Académique Documents
Professionnel Documents
Culture Documents
probability of link congestion. The objective of - The total arrival rate in bits/sec from all commodity
“minimizing” the probability of congestion is achieved services in the network:
by having balanced utilized links; selecting, as much as T = ∑ Ts = ∑ ∑ λ ps (3)
possible, routes with smaller numbers of hops; and taking s∈S p∈P s∈S
into consideration the likelihood of load congestion in - The total arrival rate in bits/sec from commodity service
the network links. s in link according to the routing decision:
Als = ∑ ( yrl λ rs x rs ) (4)
The following notation is used in the proposed model: r ∈R
- The total arrival rate in bits/sec from all commodity
P The set of the communicating source/destination services in link according to the routing decision:
pairs in the network.
S The set of the commodity services to be transmitted Al = ∑ Als = ∑ ( yrl * ∑ λ rs x rs ) (5)
s∈S r∈R s∈S
in the network (voice, data, video, etc.)
L The set of all links in the network - The utilization of link : Ul =
Al + El (6)
R The set of candidate routes. This set is provided by a Cl
route generation algorithm (e.g. double-sweep - The average delay in each link [2]:
algorithm). A route is characterized by the ordered 1
set of links (from source to destination) in the route. Dl = + dl (7)
Cl − Al
V The average number of links in the routes that belong
to R. - The average networkwide delay for commodity service s
1
Rp The set of candidate routes for communicating pairs
p. Obviously we have:
Ds =
Ts
∑ Als * Dl (8)
l ∈L
∑ ( Al + El )
1 1 1 congestion degree of each candidate route, grs. This
U avg = ∑ U l = (13)
L l ∈L L Cl l ∈L estimated degree can be calculated from the list of all
candidate routes known in advance between all
communicating pairs of nodes. For a specific route, it
Also, we can have the following approximation for the
reflects the number of other routes fighting to share its
previous summation:
links. This term is weighted by w2, the same as the
∑ (Al + El ) = T * V
utilization terms because it helps in avoiding congestion.
(14)
l ∈L
Including the estimated congestion degree, grs, as well
as the current load, El, in the model is beneficial in two
where V, as defined before, is the average number of links
perspectives. It allows the model to be applied as a
in the routes connecting all communicating pairs in the
dynamically routing tool. Also, as we will see in the case
network. T, in this case, is not only the total arrival rate of
study, using traditional optimization tools limit the
all commodities to be routed but also the load already
number of the variables in the model. This can be relaxed
routed in the network (the current load in the network.)
by solving the routing problem in multiple stages. In
every stage the model is used to route the traffic between
So, now we have Ds (Equation 8) and Ul (Equation 6)
a subset of the communicating pairs.
as a linear expression in the zero-one variable x. Also,
Uavg is a constant. This makes the proposed optimization
model to be zero-one linear programming.
w1 w
max w1 . Z1 + w2 . Z2 + . ∑ t s + 2 . ∑ tl −
Maximizing the minimum value of membership S ∀s L ∀l
w3 w2
function in fuzzy optimization models is widely used. In
the case of the routing problem and the proposed model
∑ crs . xrs − V. R . S ∑ grs . xrs
V. R . S ∀r , ∀s ∀r , ∀s
this objective is not enough. For example, if we have a
link that is no way to be used in many routes, then it will Such that
be overloaded resulting in poor or small membership Z1 − t s ≤ 0 ∀s ∈ S
Z 2 − tl ≤ 0 ∀l ∈ L
utilization function. This low membership value, with no 1
way to be improved, will make the overall optimization (ds − ds min ) * t s −
Ts
∑ ∑ dl ( yrl λ rs xrs ) ≤ −ds min ∀s ∈ S
process to stop seeking for improvement in the other l∈L r∈R
1
utilization balancing for other links (because there is no (ds max − ds ) * ts +
Ts
∑ ∑ dl ( yrl λ rs x rs ) ≤ ds max ∀s ∈ S
way to maximize the minimum utilization membership l∈L r∈R
function anymore.) 1 1
( * * T * V − Umin ) * t l −
L Cl (P4)
To solve this problem, other terms may be added to
1 E
the objective function. These terms are the normalized ∑ ∑
Cl s∈S r∈R
( yrl λ rs x rs ) ≤ l − Umin
Cl
∀l ∈ L
weighted summation of the membership functions for all
traffic services delay and links utilization. Now we can 1 1 1
make sure that the optimization model will maximize all (1 − * * T * V ) * tl +
L Cl Cl
∑ ∑ ( yrl λ rs xrs )
membership functions as well as the minimum of them. s∈S r∈R
El
This will have the effect of having homogeneously ≤ 1− ∀l ∈ L
distributed traffic in the whole network. Cl
Delay in msec
25.0
- Load deviation: It is a measure of how the load is
24.0
distributed all over the network links. It is the
23.0
standard deviation of the links utilization.
22.0
- Average propagation delay (ms): It is the propagation
21.0
delay experienced by traffic.
- Average number of hops/path: For all paths used to 20.0
LL traffic HB traffic
5.4. Results and analysis
Figure 6. Effect of the Segment size on traffic
In Figure 5 and 6, the proposed model has been used delay: W=[1 1 1] and load = 60MB/node
to route a load of 60 Mbps between the communicating
pairs. The weighting coefficients are assigned equal Figures 7 to 12 show the effect of changing the values
values of 1, i.e. W=[1 1 1]. The pairs are divided into of the weighting coefficients. W=[1 1 1] represents the
segments of size N (number of source/destination pairs model in its recommended mode. For W=[0 0 1], the
per segment). The model has been tested for different model will optimize only the number of hops in all routes
values of N (10, 15, and 20). From the figures increasing to be minimized. Finally, W=[1 0 0], will minimize the
the segment size has a little effect on the performance. delay of the services in the routes (in this case ds= 0 ms
This is due to taking into consideration while solving the for both services.)
model for the first segments the degree of estimated
congestion as well as the updated load in the links. This In Figure 7, for a load of 70 Mbps per each pair, the
means that solving the problem using a small segment throughput as a function of the load is represented. Using
size will not affect the results dramatically while it will W=[1 1 1 ] we get the highest throughput for all segment
decrease dramatically the run time. In our case having 10 sizes. This is due to the fact that balancing of the load in
pairs per segment (25% of the whole pairs) seems to be a all links as a goal while solving for the first segments will
reasonable choice. help in avoiding congestion in routing the later segments.
Hence, segment size of 10 is a reasonable choice that
compromises between being a small value and offering
40 good performance.
35
30 100
25
Percentage
90
20
15 80
Throughput
10
70
5
0 60
10 15 20
Segment Size
50
Load Deviation Average Utilization 5 10 15 20
Segment Size
Figure 5. Effect of the Segment size on load
W=[1 1 1] W=[0 0 1] W=[1 0 0]
deviation and average utilization
W=[1 1 1] and load=60 MB/node Figure 7. Effect of segment size on throughput
for different W: Load = 70 MB/node
23.50
90
22.50
70
22.00
60
21.50
50 21.00
60 80 100 120 140 40 50 60
Load per node in MB/sec Load per node in MB/sec
Figure 8. Effect of different loads/pair on the Figure 10. Effect of different loads/pair on the
throughput for different Ws: Segment Size = 10 delay of LL traffic for different Ws:
Segment size= 10
From Figure 8, for low load value the three modes have
26.00
100% throughput. For higher loads (70 Mbps and 110
25.50
Mbps), the recommended mode has the highest
25.00
throughput because with higher load, resources become Delay of HB traffic in msec
more scarce, and the recommended mode outstands the 24.50
load becomes very high (greater than 120 Mbps), the 23.00
performance of the three modes turns poor and almost the 22.50
same. The reason is that with very high loads, all links 22.00
are likely to be congested. 21.50
40 21.00
40 50 60
Load per node in MB/sec
35
W=[1 1 1] W=[0 0 1] W=[1 0 0]
30
25
delay of HB traffic for different Ws:
20
Segment size = 10
15
10 The third mode (W=[1 0 0]) has the lowest delay among
the three modes as shown in Figures 10 and 11. One of
5
the objectives of the recommended mode is to have the
0
40 50 60
delay of both traffic classes in the vicinity of their ds
Load per node in MB/sec value. That is why, for this mode, the delay of LL traffic
W=[1 1 1] W=[0 0 1] W=[1 0 0]
is low while for HB traffic is high as required.
Figure 9: Effect of different loads/pair on the The effect of using the three modes of operation on the
load deviation for different Ws:Segment size = 10 average path length (in hops) is shown in Figure 12. As
expected, the second mode of operation (W=[0 0 1]) has
Figure 9 shows that the recommended mode has the the lowest average number of hops in the routes as it is its
lowest load deviation among the three modes. The other main function. The third mode has the worst values for
two modes are almost the same because they do not take different load while the recommended mode comes in the
into consideration the balancing of the load among the middle.
network links. In this figure, 10 communicating pairs per
segment are used.