Skip to Main content Skip to Navigation

Multi-Purpose Efficient Resource Allocation for Parallel Systems

Abstract : The field of parallel supercomputing has been changing rapidly inrecent years. The reduction of costs of the parts necessary to buildmachines with multicore CPUs and accelerators such as GPUs are ofparticular interest to us. This scenario allowed for the expansion oflarge parallel systems, with machines far apart from each other,sometimes even located on different continents. Thus, the crucialproblem is how to use these resources efficiently.In this work, we first consider the efficient allocation of taskssuitable for CPUs and GPUs in heterogeneous platforms. To that end, weimplement a tool called SWDUAL, which executes the Smith-Watermanalgorithm simultaneously on CPUs and GPUs, choosing which tasks aremore suited to one or another. Experiments show that SWDUAL givesbetter results when compared to similar approaches available in theliterature.Second, we study a new online method for scheduling independent tasksof different sizes on processors. We propose a new technique thatoptimizes the stretch metric by detecting when a reasonable amount ofsmall jobs is waiting while a big job executes. Then, the big job isredirected to separate set of machines, dedicated to running big jobsthat have been redirected. We present experiment results that show thatour method outperforms the standard policy and in many cases approachesthe performance of the preemptive policy, which can be considered as alower bound.Next, we present our study on constraints applied to the Backfillingalgorithm in combination with the FCFS policy: Contiguity, which is aconstraint that tries to keep jobs close together and reducefragmentation during the schedule, and Basic Locality, that aims tokeep jobs as much as possible inside groups of processors calledclusters. Experiment results show that the benefits of using theseconstrains outweigh the possible decrease in the number of backfilledjobs due to reduced fragmentation.Finally, we present an additional constraint to the Backfillingalgorithm called Full Locality, where the scheduler models the topologyof the platform as a fat tree and uses this model to assign jobs toregions of the platform where communication costs between processors isreduced. The experiment campaign is executed and results show that FullLocality is superior to all the previously proposed constraints, andspecially Basic Backfilling.
Document type :
Complete list of metadatas

Cited literature [54 references]  Display  Hide  Download
Contributor : Abes Star :  Contact
Submitted on : Friday, January 12, 2018 - 2:19:30 PM
Last modification on : Wednesday, October 14, 2020 - 4:18:41 AM
Long-term archiving on: : Wednesday, May 23, 2018 - 8:09:12 PM


Version validated by the jury (STAR)


  • HAL Id : tel-01681424, version 2



Fernando Mendonca. Multi-Purpose Efficient Resource Allocation for Parallel Systems. Computer science. Université Grenoble Alpes, 2017. English. ⟨NNT : 2017GREAM021⟩. ⟨tel-01681424v2⟩



Record views


Files downloads