Power and Performance Modeling of Scientific Applications for Energy Optimization in High Performance Computing

Abstract

The increase in power consumption of High Performance Computing (HPC) systems have became an important concern. Many decades of race for performance has increased the total power consumption of supercomputers. Very few studies provide deep insights into the power and energy consumption of scientific applications. A detailed performance, power and energy analysis is essential to identify the most compute intensive and costly parts of an application and to develop possible improvement strategies. In this paper, we focus on power and performance modeling of various HPC benchmarks and scientific applications in order to reduce energy consumption. We study the power-performance efficiency and conserve energy using Dynamic Voltage and Frequency Scaling (DVFS) for scientific applications such as High Performance Linpack (HPL) benchmark, NAS Parallel Benchmarks (NPB), Multiple Em for Motif Elicitation (MEME), STREAM and Seasonal Forecasting Model (SFM). The HPC applications are executed in certain voltage and frequency (v/f) for CPU. The v/f used for execution of a job has a key impact on overall energy consumption. We evaluated HPC scientific applications by changing v/f during run-time for CPU and observed an average measured energy savings of 11.6% and a maximum of 14.8% with less than 5% performance degradation. The approach is to reduce energy at run-time by slowing down (applying voltage and frequency scaling) the processor during light workloads. The processor will deliver high performance whenever required, while significantly reducing power consumption during low workload periods. Our idea is to profiling and characterizing the application into multiple sections using time slicing to efficiently determine optimal frequency and voltage combinations over all the sections in the application using knowledge base. Selecting an optimal frequency and voltage by profile based technique requires several runs of the application with varying input data sizes. Another methodology is to combine power measurements and performance modeling to predict energy consumption for various frequency and voltage combinations. The findings and analysis of these high performance scientific applications can be used by energy aware schedulers to reduce energy consumption with very less or no performance degradation.