Principal Component and Cluster Analysis of SPEC CPUint2006 Benchmarks: Input Data set Selection

Abstract

Technological scaling of processor parameters has a critical limit. The Scaling advanced CMOS technology to the next generation effects improves performance, increases transistor density, and reduces power consumption of the processor. In this paper we describe the statistical analysis of SPEC CPUint2006 benchmarks workload and input data selection for microarchitectural research. Today we need a processor which can provide high performance boost for a broad spectrum. We use statistical analysis techniques, Principal Component Analysis (PCA) and Cluster Analysis (CA) for the study of benchmark workload classification using recently published SPEC CPUint2006 performance numbers of thirty Intel’s commercial processors. We calculated five most significant PCs, which are retained for 85% of the variance, PC2, PC3, PC4 and PC5 covers 11.1%, 2.9%, 0.6% and 0.1% variance respectively. We classified the CINT benchmarks in two sub groups. We found that the benchmarks 471.omnetpp, 462.libquantum 403.gcc, and 429.mcf exhibits higher memory wait time. Our results and analysis can be used by performance engineers, scientists and developers to better understand the benchmark workload and select input dataset for better microarchitecture design of the processors