Actas JP2011 - Universidad de La Laguna

Actas XXII Jornadas de Paralelismo (JP2011) , La Laguna, Tenerife, 7-9 septiembre 2011(a) 5L(a) 5L(b) 3LFig. 5. Worst Fit variants comparison for different DVS levels.(b) 3LFig. 6. WF versus DP for different DVS levels.maximum speed for diverse benchmark mixes and DVSconfigurations.As observed, migration can provide huge energy savingswith respect to no migration (WF) regardless whenmigration is applied. For instance, in the 5-level systemwith task migration mixes 2 and 3 improve their energyconsumption in a factor up to 1.33 and 2.18, respectively,when compared with their execution in the same systemwithout migrations. This trend is also followed, althoughto a lesser extent, in the 3-level system.Comparing the three WF versions with task migration,it can be observed that if migration can apply only eachtime a new task arrives instead of when a task terminates,then much higher energy savings can be achieved. Themain reason is that the inter-arrival time standard deviationis higher than that of the inter-leaving time, sinceseveral tasks reach the system at the same time. Interarrivalstandard deviation values of the mixes are 24.48,43.98, and 14.65 Mcycles for mix 1, mix 2, and mix 3,respectively. On the other hand, the inter-leaving timeis, on average, 22.50, 36.40, and 12.32 Mcycles. Finally,W F in−out offers scarce benefits over W F in since it onlyadds a low number of extra migrations.Notice that if the system implements more DVS frequencylevels (5 levels in the figure), then more energysavings can be obtained since the system can select a frequencycloser to the optimal estimated by the scheduler.However, despite this fact, an interesting observation isthat energy benefits due to migration in the 3-level systemcan reach or even surpass the benefits of having the 5-level system without migrations. For example, the energyconsumption of W F out for mix 3 in the 3-level system isaround 11% of the consumption of the baseline, whereasthe same value of WF in the 5-level system is 17%.B. Comparing DP versus WF variantsThis section analyzes the energy improvements oftwo variants of the proposed DP algorithm (DP in andDP in−out ) over the WF algorithm. For comparison pur-poses the best variant of the WF (W F in−out ) with migrationhas been also included in the plots. Figure 6 showsthe results.Results show that, regardless the mix and systemlevel,both variants of DP always consume less powerthan W F in−out . DP in−out achieves, for mixes 2 and3, energy improvements over WF in a factor up to 2.74and 1.56, respectively. Moreover, for mix 1, whereW F in−out is only able to find scarce benefits over WF,the proposed DP improves the energy consumption ofWF around 1.51.For a better understanding of the algorithms behavior,we define the migration rate metric as the number of migrationsperformed by the algorithm divided by the numberof times that the migration algorithm is executed. Forinstance, regarding the in variant of the WF and DP algorithms,the migration rates of W F in are 62%, 54%,and 45% for mix 1, mix 2, and mix 3, respectively; whilefor DP in the corresponding values are 76%, 68%, and73%. This means that the proposal performs migrationsin some cases where the WF is not able to find any candidateto migrate at all.VI. CONCLUSIONSWorkload balancing has been already proved to be anefficient power technique in multicore systems. Unfortunately,unexpected workload imbalances can rise at runtimeprovided that the workload is dynamically changingsince new tasks arrive to or leave the system. To palliatethis situation this paper has analyzed the impact on energyconsumption of task migration combined with workloadbalancing.To prevent excessive overhead, task migration has beenstrategically applied at three different execution timeswhere the workload changes (at task arrival, at task termination,and in both cases). Results with respect tothe WF algorithm showed that applying migration atarrival time can save results in a factor up to around2.18. This results can be slightly improved if migrationJP2011-189