HPC Benchmark Assessment with Statistical Analysis

Abstract

High-performance computing (HPC) benchmarks are widely used to evaluate and rank system performance. This paper introduces a benchmark assessment tool equipped with a rigorous statistical method to evaluate HPC benchmarks against a set of scientific applications. The method is based on the combination of Variable Clustering (VarCluster) and Principal Component Analysis (PCA). We built the tool upon HPC Challenge (HPCC) benchmark suite and six popular scientific applications of Kraken, a petaflop supercomputer. Experimental results show that HPCC's Fast Fourier Transform (FFT) kernel, rather than the High-Performance Linpack (HPL) on which Top500 is based, is more representative of the HPC workloads on Kraken.