In nowaday multi-core or many-core era, the most difficult problem is how to make use of the abundant hardware resources.There are three bottlenecks limiting program parallelism: indirect access, inputdependent parallelism and unpredictable control flow.Thread level speculation has been proposed and researched recently to parallelize traditional sequential applications on CMP.In this paper, a modified CMP hardware simulation system is present, providing with TLS implementation principle and a convenient programming model.With a lot of parallel tuning techniques, the three bottlenecks have been well alleviated.Gzip is parallelized from -2% to 172% on a four-core processor, and the speedup of the test applications are 7%, 22% and 126%, respectively with arbitrary, hotspot and insight speculation.Therefore, the efforts are benefit and promising for the parallel computing.
展开▼