HOT Thread Level Speculation Research Based On OpenMP

With the development of multi-processor, parallel compiling, especially the parallel compiling for loops, is becoming more and more important. TLS (Thread Level Speculation) compiling can effectively improve the parallel efficiency. But the overheads, introduced by the management of the speculative data and the failure speculated thread’s rollback, hinder the improvement of the parallel performance. On the other hand, parallel programming is too hard for programmers, because programmers need to know much about the knowledge of low-end hardware and parallel programming about tasks management and overheads, which are obstacles for the application of parallel programming. Aimed at the too big overhead of data management and thread rollback, as well as the easy parallel programming, the HOT TLS (Thread Level Speculation) framework based on OpenMP is proposed. The HOT TLS framework based on OpenMP marks the loops which can be efficiently paralleled as HOT spots, using a more eager parallel way on HOT spots to improve performance, while using conservative sequence way on NONE-HOT spots to reduce the overheads. And because it is based on OpenMP, it is parallel programming friendly. The experiment result shows that HOT TLS framework based on OpenMP improves20%performance.