HPC-Oriented Power Evaluation Method

Abstract

In the field of high performance computing (HPC), energy consumption is an increasingly important consideration. The Standard Performance Evaluation Corporation (SPEC) Power benchmark and the Green500 are two well-known power evaluation methods. However, the former focuses on data centers and the latter concentrates on compute-intensive applications. This paper focuses on the power evaluation of single multi-core HPC servers. We analyze the limitations of these existing evaluation methods and construct a novel evaluation method using the High-Performance Linpack (HPL) and NAS Parallel Benchmarks-Embarrassingly Parallel (NPB-EP) programs. We conduct experiments on three HPC servers to test the evaluation method. The results from our evaluation method differ from the results of the Green500, and are more general and close to real-world HPC applications. We also build a regression model of power to assist in the analysis. We use the HPC Challenge Benchmark (HPCC) to train the model and use the NPB to perform verification. The R^2 representing similarities for the B and C classes of the NPB are 0.634 and 0.543, indicating that the model satisfies most cases.