Comparative Study of Path Prediction Method for Speculative Loop Execution

Execution path ratio is mostly dominated by two execution paths in program loops. We have developed Two-Path Limited Speculation Method that achieves speed-up in programs using optimization of the most frequent two paths and speculative multi-thread execution of them. The path predictor used in the method predicts the next execution path in Two-Path Limited Speculation Method, and plays an important role in performance improvement of the method. In this paper, we apply several well-known branch prediction methods to the path prediction and evaluate them in terms of mis-prediction ratio. Experimental results show that the mis-prediction ratios of the path predictors vary from 10% to 45%, depending on the benchmark programs, and are 20% on average, the Gshare path predictor performs best in eight path prediction methods.

1. Introduction

It is widely recognized that most programs tend to spend the majority of their execution time in a small portion of their code, i.e., known as the 90/10 rule [1]. Most of this code is located within loop structures, and thus it is important to optimize the execution of the frequently executed loops. Recent development of multicore architecture has paved the way to their speculative multithreaded execution utilizing their regular structure. Loop iterations are defined as speculative threads. However, in order to realize speculative multithreaded execution of loop iterations, there are many issues such as the generation of optimized speculative thread, the prediction of speculative thread, the resolution of inter-iteration data-dependency and the treatment of speculative memory access operations [2].