ERFORMANCE ISSUES 535
= 4250
APPENDIX 16B PERFORMANCE ISSUES TE 432
he choice of a LAN or MAN architecture is based on many factors, but one of
the most important is performance. Of particular concern is the behavior
(throughput, response time) of the network under heavy load. In this apper
provide an introduction to this topic. A more det
[STALO9}
dix, we
led discussion can be found in
The Effect of Propagation Delay and Transmission Rate
In Chapter 7, we introduced the parameter @, defined as
Propay
tion time
Transmission time
In that context, we were concerned with a point-to-point link, with a given propaga.
n time between the two endpoints and a transmission time for either a fixed or
frame size. It was shown that a could be expressed as
Length of data link in bits,
Length of frame in bits
This parameter is also important in the context of LANs and MANs, and in
fact determines an upper bound on utilization, Consider a perfectly efficient ac
cess mechanism that allows only one transmission at a time. As soon as one trans.
mission is over, another station begins transmitting, Furthermore, the transmission,
head bits. What is the maximum possible utilization of the
network? TL can be expressed as the ratio of total throughput of the network to it
data rate
Throughput
(161
Data Rate
Now define, as in Chapter
R = data rate of the channel
maximum distance between any two stations
locity of signal propagation536° CHAPTER 16 / HIGH-SPEED LANS
The throughput is just the number of bits transmitted per unit time. A frame eon
tains L bits, and the amount of time devoted to that frame is the actual transmission
ime (L/R) plus the propagation delay (d/V). Thus
L
Throughput = Fy
But by our preceding definition of a
av _ Rd
“> TR” WV
Substituting (16.2) and (16.3) into (16.1),
Note that this differs from Equation (7.4) in Appendix 7. This is because the lat
assumed a half-duplex protocol (no piggybacked acknowledgments)
So utilization varies with a, This can be grasped intuitively by studying Fig
ture 16.15, which shows a baseband bus with (wo stations as far apart as possible
(worst case) that take turns sending frames. If we normalize time such that frame
‘ansmission time = 1, then the propagation time = a. For a = 1, the sequence of
ts is as follows:
Z
‘s ‘Start of transmission
ea
=a
oo
End of transmission Start of reception
5
atite—$— ts atite
L
End of reception i of reception
(ay Transision ime = 1; propagation time = 6 1, events 2 and 3 are interchanged. In both cases, the total time for one
turn” is 1 + a, but the transmission time is only 1, for a utilization of 1/(1 + a).
The same effect can be seen to apply to a ring network in Figure 16.16. Here we
assume that one station transmits and then wails to receive its own transmission be-
fore any other station transmits, The identical sequence of events just outlined applies
Figure 16.16 The Effect of a on Utilization Ring538. CHAPTER (0 / HIGH-SPEED
able 168 RG
presentative Values of
| (stbps) (its) Length (kim)
Typical values of a range from about 0.01 to 0.
the restriction of only one frame at
Finally, the p)
ame can be transmitted as soon
srotocol adds overhead that makes utilization wor
ding, analysis assumes a “pe
an old frame
hext subsection for token passing and CSMA/CD.
1 for LANs
O for MANs Table 168 gives some representative values for a bus top
er speed networks, utilization suffers. For this reason,
time is lifted for LANs such
sfect” protocol, for wh
In p
=
=|
pology. Asean
s FDDI
active, the MAC
Simple Performance Models of Token Passing and CSMA/CD
The purpose of this section is to give the reader some insight into
formance of the most important LAN protocols: CSMA/CD, token bus. and t
ring, by developing two simp at this exercise
Will aid in understanding the results of more rigord
For these models we assume a local network with N active
on delay of a. To simplify the
imum normalized propaga
ich station is always prep
figure of merit and does pern
ed to transmit a frame
lowed by a token asa eycle and defini
c
r
crage time for one ey
erage lime to transmit
sage time to pass a
e performance models. It is hoped
analysis, we
‘ession for maximum achievable utilization (U). Although this
Strucd to be the sole figure of merit for a local network, it isthe si
useful performance comparisons.
First, let us consider token ring, Time on the ring will alternate
frame transmission and token passing, Refer to a single instanceAPPENDIX 16 / PERFORMANCE ISSUES 539
It should be clear that the average eyele rate is just JC = 1/(T, +7
Intuitively,
(6.5)
That is, the throughput, normalized to system capacity, is just the fraction of time
that is spent transmitting data
Refer now to Fi
time equals I and pr
gure 16.16; time is normalized such that frame transmission
gation time equals a. Note that the propag
include repeater delays. For the case of a < 1, a station transmits a frame at time
fa, teceives the leading edge of its own frame at f + a, and completes transmission
at + 1. The station then emits a token, which takes an average time a/N to reach
the next station. Thus one cycle takes I + a/N and the transmission time is 1, So
U= 11+ aN
For a > 1 the reasoning is slightly diffe
pletes transmission at fy + 1, and receives the lead
tion time mus
nt. A station transmits at fy, com=
of its frame at fy + a. At
to emit a token, which takes an average time a/N to reach the
next station. The cycle time is therefore @ + a/N and U = 1/(a(1 + 1/N)).
Summarizing
that point, itis f
Token ring: | = 16.6)
For CSMAJCD, consider time on the medium to be or
length is twice the end-to-end propagation delay. This is a convenient way to view
the activity on the medium; the slot time is the maximum time, from the start of
transmission, required to detect a collision. Again, assume that there are N active
stations. Clearly, if each station always has a frame to transmit and does so, there will
be nothing but collisions on the line, So we assume that each station restrains itself
to transmitting during an available slot with probability P.
Time on the medium consists of two types of intervals, First is a transmission
interval, which lasts 1/2a slots, Second is a contention interval, whieh is a sequence
of slots with either a collision or no transmission in each slot. The throughput is just
the proportion of time spent in transmission intervals [similar to 1
Equation (16.5)}
anized into slots whose
reasoning for
To determine the average length of a contention interval, we begin by comput
ing A, the probability that exactly one station attempts a transmission in a slot and
therefore acquires the medium. This is just the binomial probability that any one
station aitempts to transmit and the others do not540. CHAPTER 16 / HIGH-SPEED LANS
A Pl — PyN
(VP
NPC. Py?
This function takes on a maximum over P when P
a=(-1jNye
We are interested in the maximum because we want to calculate the maximum
throughput of the medium. It should be clear that the maximum throughput will be
ichieved if we maximize the probability of successful seizure of the medium
Therefore, the following rule should be enforced: During periods of heavy usage.a
station should restrain its offered load to 1/N. (This assumes that each station
knows the value of V.To derive an expression for maximum possible throughput
‘we live with this assumption.) On the other hand, during periods of light usage
maximum utilization cannot be achieved because the load is too low: this region is
not of interest here.
Now we ean estimate the mean length of a contention interval, 10, in slots:
{slots ina rove with collision or no
E{w] = 32% Pe transmission followed by slot with one
We can now determine the maximum utilization, which is just the length of a trans
mission interval as proportion of a eycle consisting of a transmission and a con
tention interval:
12a
esMArCD: | U = =
[ous Tat (= Aya 7
Figure 16.17 shows normalized throughput as a function of @ for various
of N and for both token passing and CSMA/CD. For both protocols, through
‘lines as a increases. Ths is to be expected But the dramatic difference betwesAPPENDIX 1618 / PERFORMANCE ISSUES 541
CSMACD N= 2
CSMACDN = 10
Sr 3 4 567891 2 3 45678910
Figure 16.17 Throughput as a Function of @ for Token Passing and
cSMaICD
‘wo protocols is seen in Figure 16.18, which shows throughput as a function of N.
Token-passing performance actually improves as a function of N, because less time
is spent in token passing. Conversely, the performance of CSMA/CD decreases be-
cause of the increased likelihood of collision or no transmission.
Itis interesting to note the asymptotic value of U as N increases. For token ring,
1 a