Professional Documents
Culture Documents
1
se
The rate of a class s S connection on link e E
1
or the number of circuits
required by the connection on the link. Let
1
e
= {
1
1e
, . . . ,
1
|S|e
}. If rate
is link-independent, then
1
se
=
1
s
e E
1
;
K
1
e
Cost of reserving bandwdith on link e E
1
;
W
1
e
Revenue generated by reserved bandwidth on link e E
1
;
K
1
sdr
Cost of losing trac from class s S connections on route r R
1
d
. If the
cost is independent of route, then K
1
sdr
= K
1
sd
r R
1
d
;
W
1
sdr
Revenue generated by class s S connections on route r R
1
d
. If revenue
is independent of route, then W
1
sdr
= W
1
sd
r R
1
d
.
Les Cahiers du GERAD G9851 Revised 10
3.3 Variables
The sets of variables considered in most models are typically the virtual link bandwidth and
the trac oered to each virtual network route. Other possible variables include oered
trac for each virtual link or required bandwidth for routes. The notation is as follows:
C
1
e
Virtual bandwidth for link e E
1
. Let
C
1
= {C
1
e
} e E
1
;
C
1
d
Total virtual bandwidth reserved between O-D pair d D
2
;
A
1
sdr
Class s S trac oered to route r R
1
d
;
a
1
se
Total class s S trac oered to link e E
1
. Let a
1
e
= {a
1
1e
, . . . , a
1
|S|e
};
X
1
sdr
Bandwidth or number of bandwidth units required by class s S trac
on route r R
1
d
.
3.4 Performance measures
Typical performance measures employed are call blocking and packet loss probabilities,
although in some cases average packet delay is used.
Notation used is:
g
1
se
(a
1
e
,
1
e
, C
1
e
) Blocking function for class s S connections on link e E
1
B
1
se
Call blocking probability for class s S connections on link
e E
1
(upper bound: B
1
se
);
B
1
sd
End-to-end call blocking for class s S connections between
O-D pair d D
2
(upper bound: B
1
sd
);
B
1
sdr
End-to-end call blocking for class s S connections on route
r R
1
d
(upper bound: B
1
sdr
);
L
1
sdr
End-to-end packet loss for class s S connections on route
r R
1
d
(upper bound: L
1
sdr
);
H Average packet delay.
Given that call blocking probability is the network performance measure most frequently
employed, a brief introduction to call blocking functions is now given.
We consider a single virtual link e E
1
with a xed capacity C
1
e
in which a number of
calls of diering classes share the link. Class s calls arrive following a Poisson process with
rate
s
and require bandwidth
1
s
during an exponentially distributed holding time with
mean 1/
s
. If there is insucient bandwidth to accept a call, it is blocked and lost. The
Les Cahiers du GERAD G9851 Revised 11
total class s trac oered to link e is a
1
se
=
s
/
s
. If only a single service class is oered
to link e, the blocking function is simply the Erlang-B function E()
g
1
e
(a
1
e
, C
1
e
) = E(a
1
e
, C
1
e
) =
(a
1
e
)
C
1
e
(C
1
e
)!
C
1
e
i=0
(a
1
e
)
i
i!
1
. (1)
For multirate trac, calculating the exact product form solution [16] for the state
probabilities has proven to be intractable. However, a faster exact recursive solution has
been provided by Kaufman [39] and Roberts [70]. This method assumes Poisson arrivals
and allows for general holding time distributions [70]. It also limits the eective bandwidth
and link capacities to integer values. The probability P
n
of having n units of capacity
occupied in a link e having C
1
e
units of capacity is:
P
n
=
1
n
sS
1
s
a
1
se
P
n
1
s
for n = 1, . . . , C
1
e
P
0
for n = 0
0 for n < 0,
(2)
where P
0
is determined by solving the following equation:
C
1
e
n=0
P
n
= 1. (3)
The blocking probability for service class s on link e E
1
is thus given by
B
1
se
=
C
1
e
n=C
1
e
1
s
+1
P
n
. (4)
Since the Kaufman and Roberts method becomes impractical if the number of service
classes or the number of bandwidth units per link is large, several approximations have been
proposed [46, 52, 71]. In what follows, we describe the popular approximation developed by
Labourdette and Hart [46], in which the classes of oered trac are combined to create an
equivalent single class system. This approximation is exact under asymptotic conditions,
that is, when the link capacity and oered trac intensity become very large yet remain
in a constant ratio. The method also allows the eective bandwidth and link capacities to
be continuous.
The equivalent single class trac is characterized by the total oered trac M, the
variance of oered trac V, and the eective bandwidth of the equivalent trac
M =
sS
1
s
a
1
se
, V =
sS
(
1
s
)
2
a
1
se
, =
log
C
1
e
M
log
, (5)
Les Cahiers du GERAD G9851 Revised 12
where is the unique positive root of the following equation in z
sS
1
s
a
1
se
z
1
s
= C
1
e
. (6)
The approximate link blocking probability can subsequently be determined using a
continuous approximation of the Erlang blocking function:
B
1
se
=
1
s
1
C
1
e
M
E
_
M
,
C
1
e
_
if = 1
1
s
M
V
E
_
M
,
C
1
e
_
if = 1.
(7)
In this section, we have assumed that the call admission policy is a complete sharing
policy. Consequently, calls with larger eective bandwidths experience higher blocking
probabilities than calls with smaller eective bandwidths. Partial sharing or trunk reser-
vation schemes, providing for fairness, are also discussed in the literature. For a thorough
discussion on blocking probability calculations for fair systems see [69, 73].
4 Review of models
Using our typology presented in Table 1, we can identify ve distinct model types charac-
terized by their respective control and switching strategies:
1. Centralized independent connection models;
2. Centralized end-to-end path models;
3. Centralized node-to-node path models;
4. Centralized hybrid models;
5. Decentralized end-to-end path models.
For each model type, we rst give a brief introduction to the network and problem data
that is commonly given as input, the variables which must be determined by the optimiza-
tion, and the performance measures required for problem resolution. We then describe
the mathematical formulation for the problem and review the relevant literature. In Ta-
bles 27, we summarize each model type by describing the variables, objective functions,
constraints, solution approaches, and problem features for each paper reviewed. These ta-
bles also include computational problem sizes and running times for those papers in which
the information is available.
4.1 Centralized independent connection models
When independent connection switching is implemented, there is at most one virtual link
between every O-D pair in the virtual network G
1
; therefore, G
1
is not necessarily fully-
meshed. Connections are routed either over the direct virtual link (if it exists) or over a
Les Cahiers du GERAD G9851 Revised 13
Virtual links
Routes
C
e
e
sd
A
2
A-C
r
e
(A,B)
r
r
(A,B,C)
(A,C)
(A,C)
(B,C)
(B,C)
A
B
2
sd
A
B-C
Figure 5: Virtual network for the independent connection model
series of links (i.e., |r| 1 for all r R
1
d
, d D
2
). Given a trac demand A
2
sd
oered by
every O-D pair d D
2
for each class s S and the possible routes R
1
d
between each O-D
pair, the network synthesis problem is to nd the required bandwidth C
1
e
of each virtual
link e E
1
and the class s trac A
1
sdr
that will be oered to each route r between O-D
pair d.
In Figure 5, we illustrate the problem. The A
2
sd
AC
Erlangs of trac oered to O-D pair
A-C can be routed either over the direct route r
(A,C)
or over the two-hop route r
(A,B,C)
.
Trac oered to O-D pair B-C is routed over the direct link e
(B,C)
. Since there is at
most one virtual link between each O-D pair, trac from multiple O-D pairs and classes
are required to share the virtual links (e.g., trac on routes r
(A,B,C)
and r
(B,C)
share link
e
(B,C)
).
Note that models in this section assume that transmission link capacities
C
0
are large
enough to support any amount of virtual link bandwidth. Therefore, the models contain no
constraints that specically limit the virtual link capacities with respect to the transmission
link capacities.
4.1.1 Throughput-based and cost-based optimization Throughput- and cost-based
models typically play with the classical trade-o between the cost of installing or reserving
bandwidth and the revenue generated by carried trac. The problem is modeled as the
minimization of virtual link bandwidth costs Z
1
and the maximization of revenue from
carried trac Z
2
subject to routing and blocking constraints.
Minimize
Z
1
(C
1
e
) Z
2
(A
1
sdr
) =
eE
1
K
1
e
(C
1
e
)
sS
dD
2
rR
1
d
W
1
sdr
A
1
sdr
_
1 B
1
sdr
_
Les Cahiers du GERAD G9851 Revised 14
subject to
rR
1
d
A
1
sdr
A
2
sd
, s S, d D
2
(8)
B
1
sdr
B
1
sdr
, s S, d D
2
, r R
1
d
(9)
A
1
sdr
0, s S, d D
2
, r R
1
d
(10)
C
1
e
0. e E
1
(11)
For this type of model, it is assumed that there is complete sharing between all connec-
tions carried on a link. Also, by using eective bandwidth calculations, each connection
can be estimated by a xed bandwidth requirement. The network can therefore be modeled
as a circuit-switched multirate broadband network.
Assuming link independence, the end-to-end call blocking probability B
1
sdr
of class s on
route r R
1
d
between O-D pair d is
B
1
sdr
= 1
er
(1 B
1
se
). (12)
The call blocking probability B
1
se
and the total oered trac a
1
se
per service class s on
each link e E
1
are linked by the well-known Erlang xed point system [25, 41]
B
1
se
= g
1
e
_
a
1
e
,
1
e
, C
1
e
_
, (13)
a
1
se
=
dD
2
rR
1
d
: er
A
1
sdr
r{e}
_
1 B
1
se
_
. (14)
In equation (14), the trac oered to a link e for a specic route r between O-D pair d
depends on the blocking probability B
1
se
on each link e
1
se
(t), the set of routes R
1
d
(t), the oered trac A
1
sdr
(t), and the link blocking probabilities
B
1
se
(t) must be calculated. This model minimizes a linear network cost function and the
expected cost of lost trac in the event of failures under ow conservation constraints.
Minimize
eE
1
K
1
e
(C
1
e
) +
tT
P(t)
sS
eE
1
1
se
(t)
dD
2
K
1
sd
rR
1
d
(t) : er
A
1
sdr
(t)
B
1
se
(t)
subject to (8), (10) and (11).
Les Cahiers du GERAD G9851 Revised 16
Blocking constraints (9) are implicitly included in the model by associating penalty costs
K
1
sd
to lost trac.
The model is solved by decomposing the problem into two subproblems. Given xed link
capacities, the routing subproblem is solved using the Frank-Wolfe ow deviation algorithm
to obtain oered trac values. Using these trac values, the dimensioning subproblem is
then solved to obtain new capacity estimates. This process is repeated until convergence.
An advantage of this method is that it is independent of the type of call routing used by
the network.
This algorithm was tested on the 10 node network used in [56], and the results were
compared to those obtained using Sun and Girards [86] approach. Network costs and
computation times were similar, and the resulting network proved to be more robust in
the event of transmission link failures.
4.1.2 Linear cost-based optimization In the previous subsection, the relationship
between end-to-end blocking and link capacity resulted in dicult non-linear problems.
By using a trunk reservation method, Mehdi and Guptan [58] propose a simpler linear
approximation of the model by inverting the Erlang-B blocking function to estimate (o-
line) the number of bandwidth units required for each node pair d and service class s. To
simplify these calculations, bandwidth requirements
1
s
are not used as blocking function
parameters but are instead included explicitely within the model (see Equations (15) and
(16)). The bandwidth unit estimates are used in a linear multi-commodity ow model that
determines virtual link capacities C
1
e
and the number of bandwidth units X
1
sdr
required on
each route.
Minimize
Z
1
(C
1
e
) =
eE
1
K
1
e
(C
1
e
)
subject to (11),
sS
dD
2
rR
1
d
:er
1
s
X
1
sdr
C
1
e
, e E
1
(15)
rR
1
d
X
1
sdr
g
1
_
A
2
sd
, B
1
sd
_
, s S, d D
2
(16)
X
1
sdr
0. s S, d D
2
, r R
1
d
(17)
Simulations were carried out with 2 service classes on the 10 node network from [56], and
results were obtained in less than 3 minutes.
4.1.3 Delay-based optimization Since independent connection switching means that
connections are routed over a series of concatenated virtual links, a useful objective is
to minimize packet delays. One example is Wolfgangs model [89] in which the routing
problem is solved. This model minimizes average packet delay H subject to oered trac
Les Cahiers du GERAD G9851 Revised 17
Table 2: Articles presenting Centralized Independent Connection Models
ARTICLES VARIABLES OBJECTIVE CONSTRAINTS SOLUTION PROBLEM
APPROACHES FEATURES
Girard and load sharing max carried trac Erlang xed point decomposition nonlinear
Gardouh (1992) virtual link capacity min link costs end-end blocking heuristic Labourdette
Sun and load sharing max carried trac Erlang xed point approx. obj. gradient nonlinear
Girard (1998) virtual link capacity min link costs end-end blocking reduced gradient Erlang xed
Lagrange relax. point
Mitra, Morrison and load sharing max carried trac Erlang xed point approx. blocking Erlang xed
Ramakrishnan (1996) steepest ascent point
Sans` o, Girard and oered trac min lost trac multi-commodity Decomposition link failures
Mobiot (2000) virtual link capacity min link costs Flow deviation Labourdette
Medhi and oered trac min link costs multi-commodity trunk reservation linear model
Guptan (1997) virtual link capacity o-line calculations
Erlang xed point relax blocking nonlinear
Wolfgang (1995) virtual link capacity min packet delay end-end blocking penalty function
end-end delay
ARTICLES PROBLEM NO. SERVICE CLASSES PLATFORM RUNNING
SIZE TIMES (SECS.)
Sun and |V
1
| = 40, |E
1
| = 1560 4 (64, 384, 1536, 6144 Kbps) SUN Ultra 1 routing: 750
Girard (1998) routing & dim: 4350
Mitra, Morrison and |V
1
| = 8, |E
1
| = 20 6 (16, 48, 64, 96, 384, 960 Kbps)
Ramakrishnan (1996)
Sans` o, Girard and |V
1
| = 10, |E
1
| = 90, 3 (64, 384, 1536 Kbps) SUN Ultra 1 Sans`o: 162
Mobiot (2000) Sun: 183
Medhi and |V
1
| = 10, |E
1
| = 90 2 (64, 384 Kbps) DECstation 5000-200 180
Guptan (1997) running MIPS Ultrix
Wolfgang (1995) |V
1
| = 18, |E
1
| = 306 3
constraints (8) and (10), end-to-end blocking constraints (9), and the Erlang xed point
equations (13) and (14). Packet ow and delay constraints are also included in this model.
The blocking constraints and Erlang xed point equations make the model a dicult
non-convex problem to solve. However, by ignoring the eect of blocked connections when
calculating the oered trac on each route, the model becomes convex. This approximated
model is solved using a penalty function method and was tested on networks having up to
18 nodes.
The papers on centralized independent connection models that were reviewed in this
section are summarized in Table 2.
4.2 Centralized end-to-end path models
In end-to-end path models, every O-D pair is connected by at least one virtual link. More-
over, connections oered to an O-D pair are always routed over a direct virtual link joining
the pair (i.e., |r| = 1, r R
1
d
, d D
2
). Since each O-D pair can be connected by several
virtual links, a connection routing algorithm, such as xed alternate routing, is required
to select those links on which new connections will be set up.
In many models, connections from dierent service classes are mixed together over the
virtual links. However, when networks carry highly dissimilar service classes, such as both
very bursty and constant bit rate trac, separating service classes onto dierent links may
be more ecient [85].
We provide an illustration of end-to-end path models in Figure 6. There are three
virtual links between nodes A and C across which A-C connections can be routed. Given
oered trac A
2
sd
AC
from node A to C, the synthesis problem determines how much trac
is to be carried across each of the three possible links while respecting the transmission
capacity constraints.
Les Cahiers du GERAD G9851 Revised 18
Virtual links
Routes
2
sd
A
A
B
C
A-C
Figure 6: Virtual network for the end-to-end path model
End-to-end path models are typically formulated as nonlinear programs [17, 32, 54, 83],
which maximize network throughput, or as linear models [34, 37, 57], which optimize
capacity costs or revenues. Link utilization based models [22, 44], which minimize the
amount of reserved transmission link bandwidth, have also been proposed. Since using end-
to-end paths minimizes the delay in a network, delay-based optimization is not considered.
4.2.1 Throughput-based optimization In this section, we review the centralized end-
to-end path models that maximize carried trac or carried trac revenue. These models
assume that virtual links can carry connections from multiple service classes.
Given the oered trac A
2
sd
for every service class s and O-D pair d and the incidence
matrix R
0
of virtual and transmission links, bandwidth C
1
e
must be allotted to each virtual
link e, so that carried trac revenue is maximized while transmission capacity constraints
are respected.
Maximize
Z
3
(C
1
e
) =
sS
dD
2
W
1
sd
A
2
sd
_
1 B
1
sd
_
subject to
R
0
C
1
C
0
, (18)
C
1
e
0. e E
1
(19)
If xed alternate routing is implemented, there is no known exact analytical expression
for the end-to-end blocking probabilities B
1
sd
of multiple service classes sharing a group of
parallel virtual links. Consequently, in [54, 84], the model is simplied by substituting the
parallel links between each O-D pair d D
2
with a single link whose capacity is the sum
of the parallel link capacities
C
1
d
=
eE
1
d
C
1
e
. (20)
Les Cahiers du GERAD G9851 Revised 19
As a result, the routing is xed and end-to-end blocking B
1
sd
can be calculated using a
multirate approximation
B
1
sd
= g
1
sd
(a
1
d
,
1
d
, C
1
d
). (21)
In Siebenhaar [83], blocking constraints B
1
sd
B
1
are added to the model. A reduced-
gradient and quasi-Newton algorithm is used to solve the problem once the blocking con-
straints are linearized. The maximal blocking probability B
1
is determined by solving the
following minimax problem using Labourdette and Harts multirate blocking approxima-
tion and a bisection method.
Minimize
Z
4
(C
1
e
) = max
sS dD
2
{B
1
sd
}
subject to (18), (19), (20) and (21).
This algorithm was tested on a 23 node network in which each O-D pair was connected
by up to 6 virtual links. The total number of virtual links was 960. When the network
carried 5 service classes, solutions were obtained in 5 to 10 minutes on a SUN Sparcstation
10.
Logothesis and Shioda [54] also minimize Z
4
subject to constraints (18), (19), (20) and
(21) but add linear switch capacity constraints. Multirate blocking is calculated using the
Kauman and Roberts recursive formula. The authors solve the problem using a heuristic
approach in which bandwidth is successively added to paths having large blocking rates and
removed from paths having small blocking probabilities, while guaranteeing transmission
capacity constraints. The heuristic algorithm was compared to an analytic solution by
testing on networks having up to 14 nodes. The heuristic gave near optimal results in less
time.
It is not necessary to estimate network throughput using end-to-end blocking probabil-
ities. For example, in Evans [17], the expected number of calls in progress is maximized.
The author develops a simple approximation for this objective and its derivative. While
the objective is not concave, she shows that it is concave over the relevant part of the
feasible region if trac demands and available link capacities are not greatly mismatched.
A Frank-Wolfe algorithm is consequently employed to solve the problem.
Herzberg and Pitsillides [32] use a uid ow function to model the total expected
network throughput. By discretizing the throughput function, the authors reformulate the
relaxed model as a linear programming problem, which is solved using standard techniques.
4.2.2 Linear capacity-based optimization Instead of mixing service classes, some
models assume that only one service class is carried on each virtual link. This means that
the Erlang-B function can be calculated o-line to estimate bandwidth requirements and
that linear models can be easily formulated using trunk reservation schemes.
In Hui et al. [34], a formulation is given in which linear bandwidth costs are minimized
subject to bandwidth requirements and transmission capacity constraints.
Les Cahiers du GERAD G9851 Revised 20
Minimize
Z
5
(C
1
e
) =
sS
dD
2
eE
1
sd
K
1
e
C
1
e
subject to (18), (19), and
eE
1
sd
C
1
e
g
1
sd
_
A
2
sd
, B
1
sd
_
. s S, d D
2
(22)
To this model, Mehdi [57] has added the assumption that class s trac between O-D
pair d can not be bifurcated and thus is carried on only one virtual link. Given the variable
x
sde
, such that x
sde
= 1 if virtual link e E
1
sd
is chosen to carry class s trac between
O-D pair d and is 0 otherwise, the author has formulated a mixed-integer linear program
for optimizing total revenue.
Maximize
Z
6
(x
sde
, C
1
e
) =
sS
dD
2
eE
1
sd
W
1
e
C
1
e
subject to (18), (19), (22),
C
1
e
x
sde
g
1
sd
(A
2
sd
, B
1
sd
), s S, d D
2
, e E
1
sd
(23)
eE
1
sd
x
sde
= 1, s S, d D
2
(24)
x
sde
{0, 1}. s S, d D
2
, e E
1
sd
(25)
Another linear model has been proposed by Jiyapanichkul and Lambert [37], in which
the carried class s trac for virtual link e is estimated as C
1
e
/
1
se
. The model maximizes
Z
6
subject to constraints (18), (19), and
eE
1
sd
C
1
e
1
se
A
2
sd
. s S, d D
2
(26)
All models in this section can be solved by using standard linear optimization or mixed-
integer routines; however, some authors [57] have not provided any experimental results
and others [34, 37] have simulated networks that are only carrying a single service class.
4.2.3 Link utilization-based optimization When only carried trac is maximized,
all available transmission capacity is allocated to the oered trac. This results in networks
that have very little spare capacity and are thus unable to accommodate potential increased
bandwidth demands. Consequently, Gersht et al. [21] propose a model that combines a
load sharing objective with an access control objective. This model is optimized over the
admission control variables
sd
and the load sharing variables
sde
. Note that
sd
is the
Les Cahiers du GERAD G9851 Revised 21
percentage of A
2
sd
trac blocked at the source, and
sde
is the fraction of (1
sd
)A
2
sd
trac carried over link e joining O-D pair d. The load sharing variables are determined by
maximizing the minimal residual transmission link capacity Z
7
, and the access variables are
calculated by maximizing the minimal fraction of admitted trac Z
8
. Using our notation,
the objectives and constraints are written:
maximize
Z
7
(
sde
,
sd
) = min
eE
0
_
C
0
e
E
1
r
0
e,e
C
1
e
_
;
maximize
Z
8
(
sd
) = min
dD
2
_
(1
sd
)
_
;
subject to (18), (19),
C
1
e
=
sS
dD
2
eE
1
d
(1
sd
)A
2
sd
1
s
sde
, e E
1
(27)
eE
1
d
sde
= 1, s S, d D
2
(28)
sde
0, s S, d D
2
, e E
1
d
(29)
0
sd
1, s S, d D
2
(30)
sd
> 0 max
e
E
1
d
_
min
eE
0
:ee
_
C
0
e
E
1
r
0
e,e
C
1
e
_
_
= 0. (31)
Equation (31) describes the connection accessibility conditions, which only permit a
fraction of A
2
sd
trac to be blocked if no residual capacity is available between O-D pair
d. In [44], these access conditions are slightly modied so that bandwidth demand for
long duration calls is rejected until accepted bandwidth demand for regular calls is above
a preset value. This way, long duration calls do not monopolize the networks capacity at
the expense of shorter duration calls. Another variant of the problem is tackled in [22],
where possible transmission link failures are taken into consideration and the network is
designed with the guarantee that it can fully restore trac ow after any failure.
This model is tackled by solving the network equilibrium equations of the problem,
which are shown to satisfy both objectives.
Table 3 summarizes articles reviewed in Section 4.2, which use centralized end-to-end
path models.
4.3 Centralized node-to-node path models
As described in Section 2.2.2, reserving bandwidth between O-D pairs using end-to-end
paths may reduce the delay and complexity of a network, however it also wastes bandwidth
if the paths are underutilized. To compensate for this ineciency, connections may be
routed over a series of paths to give us a node-to-node path model. The virtual network is
Les Cahiers du GERAD G9851 Revised 22
Table 3: Articles presenting Centralized End-to-end Path Models
ARTICLES VARIABLES OBJECTIVE CONSTRAINTS SOLUTION PROBLEM
APPROACHES FEATURES
Siebenhaar (1994) virtual link capacity max. carried trac link capacity linearize constraints nonlinear
end-end blocking reduced-gradient Labourdette
Logothetis and virtual link capacity min max link capacity heuristic nonlinear
Shioda (1992) call blocking switch capacity Roberts
Evans (1991) virtual link capacity max # connections link capacity concave
Frank-Wolfe
Herzberg and virtual link capacity max carried trac link capacity discetize objective nonlinear
Pitsillides (1993) route bandwidth linear relaxation uid ow model
Jiyapanichkul and virtual link capacity max revenue link capacity linear formulation
Lambert (1998) max connection rate ow constraints
Mehdi (1997) virtual link capacity max revenue link capacities required bandwidth linear formulation
virtual link routing 1 route per O-D calculated o-line 1 class/path
Kheradpir, Gersht,
Shulman and virtual link capacity max min ow conservation parallel algorithm SBMS network
Stinson (1996) residual capacity
Gersht, Kheradpir virtual link capacity max min ow conservation parallel algorithm SBMS network
and Friedman (1996) residual capacity
Hui, Gursoy, Moayeri virtual link capacity min. link costs link capacities linear formulation
and Yates (1991)
ARTICLES PROBLEM NO. SERVICE CLASSES PLATFORM RUNNING
SIZE TIMES (SECS.)
Siebenhaar (1994) |V
1
| = 23, |E
1
| = 960 5 SUN Sparcstation 10 600
|V
0
| = 23, |E
0
| = 41
Logothetis and |V
1
| = 14 2 (100, 1000 Kbps) VAX-6330 (VMS) analytic 3540
Shioda (1992) heuristic 2940
Gerrsht, Kheradpir |V
1
| = 20, |E
1
| = 380 HP-730 370
and Shulman (1996)
Hui, Gursoy, Moayeri |V
1
| = 5, |E
1
| = 42
and Yates (1991)
therefore a multigraph, in which each possible route R
1
d
for an O-D pair d D
2
is a series
of links in E
1
. Note that O-D pairs are not necessarily connected by a direct virtual link
in G
1
. Since the virtual links C
1
e
may now hold trac from many O-D pairs, oered trac
A
1
sdr
to each route r R
1
d
must be included as a variable in the model formulations.
In Figure 7, we illustrate the multigraph, which represents the virtual network for a
node-to-node path network model. Connections between O-D pair A-C are either routed
over one of the direct links between A-C or across more than one virtual link (i.e., over
links (A-B) then (B-C)).
An example of node-to-node bandwidth allocation is the virtual network allocation
proposed by Dziong, Xiong and Mason [15], in which routing and bandwidth allocation
functions are separated between paths and virtual network links. Paths are used to dene
routes between every node pair, and virtual network links are used to control bandwidth
allocation.
Node-to-node path models may optimize performance by maximizing network through-
put [11, 18, 19, 49, 67]. However, cost-based models [1, 5, 12, 13, 14, 45, 59, 65, 79], which
minimize the network costs, and delay-based models [20, 50, 61, 63, 87] are also common.
4.3.1 Throughput-based optimization In this section, we review the node-to-node
path models that optimize network throughput. In Farago et al. [19], carried trac revenue
is maximized by estimating the end-to-end call blocking probabilities for each route r R
1
d
.
Given oered trac A
2
sd
and the incidence matrix R
0
of virtual and transmission links,
virtual link bandwidth C
1
e
and oered trac A
1
sdr
are determined such that transmission
capacity constraints and trac requirements are respected.
Les Cahiers du GERAD G9851 Revised 23
Routes
Virtual links
A
2
B-C
sd
sd
A
2
A-C
A
B
C
Figure 7: Virtual network for the node-to-node path model
Maximize
Z
2
(C
1
e
, A
1
sdr
) =
sS
dD
2
rR
1
d
W
1
sdr
A
1
sdr
_
1 B
1
sdr
_
subject to (8), (10), (12), (13), (14), (18) and (19).
Farago et al. employ a two-step method to solve the problem. Initially, connection
routing is xed, and the link dimensioning problem is reformulated as a convex program
by approximating the expected revenue by a concave function based on a continuous ap-
proximation of Erlangs formula. Using the solution to this rst step as an initial point,
approximations for the objective function and objective gradient are used in a gradient-
based hill climbing method to nd an improved solution. Alternatively, a neural network
approach can solve the dimensioning problem [18].
The solution approach described in [19] is tested on a 3 node network carrying 2 service
classes. Total revenue values were similar to those obtained using a more exact approach.
Cheng and Lin [11] solve this model too but add constraints on the number of concate-
nated virtual links that a connection can be routed across. They solve the program using
a two-phase iterative algorithm. In the rst phase, virtual link capacities are xed while
routing is optimized, while in the second phase, the reverse is performed.
In Lazar et al. [49, 67], a novel approach is taken to solve the bandwidth allocation
problem. Rather than maximizing overall network performance, path allocation is con-
sidered as a non-cooperative game, where each user attempts to selshly reserve as much
bandwidth as possible for its own end-to-end paths. In order to reect this competition
between users, a two-part cost function is used. The rst part accounts for the fact that
as more bandwidth is reserved for a user, its call blocking, and consequently its cost di-
minishes. The second part reects the fact that as more bandwidth is reserved by other
users of links making up a users path, the available bandwidth diminishes and the cost
of reserving bandwidth rises. Cost functions are constrained by the following: no user is
allowed to reserve all network bandwidth for itself; every user reserves some bandwidth for
itself; and total reserved bandwidth is less than the total available network bandwidth. For
single link, parallel and general network topologies, the authors show that the game has a
Les Cahiers du GERAD G9851 Revised 24
unique Nash equilibrium and is fair, which means that there exists a bandwidth allocation
strategy such that no user nds it benecial to change its paths bandwidth.
4.3.2 Cost-based optimization In work by Dziong [14] and Oki and Yamanaka [65],
virtual link capacity costs are minimized subject to blocking constraints. Since they as-
sume that transmission link capacities are very large, there are no transmission capacity
constraints.
Minimize
Z
1
(C
1
e
) =
eE
1
K
1
e
(C
1
e
)
subject to (9) and (11).
In Dziongs model [14, pp.227233], reward parameters, dened as the average reward
for carrying connections of a certain class, are added as optimization variables so that the
connection blocking constraints can be simplied. The problem is solved using an iterative
algorithm, which gives nearly optimal solutions when tested on a 4 node network. The
algorithm starts with initial bandwidth allocation and reward parameters, which then are
increased until all constraints are satised.
A similar model is proposed by Oki and Yamanaka, in which link capacity variables
are determined by minimizing link and node capacity costs. Binary decision variables,
which indicate whether or not an O-D pair is joined by a virtual link, are also decided.
This model not only includes call blocking constraints but also constraints on packet loss
L
1
sdr
L
1
and on the number of transmission hops a path can take. Since the problem
is a dicult nonlinear combinatorial program, the authors propose a heuristic algorithm.
Random initial values are assigned to the decision variables so that there are at least 2 node-
disjoint routes between every O-D pair. At each iteration, a decision variable is chosen at
random and its value is negated. This process stops when each decision variable has been
chosen at least once and when the variables have returned to their initial random values.
The solution having the lowest cost and for which blocking and packet loss constraints are
veried is chosen as optimal.
In Sections 4.1.2 and 4.2.2, trunk reservation was used to obtain linear models for solving
the network synthesis problem. This approach can also be used for node-to-node networks
[1, 12, 13, 45, 59], resulting in multicommodity ow models that determine required virtual
link C
1
e
and route X
1
sdr
capacities.
Minimize
Z
1
(C
1
e
) =
eE
1
K
1
e
(C
1
e
)
subject to (17), (18), (19),
Les Cahiers du GERAD G9851 Revised 25
sS
dD
2
rR
1
d
:er
X
1
sdr
C
1
e
, e E
1
(32)
rR
1
d
X
1
sdr
= g
1
_
A
2
sd
,
1
s
, B
1
sd
_
. s S, d D
2
(33)
In Mehdi and Lu [59], the synthesis problem is solved in conjunction with the circuit
routing and transmission dimensioning problems. The model is a complex mixed-integer
linear optimization problem, which must be decomposed into two problems: the transmis-
sion network dimensioning subproblem and the virtual network dimensioning subproblem.
The two subproblems can be solved using an iterative procedure in which the virtual net-
work subproblem is used as an input for the transmission network subproblem, and vice
versa. The authors have tested their model on a 15 node network carrying 3 service classes,
using CPLEX to solve the subproblems.
In [1, 12, 13, 45], a link ow formulation of the previous multicommodity model is
solved. Chlamtac et al. [12] solve the problem using standard linear routines, whereas
Ahn et al. [1] solve it in two steps. Initially, they determine the transmission routes for the
oered trac using a mixed integer linear multi-commodity ow formulation. Using this
trac layout, both virtual link capacities and virtual link routing variables are determined
by using a heuristic method that minimizes switching and call setup costs.
In the model proposed by Kim [45], nonlinear switching and buer costs are added to the
objective function, and an interior penalty function method is used to solve the problem.
In Chung et al. [13], a heuristic algorithm based on the local linear approximation of the
objective is proposed and is shown to give good results more quickly when compared with
the approach proposed in [45].
4.3.3 Delay-based optimization In [20, 50, 61, 63, 87], the objective is to minimize
packet loss. However, since models for packet loss are complex, average packet delay is
used as an indirect measure of packet loss probability. The average delay is calculated by
modeling each link e E
1
as an independent M/M/1 queue. The objective of the model
is to optimize required virtual link C
1
e
and route X
1
sdr
capacities, such that the average
packet delay is minimized:
Minimize
H(C
1
e
, X
1
sdr
)
subject to (32), (17), (18) and (19).
In Gerla et al. [20] and Murphy et al. [63], the objective function H is non-convex. By
generating many random initial solutions to be used as starting points for their algorithm,
the authors can select the global optimum from among the local optima having the lowest
value. In [20], a solution is found using Frank-Wolfes steepest descent method, while in
[63] a standard library routine for linearly constrained optimization problems is used.
Les Cahiers du GERAD G9851 Revised 26
Monteiro et al. [61] improve the solution provided in [20] by employing an average
packet delay objective which is convex in both C and X if the integer capacity variables
C
1
e
are relaxed. They develop an iterative algorithm that routes heavy O-D trac over
direct paths and less loaded O-D pairs over indirect paths. Finally, Tang et al. [87] have
studied the same problem but have included switch capacity constraints. They solve the
problem using a fast genetic algorithm.
Lee and Yee [50] solve a model similar to [20]; however, total packet loss is minimized
directly rather than by optimizing average packet delay. Although the packet loss function
is non-convex, the authors show that the relaxed model can be simplied to a convex
problem.
Tables 4 and 5 provide a summary of the papers describing centralized node-to-node
path models.
Table 4: Articles presenting Centralized Node-to-node Path Models
ARTICLES VARIABLES OBJECTIVE CONSTRAINTS SOLUTION PROBLEM
APPROACHES FEATURES
Farago, Blaabjetg, virtual link capacity max carried trac link capacity convex approx. 2 phase solution
Ast, Gordos and load sharing approx. obj. gradient
Henk (1995) steepest ascent
Farago, Biro, Henk virtual link capacity max carried trac link capacity convex approx. Erlang xed point
and Boda (1997) neural networks xed routing
Cheng and virtual link capacity max carried trac link capacity model decomposition Erlang xed point
Lin (1994) connection routing # virtual hops iterative algorithm
Lazar, Orda and virtual link capacity min blocking link capacity Non-cooperative
Pendarakis (1997) min capacity cost Game
Pendarakis (1996)
Dziong (1997) virtual link capacity min. link costs end-end blocking iterative algorithm
connection routing link eciency
Oki and virtual link capacity min. link costs end-end blocking
Yamanaka (1995) connection routing min. switch costs packet loss rate heuristic switch costs
virtual hops
transm. link capacities min virtual and link capacity decompose model linear model
Mehdi and Lu (1997) virtual link routing transm. link costs ow constraints
virtual link capacity
connection routing
Ahn, Tsang, Tong virtual link capacity min. max ow multicommodity heuristic
and Du (1994) virtual link routing
connection routing
Kim (1995) virtual link capacity min link costs multicommodity linear relaxation linear objectives
connection routing min control costs link capacity exterior penalty
buer capacities min buer costs integer variables
Chung, Hong, Song virtual link capacity min link costs multicommodity linearize obj. concave objective
and Kim (1997) connection routing min control costs link capacity bilinear method
buer capacities min buer costs
Chlamtac, Farago virtual link capacity min link costs multicommodity linear model
and Zhang (1996) connection routing link capacity
Gerla, Monteiro virtual link capacity min packet loss link capacity Random initial sol. non convex
and Pazos (1989) trac ows ow conservation Frank-Wolfe
Murphy and virtual link capacity min packet loss link capacity Random initial sol. non convex
Curran (1993) trac ows ow conservation library routing
Monteiro and virtual link capacity min packet loss link capacity integer relaxation convex
Gerla (1994) trac ows ow conservation Frank-Wolfe
Tang, Ko, Man virtual link capacity min packet loss link capacity genetic algorithm non convex
and Kwong (1998) trac ows ow conservation
Lee and Yee (1994) virtual link capacity min packet loss ow conservation linear relaxation non convex
connection routing total packet loss heuristics mixed integer
integer capacities
Les Cahiers du GERAD G9851 Revised 27
Table 5: Articles presenting Centralized Node-to-node Path Models
ARTICLES PROBLEM NO. SERVICE CLASSES PLATFORM RUNNING
SIZE TIMES (SECS.)
Farago, Blaabjetg, |V
1
| = 3, |E
1
| = 6 2
Ast, Gordos and
Henk (1995)
Farago, Biro, Henk |V
1
| = 4, |E
1
| = 6 2 (750, 1500 Kbps)
and Boda (1997)
Dziong (1997) |V
1
| = 4, |E
1
| = 6 2 (64, 192 Kbps)
Oki and |V
1
| = 8
Yamanaka (1995)
Mehdi and Lu (1997) |V
1
| = 15 3 (64, 384, 384 Kbps)
Ahn, Tsang, Tong |V
1
| = 20, |E
1
| = 50
and Du (1994)
Kim (1995) |V
1
| = 6, |E
1
| = 90 3 (64, 2000, 10000 Kbps)
Chung, Hong, Song |V
1
| = 6, |E
1
| = 90 3 (64, 2000, 10000 Kbps) Kim: 2400
and Kim (1997) Chung: 4
Gerla, Monteiro |V
1
| = 5, |E
1
| = 28 SUN 3/280 1800
and Pazos (1989)
Murphy and |V
1
| = 3, |E
1
| = 7
Curran (1993)
Tang, Ko, Man |V
1
| = 5, |E
1
| = 28 SUN 3/280 1800
and Kwong (1998)
Lee and Yee (1994) |V
1
| = 18 SUN 4 100
4.4 Centralized hybrid models
Hybrid allocation models route some connections over direct end-to-end paths and route
others independently in order to increase multiplexing gains and decrease processing eorts.
This approach is likely to be useful for networks in which delay sensitive, high demand
trac can be routed over end-to-end paths, while delay insensitive, low demand trac can
be routed independently. For example, in Chan and Geraniotis [10], voice calls are carried
on end-to-end paths, whereas video calls are set up as independent connections.
In Aneroussis and Lazar [3], the network revenue is maximized while satisfying process-
ing load constraints on the nodes and blocking constraints for each O-D pair. An algorithm
is proposed in which, initially, the trac between all O-D pairs is routed using either a
xed alternate or an adaptive routing scheme, with routes and alternate routes made up of
either paths, or independent connections, or of both. First, path bandwidths are gradually
increased until both blocking and signaling constraints are satised, then they are further
increased so as to maximize revenue, yet satisfy all constraints. This algorithm was tested
on a 10 node network carrying 2 service classes.
Articles describing centralized hybrid models are summarized in Table 6.
Table 6: Articles presenting Centralized Hybrid Models
ARTICLES VARIABLES OBJECTIVE CONSTRAINTS SOLUTION PROBLEM
APPROACHES FEATURES
Aneroussis and virtual link capacity max. network processing loads heuristic
Lazar (1996) revenue end-end blocking
ARTICLES PROBLEM NO. SERVICE CLASSES PLATFORM RUNNING
SIZE TIMES (SECS.)
Aneroussis and |V
1
| = 10, |E
1
| = 52 2 (64, 4000 Kbps) 5
Lazar (1996)
Les Cahiers du GERAD G9851 Revised 28
4.5 Decentralized end-to-end path models
When decentralized models are implemented, each paths bandwidth is computed by a
controller based on data acquired from some, but not all, paths in the network. Since
global information is not considered, these models give worse solutions than do centralized
models, however, they are likely to operate on a faster time scale.
Much research has been devoted to decentralized dynamic end-to-end path algorithms.
In general, very simple algorithms are used to allocate the bandwidth. For example, in
order to react quickly to trac variations, paths may have their bandwidth varied by a
xed step [78, 64, 51, 90, 40], either at each new call arrival or whenever quality of service
requirements are not met. Variations include making the increase step size larger than the
decrease step size [51] or having two possible step sizes [90].
There are some disadvantages, however, in using xed step sizes. For example, large
bandwidth steps result in increased bandwidth waste and small step sizes result in increased
processing loads [66, 90]. Furthermore, bandwidth may oscillate around a steady state if
xed step sizes are used [66]. For this reason, variable step sizes are most often employed.
When the bandwidth steps are variable, bandwidth is shared between the paths in such
a way as to optimize network performance. For example, in Gu [28] a heuristic algorithm is
used to minimize the maximum call blocking probability, and in Bolla et al. [6] a steepest
descent algorithm is used to minimize the total expected packet loss rate. Tsang and Chan
propose two heuristics methods: the rst [9] involves combining paths together in order
to minimize total bandwidth required while satisfying the QoS requirements of each class;
the second approach [88] involves having high priority classes overow into under-utilized
paths.
Larsson and Arvidsson [47] propose a distributed method where trac between each O-
D pair is divided among an arbitrary number of end-to-end paths. The paths bandwidth
is estimated using an iterative procedure based on the Erlang-B formula. In a subsequent
paper [48], bandwidth is estimated more quickly using an approximation requiring only
the number of active connections.
Bruni [7] assumes that paths have Poissonian trac demands, so bandwidth for each
period is calculated based on the oered trac and desired blocking. With only the
number of active connections known, path bandwidths are quickly determined by using
pre-calculated capacity allocation tables.
Recently, IP resource management methods based on the MPLS architecture [72] have
been proposed [4, 36] to provide QoS guarantees for delay sensitive services using constraint-
based routing label paths (CRLSPs). In [4], the edge nodes periodically increase and de-
crease CRLSP bandwidth based on estimated bandwidth needs, bandwidth use, and the
status of the transmission links carrying the CRLSP. Bandwidth changes are implemented
using the constraint-based routing label distribution protocol [2, 36].
Shioda [81] predicts the bandwidth step for two bandwidth demand cases: where band-
width demand is stationary and where it is modeled as a linear dynamic time-series.
In the algorithms by Saito [74] and Saito et al. [75], the bandwidth step is estimated
every 15 minutes based on actual trac and on trac data collected by the system on the
Les Cahiers du GERAD G9851 Revised 29
previous day during the same measurement period. This algorithm ensures that there will
be sucient bandwidth during the next control period if the largest variation in trac is
less than the variation which occurred on the previous day.
Orda et al. [66] develop an event-driven, variable bandwidth step algorithm. Each
virtual link is assigned an upper and lower threshold. Bandwidth is increased if the demand
is greater than the upper threshold, and bandwidth is decreased if the demand is less than
the lower threshold. New thresholds are calculated using Markov chains to model the path
dynamics.
Table 7 provides a summary of the articles reviewed in Section 4.5 using decentralized
end-to-end path models.
Table 7: Articles presenting Decentralized End-to-end Path Models
ARTICLES LOCAL OR BANDWIDTH UPDATE PROBLEM FEATURES
DISTRIBUTED STEP INTERVAL
Sato, Ohta, and Tokizawa (1990), local xed event driven update if connection is blocked,
Ohta and Sato (1992) equal increase and decrease steps
Gu , Sohraby and local variable periodic minimize call blocking
Vaman (1995) heuristic algorithm
Bolla, Davoli, Lombardo, local variable periodic minimize expected packet loss rate
Palazzon and Panno (1994) steepest descent method
Chan and Tsang (1994) local variable periodic combination of paths
Tsang and Chan (1998) local variable periodic priority classes
Lin, Su and Lo (1996) distributed xed event driven update if connection is blocked,
if link utilization is too high or low
or if processing load is too high
increase step larger than decrease step
Wu, Chou and Tzeng (1995) distributed xed event driven two possible step sizes
Saito, Nakagawa, Kato distributed variable event driven self-sizing network
and Ohta (1996) or periodic measurement-based bandwidth estimation
Saito (1997)
Kawamura, Hadama, Sato distributed xed periodic peak rate allocation
and Tokizawa (1994)
Larsson and Arvidsson (1998) distributed variable periodic # active calls for bandwidth estimation
Larsson and Arvidsson (1997) distributed variable periodic iterative procedure
Bruni, Mocci, Pannunzi distributed variable periodic capacity assignment tables
and Scoglio (1995)
Ash, Jamoussi, Lee distributed variable periodic IP MPLS-based networks
and Aboul-Magd (1999)
Orda, Pacici and distributed variable event-driven upper and lower thresholds
Pendarakis (1996) Markov chains to calculate thresholds
Shioda (1994), distributed variable event-driven call blocking for bandwidth estimation
Shioda and Uose (1991)
Shioda, Saito and Yokoi (1997) distributed variable event-driven time series model
5 Conclusion
An extensive review of the multi-service network literature revealed a large diversity of
models and technological choices suggested for the solution of the network synthesis prob-
lem. In this paper, we have proposed a technology-independent typology permitting clas-
sication of all these multi-service network synthesis methods. We found that there are,
at present, ve basic modeling types which we classied based on the switching and
control strategies used. To do this, we developed a unifying notation, which we then used
to analyze and compare the mathematical formulations and solution approaches currently
suggested for the multi-service network synthesis problem.
Our review reveals several possible directions for future research. These include the
development of fast heuristics to allow for real-time centralized network synthesis, the
Les Cahiers du GERAD G9851 Revised 30
development of models for dimensioning decentralized independent connection, node-to-
node and hybrid networks, and the development of models and algorithms specic to
multi-technology networks.
Acknowledgement
Finally, we would like to thank the anonymous referees for their valuable comments and
suggestions.
References
[1] S. Ahn, R.P. Tsang, S.-R. Tong, and D.H.C. Du. Virtual path layout design on ATM
networks. In Proc. IEEE Infocom, pages 192200, 1994.
[2] L. Anderson, P. Doolan, N. Feldman, A. Fredette, and B. Thomas. LDP
specication, IETF draft-ietf-mpls-ldp-01.txt. Internet Draft, available at
www.crt.umontreal.ca/bruni/archive/, August 1998.
[3] N.G. Aneroussis and A.A. Lazar. Virtual path control for ATM networks with call
level quality of service guarantees. IEEE/ACM Trans. Networking, 6(2):222236,
April 1998.
[4] G. R. Ash, B. Jamoussi, Y. Lee, and O. S. Aboul-Maged. QoS resource management
in MPLS-based networks, draft-ash-qos-routing-00.txt. Internet Draft, available at
www.crt.umontreal.ca/bruni/archive/, February 1999.
[5] M. Aydemir and Y. Viniotis. Deterministic algorithm for VP assignment in ATM
networks. Computer Communications, 19(13):10361050, November 1996.
[6] R. Bolla, F. Davoli, A. Lombardo, S. Palazzo, and D. Panno. Hierarchical dynamic
control of multiple trac classes in ATM networks. European Trans. Telecommunica-
tions, 5(6):747755, November 1994.
[7] C. Bruni, U. Mocci, P. Pannunzi, and C. Scoglio. Ecient capacity assignment for
ATM virtual paths. In Proc. ATM Hot Topics on Trac Performance from Race to
ACTS, No.14, Milan, 1995.
[8] J. Burgin and D. Dorman. Broadband ISDN resource management: The role of virtual
paths. IEEE Commmunications Magazine, 29(9):4448, September 1991.
[9] J.H.S. Chan and D.H.K. Tsang. Bandwidth allocation of multiple QoS classes in ATM
environment. In Proc. IEEE Infocom, pages 360367, 1994.
[10] W.C. Chan and E. Geraniotis. Near-optimal bandwidth allocation for multi-media
virtual circuit switched networks. In Proc. IEEE Infocom, pages 749757, 1996.
[11] K-T. Cheng and F. Lin. On the joint virtual path assignment and virtual circuit
routing problem in ATM networks. In Proc. IEEE Globecom, pages 777782, 1994.
Les Cahiers du GERAD G9851 Revised 31
[12] I. Chlamtac, A. Farag o, and T. Zhang. VP network design using a multicommodity
ow model. In J. Roberts, U. Mocci, and J. Virtamo, editors, Broadband Network
Teletrac: Performance Evaluation and Design of Broadband Multiservice Networks,
COST 242 nal report, pages 272273. Springer, 1996.
[13] S.-J. Chung, S.-P. Hong, H.-G. Song, and S.-B. Kim. Toward real-time optimal man-
agement of virtual paths in ATM-based B-ISDNs. In Proc. IEEE ICC, pages 225230,
1997.
[14] Z. Dziong. ATM Network Resource Management. Computer Communications. Mc-
Graw Hill, New York, 1997.
[15] Z. Dziong, Y. Xiong, and L.G. Mason. Virtual network concept and its applications
for resource management in ATM networks. In L. Mason and A. Casaca, editors,
Broadband Communications, IFIP-IEEE Conference on Broadband Communications,
pages 223234. Chapman-Hall, 1996.
[16] O. Enomoto and H. Miyamoto. An analysis of mixtures of multiple bandwidth trac
on time division switching networks. In Proc. 7th ITC, pages 635.18, 1973.
[17] S.P. Evans. Optimal bandwidth management and capacity provision in a broadband
network using virtual paths. Performance Evaluation, 13:2743, 1991.
[18] A. Farag o, J. B