Experimental Results

Local and External Request Rejection Rate

In Table 3.1, the mean difference of decrease in local requests rejection rate is reported along with a 95% confidence interval of the difference. We report the difference between rejection rate in two situations; First, when no preemption policy is in place, and second, when the MOML policy is used as the preemption

policy. We use a T-test to determine the mean difference between these two policies. To perform the T-test we have ensured that the distribution of differences is normal.

According to Table 3.1, local request rejection rate significantly decreased statistically and practically by applying preemption in all cases. More impor-tantly, this reduction in the local request rejection rate was achieved without rejection of more external requests. Based on Table 3.1, external request rejec-tion rate does not change significantly in any of the experiments.

Table 3.1: Mean difference and 95% confidence interval (CI) of decrease in local requests rejection rate and external requests rejection rate as a result of lease preemption in an RP of InterGrid.

Modified Parame-ter

Mean Decrease in Local Requests Re-jection Rate

CI of Decrease in Lo-cal Requests Rejection

72.0% (51.1,92.8), P-Value=0.001 Not statistically signifi-cant, P-Value=0.6 Percentage of DC

Ex-ternal Requests

54.3% (35.0,73.7), P-Value=0.001 Not statistically signifi-cant, P-Value=0.3 Percentage of Local

Requests

58.2% (40.3,75.9), P-Value<0.001 Not statistically signifi-cant, P-Value=0.6

Based on this experiment, the maximum reduction in the local request re-jection rate occurs when the percentage of best-effort external requests is higher (the first row in Table 3.1). In this circumstance, more local requests can be accommodated with preemption of best-effort leases.

Resource Utilisation

In this experiment, we measure the resources utilisation when different preemp-tion policies are applied.

In all sub-figures of Figure 3.4, it is observed that the MOV policy results in better utilisation comparing with the other policies. However, in a few points (e.g., in Figure 3.4(a) when 40% of the requests are best-effort), MOV has slightly less utilisation than MOML. The reason is resource fragmentations (i.e., unused spaces) in the scheduling queue, which leads to lower resource utilisation. Sub-figures of Figure 3.4 also demonstrates that the resource utilisation MOML lies between MLIP and MOV.

Figure 3.4(a) indicates that increasing the percentage of best-effort requests improves the resource utilisation; however, after a certain point (i.e., best-effort>20%) resource utilisation does not fluctuate significantly in different policies. Indeed, in this situation unused spaces are allocated to the preempted leases.

Best Effort Ext. Request (%)

Deadline Constraint Ext. Requests (%)

Utilization (%)

Figure 3.4: Resource utilisation results from different policies. The experiment was carried out by modifying (a) the percentage of best-effort external requests, (b) the percentage of deadline-constraint external requests, and (c) percentage of local requests.

In Figure 3.4(b) shows that resource utilisation increases by increasing the percentage of constraint requests in all policies. In fact, more deadline-constraint requests imply fewer preemptions and more resource utilisation. As expected, the MOV policy outperforms other policies due to preemption of leases that impose less overhead.

In Figure 3.4(c), it is expressed that by increasing the percentage of local requests, the number of preemption and subsequently the amount of overhead is increased. Therefore, resource utilisation decreases almost linearly in all policies.

Another reason for the reduction in resource utilisation is that local requests are not preemptable and their scheduling leads to many fragmentations in the scheduling queue.

Number of Lease Preemptions (Resource Contention)

The number of external leases that are preempted in different preemption policies indicates the amount of resource contention in the system.

Figure 3.5(a) shows that when the percentage of best-effort requests in-creases, the number of preemptions rises almost linearly. For the lower percent-ages of best-effort external requests (best-effort<30%), MOML behaves similarly to MOV, however, after that point MOML approaches MLIP. The reason is that when the percentage of best-effort leases is high, the likelihood of having a can-didate set with the minimum number of leases and not large overall overhead is high. Thus, MOML approaches MLIP.

Best Effort Ext. Request (%)

No. Lease Preemption

Deadline Constraint Ext. Requests (%)

No. Lease Preemption

Figure 3.5: Number of lease preemption resulted from different policies by chang-ing (a) percentage of best-effort external requests, (b) percentage of deadline-constraint external requests, and (c) percentage of local requests.

Figure 3.5(b) demonstrates that the number of preemptions does not vary significantly when the percentage of deadline-constraint requests is less than 40%.

In fact, in this situation there is enough best-effort requests for preemption and changes in the percentage of deadline-constraint requests does not play an im-portant role.

Figure 3.5(c) reveals the impact of number of local requests on the resource contention. It shows that in all policies the number of lease preemptions is in-creased almost linearly with the increase in the percentage of local requests.

In general, in all sub-figures of Figure 3.5, MLIP results in fewer number of lease preemptions (resource contention) and MOML operates between MLIP and MOV.

Average Response Time

In this experiment, we investigate the impact of different preemption policies on the average response time of best-effort external requests. The results of the experiment under different workloads are illustrated in Figure 3.6.

All subfigures of Figure 3.6 show that MLIP leads to smaller response time in comparison to other policies. The reason is that MLIP disregards the type of leases for preemption. This means that, comparing with MOV, it is less likely that MLIP will preempt best-effort requests. Therefore, the best-effort requests are completed earlier and their average response time is lower in MLIP.

Figure 3.6(a) demonstrates that, by increasing the percentage of best-effort requests, the average response time decreases after a certain point. When 20%

of external requests are best-effort, the average response time reaches its peak because of numerous preemptions occur. However, after that point we notice a decrease in average response time of best-effort requests. This decrease occurs due to fewer deadline-constraint requests and more opportunities for local requests to be allocated. When 10% of the external requests are best-effort, since there are not many preemptable requests in the system, many local requests are rejected and few preemption occurs. Hence, the average response time is low in that point.

Figure 3.6(b) shows that, by increasing the percentage of deadline-constraint requests, the average response time decreases. In fact, increasing the percentage of deadline-constraint requests implies fewer best-effort external requests in the system. Therefore, the average response time for best-effort external requests decreases.

Figure 3.6(c) illustrates that, by increasing the percentage of local requests in the system (and consequently increasing the number of preemptions), the average response time increases. However, the reason for stable situation in ART, when local requests are more than 50%, is that there are many local requests in the system that collide and rejected. Therefore, the number of local requests after that point does not vary significantly and the impact on ART is not substantial.

Best Effort Ext. Request (%)

Deadline Constraint Ext. Requests (%)

ART(h)

Figure 3.6: Average response time (ART) resulted from different policies. The ex-periment is carried out by altering (a) percentage of best-effort external requests, (b) percentage of deadline-constraint external requests, and (c) percentage of local requests.

3.4 Summary

In this chapter, we investigated how origin-initiated resource contention between local and external requests can be resolved in the local scheduler of a resource provider (RP) in InterGrid. For this purpose, we applied preemption mechanism to preempt external leases in favour of local requests. We observed that preemp-tion of leases substantially decreases the rejecpreemp-tion of local requests (up to 72%) without increasing external requests rejection rate. Furthermore, we investigated the side-effects of the preemption mechanism when VMs are utilised for resource provisioning. Specifically, we modelled the overhead of suspension and migration operations on VMs of leases.

Then, we proposed 3 policies to decide which lease(s) are better choices for preemption. The MOV policy aims at minimising the imposed overhead time and improving resource utilisation. The MLIP policy results in less resource contention and increases user satisfaction. However, it does not lead to a high

re-source utilisation. Finally, the MOML policy makes a trade-off between rere-source utilisation and resource contention.

This chapter tackles the problem of resolving resource contention using pre-emption mechanism at the local scheduler level through prepre-emption policies. In the next chapter, we investigate how resource contention can be avoided by proac-tive scheduling of external requests in the meta-scheduler level of InterGrid (i.e., in the IGG).

Proactive Resource Contention