Exploiting Energy Saving Opportunity of Barrier Operation in MPI Programs

Abstract

In high performance parallel computing, energy consumption has become a more and more important problem. Under the support of the voltage scaling- available processors, reducing CPU voltage/frequency in the non-critical path is an efficient energy reduction method. This paper exploited the energy saving opportunities in MPI_Barrier operation. Through testing NPB3.2-MPI programs on 256-process cluster system, we found that the busy time of each process has the big difference each other when the cluster is connected by a 100 Mbps network. We proposed two low-power methods: one is shut-down based MPI_Barrier (SDMB) and the other is dynamic voltage/frequency scaling based MPI_Barrier (DVSMB). Experimental results show that average 4.95% energy saving is obtained through SDMB and average 12.28% energy saving is obtained through DVSMB. SDMB avoids the performance loss through pre-activation strategy and the performance loss of DVSMB is less than 1%.