Research on performance tool in HPC and grid computing

Abstract

Research on performance tool in HPC (high performance computing) usually focuses on two areas: the collecting of performance data, and the analysis of performance data. Furthermore, a proper representation of performance data in grid is also required. We explored all these three areas, and three innovation ideas are proposed. This paper explores these three aspects, and presents THGPT, a Web-and-grid-oriented performance tool, which can collect performance data about both resource utilization and the program behavior during runtime in a multithreading manner with low cost, and represent all types of performance data in a uniform ML format, and visualize the performance data in browser with JAVA applets in multiple filters and multiple views. Furthermore, a quantifying algorithm measuring the load balance of a parallel system is proposed, which considers the load balance in each time duration instead of during the whole application execution. THGPT can help the programmers to analyze performance data for improvement of the application performance in various ways with friendly interface in browsers across Internet.