Study on Granularity Tuning Mechanism in Fine-Grain Parallel Processing

点击翻译按钮获取中文摘要

This research addresses to give granularity tuning mechanism for obtaining the maximum speed in parallel program execution on a distributed memory-type parallel computer. The mechanism requires to exploit parallelism in a program and divide the computation into subprograms (threads) of suitable granularity. Performance improvement can be achieved to allocate the suitable-size threads statically to each processing element of a parallel computer by efficiently utilizing its hardware's potential ability.(1) Parallelizing compiler. We proposed a novel granularity tuning mechanism based on a LogP model, which is an abstract parallel computer model for analyzing execution time of a parallel program. The mechanism was experimentally installed in a SISAL compiler for a dataflow computer EM-X,and some evaluation study was carried out by executing benchmark programs with do-all loops and do-across loops.(2) Performance debugging for parallel programs : A performance debugging system for a parallel program was developed, which shows an execution process of a parallel program as a form of a Gantt chart. A programr can find performance bottleneck by carefully checking the chart, and interactively change its execution sequence, as a result the parallel granularity, in order to obtain better performance. We proposed 3D animation technique based on a dynamical model, which is adopted to visualization of a process of parallel discrete event. A 3D visualization tool was developed and the utilization for parallel debugging was investigated.