Design Of PBuster-a Performance Debugging Environment For Message-Passing Parallel Programs

The ultimate goal of parallel computing is to achieve higher performance, but, current parallel machines achieve poor performance, which only reach 10%~20% of the peak performance. The processors, memory and network bandwidth are all far from full utilization. Performance debugging for parallel program aims to help user to achieve as high performance as possible with existing parallel model and hardware, so that all kinds of resources could be fully utilized. The focus of this thesis is Message-Passing parallel program, which excels other kinds of parallel model in universality, expressivity and performance.The major work of this thesis is as below:The definition and model of the parallel performance debugging environment is clarified: collect performance data——performance data visualization and analysis——modify the source program, such process is repeated until satisfactory performance is achieved. The key issue of performance debugging is concluded to be nondeterminacy, disturbance and performance data visualization.As a major component of the parallel performance debugging environment for clusters, a performance debugging tool—PBuster. Is designed. PBuster is based on data measurement, which adopts interactive debugging mode and state-freezing control manner, has some unique feature in data visualization.Based on analysis of the operating model of MPI programs, a model is proposed that MPI programs can run under the control of PBuster and meanwhile performance data are collected and visualized.PBuster adopts dynamic code instrumentation to collect performance data, to realize on-line debugging mode and performance metrics tailoring, so as to achieve interactive debugging mode.In performance data visualization, a procedure time-gram is brought forward on that base of process time-gram. Procedure time-gram can locate exactly where a event happens, advantageous for users to analyze.Critical path analysis is brought forward for parallel programs. the clock problem in message-passing parallel programs is discussed and analyzed. According to the initializing and finishing mechanism, the process of parallel program is abstracted to be an AOE network, so that it can be analyzed by critical path algorithm.