A User Friendly Phase Detection Methodology for HPC Systems' Analysis

Abstract

A wide array of today’s high performance computing (HPC) applications exhibits recurring behaviours or execution phases throughout their run-time. Accurate detection of program phases allows reconfiguring the system for a better power/performance trade off; and can reduce the simulation time of programs by identifying regions of code whose performance is critical to the entire program. Program phases are also reflected in different behaviours the system goes through or system phases, which can be used as an alternative means of program phase detection for users lacking expertise. In this paper, we present an execution vector based (EV-based) phase detection, which is an on-line methodology for detecting phases in the behaviour of a HPC system and determining execution points that correspond to these phases. We also present a methodology for defining a small set of EVs representative of the system’s behaviour over a fixed period of time and show that EV-based phase detection identifies recurring phases. Our methodology is illustrated with benchmarks and a real life application.