A HOT PATH BASED THREAD PARTITIONING TECHNIQUE FOR THREAD PIPELINING MODEL

摘要

We have developed a software system that can translate sequential binary codes to the multithreaded ones at the binary code level. This system parallelizes frequently executed loops within programs to improve the performance by exploiting loop-level parallelism for numerical application programs. However, many loops within non-numerical application programs are hard to be parallelized because of their small amount of loop-level parallelism and complex program structures, and it is difficult to attain speedup by exploiting loop-level parallelism. Therefore, it is necessary to generate multithreaded codes, taking more general program structure than loop structure within program into consideration. In this paper, we propose a thread a hot path (the most frequently executed control sequence of basic blocks), taking more general program structure than loops into consideration. And we apply the proposed technique to three non-numerical application programs of SPEC CINT95 benchmark, and evaluate the performance improvement by simulation. The evaluation result shows that the proposed technique can improve the performance of non-numerical application program, whose performance cannot be improved by exploiting loop-level parallelism.