A detailed study on phase predictors

Abstract

Most programs are repetitive, meaning that some parts of a program are executed more than once. As a result, a number of phases can be extracted in which each phase exhibits similar behavior. These phases can then be exploited for various purposes such as hardware adaptation for energy efficiency. Temporal phase classification schemes divide the execution of a program into consecutive (fixed-length) intervals. Intervals showing similar behavior are grouped into a phase. When a temporal scheme is used in an on-line system, phase predictors are necessary to predict when the next phase transition will occur and what the next phase will be. In this paper, we analyze and compare a number of existing state-of-the-art phase predictors using the SPEC CPU2000 benchmarks. The design space we explore is huge. We conclude that the 2-level burst predictor with confidence and conditional update is today's most accurate phase predictor within reasonable hardware budgets. © Springer-Verlag Berlin Heidelberg 2005.