Introducing Entropies for Representing Program Behaviors and Branch Predictor Performances

Predictors are inevitable components in the state-of-theart microprocessors and branch predictors are actively discussed from many aspects. Performance of a branch predictor largely depends on the dynamic behavior of the executing program, however, we have no effective metrics to represent the nature of program behavior quantitatively. In this paper, we introduce an information entropy idea to represent program behavior and branch predictor performance. By simple application of Shannon’s information entropy, we introduce new entropy, Source Entropy, that quantitatively represents the regularity level of program behavior. We show that the entropy also represents prediction performance independent of prediction mechanisms. We further discuss stereoscopic view of branch predictor performance from two entropy viewpoints, and introduce two entropies, Reference Entropy and Transition Entropy. The latter entropy offers theoretically maximum prediction performance when a predictor has table-based organization. We evaluated the proposed three entropies and prediction performance in various situations. Artificially generated branch patterns, as preliminary experiments, show overview of the entropies and prediction performance. Comparison with 2nd Championship Branch Predictor competition results show high potential of our entropy. Finally, application results to SPEC CPU2000 benchmarks show actual view of our entropies and prediction performance.