Scalability Aware Preformance AutoTuning for OpenMP Applications
Shajulin Benedict, R. S. Rejitha, A. Alex Suja · IEEE International Conference on High Performance Computing, Data, and Analytics · 2015
Performance issue, including energy consumption, is a primordial challenge for HPC application developers and HPC resource providers, including cloud providers. On the path to exa-scale computing, autotuning would find optimal solutions considering performance parameters, such as, energy efficiency, load balance, memory access time, and so forth. This paper proposed Scalability-aware Energy AutoTuning (SCALE-EA), a dynamic runtime approach, that automatically identifies the suitable number of threads for individual parallel regions of OpenMP applications in a multi-core environment. The proposed approach was tested in a sandybridge processor machine using NAS-BT benchmark. SCALE-EA using random search provided 82.7 percentage performance improvement over exhaustive search mechanism; it manifested the need for such an autotuning mechanism when OpenMP-based HPC applications were executed on multi-core machines.