Exploring thread-level parallelism based on cost-driven model for irregular programs

Abstract

Speculative Multithreading (SpMT) technology is an effective mechanism for automatic parallelization of irregular programs. For a sequential program, they can be executed speculatively in parallel by speculating that many data dependences are unlikely during runtime. Although speculative parallelization can potentially deliver significant speedup, several overheads associated with this technique can limit these speedups in practice. To take full advantage of SpMT technology, a sequential program should be programmed or compiled according the SpMT execution model. In this paper, we propose a comprehensive cost-driven compilation method to predict the resulting performance. 1)According the cost model, we can predict the resulting performance for speculative thread partitioning; 2) based on the thorough analysis of the main speculative parallelization overheads, we attempt to compress the speculative thread solution space by combining the heuristic rules. Different from prior methods that only qualitatively estimate the benefits of speculative multithreaded execution, this method also produces a quantitative estimate of the speedup in theory. Experimental results show that the proposed method is effective. we can gain 10.2% performance improvement on Olden benchmark suits.