The Changes of the Virtual Slices (Topology or Load)

The objectives for the fifth category of the tests are: First, showing that the COVN placement algorithm can react to the topology changes or the load changes for every VN. Second, proving the ability of the COVN placement algorithm for providing a good controller placement for all VNs of SD-WAN. These tests are performed in the following procedure:

 The number of the active physical controllers should be, three in high load and two in low load.

 Creating two different VNs for every topology:

1. Horizontal VN: includes only a group of the physical nodes which extends/distributes horizontally for hosting the virtual switches.

2. Vertical VN: includes only a group of the physical nodes which extends/distributes vertically for hosting the virtual switches.

 For every VN, implementing the COVN placement algorithm on four various statues, which are:

1. _{Wide extension: the VN includes all the nodes of this VN.} 2. Shrink extension: the VN include 65% of its nodes.

3. Heavy load: all nodes of VN work at the highest expected load (requests per second).

Page | 162  The COVN placement is applied on these VNs with the optimal weights of closeness

and reliability that were identified in the weight tests for every topology, as shown in Table 6.12.

 _{Presenting and comparing the graphs and the clusters' size or load for eight} statuses (2 VNs* 4 statuses).

 _{The maximum, average and minimum values of the connection control-path} latency and controller-connectivity ratio are compared for the eight tests to prove that the COVN placement provides a good controller placement in various statuses.  Implementing the same steps over three networks (Generated-1, Internet2 and

India), to notice how these statuses could differ for every network.

A) The VNs of Generated-1 topology: This section illustrates the implementation of the

COVN placement to adapt the control layer in reaction to the changes of the VNs. This is performed by presenting the four different statuses of virtual controller placement for every one of the two created VNs, as shown in Figure 6.34 below.

The first four graphs (A-1 to 4), introduce the different controller placements of the first VN, while the last four graphs (B- 1 to 4) exhibit them for the second VN. Graphs (A-1) and (A-2) show that the number of the active virtual nodes are minified from 24 to 16. Consequently, the number of controllers is reduced to optimise the controller utilisation regarding maintaining the same number of nodes for every virtual controller. Graphs (B- 1) and (B-2), present a similar procedure but this time, it is for the vertical VN. Both examples demonstrate that the COVN placement achieves a good load balancing in the number of nodes per controller. After that, graphs (A-3) and (A-4), display that the horizontal VN keeps the same topology, but it only changes the load of the switches. Similarly, this occurs in graphs (B-3) and (B-4) of the vertical VN. The last four graphs illustrate that the nodes' load is fairly distributed among the controllers.

Topology Closeness weight Reliability weight

Generated-1 0.3 0.7

Internet2 0.5 0.5

India 0.0 1.0

Page | 163 (A-1) Horizontal VN- Wide extension. (A-2) Horizontal VN- Shrink extension.

(A-3) Horizontal VN- Heavy load. (A-4) Horizontal VN- Light load.

(B-1) Vertical VN- Wide extension. (B-2) Vertical VN- Shrink extension.

(B-3) Vertical VN- Heavy load. (B-4) Vertical VN- Light load.

C1 8 C2 8 C3 8 T 24 C1 8 C2 8 T 16 C1 8619 C2 8347 C3 8485 T 25451 C1 8936 C2 8572 T 17508 C1 6 C2 7 C3 7 T 20 C1 6 C2 5 T 11 C1 7582 C2 7268 C3 6682 T 21532 C1 7423 C2 6928 T 14351

Figure 6.34: The possible changes with the reaction of COVN algorithm for two different VN's of Generated-1 topology.

Page | 164 The next step is comparing some of the resultant metrics of all the controller placements which are presented in Figure 6.34 to investigate the second objective. The connection control-path latency is compared for the eight statuses because it summarises the effect of changing the controller placement, as shown in Figure 6.35 (A). The figure discloses that the average value of the connection control-path latency is stable, while the maximum and minimum values of this latency occupies a small variation (about 1 ms) when changing the controller placement.

Also, Figure 6.35 (B) presents the variation of the controller-connectivity ratio for the eight tests. It shows that this variation is restricted within the acceptable range of connectivity ratio (between 52 and 42).

B) The VNs of the of the Internet2 topology: The second implementation of the eight

changes of the virtual topology is applied over the Internet2 topology as exposed in Figure 6.36 below.

The eight graphs in Figure 6.36 present the eight cases of the VNs, which are similar to those presented in the previous section. This figure reveals that the COVN placement algorithm acquires a good balance between clusters, whether it is by optimising the number of nodes or the load of the switches.

(A) Comparing the connection control-path latency (B) Comparing the controller-connectivity ratio Figure 6.35: The comparisons of some latency and reliability metrics for the eight different statuses

Page | 165 (A-1) Horizontal VN- Wide extension. (A-2) Horizontal VN- Shrink extension.

(A-3) Horizontal VN- Heavy load. (A-4) Horizontal VN- Light load.

(B-1) Vertical VN- Wide extension. (B-2) Vertical VN- Shrink extension.

(B-3) Vertical VN- Heavy load. (B-4) Vertical VN- Light load..

C1 8 C2 8 T 16 C1 8 C2 7 C3 7 T 22 C1 8202 C2 7604 C3 7930 T 23736 C1 7960 C2 7860 T 15820 C1 6 C2 5 C3 5 T 16 C1 5 C2 5 T 10 C1 5760 C2 5520 C3 5846 T 17126 C1 5999 C2 5416 T 11415

Figure 6.36: The possible changes with the reaction of COVN algorithm for two different VN's of Internet2 topology.

Page | 166 The comparison of the connection control-path latency in Figure 6.37(A), expresses that this latency varies by about 10 to 20 ms up or down when changing the VN status and the placement of the virtual controllers. The amounts of change in latency is not large compared to the long length of links for this topology, which is in thousands of km, whereas the previous topology produces latency changes around 1 ms when the length of links is limited to hundreds of km only. The abnormal behaviour of the average value of this latency at seventh test (V-shrink) resulted from shrinking the nodes of vertical VN towards the network boundaries (right side), which makes them farther from the virtual controllers (see Figure 6.36(B-2) above). However, the increment does not exceed the maximum latency of this topology. For the similar reason, the latency of the first four test (Horizontal VN tests) is lower than the last four tests (vertical VN tests). The nodes of the horizontal VN are closer to the virtual controllers than the nodes of the vertical VN.

Also, Figure 6.37 indicates that the controller connectivity ratio alternates between 45 to 59, which is good connectivity for controller nodes.

C) The VNs of India topology: The third tests of VNs are applied on India network, and

the resulted graphs are presented in Figure 6.38 below.

(A) Comparing the connection control-path latency (B) Comparing the controller-connectivity ratio Figure 6.37: The comparisons of some latency and reliability metrics for the eight different statuses

Page | 167 (A-1) Horizontal VN- Wide extension. (A-2) Horizontal VN- Shrink extension.

(A-3) Horizontal VN- Heavy load. (A-4) Horizontal VN- Light load.

(B-1) Vertical VN- Wide extension. (B-2) Vertical VN- Shrink extension.

(B-3) Vertical VN- Heavy load. (B-4) Vertical VN- Light load..

C1 26 C2 26 T 52 C1 30 C2 29 C3 29 T 88 C1 3661 C2 3431 C3 3131 T 10223 C1 3598 C2 3200 T 6798 C1 30 C2 30 C3 30 T 90 C1 28 C2 28 T 56 C1 3701 C2 3420 C3 3170 T 10291 C1 3595 C2 3247 T 6842

Figure 6.38: The possible changes with the reaction of COVN algorithm for two different VN's of India topology.

Page | 168 Figure 6.38, displays the wide extension, shrink extension, heavy load and light load statuses for the horizontal VN in the first four graphs, then present the same statuses for the vertical VN in the last four graphs. The graphs show that the COVN placement algorithm effectively responds to the changes of the VNs and successfully optimises the controllers' load with the equitable balance of the load. This proves the validity of the first objective of this category of tests.

The second objective is proven based on the comparison below in Figure 6.39. The comparison of the connection control-path latency in Figure 6.39(A), demonstrates that this latency settles with a narrow range of variation (less than 1 ms) during all the occurring changes. However, this latency falls in the second test (H-shrink) because the VN shrinks toward the centre of the network (nearer to the virtual controllers), which reduces the value of this latency. This decrease in the connection control-path latency is beneficial in this case.

Also, the comparisons of the controller-connectivity ratio in Figure 6.39(B), illustrates that their three presented values continue in the same range of connectivity, which regulates the reliability of the network.

The findings of the fifth category of tests:

 The COVN placement algorithm can maintain a good balance within a number of

nodes or a load of these nodes when re-clustering the VNs.

 _{The changes of the VN topology produces unavoidable changes in the connection} control-path latency and the connection control-path ratio. However, the COVN (A) Comparing the connection control-path latency (B) Comparing the controller-connectivity ratio Figure 6.39: The comparisons of some latency and reliability metrics for the eight different

Page | 169 placement algorithm is designed to limit these changes to smallest values keeping sufficient performance for all VNs of SD-WAN. This is proved by the results of the three topologies when considering the following issues:

1. The latency variation of Generated-1 and India topologies is about 1 ms, while the variation of the Internet2 topology is about 20 ms due to the long length of these links, which is not related to the performance of the COVN algorithm. 2. The VN gains lower latency when it extends towards the centre of the network,

which is not lower than the minimum latency of the complete topology (see India tests).

3. The VN obtains higher latency when it extends towards the boundary of the network, but it does not exceed the maximum latency of the complete network (see the tests of Internet2).

 The controller-connectivity ratio is affected by the number of the used controllers and the number of links of the network, but it is not affected by the location of the VN.

Finally, all the mentioned variations of latency are restricted to its maximum value. The maximum value of latency does not exceed the latency of the connection when its length is equal to half of the network diameter. Also, the controller-connectivity ratio is limited by the minimum value of connectivity for the location of physical controllers, which are in the optimal locations of the network. These restrictions regulate the performance of all VNs at a good level and do not produce unexpected changes that demolish the operation of the VNs.