Tuning Pipeline Granularity Based On Feedback Directed Framework

Parallel program languages and auto-parallelizing compiling technology are primary approaches to develop parallel software. One of the most important roles of a modern compiler is to exploit and optimize parallelism of applications whatever which program design mode will be taken. Advanced compilers for parallel program languages optimize the parallelism specified by users while auto-parallelizing compiler can exploit the hidden parallelism of a serial program and translate the serial program into SPMD(single program multi data) program automatically. Code reuse could be well done by auto-parallelizing compiler, and a viable way to parallelize a mass of existed serial applications is supplied.However, because of the complexity and diversity of applications and computation plants, auto-parallelizing compiler couldn't get the perfect parallel performance as people had thought, especially on distributed memory machines. Performance tuning for different computation plats is essential to be done. Tuning by hands is a tedious process and always need a long period. The selection of optimization strategy or the value of key factors is related with the behaviors of applications and characters of different computation plats. Feedback directed technology can get dynamic execution information of applications and tune performance automatically.We select how to optimize pipeline computations based on feedback directed framework and talk about the feasibility of adaptive optimization. Pipeline computation is a popular solution for DOACROSS loops and pipeline granularity is the key to obtain the good parallelism performance. The paper focuses on:1) Comparing the efficiency of pipeline parallelism based on different communication mode, we built a cost model based on non-blocking send and blocking receive mode. The total pipeline computation time is estimated by runtime characters of some blocks.2) Feedback directed approach is presented and realized to compute the pipeline granularity. Runtime information of the pipelined loop gained at the profiling-running phase will improve the veracity. The selections of initial block size and critical block size are used to get the appropriate pipeline granularity and reduce the cost of profiling-running phase.We performed three applications of NPB1.0 benchmarking on different computation plats and the results proved that the pipeline granularity achieved by feedback guided framework has good adaptability and speedup of the execution time of pipelined loop. Meanwhile, the veracity of the cost model was proved.