Dynamic Code Region-based Program Phase Classification and Transition Prediction

Abstract

Detecting and predicting a program's execution phases is crucial to dynamically adaptable systems and dynamic optimizations. Program execution phases have a strong connection to program control structures, in particular, loops and procedure calls. Intuitively, a phase can be associated with some dynamic code regions that are embedded in loops and procedures. This paper proposes off-line and on-line analysis techniques could effectively identify and predict program phases by exploiting program control flow information. For off-line analyses, we introduce a dynamic interval analysis method that converts the complete program execution into an annotated tree with statistical information attached to each dynamic code region. It can efficiently identify dynamic code regions associated with program execution phases at different granularities. For on-line analyses, we propose new phase tracking hardware which can effectively classify program phases and predict next execution phases. We have applied our dynamic interval analysis method on 10 SPEC CPU2000 benchmarks. We demonstrate that the change in program behavior has strong correlation with control transfer between dynamic code regions. We found that a small number of dynamic code regions can represent the whole program execution with high code coverage. Our proposed on-line phase tracking hardware feature can effectively identify a stable phase at a given granularity and very accurately predict the next execution phase.