On the energy footprint of task based parallel applications

From HPC systems to embedded devices, energy consumption is becoming a dominant factor in managing costs. With Chip multiprocessors becoming the platform of choice in almost all ICT segments, software developers need to employ parallel programming to fully exploit this architecture. However, parallelization adds a management overhead to the execution of an application. In this paper, we study parallel applications implemented using two TBP libraries, Wool and Intel TBB. We explore both compute and memory bound executions. We also investigate the energy footprint of the parallelization overhead and the effect it has on the energy-efficiency of the executing system. Our study looks into the behavior of some basic parallelization operations like task spawning, task synchronization and task stealing. We encountered situations when the number of task stealing operations grows exponentially with the core count increasing execution time. This behavior drastically reduces the energy-efficiency of those executions. Avoiding such behavior is crucial if future parallel systems are to reach their performance potential. Our results also show that the energy efficiency for compute intensive applications improves with the core count while for memory intensive application that is not the case.