Envejecimiento de la población

We evaluate the scenario in which the number of VLs configured in the system is equal to or greater than the number of applications that are being executed on the system.

For this scenario, we evaluate the benefits of segregating applications into VLs in such a way that each application is assigned to an independent VL exclusively for its entire execution. Also, applica-tions are assigned to VLs in the following order, the first application in the workload indicated on the x-axis is mapped to VL0, the second to VL1, etc. Here we assume all VLs are of low priority unless specified differently.

Uniform bandwidth allocation

In this part we will evaluate the case when the weights assigned to all VLs are the same and allow each VL to send one packet in each turn (we annotate it as vl0x1, vl1x1). We refer to this as interleaving of VLs. We compare the results of segregating with those achieved by assigning all applications to the same VL (vl0). Note that the total buffer space should remain the same size in the case of using one VL and using several VLs. Table 5.2 provides the details on per VL buffer size for each set of experiments.

Figure 5.8 shows the total inter-application contention time for the case of using only one VL and the case of segregating applications on two VLs for the two-application mixes. As can be seen, by segregating applications into two VLs a substantial improvement is achieved. In particular, reduc-tion of the contenreduc-tion time ranges from 7% for two FT applicareduc-tions executed together to 59% when FT shares the network with BT. This result indicates that in the case where there are as many VLs as applications, each application should be mapped to a different VL. The case of CG and BT where the total contention time is increased upon segregation is most probably caused by dividing up the available buffer size. Namely, when CG runs with BT on VL0 they share 64KB buffer. As BT com-municates much less then CG, CG can use the whole buffer most of the time. When we segregate two

applications on VL0 and VL1 each can use only half of the buffer space - 32KB.

Figure 5.9 shows the normalized execution time for each application for the same mixes as before.

We can see that segregating onto different VLs significantly improves the performance of both ap-plications in most cases. Also, there are some cases in which the performance of one application is substantially improved at the cost of slightly degrading the other one. In particular, when sharing the network with FT, CG may be improved by 8% through segregation, and FT is only degraded by 1%.

On the other hand, we can obtain improvements for both applications by segregation when the two are the same applications (e.g., FT with FT, CG with CG).

Figure 5.10 shows the total contention time for the three-application mixes. When increasing the number of applications in the workload we again obtain substantial reductions in the contention time owing to the segregation. Specifically, this reduction is 22% for the case of a mix of FT, CG, and BT. Moreover, a workload comprising four applications increases the individual contention time of each application and thus total contention time. However, also in this scenario segregating achieves a reduction in inter-application contention. In particular, a reduction of 32% is observed compared with the case of non-segregating.

Figure 5.11 shows the impact of segregation on each application’s overall performance in the three-applications mix (on the left) and the four-three-applications mix (on the right). The improvement in per-formance is seen for all applications but one - FT. However, degradation of FT, around 1% is less than the improvement of any other application in both three- and four-application mixes.

Non-uniform bandwidth allocation

Figures 5.12, 5.13 and 5.14 show the total contention time for the FT+CG, FT+BT, and CG+BT mixes, respectively, when assigning different weights to VLs as explained in Sec. 5.1.2. Note that we are as-suming full segregation of applications into VLs. Also, the mapping of applications to VLs is constant (vl0 - blue, vl1- red). Additionally, we show the two extreme cases of assigning full link bandwidth to VL0 and VL1 which corresponds to assigning a high priority to these VLs. The inter-application con-tention gradually diminishes in the application as its bandwidth allocation increases, but at the same

FT+CG FT+BT CG+BT FT+FT CG+CG BT+BT 0

1 2 3 4 5

6 vl0

vl0 x1, vl1 x1

Total contention time (s)

Figure 5.8:Total contention time when using one and two VLs for the two-application mixes.

FT CG FT BT CG BT FT FT CGCG BT BT

1 1.05 1.1 1.15 1.2 1.25 1.3

1.35 vl0

vl0x1, vl1x1

Normalized execution time

Figure 5.9:Impact on the execution time of each application when using one and two VLs for the two-application mixes.

time increases the contention of the other application. In other words, the ratio between individual application contention times is changing when favoring one VL (vl0, on the left) or another VL (vl1, on the right). However, a sweet spot is reached around the middle of the graphs on where the total

FT+CG+BT FT+CG+BT+MG 0

0.5 1 1.5 2 2.5

vl0

vl0x1,vl1x1,vl2x1,vl3x1

Total contention time (s)

Figure 5.10:Total contention time when using one and three/four virtual lanes for the three/four-application mixes. In FT+CG+BT, FT on VL0, CG on VL1 and BT on VL2. In FT+CG+BT+MG, FT on VL0, CG on VL1, BT on VL2 and MG on VL3.

FT CG BT FT CG BT MG

1 1.02 1.04 1.06 1.08 1.1 1.12 1.14

1.16 vl0

vl0x1, vl1x1, vl2x1, vl3x1

Normalized execution time

Figure 5.11:Impact of segregation on the execution time of each application in three/four application mixes.

contention time is lowest.

In particular, for the FT+CG mix the optimal strategy is achieved when FT (vl0) sends one packet and CG (vl1) sends two packets per turn (vl0 x1 vl1 x2). The reduction of total contention time is around 7% compared to case of equal-weight interleaving (each VL sends one packet in a turn). In application classes terms, this is the case of moderate-intensity application (FT) sharing resources with

another moderate-intensity application (CG). By assigning less bandwidth to application of higher intensity we are, in fact, performing regulation of FT’s flow. Similar behavior is also observed for the mixes FT+BT where the minimum total contention is observed at (vl0 x1 vl1 x4). Since BT is of lower intensity than CG the mutual impact is lower than in the case of FT+CG. The case of CG and BT sharing the network reaches minimum contention at (vl0 x1 vl1 x8), but overall the impact of their interference on the performance is not too high. This result suggest that the total contention time is smaller when the application that has lower communication demands is favored. Note that this bandwidth allocation strategy improves the performance of one application without excessively degrading the other one, thereby resulting in an overall decrease of the total contention time.

The case of three-application mixes is shown in Figure 5.15. For this case a similar behavior is ob-served. The largest reduction in contention is achieved when the bandwidth allocation favors the application with lower communication demand and restricts the bandwidth of applications of higher intensity (FT). In this case we can see that the optimum is achieved for vl0 x1 vl1 x2 vl2 x4 which gives more weight to the lowest communication demanding application (BT), then to the second lowest (CG) and finally, the least weight to most demanding application (FT). The improvement observed is 9% with respect to the case of equal-weight interleaving (vl0 x1 vl1 x1 vl2 x1). Note that any deviation from this bandwidth allocation leads to a suboptimal QoS strategy. For example, applying the oppo-site strategy, i.e., giving more bandwidth to more demanding applications, leads to an increase of 23%

in total contention time compared to interleaving.