

collector to the first m sources in the list. When a connection finishes retrieving
packets from the source to which it connects, another connection is opened
with the next available source in the list. The application continues in this
way until reaching the last source in the list. We believe this should provide a
fair comparison with between TCP and TICP.
m-TCP connections (500 sources, 200 packets/source)
TICP (500 sources, 200 packets/source)

inria-00324121,
tel-00408745, version
version 12 -- 324Aug
Sep2009
2008

Collect session duration (in s)

2000
1800
1600
1400
1200
1000
800
0

50

100

150
200
250
300
350
Number of TCP connections (m)

400

450

500

Fig. 13. Parallel TCP connections vs. TICP

We implement this parallel TCP based collection application in ns-2 and we

generate topologies similar to those used earlier in this paper. We choose
500 sources and a collector randomly among the nodes of the topology. Each
information source is supposed to have 200 packets of 1000 bytes each to send
to the collector. First, we run TICP under this configuration. Then, we run
the collection application for different values of the number of parallel TCP
connections m. Fig. 13 plots two curves in the same space. The first curve
is for the TCP based collection application and the second one is for TICP
and is present only for comparison. In the figure, the x-axis represents the
number of parallel TCP connections (m) for the collection application using
TCP. The y-axis represents the collection session duration. As TICP opens
one connection to all sources, the duration of its session is independent of the
value of m on the x-axis. As expected, we can see that a small number of
parallel TCP connections gives the optimal collection session duration. But
surprisingly, the figure also shows that the collection session duration for TICP
is shorter than the collection session duration for any number of parallel shortlived TCP connections. This proves the efficiency and the importance of TICP
for the collection of several packets of data from a large number of sources.
We interpret this result by the fact that in case of TCP, at least one slow start
phase is needed by each connection, which wastes network bandwidth. In case
of TICP, these slow starts are avoided since the collection is done in parallel
from all sources with only one congestion window. The other explanation is
that TICP collects data in parallel from all sources, which allows an efficient

utilization of network resources. In case of parallel TCP, we are obliged to open

connections with a limited number of sources and collect from them entirely
before moving to other sources. This is suboptimal since only few bottlenecks
are utilized while others are not.

inria-00324121,
tel-00408745, version
version 12 -- 324Aug
Sep2009
2008

Collection session duration (in s)

100000

TCP - 5 connections- 500 sources

TCP - 50 connections- 500 sources
TICP

10000

1000

100

10

10

100
1000
Number of packets per source

10000

Fig. 14. Impact of the number of packets per source

To study the impact of the number of packets on the performance of both

TICP and parallel TCP connections, we do further simulations varying the
number of packets per source. For the TCP-based collection, we consider the
two cases of 5 and 50 connections. Figure 14 plots on a log-log scale the collection session duration as a function of the number of packets for TICP, TCP-5
connections and TCP-50 connections. As expected, this figure shows the outperformance of TICP when the number of packets to collect is relatively small
(from 1 to 500 packets here). We can notice a gain of one order of magnitude
due to avoiding the slow starts phases of the individual TCP connections. For
large files however, the gain reduces and both TICP and parallel TCP perform
equally well. This is because large files are able to bring the individual TCP
connections to their stationary regimes, hence reducing the gain introduced
by the traffic multiplexing in TICP. For small files this stationary regime is
barely reached making the slow start phases dominant and network ressources
under utilized if TICP multiplexing across sources is not used.

Parallelizing the collection using proxy collectors

Even though TICP has been shown to be efficient, its performances are still
limited by the fact that there is only one collector that probes sources distributed worldwide. In this section, we study the impact of delegating information collection to a set of proxy collectors. The goal of this delegation is

inria-00324121,
tel-00408745, version
version 12 -- 324Aug
Sep2009
2008

to decrease the collection session duration. Proxy collectors are specific welldimensioned machines responsible of collecting packets from sources in their
neighborhood. We assume the more general scenario where proxy collectors
are sources of information themselves. The main collector in its turn is responsible of the global collection by gathering packets from proxy collectors. First,
we present the method we use to choose the proxy collectors among information sources. Then, we explain how we extend TICP to support delegation.
Finally, we discuss the simulation results.
Note that in some cases the delegation of collection can influence the collected
data itself. This can be seen for example when part of the collected information
is related to the characteristics of the path between the collector and the source
of information as the delay and the bandwidth. The delegation in this case
should be done with further attention and its impact on the data itself is
to be studied carefully before deploying proxies. These specific cases are out
of the scope of this work. We only deal here with the network load and we
suppose that the collected information is the same whether it is probed from
machine A or machine B. Our optimization only aims to reduce TICP traffic
and fastens the collection while collecting the entire data. The specific cases
where the data to collect is function of the way the collection is done is an
important problem to consider but is somehow application specific, so to stay
general we leave it for a future research.

5.1 Choice of proxy collectors

The idea is to choose p proxy collectors among the sources in a way to minimize
the sum of the latencies from each source to its closet proxy collector. This
way we minimize the length of network paths crossed by probes and reports.
The choice of proxy collectors is a particular instance of the well-known pmedian problem. The p-median problem is defined as follows. Given a set F of
l potential facilities, a set U of n users, a distance function d: U x F-> R, and a
constant p l, determine which p facilities to open so as to minimize the sum
of the distances from each user to its closest open facility. This is a well-known
NP-hard problem. In our case, the set of information sources represents the
facilities as well as the users. The distance function is the Euclidean distance
computed using sources coordinates given by GNP.
To solve this problem, we use the hybrid heuristic proposed in [10]. It is a
multi-start iterative method where each iteration consists of a randomized
greedy construction of a solution, which is then submitted to local search.
A pool of best solutions found in iterations is kept. In each iteration, the
solution found with local search is combined with one of elite solutions. After
all iterations, there is a post-optimization phase in which elite solutions are

combined with each other. This method gives good solutions in short running
time.

inria-00324121,
tel-00408745, version
version 12 -- 324Aug
Sep2009
2008

5.2 TICP with collection delegation

Three types of actors take part of a TICP session with collection delegation:
the main collector, the p proxy collectors and the n-p ordinary information
sources. n is the total number of sources. A collection session starts with
choosing p information sources as proxy collectors using the above heuristic.
Then, the main collector runs a TICP session between itself and the proxy
collectors. In its first probe packets to a proxy collector, the main collector
sends the identifiers of ordinary sources that are under the responsibility of
this proxy collector. Once a proxy collector receives the list of sources for
which it is responsible, it updates its data structure to begin immediately a
local TICP collection session. During this local session, it plays the role of a
regular TICP collector towards ordinary sources in its charge. This behavior
is described with details in Section 4.
Now from the point of view of the main collector, a proxy collector behaves
like an information source. It considers the packets issued by ordinary sources
under its responsibility as its own packets. First, it starts by sending its own
packets to the collector. Then, it sends the packets of ordinary sources in the
order of their arrivals. It may happen that the main collector asks the proxy
collector to send a new packet and that the proxy collector is waiting for
new packets to arrive from ordinary sources. In this case, the proxy collector
replies with a persistency packet instead of the requested packet to tell the
main collector that it has nothing to send right now but that more packets will
come later. We add this persistency packet in order for the main collector not
to consider the absence of a response as an indication of network congestion.
Such a persistency packet has the same sequence number as the requested
packet but with a MORE field set to 2.
We run simulations in ns-2 over network topologies described in Section 3 in
order to evaluate the gain obtained when delegating the collection. We vary
the number of proxy collectors and we compute for each number, the average
duration of the collection session over several simulations. Fig. 15 plots the
results for 500 randomly distributed sources, with the x-axis showing the number of proxy collectors. We can clearly observe in the figure an improvement
in the performances of TICP when we delegate to the first proxies, then a
deterioration in the performances as long as we add more proxies. There is an
optimal number of delegated proxies that leads to the shortest session (a gain
by two in our settings). Indeed, for a small number of proxy collectors, there
is a waste of time and resources since the proxy collectors are in charge of

many sources and so are overwhelmed and not able to answer in real time the
probes of the main collector. If we continue increasing the number of proxies,
TICP will be able to collect with a high speed from the proxy collectors who
themselves are collecting at high speed from the ordinary sources, which explain the best performances. Further increase in the number of proxy collectors
lengthens the collection session since the main collector becomes in charge of
many proxy collectors which prohibits it from collecting from them at high
speed. For the extreme case where the number of proxy collectors is equal
to the total number of sources, the proxy collectors have no ordinary sources
in their responsibilities, so the collection operation is similar to an ordinary
TICP collector gathering data from all sources. Generally speaking, delegating
to everyone is equivalent to not delegating at all. The collect session duration
for 500 proxy collectors (the point on the very right hand side of the figure)
is equal to that for zero proxy collector (the point on the very left hand side
of the figure).

Collect Session Duration (in s)

inria-00324121,
tel-00408745, version
version 12 -- 324Aug
Sep2009
2008

TICP with delegation

1000

800

600

400

200

50

100

150
200
250
300
350
Number of proxy collectors

400

450

500

Fig. 15. TICP with delegation for 500 sources

Given the presence of an optimal value for the number of delegated proxies,
we want to check how this optimal value varies with the number of ordinary
sources. For this, we vary the number of sources and record the number of
proxy collectors giving the minimal collection session duration. The results
are average over several simulations and are plotted in Fig. 17. We can see
that when the number of sources is less than some threshold (140 in our settings), it is better to delegate to all sources (or equivalently not to delegate
at all).This observation is illustrated in Fig. 16 where we plot the collection
session duration as a function of the number of proxies for 100 sources randomly distributed in the network. As the number of sources is less than the
threshold, the performance keeps improving by adding more proxy collectors.
In this case, the population is small and one TICP session is able to do a good

job. For a number of sources larger than the threshold, we notice that the
optimal number of proxy collectors to consider is almost constant equal to the
threshold (140 in our settings). Our interpretation for this behavior is that
there is a maximum capacity for the network that can be achieved by some
delegation and any further increase of sources beyond this delegation will be
equivalent to adding more data per sources and so will not impact the number
of delegated proxies. This maximum number of delegated proxies to be used
is topology dependent. Our future research will focus on its calculation for
different scenarios.
700

TICP with delegation for 100 sources

Collection session delegation (in s)

600
550
500
450
400
350

10

20

30
40
50
60
70
Number of proxy collectors

80

90

100

Fig. 16. TICP with delegation for 100 sources

160

Optimal number of proxy collectors

inria-00324121,
tel-00408745, version
version 12 -- 324Aug
Sep2009
2008

650

Optimal p

140

120

100

80

60
0

100

200

300

400
500
600
Number of sources

700

800

900

Fig. 17. Optimal number of proxy collectors

1000

Related work

inria-00324121,
tel-00408745, version
version 12 -- 324Aug
Sep2009
2008

There is actually no general stand-alone transport solution to collect information from a large number of network entities on an end-to-end basis while
providing congestion and error control. Most if not all existing solutions are
application specific that do not answer all the questions we rise in this work.
Some solutions provide partial collection because the applications dont need
it, others require the collaboration of intermediate nodes and routers and so
are not general, and there are some solutions close to ours but use simple
congestion control as periodic or exponential probing at a low rate in order
not to congest the network.
TCP is a protocol to collect one file from one machine. Ensemble-TCP [20] and
HTTP 1.1 [21] are two protocols to collect many files from one machine using
one single TCP connection. SNMP [15] and measurement infrastructures as
Skitter [17] implement periodic and exponential probing at a low rate to collect
information from routers and to probe machines for the purpose of delay and
topology measurements. When it comes to data disseminated through the
Internet, there is no clear solution to achieve this collection on an end-toend basis. Though some protocols like Concast [14] use the principle of active
networks to program routers so that they can participate to the collection.
There has been some effort in the literature to develop collection solutions
for reliable multicast applications. Indeed, in reliable multicast, sources that
did not receive a packet need to send a NACK asking the collector for a
retransmission of this packet (we keep the terms collector and sources even
though the collector in this context is transmitting data and the sources are
listening). Sending many NACKs may cause congestion in the network or at
the collector. The problem is called NACK implosion and several solutions
have been proposed in the literature to collect these NACKs while avoiding
congestion. The difference from our context is that the NACK information can
be safely filtered; there is no need that a host sends a NACK if another host has
already sent a NACK for the same packet. Thus, it was proposed to aggregate
NACKs either along a tree that connects all sources or using multicast itself.
In [12], it is proposed that leaf sources send their NACKs to a parent source
(called designated receiver), which aggregates this information and sends it to
its parent until it reaches the collector. In [13,16], a source waits for a random
time before sending a NACK, and listens at the same time if another source
has sent a NACK for the same packet. If so, the former source cancels its
request, otherwise it sends it when the timer expires. Another approach for
NACK aggregation is to use the principle of active networks to program nodes
of a multicast tree as advocated by [14].
The reliable collection of information has also its application in sensor net-

inria-00324121,
tel-00408745, version
version 12 -- 324Aug
Sep2009
2008

works. Sensors are sources of information that wake up generally when an

event happens and send information about this event to some collecting point
called the sink. Some protocols exist in the literature for a reliable collection,
e.g., [18], [19]. These protocols agree on that end-to-end transport solutions
lead to poor performance in this case given the noisy nature of wireless links
connecting the sensors, and the absence of permanent routes caused by the
intermittent wake up of sensors and their limited lifetime. Per-hop transport
protocols have been advocated for sensor networks. The information is proposed to be reliably sent from one sensor to another until it reaches the sink,
with retransmissions done on a hop-per-hop basis. Clearly, such solutions are
not optimal in the wired Internet where permanent routes exist and are provided by the IP protocol. Moreover, the round-trip time in the Internet is
usually in the order of hundreds of milliseconds and links are of good quality.
All this make an end-to-end solution the most appropriate for the Internet,
which is what TICP provides.

Conclusions and perspectives

TICP is a transport protocol to collect information from a large number of

network entities. It aims to control the congestion of the network and to minimize the collection session duration. This work explains and validates three
main components of TICP. First, to ensure a smooth variation of the congestion control parameters, we add to TICP a mechanism to cluster information
sources. Simulation and experimental results show that this mechanism ameliorates the performance by reducing the loss ratio of reports, and therefore
shortens the collection sessions. Second, we extend the protocol to support
collecting several packets of data from each source. We prove by simulations
that our approach is scalable and that TICP performs better than a collection
application using parallel TCP connections. Lastly, we enhance the collection
further by delegating it to some intermediate nodes called proxy collectors. We
modify TICP to support this delegation and to choose optimally the proxy
machines. The simulations showed that for a well-chosen number of proxy collectors, TICP is able to collect faster data from sources. The optimal number
of proxies to be used seems to be topology dependent and not function of the
number of sources of information.
Our future research will be about the experimental evaluation of TICP in
the cases of several packets of data per source and delegated proxies. We
are also working towards the integration of TICP in Internet measurement
infrastructures for the purpose of controlling the rate of probing packets and
for the collection of passive measurements carried out at different links inside
the network.

References
[1] M. Allman, H. Kruse, S. Ostermann, An Application-Level Solution to TCPs
Satellite Inefficiencies, Proceedings of the First International Workshop on
Satellite-based Information Services (WOSBIS), New York, Nov. 1996.
[2] M. Allman, V. Paxson, W. Stevens, TCP Congestion Control, RFC 2581,
Apr. 1999.
[3] C. Barakat, M. Malli, N. Nonaka, TICP: Transport Information Collection
Protocol, in Annals of Telecommunications, vol. 61, no. 1-2, pp. 167-192, Jan.
2006.

inria-00324121,
tel-00408745, version
version 12 -- 324Aug
Sep2009
2008

[4] Mohamed Karim Sbai, Chadi Barakat, Transport Information Collection

Protocol with clustering of information sources, in proceedings of NTMS 2007
(Conference on New Technologies, Mobility and Security), Paris, May 2007.
[5] B. Donnet, P. Raoult, T. Friedman, M. Crovella, Deployment of an Algorithm
for Large-Scale Topology Discovery, in IEEE Journal on Selected Areas in
Communications (JSAC), Dec. 2006, vol. 24, no. 12.
[6] T. S. Eugene Ng and H. Zhang, Towards Global Network Positioning, ACM
SIGCOMM Internet Measurement Workshop 2001, San Francisco, Nov. 2001.
[7] The Network Simulator ns-2, http://www.isi.edu/nsnam/ns/
[8] V. Paxson, M. Allman, Computing TCPs Retransmission Timer, RFC 2988,
Nov. 2000.
[9] The PlanetLab platform, http://www.planet-lab.org/
[10] M. G. C. Resende and R. F. Werneck, A hybrid heuristic for the p-median
problem, Journal of Heuristics, vol. 10, pp. 59-88, 2004.
[11] E. W. Zegura, K. Calvert and S. Bhattacharjee, How to Model an
Internetwork, IEEE Infocom 96, San Francisco.
[12] Lin (J.C.), Paul (S.), RMTP: A Reliable Multicast Transport Protocol.IEEE
INFOCOM, March 1996.
[13] Floyd (S.), Jacobson (V.), McCanne (S.), Liu (C.), Zhang (L.), A reliable
multicast framework for light-weight sessions and application level framing.
ACM SIGCOMM, August 1995.
[14] Calvert (K. L.), Griffioen (J.), Mullins (B.), Sehgal (A.), Wen (S.), Concast:
Design and Implementation of an Active Network Service. IEEE Journal on
Selected Area in Communications (JSAC), vol. 19, no. 3, March 2001.
[15] Case, J., M. Fedor, M. Schoffstall, and J. Davin. 1990. A Simple Network
Management Protocol (SNMP). RFC 1157, May.

[16] Lacher, M. S., Nonnenmacher, J., and Biersack, E. W. 2000. Performance

comparison of centralized versus distributed error recovery for reliable
multicast. IEEE/ACM Transactions on Networking. 8, 2 (Apr. 2000), 224-238.
[17] Skitter
infrastructure
by
http://www.caida.org/tools/measurement/skitter/

CAIDA,

[18] Stann (F.), Heidemann (J.), RMST: Reliable Data Transport in Sensor
Networks. IEEE International Workshop on Sensor Net Protocols and
Applications (SNPA), May 2003.
[19] Wan (C.), Campbell (A.), Krishnahmurthy (L.), PSFQ: A Reliable Transport
Mechanism for Wireless Sensor Networks. ACM International Workshop on
Wireless Sensor Networks and Applications, September 2002.
[20] Eggert, L., Heidemann, J., and Touch, J., Effects of Ensemble TCP, ACM
Computer Comm. Review, January 2000.

inria-00324121,
tel-00408745, version
version 12 -- 324Aug
Sep2009
2008

[21] RFC 2616, Hypertext Transfer Protocol http://tools.ietf.org/html/rfc2616

tel-00408745, version 1 - 3 Aug 2009

!"#$!

!""! "#$%! &'(%!)"! *) !)%!+,-! .! %/-*"0/)% &/*' *)! +("'/-/10! !2343! .! -56)"!')!"7 84 +("'/-/10! .!
-56)"!')!" !%" *) ./+40)! !) &-!0)! (9/-*"0/) ("4)" ./))(! -4 3'/0%%4)3! !:&/)!)"0!--! .* '(%!4* !) "!'+!% .!
+43#0)!%; '(%!4*:; !" 4&&-034"0/)%; 3/*&-(! < -54,%!)3! .! %/-*"0/)% %"4).4'. &/*94)" '(&/).'! < "/*% -!% ,!%/0)%
.!% *"0-0%4"!*'% !" .!% /&('4"!*'%7 =/"'! /,>!3"0? -! -/)1 .! 3!""! '!3#!'3#! 4 ("( -! .(9!-/&&!+!)" .!% %/-*"0/)% @*0
&4%%!)" < -5(3#!--! !" @*0 '(.*0%!)" -4 3#4'1! %*' -! '(%!4* %4)% &/*' 4*"4)" 3/+&'/+!""'! -4 &'(30%0/) .!% '(%*-"4"%
/,"!)*%7 =/*% 49/)% 410 &/*' 3!-4 %*' .!*: ?'/)"%7 A*' -! ?'/)" .!% +!%*'!% &4%%09!%; )/*% 49/)% &'/&/%( *)!
%/-*"0/) &/*' -4 34&"*'! .* "'4B3 < "'49!'% *) '(%!4* .5/&('4"!*'7 C* -0!* .! 3/--!3"!' "/*" -! "'4B3 !" .! -54)4-D%!'
*-"('0!*'!+!)"; )/"'! %/-*"0/) &!'+!" 4*: /&('4"!*'% .! .(B)0' *)! "E3#! .! +!%*'! !" .! &'/1'4++!' -!% '/*"!*'%
.! "!--! %/'"! @*! 3#43*) .5!*: 3/--!3"! -4 @*4)"0"( .! "'4B3 )(3!%%40'! %4)% @*! -! 9/-*+! "/"4- .* "'4B3 3/--!3"(
.(&4%%! *) 3!'"40) %!*0-7 !""! '(.*3"0/) .* 9/-*+! .* "'4B3 3/--!3"( %5!F!3"*! &4' -! ,040% .! -5(3#4)"0--/))41!
%&4"04-! !" "!+&/'!--!7 85(3#4)"0--/))41! 4 *) 0+&43" )(14"0? %*' -4 &'(30%0/) .! -4 +!%*'!; *) 0+&43" @*0 &!*"
G"'! '(.*0" &4' .!% "!3#)0@*!% .! B-"'41! !" .!% %/-*"0/)% &'/"/3/-40'!%7 =/*% 49/)% 0--*%"'( 3!30 %*' .!*: !:!+&-!%
3/)3'!"%; -! &'!+0!' 3!-*0 .! -4 .("!3"0/) .!% 1'/% H/"% *"0-0%4"!*'% !" -! .!*:0$+! 3!-*0 .! -4 +("'/-/10! .!% 1'4).!%
&/&*-4"0/)% 94'04)" 4* 3/*'% .* "!+&%7 I/*' -!% .!*: !:!+&-!% )/*% 49/)% +!)( *)! ("*.! &/*' (94-*!' -50+&43"
.! -5(3#4)"0--/))41! %*090 &4' .!% %/-*"0/)% '(.*0%4)" -5!''!*' -/'% .! -50)9!'%0/) .!% 0)?/'+4"0/)% (3#4)"0--/))(!%7
J)%*0"! )/*% 49/)% 4,/'.( -54&&'/3#! 43"09! .4)% -4 +("'/-/10! !" 49/)% &'(%!)"( *)! %/-*"0/) 4&&!-(! K6 I
@*0 &!'+!" .! 3/)"'L-!' -! .(,0" .!% &4@*!"% %/).!% < 0)>!3"!' .4)% -! '(%!4* !" .54%%*'!' -4 B4,0-0"( @*4). 0%5410" .5*)! 4&&-034"0/) %/).4)" *) 1'4). )/+,'! .! +43#0)!% .0%"'0,*(!% < "'49!'% -! '(%!4*7 K6 I &!*" G"'! 9*
3/++! *) +*-"0&-!:41! .5*) 1'4). .! 3/))!:0/)% K I &4'"4)" .5*)! +G+! +43#0)! !" .!%"0)(!% < .!% +43#0)!%
.0F('!)"!% !) *)! %!*-! 3/))!:0/)7 K/*"!% )/% %/-*"0/)% /)" ("( 94-0.(!% &4' .!% %0+*-4"0/)% !" .!% "'43!% '(!--!%7

!"#$%&'#( M("'/-/10! .! -56)"!')!"; &'/3!%%*% %"/3#4%"0@*!%; "'43!%; (3#4)"0--/))41!; +!%*'!% &4%%09!% !" 43"09!%

%&"'()*'

6) "#0% "#!%0% N! &'!%!)" 4 %!" /? %/-*"0/)% ?/' 4 ,!""!' +/)0"/'0)1 /? "#! 6)"!')!"7 K#0% 0% 4) 0)"!'!%"0)1
4'!4 /? '!%!4'3# !9/-90)1 @*03O-D "/ 3/&! N0"# "#! 0)3'!4%! 0) "#! %0P! 4). #!"!'/1!)!0"D /? "#! 6)"!')!" 4). "/
B-- "#! 14& 34*%!. ,D "#! 4,%!)3! /? 1-/,4- 4). %"4).4'. +/)0"/'0)1 0)?'4%"'*3"*'! ?/' "#! 6)"!')!"7 Q*' +40)
/,>!3"09! .*'0)1 "#0% '!%!4'3# 0% "#! .!%01) 4). 94-0.4"0/) /? !230!)" +/)0"/'0)1 %/-*"0/)% "#4" 4'! 4,-! "/ %34-!
4). "/ '!.*3! "#! -/4. /) "#! )!"N/'O N0"#/*" 3/+&'/+0%0)1 "#! *%!?*-)!%% 4). @*4-0"D /? "#! '!%*-"%7 R! N/'O
0) "N/ .0'!3"0/)%7 6) /)! .0'!3"0/) N! 4..'!%% "#! &4%%09! +/)0"/'0)1 4&&'/43# 4). N! &'/&/%! 4 ?'4+!N/'O ?/'
3#//%0)1 "#! '/*"!'% "#4" %#/*-. 3/)"'0,*"! "/ %/+! +/)0"/'0)1 "4%O 4). "#! 4+/*)" /? "'423 "/ ,! 3/--!3"!. ,D
!43# '/*"!'7 K#0% 0% ./)! ,D N'0"0)1 "#! &4%%09! +/)0"/'0)1 &'/,-!+ 4% 4) /&"0+0P4"0/) &'/,-!+ 4). 3/)"'/--0)1
"#! &4'"030&4"0/) /? '/*"!'% ,D "#! +!4)% /? %&4"04- 4). "!+&/'!- "'423 %4+&-0)17 S)?/'"*)4"!-D "'423 %4+&-0)1
34))/" ,! ./)! N0"#/*" 4 )!14"09! 0+&43" /) "#! @*4-0"D /? +!4%*'!+!)"%7 K#0% 0+&43" 0% "/ ,! !94-*4"!. 4). 4)
/&"0+0P!. 0)9!'%0/) +!"#/. 0% "/ ,! .!B)!. ?/' !43# +!"'03 /)! N4)"% "/ 3/)%0.!'7 R0"# "#! #!-& /? "N/ 3/)3'!"!
!:4+&-!%; N! !94-*4"! "#! 0+&43" /? "'423 %4+&-0)1 4). 0--*%"'4"! #/N "#! 0)9!'%0/) /? "#! %4+&-!. 0)?/'+4"0/)
0% "/ ,! ./)!7 Q)! !:4+&-! 0% "#! .!"!3"0/) 4). '4O0)1 /? "#! -4'1!%" H/N% 0) 4 "'423; 4). "#! %!3/). !:4+&-!
0% "#! 3/*)"0)1 4). "'43O0)1 0) "0+! /? "#! %0P! /? 4 -4'1! 4). 3#4)10)1 &/&*-4"0/)7 6) &4'4--!-; N! N/'O /) "#!
43"09! 4&&'/43# 4). N! &'!%!)" 4 &'/"/3/- %/-*"0/) 34--!. K6 I ?/' '!1*-4"0)1 "#! '4"! /? &'/,!% "/ ,! 0)>!3"!. 0)
"#! )!"N/'O 4). ?/' 3/)"'/--0)1 "#! '!-04,0-0"D /? "#! +/)0"/'0)17 K6 I 34) ,! %!!) 4% 4 %/-*"0/) "/ +*-"0&-!: "#!
+4)D K I 3/))!3"0/)% /)! )!!.% "/ *%! ?/' 4 )!"N/'O N0.! 43"09! +/)0"/'0)1 0) /)! 3/))!3"0/)7 C-- /*' %/-*"0/)%
.*'0)1 "#0% "#!%0% 4'! %*&&/'"!. ,D 4)4-D"034- +/.!-% 4). 94-0.4"!. /) '!4- "'423 "'43!%7 K#! "#!%0% &'/90.!% 4)
/9!'90!N 4). 4 .0%3*%%0/) /? "#! %"4"! /? "#! 4'" 4). &!'%&!3"09!% /) /*' ?*"*'! '!%!4'3#7

)*+,!-.#( 6)"!')!" +/)0"/'0)1; %"/3#4%"03 &'/3!%%!%; "'43!%; %4+&-0)1; &4%%09! 4). 43"09! +!4%*'!+!)"%