Sei sulla pagina 1di 5

A Lagrangian-based Heuristic for

Traffic Grooming in WDM Optical Networks


Zenilton K.G. Patrocnio Jr. and Geraldo Robson Mateus
Depart. de Ciencia da Computaca o Universidade Federal de Minas Gerais
30210-150 Belo Horizonte Minas Gerais Brazil
Email: {zenilton, mateus}@dcc.ufmg.br

Abstract Traffic Grooming Problem (TGP) deals with efficiently combining low-speed traffic streams into high-capacity
wavelength channels in order to improve bandwidth utilization
and minimize network cost. In this paper, we investigate TGP
in WDM optical networks regardless of underlying physical
topology. The problem is formulated as an integer linear program
(ILP) and a Lagrangian-based heuristic is proposed. Numerical
results for ring and mesh networks are presented and analyzed.

This paper is organized as follows. Section II presents a


layered graph adopted to represent a WDM network. Then a
mathematical formulation of TGP is presented at Section III.
Section IV presents our Lagrangian-based heuristic, while
Section V shows the results obtained by our approach. Finally,
Section VI concludes the paper.
II. L AYERED G RAPH R EPRESENTATION

I. I NTRODUCTION
Wavelength-division multiplexing (WDM) allows an enlargement of transmission capacity over optical links, because
speed limitations of electronics can be bypassed using different
wavelengths to establish independent channels within the
same optical fiber. However most capacity requirements of
customers connections are lower than the full wavelength capacity (e.g., OC-48, OC-192, and OC-768 in the near future).
So, in order to utilize network resources more efficiently and to
maximize revenue from available capacity, it is very important
for the network operator to be able to combine low-speed
traffic streams into high capacity wavelength channels. This
problem has been called Traffic Grooming Problem (TGP) in
the literature and has been addressed by many authors.
SONET/WDM ring networks were the main focus of
most previous research works [1][3]. These works have
explored the use of heuristic techniques in order to solve
TGP. More recently WDM mesh networks have gained more
attention because they are more flexible and scale better than
SONET/WDM ring networks. TGP in WDM mesh networks
was explored in [4], [5].
In this paper, TGP in WDM optical networks is explored
regardless of underlying physical topology. Our main concern is to obtain optimal (or near-optimal) solutions for this
problem. In order to do that, TGP was formulated as an
integer linear program (ILP). We also present a layered graph
representation of the problem which was used to formulate
the problem. In order to solve the problem, we proposed and
tested a Lagrangian-based heuristic, which uses Lagrangian
relaxation to generate lower bounds for TGP along with valid
cut inequalities that are added to strength the formulation and,
consequently, to achieve tight bounds. Our approach employs
also an efficient heuristic to generate upper bounds (feasible
solutions) based on lower bound solutions and a subgradient
search procedure to maximize the lower bound value.

GLOBECOM 2003

Although an undirected graph is enough to describe basic


information about network topology, it does not allow the
representation of internal components of network elements
and their relationships. This kind of extended representation
is necessary for modelling TGP. So, each node of the original
network will be divided in several others (Figure 1). Associated with each node of the original network there will be
always a node (a.k.a. add/drop node) to represent the add/drop
mechanism of each network element, e.g., the electronic part
of it. Moreover, for each I/O port of a node in the original
network, a pair of nodes will be added (one standing for the
input port and the other for the output port).
These nodes are called optical nodes, and all optical nodes
that belong to the same network element will be connected
to the add/drop node of the same element. These connections
will be used to represent the add/drop multiplexers and will
be shown as solid lines leaving or entering the add/drop node
in Figure 1. Finally, every optical input node belonging to
the same network element is also connected to every optical
output node (except to the one that represents the same I/O
connection), in order to represent the capacity of routing
an wavelength without any OEO (optical-electro-optical)
conversion. In Figure 1, these connections are shown as dotted
lines between optical input and output nodes.
Based on that extended representation of each network
element, a layered graph will be constructed as follows. Each

- 2767 -

Optical
Input Node

Network
Node

Optical
Output Node
Add/Drop
Node

Optical
Output Node

Fig. 1.

Optical
Input Node

Node splitting

0-7803-7974-8/03/$17.00 2003 IEEE

(a) Two-node network

(b) Extended representation

Actually, P is a multiset, i.e., P can contain equal elements.


So, there may be many distinct products (or traffic streams)
which share the same origin and destination, and also have
the same volume and maximum number of channels per
wavelength. Moreover, the following variables are used in TGP
formulation :
fp amount (in channels) of product tp P which
passes through arc  A;
w indicates the use of arc  A.

k=1

Add-Drop
Node

Add-Drop
Node

k=2

(c) Layered representation


Fig. 2.

Layered representation example

wavelength will be represented by a distinct layer. The optical


nodes of the extended representation of all network elements
will be replicated in each layer, as well their connections (arcs
among them). On the other hand, each add-drop node will be
kept unchanged and will not be associated with any layer in
particular. The links between these nodes and the duplicates
of optical nodes will be established as follows : for each
arc from an add-drop node i to an optical node j in the
extended representation there will be a distinct arc between
the correspondent add-drop node of the layered representation
and each duplicate of the optical node j. A similar procedure
will be adopted for each arc from the optical node j to the
add-drop node i in the extended representation.
Consider a simple topology (undirected graph) for a twonode network shown in Figure 2(a). After adding optical
nodes and their links, the extended representation shown in
Figure 2(b) is obtained. If the number of available wavelengths
is 2, one should duplicate the set of optical nodes and
their interconnections, as well, in order to obtain a layered
representation. Moreover, add-drop nodes should be connected
to the respective optical nodes in each layer. Figure 2(c)
shows the layered representation generated from the extended
representation of Figure 2(b) considering an availability of 2
wavelengths. In that figure, optical nodes belonging to the
same layer are grouped by a dotted rectangle.
III. M ATHEMATICAL F ORMULATION
The following notation will be used in the mathematical
formulation of TGP associated with a layered representation
G(N, A) :
N
set of nodes, N = N e N o , N e is the set of
add/drop nodes and N o is the set of optical nodes;
A
set of arcs, A = Ae Ao , Ae is the set of arcs
between add/drop nodes and optical nodes and Ao
is the set of arcs between optical nodes;
P
set of products, P = {tp = (sp , dp , fp , mp ), 1
p np }, np is the total number of products;
sp source node of product tp P ;
dp destination node of product tp P ;
fp demand (in number of channels) of product tp P ;
mp maximum amount (in number of channels) of product tp P that can be transported by a wavelength;

GLOBECOM 2003

allocation cost of arc  A;


wavelength capacity, = lcm1 { mp | tp P };
contribution (i.e., occupation) associated with a
flow unit in a wavelength, p = /mp , tp P .

c

Finally, A+ (i) is used to represent the set of arcs which


leave from a node i N , while A (i) represents the set
of arcs which enter in a node i N . So, an integer linear
program for TGP will be given by :

c w
(1a)
(T GP )
min
A

subject to:


fp
fp =
A+ (i)

, tp P, i N (1b)

, i N o

(1c)

p fp

w

,  A

(1d)

fp
w

0
{0, 1}

, tp P,  A (1e)
,  A
(1f)

integer

, tp P,  A (1g)

A (i)

w

A+ (i)

bpi

w =

A (i)

tp P

fp
in which bpi is given

fp
fp
bpi =

by :
, if i = sp
, if i = dp
, if i =
 sp = dp

, tp P, i N .

The objective function (1a) minimizes the total allocation


cost. Constraints (1b) guarantee flow conservation for adddrop nodes and optical nodes. Constraints (1c) guarantee
wavelength conservation (or continuity). Constraints (1d) set
an upper bound over the total amount of traffic transported
in a wavelength over an arc  A, i.e., a limit over the
wavelength capacity. Finally, constraints (1e), (1f) and (1g)
define flow variables (fp ) as non-negative integer ones, while
decision variables (w ) are defined as binaries.
IV. L AGRANGIAN - BASED H EURISTIC
A. Lower Bound
A lower bound for TGP formulation was obtained by using
Lagrangian relaxation. So, a vector of Lagrange multipliers
,  0,  A, was associated with the constraints that
limit arc capacity constraints (1d) which was then relaxed.
1 least

- 2768 -

common multiplier.

0-7803-7974-8/03/$17.00 2003 IEEE

This results on the elimination of the link between flow


and decision variables. Moreover, flow variables of distinct
products become also independent.
So, a lower bound, L, could be formulated as :
L() = Lw () + Lf ()

Lw () = min

(c  ) w

A

subject to: (1c) and (1f).



p  fp
Lf () = min
tp P A

subject to: (1b), (1e) and (1g).


The value of first subproblem, Lw , could be obtained by
solving a minimum cost flow network problem. But first, one
should add the following constraints valid cut inequalities
to the formulation of subproblem Lw :


w i , i N e

A+ (i)

w i , i N e

A (i)

where i and i , i N e , represent the minimum number


of arcs needed to transport all the traffic that goes out and in,
respectively, of an add-drop node, whose values are given by :


f
f
p
p
, i =
, i N e
i =

tp P |sp =i mp
tp P |dp =i mp

A

(2)

tp P

 Cp

B. Upper Bound

in which :

Finally, for a given set of Lagrange multipliers , the total


value of the lower bound will be given by :




(6)
(c  ) +
p fp

L() =

Let w
 be the value of decision variables obtained by solving
this minimum cost flow network problem, and A = {
A|w
 = 1}. So the value of the first subproblem, Lw , will
be given by :

(c  )
(3)
Lw () =

A

On the other hand, the second subproblem, Lf (), could be


further decomposed into np independent subproblems, i.e. :

Lf () =
p Lpf ()
(4)
tp P

in which, for each product tp P , Lpf () could be solved by


shortest path calculation between the source sp and destination
dp of each product. Let C p be the set of arcs of the shortest
path between sp and dp , calculated using as node distances
(or arc costs). Therefore, a solution for the second subproblem
is obtained by routing the traffic fp of each product tp P
through its shortest path C p . Let fp be the values of flow
variables obtained by this way. So the value of the second
subproblem, Lf (), will be given by :





p
p
 f =
p fp

(5)
Lf () =
tp P

A

GLOBECOM 2003

tp P

 Cp

In order to obtain an upper bound (a feasible solution)


for TGP, an heuristic was developed. Actually, it is a multicommodity extension of Busacker-Gowen method for solving
minimum cost network flow problems.
Let A0 be the set of arcs used to generate an upper bound.
Initially it could be set to A (arc subset selected during the
 = 1} (or
lower bound computation), i.e., A0 = { A | w
set by another heuristic). First, each product will be routed
through the shortest path between its source and its destination,
using only those arcs belonging to A0 . This should be done in
descending order of the ratio fp /mp . During this procedure
arc capacities are updated, i.e., reduced. Actually, for each
product the shortest path with non-zero capacity between its
source and its destination is calculated; and used for routing.
If the path capacity is not enough to transport all the traffic,
another path is calculated and used. These steps are repeated
either until all traffic has been routed or until one cannot find a
non-zero capacity path between the source and the destination
of a product. In the latter case, the procedure fails and the
heuristic is interrupted without producing a feasible solution.
If an initial routing is found, it may violate wavelength
conservation constraints (1c) and, consequently, it may not
be feasible. In order to transform an initial unfeasible routing
to a feasible one, one should notice that only one arc leaving
from an optical input node could transport any traffic. So, for
each optical input node that violates those restrictions, one
should route all its traffic towards the corresponding add-drop
node before allowing it to continue to its final destination. An
equivalent reasoning can be made for optical output nodes,
and a similar solution can be adopted.
This heuristic is based on the resolution of several shortest
path problems and, therefore, it is very fast. But it can produce
unsatisfactory results because of excessive use of add-drop
nodes when it is trying to reach feasibility. So, complementary
steps are needed in order to enhance upper bound quality (i.e.,
to reduce its cost). This can be done in a grid approach, e.g.,
trying to merge/eliminate lightpaths and/or exchange their
colors (i.e., wavelengths).
Let (w
 , fp ) be the feasible solution obtained at the end of
the whole process, and A = { A | w
 = 1}. So, the value
is given by :
of the upper bound, L,

=
L
c
(7)

A

C. Subgradient Search Procedure


For each set of Lagrange multipliers, a new lower bound
can be calculated. In order to solve the problem, one should

- 2769 -

0-7803-7974-8/03/$17.00 2003 IEEE

maximize the value of the lower bound, i.e., one should solve
the Lagrangian dual problem P D :
(P D )

max L()

(8)

Subgradient method is a common approach for solving


Lagrangian dual problems. It can be described as follows.
Given a set of Lagrange multipliers, the relaxed problem is
solved, generating a lower bound, and subgradients associated
with relaxed constraints are calculated. Then, subgradients are
used to update the set of Lagrange multipliers, in order to
obtain a better lower bound. This procedure is repeated until
lower and upper bounds are equal (or almost equal) or a duality
gap is detected. For TGP, given a relaxed problem solution
(w
 , fp ), subgradients can be calculated as :

p fp w
 ,  A
(9)
 =
tp P

During k-th iteration, a new set of Lagrange multipliers is


obtained by :


(10)
k+1 = max 0 , k + tk dk
in which tk is a positive scalar (step size) and dk is a
search direction generated from . In this work, the following
adaptation of Crowder rule was used to generate a search
direction :
k + dk1
with = 0.7
(11)
d =
1+
The convergence of subgradient method is closely related
to the step size, tk . And, in this work, we have adopted a
traditional approach :
k

tk = k

L Lk
2

 k 

(12)

in which k is a scalar 0 < k 2, L is an estimative of the


Lagrangian dual problem solution, and Lk represents the value
of lower bound calculated in k-th iteration. Finally,  k  is
a norm, usually euclidian norm, of subgradient k .
The value of k should be updated during the procedure.
First k is set to 1.1 and, if there is no lower bound improvement in 30 iterations, k is divided by 2. But every time
that an upper bound improvement is found, k is set to 1.1
again. Moreover, a simplified procedure was used to obtain a
estimative of the Lagrangian dual problem solution :
k

= Lmin + L
(13)
L
2
in which is a scalar slightly greater than 1 (e.g., = 1.05)
min is the best upper bound found at the moment.
and L
Let gap be the measure of the difference between the best
min
upper bound and the best lower bound, i.e., gap = (L
min , in which Lmax is the best lower bound found
Lmax )/L
so far. Subgradient search procedure terminates when gap
gapmin ,  k  <
1 , k <
2 or k > max iter. If one of the
first two conditions is satisfied, the best upper bound found so
far is considered optimal (or near-optimal) and no further

GLOBECOM 2003

actions are taken. But, if the value of k becomes too small


or the maximum number of iterations (max iter) is reached,
a branch-and-bound framework has to be used.
V. N UMERIC E XPERIMENTS
We have tested our approach on 150 test problems, which
are divided in three sets. Set I contains 90 instances of ring
topologies with 4, 5, 6, 9, 10 and 16 nodes. For each ring,
non-uniform traffic patterns were randomly generated with 40,
80 and 120 products (five instances for each product amount).
All products have unitary demand values (i.e. just one channel)
and the maximum number of products per container is set to
63, which is equivalent to use a STM-1 signal (155.52 Mb/s) to
transport E-1 traffic streams (2.048 Mb/s) in a SDH network.
For a SONET network, we would have set the maximum
number of products per container to 84, in order to represent
the capacity of a OC-3 signal (155.52 Mb/s) to transport T-1
traffic streams (1.544 Mb/s).
We set max iter = 200,
1 =
2 = 1010 , and gapmin =
3
10 . In order to minimize lightpath number and also to
reduce average hop number, the following cost function was
used in our tests :

1000 + k , (i, j) Ae
cijk =
, k C, (i, j) A
10 + k , (i, j) Ao
Our approach was able to find the optimal solution of 85
instances by applying only the subgradient search procedure to
the original problem without doing any branching. So we will
show here only the results of that, as no further improvement
may be obtained by a branch-and-bound framework.
The results obtained by our approach (using a 900 MHz
AMD Athlon processor with 240 MB RAM) are compared
with those attained by the commercial optimization package
CPLEX version 7.0.0 (running in a SunBlade 100 workstation with a 500 MHz RISC processor and 1 GB RAM)
and they are shown in Table I (only for 6-, 10- and 16-node
rings). The first column of that table presents the number of
vertices in the original ring network, while the number of
nodes and arcs of the extended representation associated with
each topology are shown in the second and third columns,
respectively. Fourth column presents the number of products
for each subset. The mean number of iterations and mean time
that CPLEX needed to solve each subset are presented in the
next two columns, followed by the mean number of iterations,
time and gap obtained by our approach (using Lagrangian
relaxation).
As one can easily seen, our approach presents a quite stable
performance, except for some heavily-loaded small instances
(e.g. 6-node ring with 120 products) and some lightly-loaded
large instances (e.g., 16-node ring with 40 products). This
situation seems to be associated with the set of valid cut
inequalities used to improve the value of the lower bound,
whose quality is strongly associated with those cuts. In these
cases, they are not strong enough to promote a great lower
bound improvement. We should also mention that CPLEX was
not able to find optimal solution for most heavily-loaded large

- 2770 -

0-7803-7974-8/03/$17.00 2003 IEEE

TABLE I
R ESULTS FOR RING TOPOLOGIES ( MEAN VALUES )
|V |

|N |

|A|

|P |

CPLEX
Time
(sec)
1763
9.36
3329
39.68
56959
518.03
4545
31.48
6510
85.64
1250768 8956.44
146905
798.73
13882
181.18
141970
(7.18%)
Iter

30

48

10

50

80

16

80

128

40
80
120
40
80
120
40
80
120

TABLE III
R ESULTS FOR PAN -E UROPEAN N ETWORK

Lagr. Relaxation
Iter
Time
Gap
(sec)
(%)
16
0.63
0.03
125
10.39
0.07
192
27.41
5.98
18
1.41
0.02
97
15.61
0.08
183
46.48
0.58
133
19.64
10.50
115
35.56
0.08
173
83.73
2.83

instances (16-node with 120 products). Actually, we use a time


limit of 12 hours and no improvement was obtained after the
first hour. So we decided to show the results obtained after the
first hour, and the value enclosed with parenthesis represents
the mean gap after that time.
Set II contains 50 instances of mesh topologies with 6,
10, 15 and 20 nodes. The six-node and fifteen-node mesh
topologies have first appear in [5], while the twenty-node one
was used in [6]. Non-uniform traffic patterns were randomly
generated in the same way as we did for rings and we set
max iter to 400 and 600 for 15 and 20-node meshes with
120 products, respectively. Table II presents the results for
mesh topologies. It has the same columns as Table I, except
that in its first column the number of vertices and edges in
the original mesh topology are shown together (one above
the other). Again, the results are quite good, even for large
instances as 20-node mesh with 120 products. For this set, our
approach was able to find the optimal solution of 31 instances,
while it found near-optimal solutions (gap 1%) for the
remaining instances (19 problems). Heavily-loaded large mesh
instances seem be to very hard and CPLEX failed trying to
solve them. Actually, CPLEX was not even able to find a
feasible solution for most of those instances in the first hour
of processing (gap = ).
Finally, set III contains 10 instances based on a real-life
topology of a pan-European optical transport network used
in [6]. Table III presents detailed results for each of those
instances and also the mean values for each product subset.
TABLE II
R ESULTS FOR MESH TOPOLOGIES ( MEAN VALUES )
|V |
|E|

|N |

|A|

6
8

38

76

10
24

58

108

15
21
20
48

99

230

116

216

|P |

CPLEX
Time
(sec)
5458
37.24
10817
158.99
330777 3317.78
8358
72.15
33238
547.33
430800 5478.31
108941 2633.58
108620
()
141153
(2.19%)
74665
()
Iter.

40
80
120
40
80
120
80
120
80
120

GLOBECOM 2003

Lagr. Relaxation
Iter.
Time
Gap
(sec)
(%)
28
1.58
0.07
121
13.70
0.09
200
39.30
0.26
36
3.97
0.05
156
30.52
0.08
189
62.70
0.56
150
78.88
0.10
319
255.64 0.33
194
123.28 0.33
481
431.79 0.13

|V |

|E|

|N |

|A|

|P |

20

78

176

528

80

120

Number
of
Instance
1
2
3
4
5
Mean
1
2
3
4
5
Mean

Lagr. Relaxation
Iter.
Time
Gap
(sec)
(%)
199
278.31 0.98
70
107.61 0.91
53
88.64 0.97
68
106.21 0.93
69
114.69 0.92
92
139.09 0.94
77
225.07 0.90
600
1423.82 5.01
69
176.67 0.94
137
332.41 0.41
346
748.98 1.00
246
581.39 1.65

CPLEX was not able to find any feasible solution after one
hour of processing, so its results are omitted. Once more, the
results are quite impressive. Although those instances have a
huge number of variables and constraints, most solutions are
near-optimal and, except for one instance, the time spent by
our approach was not so significant.
VI. C ONCLUSION
In this paper, TGP in WDM optical networks is explored
regardless of underlying physical topology. We present a
Lagrange-based heuristic to solve that problem, which was
able to find optimal (or near-optimal) solutions for most of
the test problems. Actually, optimal solutions were found for
144 instances (32 near-optimal) with no need of a branchand-bound framework.
We should also mention that some heavily-loaded small
ring instances and some lightly-loaded large ring instances
presented a large gap (5 25%). This situation seems to
be related to the fact that lower bound quality is strongly
connected to that set of cut inequalities used to improve its
value. In those cases, cut inequalities used were not tight
enough to improve lower bound value.
As future work, it seems to be interesting to investigate the
use of lifting techniques (or even the adoption of other sets of
cut inequalities) in order to cope with those situations.
R EFERENCES
[1] O. Gerstel, R. Ramaswami, and G. Sasaki, Cost-effective traffic grooming in WDM rings, IEEE/ACM Transactions on Networking, vol. 8, no. 5,
pp. 618630, 2000.
[2] W. Cho, J. Wang, and B. Mukherjee, Improved approaches for costeffective traffic grooming in WDM ring networks: Uniform-traffic case,
Photonic Network Communications, vol. 3, no. 3, pp. 245254, 2001.
[3] R. Dutta and G. Rouskas, On optimal traffic grooming in WDM rings,
IEEE Journal on Selected Areas in Communications, vol. 20, no. 1, pp.
110121, 2002.
[4] A. Lardies, R. Gupta, and R. Patterson, Traffic grooming in a multi-layer
network, Optical Networks Magazine, vol. 2, no. 3, pp. 9199, 2001.
[5] K. Zhu and B. Mukherjee, Traffic grooming in an optical WDM mesh
network, IEEE Journal on Selected Areas in Communications, vol. 20,
no. 1, pp. 122133, 2002.
[6] T. Cinkler, D. Marx, C. Larsen, and D. Fogaras, Heuristic algorithms for
joint configuration of optical and electrical layer in multi-hop wavelength
routing networks, in Proc. IEEE INFOCOM, 2000.

- 2771 -

0-7803-7974-8/03/$17.00 2003 IEEE

Potrebbero piacerti anche