AN AUTOMATIC LOGFILE ANALYZER FOR PARALLEL PROGRAMS

Abstract

Detecting communication errors in parallel programs has remained a very challenging research and application area. In the thesis, we will present a new approach to detect communication errors in parallel programs. We also implement a complete tool to achieve this goal. The Logfile Analyzer is a tool for automatically analyzing the logfiles generated during a parallel program execution. The purpose of the tool is to check the communication consistency of parallel programs using MPI. It can help the programmer to detect the communication errors, improve the parallel program reliability, and give the programmer an overall picture of the communication sequence. Along with the Analyzer, the tool also provides the Logger, the Wrapper and the Preprocessor, enabling the automatic generation of the logfiles from MPI programs. The logging procedure is hidden from the user. The logfiles contain all information needed for analysis. Currently, the tool supports all major MPI functions and can be run on UNIX and Windows systems. The tool is implemented in C++. It is designed to be extensible and reusable. In the thesis, we will discuss the design idea and implementation details. We also provide a substantial number of testing cases to prove tool utility.