Applying evolutionary programming to improve branch classification in the hybrid branch prediction method using Switch-Counter

Abstract

Providing accurate branch prediction is critical to exploit instruction level parallelism effectively. This paper shows that applying evolutionary programming in the hybrid branch prediction method using Switch-Counter can improve branch classification, thus increase branch prediction accuracy. In our study, various EP strategies and algorithms are applied to search the optimum branch classification. Using trace-driven simulation on SPEC2000, SPEC95, and MediaBench benchmarks, we measured the branch prediction accuracy of the hybrid prediction method using Switch-Counter both applying EP and without applying EP. The empirical results show that EP could gain impressive improvements of the branch classification so that the hybrid method achieved higher prediction accuracy comparing with that without EP. The contributions for the improvements by various genetic operators are evaluated as well. The empirical results also show that the EP algorithm applied in branch classification converges fairly fast. This limits the increase of compilation time. The attempt at applying EP to improve branch classification is an innovation in branch prediction. The results are quite encouraging.