Actas JP2011 - Universidad de La Laguna

Actas XXII Jornadas de Paralelismo (JP2011) , La Laguna, Tenerife, 7-9 septiembre 20111: imbalance ← max core utilization − min core utilization2: target utilization ← imbalance/23: minimum difference ← MAX V ALUE4: for all task in most loaded core do5: if |U task − target utilization| < minimum difference then6: minimum difference ← |U task − target utilization|7: candidate ← task8: end if9: end for10: new max core utilization ← max core utilization − U candidate11: new min core utilization ← min core utilization + U candidate12: new imbalance ← |new max core utilization − new min core utilization|13: if new imbalance < imbalance then14: migrate(candidate)15: end ifFig. 3. Migration Attempt algorithm.B. Dynamic PartitionerThis subsection presents the proposed Dynamic Partitioner(DP). As done by the WF algorithm, DP also arrangesthe tasks arriving to the system by decreasing utilizationorder. However, before assigning any incomingtask to a given core, DP checks how the workload balancingwould become if the incoming task was assignedto the first core. Then, it also calculates the effect ofperforming a migration attempt (as shown in Figure 3).These testings are performed for each core in the system.Finally, the core assignment that provides the best overallbalance is applied. Two versions of DP are considered:DP in and DP in−out . DP in refers to the described DP algorithm,where a migration can be performed only whena task arrives to the system, while DP in−out also performsa migration attempt when a task leaves the system.Figure 4 depicts an example where the DP in heuristicimproves the behavior of W F in . The latter allocates theincoming task to core 0, and then performs a migrationattempt, but in this case, there is not any possible migra-Fig. 4. W F in vs DP in .tion enabling a better workload balancing. Thus, the finalimbalance becomes 20% (i.e., 90% − 70%). In contrast,when DP in is applied, it also checks the result of allocatingthe new task to core 1 (DP in B arrow) and thenconsidering one migration. In this case, the migrationenables a better balance since both cores remain equallyloaded with 80% of utilization, which will be the distributionselected by DP in .To sum up, the main difference between W F in andDP in is that the former selects only one core and performsa migration attempt, whereas the proposed heuristicchecks different cores, and choses the best option interms of workload balance.V. EXPERIMENTAL RESULTSExperimental evaluation has been conducted by extendingthe Multi2Sim simulation framework [20], tomodel the system described in Section III. As stated before,experiments considered a two-core processor implementingthree hardware threads each. Internal corefeatures have been modeled like an ARM11 MPCorebased processor, but modified to work as a coarse-grainmultithreaded processor with in-order execution, twoinstructionissue width, and a 100-cycle memory latency.Benchmarks from the WCET analysis project [21]were used to prepare real-time workload mixes. Thesemixes have been designed taking into account aspectssuch as task utilization, number of repetitions (task periodicity),and the sequence of active and inactive periods.The global system utilization varies in a single executionfrom 35% to 95%, in order to test the algorithms behavioracross a wide range of situations. In addition, all resultsare presented and analyzed for a system implementingthree and five voltage levels.A. Impact of Applying Migrations at Different Points ofTimeThis section analyzes the best points of time to carryout migrations focusing on the standard WF algorithm(no migration is supported) and its variants supportingmigration (W F in−out , W F in , W F out ). Figure 5 showsthe relative energy consumption compared to the energyconsumed by the system working always at theJP2011-188