Analysis Of Optimisation Issues In Multi-Period DWDM Network Planning

Neil Geary1,2, Andreas Antonopoulos1, Elias Drakopoulos1, John OReilly2

1

Lucent Technologies, 101 Wigmore Street, London W1H 9AB, UK.

Communications Engineering Doctorate Centre, Department of Electronic & Electrical Engineering,
University College London, Torrington Place, London WC1E 7JE, UK.

Abstract-- Techniques for the planning of multi-service

telecommunications networks over a multi-period scenario are
proposed, based on routing optimisation to reduce network cost
by removing DWDM transmission systems. A model of a
medium-size national network is optimised using the various
techniques, and the results presented in terms of the total
number of DWDM systems required and the amount of
wavelength-hops (giving an indication toward the number of
optical transponders required). Traffic is taken as a mix
between unprotected and 1+1 SNCP optical path protected
demands. Traffic growth is applied over the 5 year planning
horizon to give a tenfold increase in total traffic volume for the
final year. A greenfield situation is assumed, yet since
deployment occurs over several years, a Present Value
calculation is used to estimate total cost, assuming that
transponders and systems are only installed when they are due
to be lit.
Index termsOptical networks, multi-period planning

A.

INTRODUCTION

Network operators are currently deploying DWDM-based

(Dense Wavelength Division Multiplexing) optical transport
networks to meet the tremendous bandwidth requirements
caused by the explosive growth in Internet traffic, increased
competition forcing down tariffs for voice calls, and corporate
Virtual Private Networks spanning the globe.
DWDM is seen as an effective solution for this rush for
bandwidth. Optical channels can be set up between nodes
based on wavelengths within the transmission system.
Different rate client signals can also be accommodated, e.g.
2.5GHz and 10GHz, on the same network, as well as very
different service types, e.g. SDH and Gigabit Ethernet. A
consensus has been reached to implement a channelassociated overhead, to be transmitted as a digital TDM
wrapper [1]. This framing means that Optical Transport
networks will not be transparent, as originally thought, but
gives opportunity for powerful channel management using the
optical header, and strong forward error correction to increase
the maximum fibre reach and reduce the Optical Amplifier
requirement.
Current commercially available systems can offer up to 80
optical wavelengths on a single fibre.

Contact e-mail : n.geary@ee.ucl.ac.uk

Many DWDM line systems include modular components,

allowing an incremental growth of channels, spreading network
investment over the lifetime of the network.
Therefore careful optimisation of the network design is
required over the typical five year planning horizon to give a
cost-effective network meeting any budgetary restrictions for
the installation or upgrading of equipment in each annual
period. This results in the term Multi-period planning.
Towards the end of a planning horizon, the traffic forecasts
are at best loose estimates. Flexibility of a network in terms of
traffic growth and churn is an important feature.
Availability of the network is a key factor to allow the
provisioning of service level agreements. Optical layer
protection and restoration give an extra option to the
traditional client layer protection/restoration protocols such as
IP router topology reconfiguration and SDH self-healing rings.
Attractive optical layer survivability techniques proposed
include: 1+1 Optical Channel layer path protection, Optical
Multiplex Section shared protection rings, and Optical Channel
layer Protection rings (with either Shared or Dedicated
protection), and Restoration using Optical Cross Connects [2].
Lower priority traffic may be routed unprotected in the optical
layer.
In this paper though, we simplify the traffic composition to a
50-50 mix of unprotected wavelength demands, and 1+1 optical
path protected demands. Work is ongoing to include SDH
layer traffic and Optical ring architectures in future analyses.
The traffic is assigned to a typical mid-size national network,
assuming a traffic growth of x10 over 5 years.
Inputs to the planning process include the node locations,
available duct routes between nodes (links), and the traffic
forecast for each period.
B.

M ULTI-LAYER ISSUES & OPTICAL RINGS

To fully optimise the network planning, it is necessary to

consider the most widespread client signal, SDH (or SONET)
traffic, and to include survivability mechanisms based on
optical ring architectures.
1)

SDH/SONET traffic demands

SDH self-healing rings (e.g. MS-SPRings) reserve 50% of

capacity for automatic protection switching mechanisms. It is
sensible then, to avoid both duplicated spare capacity, and
multi-layer protection contention in the event of a failure, by
running SDH-layer traffic unprotected in the optical transport
network between the nodes which add/drop the SDH traffic.

wkg

Figure 1 SDH over DWDM planning issues: The SDH

ring must be routed in the optical layer as a distinct
untwisted ring too. Distances shown in km.
To maintain survivability for the SDH layer, it is vital that an
SDH ring is routed through the optical layer in a non-twisted
cycle, and that no optical link is used more than once for the
entire SDH ring. Single points of failure must be avoided. In the
example below in Fig. 1, the SDH ring is specified as dropping
at nodes A, B, C, E, and then returning to A (active nodes
marked in grey).
Routing shortest path for each SDH hop causes the path
to loop back to A before it has reached E (shown in dashed
blue). The constraint must be obeyed so that a cycle is used in
the optical layer as well as the client SDH layer. Now the path
will follow the longer outlined route to visit E before returning
to A.
Another solution to this is the joint optimisation of SDH and
DWDM layers simultaneously, rather than specifying the SDH
traffic routings as an input to the DWDM layer tools. Possible
solutions include an iterative sequential method [3], or a
wavelength packing method [4].

2)

Optical rings and shared protection

The key benefit of using optical ring architectures is that

simpler Wavelength Add-Drop Multiplexers (WADMs) can be
used in nodes, rather than the expensive DWDM end terminals
or Optical Cross Connects.
A potential capacity saving benefit appears in the form of
Optical Channel-layer Shared Protection rings (OChSPRings). Other forms of optical layer protection require at
least 50% spare capacity to be reserved. For OMS-SPRings
and OCh-DPRings the figure is exactly 50%.
OCh-SPRings allow the working channels to be reused by
non-overlapping working paths, resulting in a significant
capacity saving of 20-40% [5]. However, as with SDH MSSPRings, the actual figure depends heavily on the traffic
distribution, and also on the number of nodes in the ring. Ring

Figure 2 - IP over DWDM planning issues : Delay

(distance) may become a factor.
planning should therefore focus on generating large rings with
at least 8 nodes, and with as many demands routed on them as
possible (it is the optimal fitting of adjacent-type demands that
gives the best results).
OCh-SPRing network architectures allow the protection of
demands on a per-channel basis, rather than the previous ring
philosophy of switching an entire multiplex section for
survivability. In a mesh network composed of many embedded
and overlaying OCh-SPRings, this survivability technique is
one step towards mesh restoration.
One drawback of this technique may cause problems in
future IP-dominated networks. Corporate class IP traffic may be
protected in the optical layer to increase availability. If two IP
routers are connected using an OCh-SPRing, and the optical
layer has been optimised to save capacity using the above
recommendations (long rings for adjacent traffic), the distance
ratio between protection and working paths will be quite large.
In Fig. 2 an IP client demand from B to D is protected using a
long ring via A, N, M, and H. This demand is ideal to help
improve the capacity utilisation of the OCh-SPRing.
The working path is 271km and the protection path around
the ring is 962km. In the event of a failure on the B to D link,
protection switching will activate the longer path. Any TCP/IP
connections established will be severely impaired as the TCP
transmission windows will decrease due to the longer
propagation delay (latency) [6,7]. The amount of bandwidth
available to TCP/IP traffic on this router hop would divide by a
factor of 962/271 = 3.5 approx. Moreover, this impairment
would persist until the cable cut was repaired.
Circuit-switched traffic under ring-based protection recovers
100% of the communication after switching. For TCP/IP based
traffic running over DWDM channels, we may now be only
able to recover a percentage of the bandwidth available to end
users. It will then be up to the IP layer restoration to adjust its
routing tables to route around the lower capacity link.
However, this may cause performance degradation in other
parts of the IP backbone. Note that this performance

degradation will also apply to routers interconnected by 1+1

path protection, but the effect is likely to be less problematic
since it is expected that the protection path will be the 2nd
shortest path between the two nodes.
C.

M ULTI-PERIOD PLANNING TECHNIQUES

Multi-period network planning is planning across several

years, with a view to producing incrementally a network
capable of carrying all traffic predicted up to the end of the
planning horizon, in a cost-effective way.
A greenfield scenario is assumed prior to Year 1, i.e. there is
no previously installed equipment. This is a valid assumption
for new entrant operators, or for incumbent telcos who are
deploying DWDM in a major way for the first time, since
previous SDH equipment may well become the first client
signals of the new optical transport network. However we
assume a set of candidate links which are available fibre routes
as input to the process. This is a sensible assumption for
practical network planning, even if it does not adhere
completely to a greenfield approach.
Optimising the network involves re-routing traffic away from
shortest paths in order to avoid or delay the deployment of
certain optical line systems on a link. It is also known as Load
Balancing.
Reference [8] summarises the literature for both theoretical
and practical multi-period planning. In [3] several different
optimisation goals are proposed for practical multi-period
optical network planning:

Firstly, an approach to optimise network costs at the

same time for all periods in the plan. This technique has
been called Worst case planning in [9] and would
seem to be very sensible, yet it is highly complicated
since previous years routings affect later years due to
the fact that wavelength reconfiguration is not allowed
for reasons of continuity of service. In this paper
heuristic methods are used to optimise the plan
resulting in a network design for the final year with all
the cumulative traffic. Working backwards from the
plan, the equipment needs for each year can be
determined. Line systems are only installed when they
are due to be lit. Assuming a 5-year horizon, we denote
this method O5.
Another approach is to optimise year-by-year the
network, effectively ignoring knowledge about future
traffic. However, since demands which have been
provisioned in a previous year cannot be re-routed in
order to maintain continuous service, only the new
incremental traffic is available each year to be routed at
minimum cost through load balancing. New systems are
only deployed when absolutely necessary. This
technique shall be called O1+, for iterative optimisation

from year 1 onwards. It is an incremental build-up

approach.
Y1 plan
Backwards
fill earlier

Year 1 traffic

Y2 plan
Y3 plan

Optimise
Year 3

Year 2
Year 3
Year 4 traffic

Optimise
Year 4

Y4 plan

Y5
plan
Year 5 traffic

Optimise
Year 5

Figure 3 - Example of method O3+ for a five-year plan

These techniques will be complemented with a hybrid
approach between that of O1+ and O5. The technique On+,
where n represents a year between 2 and 4, indicates that for
years up to Year n inclusive, the O5 technique is to be used (all
traffic for these years to be routed and load balanced at the
same time). Then, for each year after Year n, the iterative O1+
method shall be applied. Wavelength demands already
deployed from previous years are not permitted to be re-routed
for load balancing purposes, they are considered live and
thus cannot be modified.
For example, the option O3+ will involve 3 main optimisation
steps (see also Fig. 3):
1)
Route and Load balance traffic from years 1, 2 and 3
simultaneously. This gives a plan for Year 3.
2)
Assuming the Year 3 plan to be fixed, introduce the
Year 4 traffic and load balance these demands to give an
optimised Year 4 plan.
3)
Assuming the Year 4 plan to be fixed, introduce the
Year 5 traffic and load balance these demands to give an
optimised Year 5 plan.
4)
Final step: Based on the routings for Year 3, apply
the routings for Year 1 and Year 2 to give plans for the early
years. The systems chosen for Year 3 need to be installed, but
if a system is not due to be lit until Year 3 it can be removed
from the Year 1 and 2 plans.
To summarise, we will test methods O1+, O2+, O3+, O4+ and
O5, but also using a Present Value calculation in all cases since

Total traffic volume

Wavelengths terminated

700
600

598

500
400

338
300
200

190
107

100

60
0
1

Year

Figure 4 - Case study network with link distances in

km

Figure 5 - Case study traffic volume evolution

from previous
period in plan, if any

equipment is deployed over several years. This will give a

fairer representation of total network cost, and these costs will
be compared for the various options.
Inputs

D.

CASE STUDY NETWORK

A case study network was determined using a random

placement of nodes. The distances involved and the level of
traffic indicate a medium-sized national backbone network. See
Fig. 4 for the topology.

Nodes, links,
demands

Outputs
Routings

Allocate demands
to links
Systems, routings

Systemallocations
If current DWDM system
capacities have been
exceeded on a link, upgrade
or stack another system.

The hub node is A, assumed to be the capital city. The

nodes then decrease in order of importance down to node P in
alphabetical order, giving a total of 16 nodes.

Systems, routings, Systems

deployment rules

The designation of suitable links was created by hand.

There are 25 links, and all nodes have a connectivity of at least
2 so that a survivable network can be realised.

Systems, load
Sorted list of
balancing threshold systems

To simplify the analysis, the traffic to be routed composes of

50% unprotected lightpath demands, and 50% 1+1 optical path
protected demands.
The traffic distribution is weighted by the relative
importance of the nodes. A tenfold increase in traffic occurs
over the five years of the scenario. This is equivalent to a 77%
compound annual growth rate percentage, taking the traffic
volume from 60 wavelengths in year 1 to 598 in year 5, as
shown in Fig. 5.

A LGORITHMS EMPLOYED

The analysis is performed using a Java-based optical

network design tool. The network topology is specified either
manually using the GUI or by importing tables. For each
optimisation period required as described in section C, the
algorithm runs as in Fig. 6. The required traffic is introduced,

Look at all systems in

order of increasing
utilisation, up to limit of
LoadBalancing
threshold.

Can next system be

removed or downgraded
by rerouting some
traffuc?

Routings, systems Optimisedroutings

and systems

finished

yes

Rerouterelevant
traffic. Observe
maximum cycle limit.

Remove or
downgrade
system

Optimisedroutings Transponder count

(Non-discounted)

E.

Route traffic using Dijkstra

for unprotected demands,
or minimum cycle for path
protected demands.

Estimate number of
transponders by
examining routings Assumption :
and demand hop count # Transponders per demand = hops - 1

Estimate total

Optimised systems, Non-discounted

system cost taking
system cost
cumulative system
into account system
constants
cost for this period

weightings

Advance to
next period
in design

Figure 6 - Single period design and optimisation step

and routed using the Dijkstra shortest path algorithm for the
unprotected demands, and a minimum cycle implementation for
the protected demands.
A limit on the maximum cycle length is imposed. If a
protected demand exceeds this distance, the demand is routed
link disjoint only if possible, thus pinchpoint nodes are used,
and the number of terminating transponders is increased. For
the case study network, the maximum cycle length was taken to
be 1500km. Carriers often impose a maximum cycle length to
raise the availability figure since long cycles are more likely to
experience dual failures.
DWDM systems are allocated based upon the link
allocations. Systems can be stacked using different fibres in
the link. Three types of system are assumed, of different
capacities: 40, 80 and 160 wavelengths. An in-service upgrade
is also permitted to migrate from 40 to 80 channel systems
when performing a multi-period design exercise. The system
allocation is based on simple lookup table heuristics.
These systems allocations are then questioned by the Load
Balancing module which tries to remove (or downgrade)
systems by re-routing traffic. Systems are considered if their
utilisation is under a certain threshold, or if they are within this
threshold of being downgraded to cheaper systems. Systems
closest to being removed/downgraded are examined first.
For example, using a 30% load balancing threshold, an 80
wavelength system will be considered for removal if less than
24 wavelengths are lit. But, if between 40 and 64 wavelengths
are lit, the system will be considered for downgrading to the
40-channel system.
A persistent re-route heuristic is used to try and re-route all
of the demands on the candidate system using other routes in
order to achieve the removal or downgrading of the system in
question.
The choice of the threshold is important since it allows
network planners to leave unused capacity in a network for
growth in subsequent years. If too high a value is chosen,
many demands will be re-routed on paths longer than the
shortest path. This causes an increase in the utilisation of the
systems carrying the re-routed demands, which may in
following years require an earlier upgrading than would have
been necessary using a lower threshold. Therefore an overoptimised network at Year N may mean that the later years
require several extra line systems.
A qualitative idea of total systems cost is calculated based
on simple rules of thumb for transmission systems. However,
since deployment happens over several years, a pay as you
grow scenario will apply, and a Present Value calculation is
used to compare the various options. Table 1 indicates the
relative costs taken for the DWDM systems. The in-service
upgrade cost is taken as the difference between the 40 and 80
channel costs.

Table 1 - DWDM system relative costs taken

System capacity
40 wavelengths

Relative cost
1

80 wavelengths

2.5 = 1.58

160 wavelengths

2.5

The effect of the load balancing on the number of optical

transponders is also investigated. Since some demands travel
along a path other than the shortest, the average number of
hops increases and so the number of optical pass through
transponders will increase. In a real network this will be
complicated by the fact that some nodes which do not
terminate many wavelengths will be eligible to become
Wavelength Add-Drop Multiplexers (WADMs), a simpler
node architecture where pass through transponders are not
required. In this study though, the levels of traffic and the
rapid growth evolution precluded the consideration of
WADMs due to their limited wavelength add/drop flexibility.
So a simplification of just counting the total network-wide pass
through wavelengths was used to estimate the transponder
requirement.
F.

RESULTS

The five-year scenario has been analysed 15 times since

there are 5 optimisation methods, and for each one, load
balancing thresholds were tried at 10%, 30% and 100% (which
implies that all systems are at least considered for
removal/downgrading).
However, using the O2+ method, the results are very similar
to O1+ and are omitted for clarity.
Initially, a discount factor of 10% per annum is taken.
1)

DWDM Systems requirements

Using the weighted cost scheme of Table 1, the following

table shows the total line system cost:
Table 2 System Cost (PVs with discount rate of 10%)
Year / Load
method bala. %
O1+
10
30
100
O3+
10
30
100
O4+
10
30
100
O5
10
30
100

Y1
24.0
22.0
22.0
24.0
24.0
24.0
35.5
34.5
34.5
35.5
32.5
32.5

Y2
0.0
0.9
0.9
1.1
1.1
2.1
0.0
0.0
0.0
0.0
0.0
0.0

Y3

Y4

Y5

4.7
4.7
3.8
2.4
2.4
0.5
0.5
1.0
0.5
10.3
11.3
11.3

9.5
9.5
10.7
9.9
9.9
10.5
1.7
1.3
1.7
3.2
5.5
5.1

18.7
19.3
19.3
18.7
18.7
18.4
17.1
17.8
16.7
6.2
3.3
3.7

Total
(PV)
56.8
56.4
56.7
56.1
56.1
55.5
54.8
54.6
53.4
55.2
52.6
52.5

The later the main optimisation period (e.g. O5) the greater
the initial investment required in Year 1. This is because higher
capacity systems have been installed in lieu of the 40 channel
systems, in anticipation of the future traffic volumes.
The load balancing step causes an increase in the average
link utilisation. If this causes a system to be operating close to
the capacity limit in one year, the following years demands will
require a new system. This is why there are still significant
systems requirements in the subsequent years following
optimisation, e.g. in year 5 for the O4+ method.
However, these figures are extremely sensitive to the
discount factor chosen, which is by nature a subjective
quantity.
Fig. 7 illustrates using the 30% load balancing threshold,
how the cheapest option varies with the discount factor
chosen, and which option has the lowest cost at that rate is
indicated.
So for discount factor rates less than 17.5% in this case,
optimising the network over the whole time scale of the project
(method O5) gives the lowest overall cost network
implementation. With a discount rate greater than 17.5%, a
year-by-year iterative optimisation of the network gives the
lowest cost network (method O1+). This is similar to the results
found in [10] which suggest that frequent deployment (and
also therefore upgrading) of systems is optimal when in an
environment of large reductions in the price per bit over the
lifetime of the plan.
However, a frequent deployment & upgrade approach such
as O1+ will result in higher Operations and Management costs
since there are more technologies and systems being used
concurrently. It is clearly a trade-off between capital
investment and operating costs.
Case study
Higher C
Lower C

Discounted total system cost

60

Best = O5

55

Best = O1+

50

Best=O5

45
Cost

Best=O1+

40

Best=O5
35

Best
=
O4+

Best=O1+

30
25
0

10
12
14 16
Discount rate (%)

18

20

22

24

Figure 7 - Lowest cost optimisation method as discount rate

varies. (Load balancing threshold = 30%)

Two other networks have also been analysed with the same
volume of traffic. They have different connectivities and traffic
patterns, but similar results are shown in Fig. 7. The higher
connectivity network has a region where the O4+ method
results in the lowest cost. The different gradients and
changeover points are highly dependent on the topology,
connectivity and traffic distribution, as well as the complex
relation between the traffic loads and the DWDM system
capacities. No obvious pattern has been found.
For all these solutions, the network capacity utilisation after
Year 5 is in the range 85-94%.

2)

Optical transponder requirements

Also contributing to the total network investment cost are

the optical transponders required, simplified here by just
looking at the number of pass-through wavelengths, totalled
for all nodes. A Present Value technique is also used since
transponders are only deployed as wavelengths are lit.
Table 3 Transponder cost PVs (10% discount factor)
Year / Load
method bala. %
O1+
10
30
100
O3+
10
30
100
O4+
10
30
100
O5
10
30
100

Y1

Y2

Y3

Y4

Y5

192
209
209
195
195
202
194
195
195
202
213
213

135
143
143
131
131
141
134
145
145
140
145
145

235
242
255
233
233
256
232
269
236
232
242
242

374
424
426
371
371
362
363
336
366
363
361
361

686
686
698
686
686
744
688
678
712
678
678
678

Total
(PV)
1622
1704
1731
1616
1616
1706
1611
1624
1654
1615
1640
1640

The higher the load balancing threshold percentage, the

greater the number of transponders required, due to more
demands being routed on longer paths. This is the penalty
paid for removing line systems. These results though are much
less sensitive to the discount rate used, due to the stronger
spreading over the years since transponders need to be
installed in each period to accommodate the newly deployed
optical channels. In comparison, the line systems require a
step-function in investment each time one is deployed or
upgraded.
The O4+ technique in this example requires the lowest
number of transponders. This is because the volumes of traffic
at Year 4 compared to the system capacities are such that not
many systems can be removed and so less re-routing occurs.

G.

CONCLUSIONS

The network operators investment strategy, including any

annual budgetary limits, dictates the optimisation technique.
We have shown how optimisation centred on a particular year
in a network planning horizon can affect the DWDM line
systems and transponder requirements. It will be possible to
optimise networks according to aimed investment strategies by
choosing the optimisation year and adjusting the load
balancing thresholds.

[4]
[5]

[6]

[7]

Different traffic distribution and volume evolution

projections will strongly affect the results obtained.
The discount rate percentage used heavily influences the
most cost-effective optimisation option in terms of the line
systems. Unless a very aggressive stance is taken in terms of
the discount factors, it seems worth considering optimising the
network centred on the later years in the plan, e.g. the methods
O4+ and O5 described above. However in the current market
climate of a rapidly decreasing price of bandwidth, it may be
prudent to optimise the network design around the earlier
years of a plan. This approach will also be less risk averse in
terms of the uncertainty of future traffic demands.
Transponder requirements are less sensitive to the discount
rate since they are installed on a per-channel basis compared
to the line system step-function capacity granularity.
H.

FURTHER W ORK

Multi-layer networks complicate the multi-period network

planning process further by introducing constraints, e.g. for
SDH ring routing, or through other issues that cause tradeoffs
for certain desired factors, e.g. an IP network might suffer
unacceptably when an optimised optical layer ring switches
direction in the event of a link failure. A complete multi-layer
and multi-period network optimisation has to be considered to
yield practical network solutions.
I.

A CKNOWLEDGMENTS

This work is part funded by EPSRC via the UCL

Communications Engineering Doctorate Centre. Thanks also to
the reviewers and John Mitchell at UCL who advised on how
to improve the paper.
J.
[1]

REFERENCES

K. Struvye, N. Wauters, P. Falcao Application, Design, and

Evolution of WDM in GTSs Pan-European Transport Network,
IEEE Communications Magazine vol. 38 no. 3, March 2000.
[2]
J. Meijen, E. Varma, R. Wu, Y. Wang Multi-Layer
Survivability (1999), a white paper available at http://www.lucentoptical.com/resources/.
[3]
Planning of Optical Networks, EURESCOM Project P709
Deliverable
3,
March
2000.

[8]

[9]

[10]

http://www.eurescom.de/Public/Projects/p700series/P709/P709.HTM
O. Gerstel, P. Lin, G.Sasaki, Combined WDM and SONET
Network design, IEEE Infocom 1999.
P. Arijs, M. Gryseels, P. Demeester Planning of WDM Ring
Networks, Photonic Network Communications, vol. 2 no. 1,
March 2000.
C. Coltro Implementing a TCP/IP aware survivable
transmission infrastructure, 2nd International Workshop on the
Design of Reliable Communications Networks, Munich, Germany,
10-12 April 2000.
L. Aguirre, C. Phillips, M. Duff, P. Lane, L. Sacks IP over
Optics: Performance Modelling of Optical Shared Protection
Rings, London Communications Symposium, 26th-27th July
1999.
Available
at
http://www.ee.ucl.ac.uk/~tcg/conference/papers99/index.html.
M . Pickavet, P. Demeester, Long-term planning of WDM
networks: A comparison between single-period and multi-period
techniques, Photonic Network Communications, vol. 1 no. 4,
December 1999.
T. Wu, R. Cardwell, M. Boyden, A Multi-Period Design Model
for Survivable Network Architecture Selection for SONET
Interoffice Networks, IEEE Transactions on Reliability, vol. 40,
no. 4, October 1991.
S. Lanning, D. Mitra, Q. Wang, M. Wright Optimal Planning
For Optical Transport Networks, Phil. Transactions of the Royal
Society (London) vol. 358 no. 1773, August 2000.