network dimensioning in wdm-based all-optical networks
TRANSCRIPT
Photonic Network Communications, 2:3, 215±225, 2000
# 2000 Kluwer Academic Publishers. Manufactured in The Netherlands.
Network Dimensioning in WDM-Based All-Optical Networks
Chun-Ming Lee, Chi-Chun R. Hui, Frank F.-K. Tong, Peter T.-S. Yum
Department of Information Engineering, The Chinese University of Hong Kong, Shatin, NT, Hong KongE-mail: [email protected], [email protected], [email protected]
Received September 17, 1999; Revised February 2, 2000
Abstract. The problem of network dimensioning, which involves optimal network designs using minimal resources for a given well-predicted
traf®c between individual nodes but without a pre-determined network topology, is analyzed in details in this paper. Such optimal designs are
critical as the network resources are directly related to the cost of implementation. The problem complexity increases as the traf®c between
every node pair varies in a periodic manner. Furthermore, the presence of wavelength con¯icts makes the network-dimensioning problem
distinct and more complicated than that for traditional circuit-switching networks. Here, we adopt the approach based on a multi-commodity
¯ow problem. A general cost function covering the resources of all system components is formulated, and two solution approaches are
presented; one based on integer programming and one on heuristics.
Keywords: WDM networks, network dimensioning
1 Introduction
Wavelength routing networks [1,2] that are based on
wavelength division multiplexing (WDM) has
become one of the most viable architectures to
achieve a scalable, high-capacity all-optical network.
Although suffering from complex architecture and
management control, the wavelength routing network
has many advantages over the simple broadcast-and-
select network [3] architecture in wavelength utiliza-
tion, where the same wavelength can be used for
multiple connections in spatially disjoint parts of the
network. The concept of wavelength reuse not only
can expand the network capability in its scalability in
number of nodes and distances, but the small limited
number of wavelengths required relaxes the require-
ments on individual electro-optics components,
making the all-optical network more realizable.
Already, several large-scale WDM networks are
currently being implemented in U.S. [4,5], Europe
[2], and Japan [6].
While there is substantial progress and many
studies on different aspects of wavelength routing
networks are made, the network dimensioning
problem, i.e., how to maximize the use of the limited
wavelength resources for different traf®c patterns in
different time segments, is not completely understood.
A core problem in network dimensioning is optical
path assignment. Optical path assignment algorithms
based on integer linear programming for some
wavelength-routing-network con®gurations were pre-
viously reported [7,8]. The network con®guration of
our interest is a variation of the considered wave-
length routing networks [1±6] with an intention to
reduce the size of the wavelength cross-connects and
to improve the network ef®ciency. In addition, results
derived from the network dimensioning analyses yield
essential information on optimal network planning.
This can be translated to minimal network imple-
mentation cost.
Two solution approaches, one based on integer
programming and the other based on heuristic, will be
presented. In our analyses, we only assume the
knowledge of the number of required network nodes
and the inter-node traf®c patterns. Although both
approaches can be applied to wavelength routing
networks with and without wavelength conversions,
we restrict our analyses to the latter con®guration
because the former is a simpli®ed case of the latter [7].
Due to the immense size of the problem, only outlines
of the solution are given for the integer programming
approach based on a multi-commodity ¯ow model.
An algorithm based on minimum variance is
established for the heuristic approach, and its results
are compared with another heuristic based on shortest
path. As will be discussed in a later section, the
variance considered here refers to the number of
wavelengths used between any two connecting nodes.
We show that the algorithm based on minimum
variance performs considerably better than that
provided by the shortest path by reducing the
wavelength con¯icts along various paths in the
network. An illustrative example is also given for
the comparisons of the two algorithms.
This paper is divided into the following sections.
The wavelength routing network architecture is
brie¯y outlined in Section 2. A detailed presentation
of the network-dimensioning problem is given in
Section 3. Both integer-programming- and minimum-
variance-algorithm (heuristic) approaches are
described in Section 4 and an illustrative example is
given. Section 5 summarizes this work.
2 Wavelength Routing Network Architecture
The wavelength-routing network consists of a set of
switching nodes interconnected by ®ber links com-
posed of multiple single-mode optical ®bers
comprising a number of wavelength channels. In
addition to the standard con®guration where all input
and output ®bers are connected to wavelength cross-
connects (WCC) [9], we also consider a variation of
this con®guration where some input ®bers are
terminated directly at the destination nodes without
going through the WCC. Thus, there are two types of
®ber links, transit and direct. Transit-®ber links
terminate at the WCC, while direct-®ber links
terminate at the receiver module [10] of the
destination node. Wavelength signals can also be
directly sent from the transmitter module [11] to the
WCC of the adjacent nodes. Direct links are justi®ed if
there is little variation in the traf®c pattern between
two adjacent nodes. The main advantage of direct
®ber links is a reduction in the dimension or the
complexity of the WCC. For w incoming ®bers
consisting of w1 transit and w2 direct ®bers
�w1 � w2 � w�, the dimension of the WCC in our
proposed con®guration is w21, instead of �w1 � w2�2 of
the standard con®guration. The node architecture is
given in Fig. 1. As shown in the ®gure, signals carried
by various wavelengths l1 . . . lm in the transit ®bers
are switched and routed by a WCC to speci®c
outgoing ®ber links. Any switching between a transit
®ber to a direct ®ber can be assisted by an optical
switch.
Optical connections between any two nodes are
formed by a set of concatenated common wave-
lengths, or lightpaths, between the nodes. For example
in Fig. 2, a lightpath between nodes 1 and 5 can be
Fig. 1. Proposed node architecture. Tx is a ®xed-tuned transmitter module and Rx is a ®xed-tuned receiver module, and SW is an optical switch.
C. Lee, C.R. Hui, F.-K. Tong, P.T.-S. Yum/Network Dimensioning216
established by occupying wavelength lk on transit
®ber links (1,3), (3,4) and direct ®ber link (4,5),
assuming the wavelength lk is available along all
these links. Nodes 3 and 4 serve as the switching
nodes, and nodes 1 and 5 are the corresponding source
and terminal nodes.
3 Network Dimensioning Problem
3.1 Network Dimensioning IssueThe traf®c pattern in a network is usually periodic in
nature. Within a certain period, we can divide the time
into contiguous segments such that the traf®c in each
segment is approximately constant. A typical period
could be a 24-hour day or a seven-day week.
Assuming we are free from a given network topology
(e.g., network planning), the network dimensioning
problem can then be stated as follows: for a given set
of traf®c patterns (one for each time segment),
determine the lightpath assignments for different
time segments such that the utilized resources are
kept at the minimum. The network resources include
wavelength channels, optical ®bers, optical ampli-
®ers, wavelength cross-connects, ®xed tuned
transmitters and ®xed tuned receivers, optical
switches, and other components necessary in the
network, all of which can be translated to the system
implementation cost. The complexity of this problem
can be reduced by solving the two following inter-
connecting problems:
1. Fiber required to be installed between adjacent
nodes;
2. Optical path assignment for different traf®c
pattern on different time segments.
The lightpath assignment is further being complicated
by the presence of a wavelength con¯ict, which
occurs when there is no common wavelength
available along all links connecting the source and
destination nodes. This wavelength con¯ict problem
can be resolved for networks with wavelength
conversion, which is a simpli®ed case of the network
without the conversion capability.
3.2 Problem settingFig. 2 shows a 5-node wavelength routing network.
Let the traf®c rate between the node pair �i; j� in time
segment h be tij�h� and the corresponding traf®c
matrix between all node pairs be T�h� � �tij�h��. Let Hbe the number of time segments in a period and let
fT�h�g be the set of H traf®c matrices. Given the
location of the switching nodes and the set of traf®c
matrices fT�h�g, a general way of expressing the
network dimensioning problem is to assign capacities
to a set of candidate links of the network such that the
blocking performance requirements are satis®ed
throughout the period at minimum resources.
An exact treatment of this problem is very
complicated. To circumvent the complexity of the
problem, we adopt the static approach which allows
®xed lightpaths be established between the node pairs
for the whole time segment. We assume the arrival of
lightpath connection requests is a known random
process and the connection holding time is a known
random variable. Therefore, for a given blocking
performance requirement, the set of traf®c matrices
fT�h�g can be transformed to a set of lightpaths
requirement or demand matrices fD�h�g where
D�h� � �dst�h�� and dst�h� is the number of lightpaths
required between node pair �s; t� in time segment h.The transformation can be performed analytically if
the arrival process and/or the connection holding time
are of the Markovian type. Otherwise, computer
simulation would give very accurately the number of
lightpaths required for a given blocking performance
requirement. Two solution approaches, one based on
integer programming and the other on heuristic, are
described in the following section.
Fig. 2. An example of a 5-node wavelength routing network.
C. Lee, C.R. Hui, F.-K. Tong, P.T.-S. Yum/Network Dimensioning 217
4 Two Solution Approaches
4.1 Integer Programming FormulationLet N � f1; 2; 3; . . . ; ng be the set of switching nodes.
To facilitate the representation of the transit ®ber links
and the direct ®ber links to be set up in the network, an
augmented network model is created by incorporating
a set of conjugate nodes N� � fn� 1; n� 2; . . . 2ng,where the conjugate node of node i is denoted as node
n� i. Conjugate nodes are used as the termination
points of direct ®ber links whereas regular switching
nodes are those connected by transit ®ber links. An
example of the augmented network for the network of
Fig. 2 is shown in Fig. 3.
The augmented network allows us to conveniently
formulate the network-dimensioning problem as a
multi-commodity-¯ow problem. To see this, we may
think of the m different wavelengths carried in each
®ber used for the network as m types of commodities.
Establishing a lightpath from node s to node t using
wavelength lk is equivalent to sending a type kcommodity from node s to node t. To establish dst
lightpaths is equivalent to sending dst commodities
from node s to node t, where these dst commodities
can be of any combinations of these m commodity
types as long as the total is dst.
To develop the mathematical model, we ®rst de®ne
Xkij;st�h� as the number of wavelengths lk used (can be
carried by multiple ®bers) from node i to adjacent
node j while establishing lightpaths between the
source node s and the destination node t in time
segment h. The set of the number of wavelengths,
fXkij;st�h�g, has to satisfy the following constraints,
denoted by c.
Xm
k� 1
Xn
i� 1i 6� tÿ n
Xkit;st�h� � dst�h�; �1�
Xm
k� 1
X2n
i� 1i 6� s
Xksi;st�h� � dst�h� �2�
X2n
l� 1l 6� s
Xkil;st�h� �
Xn
l� 1l 6� s
Xkli;st�h� i [N
and i 6� s; 1 � k � m; �3�Xk
ij;st�h� � 0 V n� 1 � j � 2n
and j 6� t; i [N; 1 � k � m; �4�Xk
ij;st�h� is a non-negative integer �5�
where
s [N; t [N� and 1 � h � H:
These constraints can be interpreted as below:
1. Constraint 1 states that in time segment h, the
total number of wavelengths assigned to a
source-destination pair �s; t� on all direct ®ber
links terminating at the receiver modules of
node t should be greater than or equal to dst�h�,the required number of lightpaths between node
s and node t in time segment h.2. Constraint 2 states that for a source-destination
pair �s; t�, the total number of assigned
wavelengths on all output ®ber links of node sshould be greater than or equal to dst�h�.
3. Constraint 3 states that on a transit node, the
number of input wavelengths lk assigned for a
source-destination node pair �s; t� should be the
same as that of the output.
4. Constraint 4 states that no wavelength will be
assigned for a node pair �s; t� on direct ®ber
links other than those originating from node sand ending in node t.Fig. 3. An example of an augmented network.
C. Lee, C.R. Hui, F.-K. Tong, P.T.-S. Yum/Network Dimensioning218
Next, we are ready to derive the ®ber link
requirements of the network in terms of fXkij;st�h�g.
Any set fXkij;st�h�g satisfying the constraint set c, i.e.,
for any lightpath connection plan satisfying the
demand fD�h�g, the number of wavelengths lk used
for establishing lightpaths between adjacent nodes iand j, is given by
Fkij � max
h
X�s;t� [Oij
Xkij;st�h�
8<:9=;1 � i � n; 1 � j � 2n;
i 6� j; 1 � k � m;
where
Oij � f�s; t�js [N; t [ N�; s 6� j; t 6� i; t 6� n� sg:
Thus, the number of ®ber links Lij needed between
nodes i and j is
Lij � maxkfFk
ijg:
Now, we formulate the minimum required transmitter
and receiver modules for each node. For a given node
i, the number of transmitter modules Si needed
(number of local ®ber link inputs to the WCC plus
the number of ®bers directly going out, see Fig. 1)
while satisfying fD�h�g is given by
Si � maxh
maxk
X2n
i� 1
X2n
t� n� 1
Xkij;it�h�
( )( )i [N;
1 � k � m; 1 � h � m:
Furthermore, we obtain the following ®ber require-
ments. The number of transit ®ber links, Ti, passing
through node i is
Ti �Xn
k� 1k 6� i
Lki i [N:
The number of direct ®ber links Ri terminating at node
i equals to the number of receiver modules required,
and is given by
Ri �Xn
k� 1k 6� i
Lki i [N�:
The total number of outgoing transit and direct ®ber
links Oi is
Oi �X2n
k� 1k 6� i;k 6� i� n
Lik i [N;
from which the dimension of the WCC is obtained as
�Si � Ti�6Oi in order to satisfy fD�h�g.The cost function Z for network resource optimiza-
tion consists of ®ber-link cost C�ij�L linking between
node pair �i; j�, WCC cost given by CS Si���Ti�6Oi�, transmitter-module cost CT , receiver-
module cost CR, and optical ampli®er cost CA. Note
that the installation component in the ®ber link cost is
substantially large part of the total cost of each
installation. Careful planning is needed to avoid
repeated ®ber installation. Minimizing Z yields
Minimize Z �Xn
i� 1
X2n
j� 1j 6� i; j 6� i� n
fC�ij�L � LijCAg
�Xn
i� 1
Cs��Si � Ti�6Oi�
�Xn
i� 1
SiCT �X2n
i� n� 1
RiCR
subjected to constraint set c. Unfortunately, the exact
solution of the above optimization problem is
formidable, even for very small network sizes. A
®rst approximation can be obtained by relaxing the
integer requirement on fXkij;st�h�g and treating them as
non-negative real variables. We may then apply the
standard non-linear programming technique to obtain
a solution and round-off the real numbers to integers.
However, this approach still requires handling of
a large number of variables. Besides, the number
of ®ber links between two nodes is usually too small
for a signi®cant reduction of the rounding error.
A second approximation is to solve the dimen-
sioning problem H times, once for each D�h� matrix,
thereby reducing the number of variables by H times.
The largest number of ®ber links is then chosen
among the results generated. But this ``time local''
approximation could generate grossly inferior solu-
tions. Given this as background, we propose a more
realistic heuristic solution as follows.
C. Lee, C.R. Hui, F.-K. Tong, P.T.-S. Yum/Network Dimensioning 219
4.2 Heuristic Solution
4.2.1 Minimum Variance Algorithm (MVA)We assume the total number of routes Ro is an
exponential function of the number of nodes n in the
network. To start with, we choose Ro � a between any
node pair. The total number of the candidate routes is
then restricted to no more than an�nÿ 1� for an n-node
network. Let us denote this set of candidate routes as R.
For a given source and destination node pair �s; t�with traf®c matrices D�h�, we begin by putting an
optical ®ber between each connecting node pair for
every possible route connecting node pair �s; t�. The
total ®ber-link cost is calculated for each of these
routes. For as long as the traf®c demand D�h�st40, we
will keep on assigning lightpaths with unused
wavelengths connecting node pair �s; t� by applying
the Minimum Variance Subroutine (MVS) as
described below in Section 4.2.2. For each possible
route, the average cost of setting up a lightpath is then
evaluated by dividing the total cost of ®ber links on
that route by the number of lightpaths that can be
established by the subroutine. We then choose the
route that yields minimum average cost. This
procedure is repeated until demands in all time
segments fD�h�g are satis®ed.
The following is the pseudo code for MVA.
Initialization
* Construct R, the set of all candidate routes linkingthe node pairs with dst�h�40 of the network.* Set Lij � 0 for all i, j* Set X k
ij;st�h� � 0 for all i, j, s, t, k, h* Set est�h� � dst�h� for all s, t, h, where est�h� is thenumber of unsatis®ed lightpaths between node pair�s; t� in time segment h.
Repeatf 1. For each route r[R,
fa) Increase the number of ®ber links along
each segment of r by one,
L#ij � Lij � 1 if segment�i; j� [ r:
Calculate the incremental cost, DZ, byadding the ®ber links on route r as
DZ � Z�L#ij � ÿ Z�Lij�:
b) Apply MVS with fXkij;st�h�g; fL#
ij g and fest�h�g asinput to obtain an updated lightpath assignment plan.
c) Add up the number of new lightpaths established inall time segments from the wavelength assignmentplan and denote it by u.d) Calculate the cost per lightpath, c � uÿ1DZg
2. Denote the route that gives minimum c as r�
3. Update the link count on r�, Lij/Lij � 1 if�i; j� [ r�, the cost Z � Z � DZjr� and fXk
ij;st�h�g;fest�h�g based on the lightpath assignment plan for r�.4. Identify all source-destination node pairs that havetheir lightpath requirements satis®ed, i.e., those �s; t�pairs with est�h� � 0; � h � H and denote them bythe set of Y.5. Remove all routes belonging to Y from R.g Until all est�h� � 0.
4.2.2 Minimum Variance Subroutine (MVS)The wavelength utilization pro®le of the ®ber links
between all adjacent node pairs �i; j� in time segment
h is given by fXkij�h�g where Xk
ij�h� is the number of
assigned lightpaths between �i; j� using lk in time
segment h, and 1 � k � m. Xkij�h� equals to
Xkij�h� �
Xn
s� 1
X2n
t� n� 1
Xijij;st�h�;
where Xkij;st�h� equals to Xk
ij;�h� under the constraint
that the source and terminal node pair must be �s; t�.The variance of the wavelength utilization pro®le,
V�rst; k; h�, is then
V�rst; k; h� �X�i;j� [ rst
(Xm
i� 1l 6� k
Xlij�h� ÿ Xij�h�
� �2
� Xkij�h� � 1ÿ Xij�h�
� �2);
with respect to k and rst. Here,
Xij�h� � mÿ1 1�Xm
k� 1
Xkij�h�
!is the mean of the wavelength utilization on path
segment �i; j� in time segment h for every new
lightpath assigned. This variance function is the
measure for the uniformity of wavelength utilization
in the network. It can be argued that the more uniform
the pro®le, the fewer the chance the wavelength
con¯ict occurs, thus requiring fewer number of ®ber
C. Lee, C.R. Hui, F.-K. Tong, P.T.-S. Yum/Network Dimensioning220
links established between nodes. A route rst between
node pair �s; t� is feasible if it has a common set of
free wavelengths on all its links of the route and can
be expressed mathematically as:
xkij�h� � 1 � L#
ij for all link segments �i; j� on rst:
Let Rst denote the set of all feasible routes for node
pair �s; t�. We then extend the search for all possible
�s; t� and lk, and determine the route with minimum
variance denoted by fst�h�. This process continues
until no more lightpaths can be setup in any time
segments.
The idea behind this algorithm is that a good
wavelength-path connection plan should make full
use of all the wavelengths on all ®ber links. The ideal
case is that all wavelengths on the ®ber links are used
up for establishing lightpaths, which corresponds to
zero variance in the wavelength utilization pro®les.
The following is the pseudo code for the subroutine
MVS:
Input: fL#ij g Fiber link count between node i
and node jfXk
ij;st�h�g Lightpath connection plansfest�h�g Number of unsatis®ed lightpaths
between nodes s and tfor � h � H dofConstruct U as the set of source-destination nodepairs with dst�h�40
Repeat ffor every node pair �s; t� in U do
f Construct Rst the set of all feasible routes fornode pair �s; t�.For a non-empty Rst, obtain
fst�h� � minVrst [Rst
min1� k�mfV�rst; k; h�g
and the corresponding k and rst valuesg
Select the node pair that has minimum fst�h� forestablishing a new lightpath using the correspondinglk and rst. If the lightpath requirements for this nodepair are satis®ed, remove it from U.Update fXk
ij;st�h�g and fest�h�g.g until no more feasible routes exist for all
node pairsg
A simple counting of the nested loops in the greedy
algorithm and the MVS shows that the complexity
equals to Ha2�n�nÿ 1��3, or O�n6�. Typically for an
optical backbone network of 5±20 nodes, the
dimensioning problem can be solved within a reason-
able time.
4.2.3 Shortest Path Algorithm (SPA)In order to demonstrate how the MVA can maximize
the wavelength utilization in an optical network, we
compare it to another algorithm based on shortest
path. Being itself also a greedy algorithm, the SPA
pseudo codes are similar to that of the MVA. It
establishes the lightpath based on the shortest distance
between the source and destination. Using SPA, the
cost of setting up a lightpath is proportional to the sum
of the costs of ®ber installation including all the links
that the lightpath transverses. Since the cost of
installing ®bers is largely proportional to the distance
between two nodes, the lightpath with the shortest
distance also yields a local minimum cost.
Being itself also a greedy algorithm, the SPA
pseudo codes are identical to that of MVA as
described in Section 4.2.1, with the exception that
the minimum variance subroutine is replaced by the
shortest path subroutine. The SPA pseudo codes are
given as follows:
Input: fL#ij g Fiber links count between node i
and node jfXk
ij;st�h�g Lightpath connection plansfest�h�g Number of unsatis®ed lightpaths
between nodes s and tfor � h � H do fConstruct U as the set of source-destination nodepairs with dst�h�40
Repeat fFor every node pair �s; t� in U do fConstruct Rst the set of all feasible routes fornode pair �s; t�.
If Rst is non-empty, for all rst on which alightpath can be set up with any wavelengthchannel, obtain
fs; t � minV
rst[R
st;k
Pathlen�rs;t�
and the corresponding k and rst, where thefunction Pathlen�rs;t� returns the length ofthe path rs;t.
gSelect the route that has minimum fst�h� for
C. Lee, C.R. Hui, F.-K. Tong, P.T.-S. Yum/Network Dimensioning 221
establishing a new lightpath using the corre-sponding lk and rst. If the lightpathrequirements for this node pair are satis®ed,remove it from U.
Update fXkij;st�h�g and fest�h�g.
g until no more feasible routes exist for allnode pairs
g
4.2.4 An Illustrative ExampleThe minimum cost achieved by the SPA is only a
local optimization and the resources may not be fully
utilized for setting up all the lightpaths. The MVA
can maximize the number of commonly available
wavelengths in the network and thus the chances of
wavelength con¯icts are reduced. Consequently,
MVA maximizes the wavelength utilization in the
network and achieves an overall more ef®cient and
better cost-saving con®guration than that provided by
SPA.
The main advantage of lightpath assignments by
MVA is illustrated by the following example. As
shown in Fig. 4(a), there are altogether ®ve nodes (N1,
N2, . . . N5) interconnected in our illustrated network.
We deliberately set the node distances N3-N2-N5 to be
shorter than N3-N4-N5. We assume that each
interconnecting node pair is linked by one single
®ber with a capacity of carrying up to m � 4
wavelengths and only a single time segment, i.e.,
H � 1, is considered. Table 1 shows the traf®c
pattern.
The wavelengths that ful®ll existing traf®c
demands between node pairs (N2, N3), (N3, N4),
(N4, N5) and (N2, N5) are represented by the number
of occupied wavelengths �l1; l2; l3�, �l1; l2; l3�,�l1; l2; l3� and �l1� respectively, as shown in the
®gure. The demand requests for (N3, N5) and
(N1, N5) in Table 1 are yet to be satis®ed.
To setup an additional lightpath between N3 and
N5, we can select l4 transversing either on route r325
(via nodes N3-N2-N5), or on route r345 (via nodes N3-
N4-N5). Using SPA, the path with the shortest
distance, i.e., r325, will be chosen. Using MVA, the
variances of the two lightpaths r325 and r345 are
calculated, and the path with the smallest variance,
i.e., r345, will be selected. The computations of the
variances are shown in Table 2.
The path assignments by the MVA and the SPA are
given respectively in Fig. 4(b) and Fig. 4(c). Note that
(c)
Fig. 4. (a) Initial wavelength assignments for our illustrative network. (b) Wavelength-assignment after a lightpath is setup by using l4 on the
path r345 when MVA is used. (c) Wavelength assignment after a lightpath is setup by using l4 on the path r325 when SPA is used.
(a) (b)
Table 1. Traf®c patterns for our 5-node illustrative example.
Node Pair Traf®c Demand
N1-N5 3
N2-N3 3
N2-N5 1
N3-N4 3
N3-N5 1
N4-N5 3
C. Lee, C.R. Hui, F.-K. Tong, P.T.-S. Yum/Network Dimensioning222
by selecting path r345, the capacity limit in (N3, N4)
and (N4, N5) is reached and no more wavelengths can
be inserted into these links. There is, however, extra
wavelength capacity available between nodes
(N2, N3) and (N2, N5). For the SPA case, a new
®ber has to be installed to setup three lightpaths
between nodes (N2, N5) since there are only two
commonly available wavelengths l2 and l3 on the
path r125. The assignment of the lightpath by using
MVA may not necessarily be the shortest path, but the
algorithm reserves more wavelengths on the links
connecting N2 and N5, thereby increasing the chance
of ®nding common available wavelengths in future
lightpath assignments.
4.3 Performance ComparisonsWe assume our network has 9 interconnecting nodes.
Different combinations of randomly generated cost
and demand matrices are inputted into MVA and SPA
and the resulting costs are compared and analyzed.
There are altogether three cost matrices C1, C2 and
C3. The traf®c loads of the demand matrices vary
from 200 to 3000, where the loads are de®ned as the
total number of lightpaths required to be set up over
the network. The number of wavelengths in the ®bers,
m, is taken to be 4, 8 and 16. The range of traf®c
demands (number of lightpaths) and normalized ®ber
installation costs between any node pair is between 0
and 10. The costs of components other than optical
®bers are absorbed by the normalized ®ber installation
cost and will not be treated as separate entities. The
number of time segments H is taken to be 3. In our
computation, the cost function of installing ®bers is
similar to the typical installation cost shown in Fig. 5.
Fig. 6 shows the normalized cost as a function of
the traf®c load predicted by the MVA and SPA for
m � 4, 8 and 16 and using cost matrix C1. In general,
both MVA and SPA give a linear increase in cost as
the traf®c demand increases. Also, the cost decreases
as m increases. Over the range of traf®c loads studied,
the projected cost calculated by MVA is uniformly
lower than that predicted by SPA. Similar results are
obtained for different cost matrices C2 and C3.
Table 3 further compares the two algorithms by
evaluating the ratio of the average projected costs q,
q � (average projected cost by MVA)/(average pro-
jected cost by SPA), for different cost matrices and
number of wavelengths per ®ber. The average cost is
generated over different traf®c matrices. In general, qdecreases as m increases. The results are similar to
those obtained in Fig. 6.
For a given traf®c matrix, the costs predicted by
both MVA and SPA differ substantially. The much
larger cost projection by SPA over MVA (20±60%,
see Table 3) suggests that the total number of installed
®bers is also larger in the SPA case. Consequently, the
maximum network capacity scales accordingly, even
though the number of requested lightpaths remains the
same for both cases. This also implies that there are
more unused wavelengths in the lightpath assignment
generated by SPA. We can therefore argue that the
wavelength con¯ict is more serious in the lightpath
assignment generated by SPA than that in MVA.
Table 2. The variances of the two lightpaths.
X12;3 � 1 X2
2;3 � 1 X32;3 � 0 X4
2;3 � 0
X12;5 � 1 X2
2;5 � 0 X32;5 � 0 X4
2;5 � 0
X13;4 � 1 X2
3;4 � 1 X33;4 � 1 X4
3;4 � 0
X14;5 � 1 X2
4;5 � 1 X34;5 � 1 X4
4;5 � 0
X2;3 � 0:75 X3;4 � 1 X2;5 � 0:75 X4;5 � 1
V�r325; 4� � 0:75 V�r345; 4� � 0
Fig. 5. Typical ®ber installation cost as a function of ®ber num-
ber w.
C. Lee, C.R. Hui, F.-K. Tong, P.T.-S. Yum/Network Dimensioning 223
We further investigate how the projected cost
changes as the number of wavelengths increases for
a given ®xed number of installed optical ®bers. Here,
we only consider the case in which the cost is
predicted by MVA, since we have already demon-
strated, MVA is superior than SPA. As the number of
wavelengths (or lightpath) doubles, the traf®c capa-
city of the optical network will also double.
Conversely, the ®ber installation cost will be reduced
by a maximum of 50% under the same traf®c load.
This effect is demonstrated in Table 4, showing that
the projected cost by MVA (about 30±40% cost
reduction) using the cost matrix C1. Similar results
are obtained for cost matrices C2 and C3.
5 Summary
This paper studies the network-dimensioning problem
for an all-optical wavelength routing network. Two
solution approaches, one based on integer program-
ming and the other based on heuristic, are presented.
Fig. 6. Network installation cost comparison for the shortest path (SPA) and the minimum variance (MVA) algorithms using cost matrix. C1
assuming 4, 8, and 16 wavelengths. SPA(4) � r; MVA(4) � ;SPA(8) � m;MVA(8) � j;SPA(16) � * ; MVA(16) � d.
Table 3. The average cost ratio for different combinations of cost matrices and number of wavelengths.
Cost Matrix
Number of Wavelengths C1 C2 C3
4 0.743923 0.494199 0.821689
8 0.646571 0.454908 0.779649
16 0.587841 0.448158 0.756901
Table 4. Assignment results for different combinations of number of wavelengths and traf®c loads using cost matrix C1.
Number of Wavelengths
Traf®c loads 4 8 16
1200 110.6 66.5 44
1500 132.6 77.3 51.2
1800 155 89.3 56.2
C. Lee, C.R. Hui, F.-K. Tong, P.T.-S. Yum/Network Dimensioning224
Solutions of the integer-programming approach
should yield a con®guration with minimum imple-
mentation cost. Due to the immense size of the
problem, only outlines of the solution are given for the
integer programming based on a multi-commodity
¯ow model. An algorithm based on minimum
variance is established for the heuristic, and its results
are compared with another heuristic based on shortest
path. We show that the algorithm based on minimum
variance performs considerably better than that
provided by the shortest path by reducing the
wavelength con¯icts along various paths in the
network.
References
[1] S. B. Alexander, et al., A scalable multiwavelength multihop
optical network: A proposal for research on all-optical
networks, IEEE Journal of Lightwave Technology, vol. 11,
no. 5/6, (May/June 1993), pp. 736±754.
[2] G. R. Hill, et al., A transport network layer based on optical
network elements, Journal of Lightwave Technology, vol. 11,
no. 5/6, (May/June 1993), pp. 667±679.
[3] N. R. Dono, et al., Wavelength division multiple access
networks for computer communications, Journal of Selected
Areas in Communication, vol. 8, no. 6, (August 1990),
pp. 983±994.
[4] A. Robinson, M. J. O'Mahony, Performance of an optical
backplane for switch interconnection, Journal of Lightwave
Technology, vol. 143, no. 4, (April 1996), pp. 237±243.
[5] G. K. Chang, M. Z. Iqbal, C. A. Brackett, Multiwavelength
recon®gurable WDM/ATM/SONET network testbed, C, vol.
14, no. 6, (June 1996), pp. 1320±1340.
[6] S. Okamoto, A. Watanabe, K. I. Sato, Optical path cross-
connect node architecture for photonic transport network,
Journal of Lightwave Technology, vol. 14, no. 6, (June 1993),
pp. 1410±1422.
[7] R. Ramaswami, Multiwavelength lightwave networks for
computer communication, IEEE Communication Magazine,
vol. 31, no. 2, (February 1993), pp. 78±88.
[8] N. Wauters, P. Demeester, Design of the optical path layer in
multiwavelength cross connected networks, IEEE
Communication Magazine, vol. 14, no. 5 (June 1996),
pp. 881±892.
[9] Journal of Lightwave Technology, vol. 11, no. 5/6, (May/June
1993), Entire Issue.
[10] F. Tong, et al., A 32-channel tunable receiver module for
wavelength-division multiple-access networks, IEEE
Photonics Technology Letters, vol. 9, no. 11, (November
1997), pp 1523±1526.
[11] C. E. Zah, et al., High power 10 wavelength DFB laser arrays
with integrated combiner and optical ampli®er, 15th IEEE
Semiconductor Laser Conference (1996), pp. 131±132.
Ronald C. C. Hui received his Master
Degree from the Information
Engineering Department in 1998 from
the Chinese University of Hong Kong.
His current research topics include QoS
routing, policy architecture and high-
speed packet scheduling.
Frank Tong received his Ph.D. from
Columbia University in 1987 with his
thesis work carried out at MIT Lincoln
Laboratories. After graduation, he joined
IBM Almaden Research Center where he
worked on short wavelength lasers for
optical recording. From 1989±1996, he
worked on optical networking technolo-
gies under Paul Green at IBM T. J.
Watson Research Center. He is currently
an Associate Professor with the
Information Engineering Department of
the Chinese University of Hong Kong. Frank has received many
awards from IBM including the IBM Outstanding Innovation Award
in 1995.
Peter T. S. Yum worked in Bell
Telephone Laboratories, USA for two
and a half years and taught in National
Chiao Tung University, Taiwan, for two
years before joining the Chinese
University of Hong Kong in 1982. He
has published original research on packet
switched networks with contributions in
routing algorithms, buffer management,
deadlock detection algorithms, message
resequencing analysis and multiaccess
protocols. In recent years, he branched out to work on the design and
analysis of cellular network, lightwave networks and video
distribution networks. He believes that the next challenge is
designing an intelligent network that can accommodate the needs
of individual customers. Professor Yum's research bene®ts a lot
from his graduate students. Six of them are now professors at local
institutions. His recent hobby is reading history books. He enjoys
playing bridge and badminton.
C. Lee, C.R. Hui, F.-K. Tong, P.T.-S. Yum/Network Dimensioning 225