Performance Measurement and Analysis Tools for Parallel and Distributed Programs

点击翻译按钮获取中文摘要

The area of research is in the tools and methods for performance measurement and analysis of parallel and distributed programs. The research is based on two main principles. The first principle is that programmers should be supplied with the maximum information about their program. This information should be available from all levels of abstraction. To prevent the programmer from being inundate with irrelevant details, there must a logical and intuitive organization to this data. The second principle is that programmers should be supplied with answers, not numbers. The performance tool should be able to guide the programmer to the location of the performance problem, and describe the problem in terms of the source program. The current research is in techniques for automatically guiding the programmer to performance problems. The techniques for identifying critical resource utilization, determining interaction and scheduling affects, and detecting program time phase behavior will be explored. The investigators will include user interface issues to relate performance results and measurement specification to the program source code. The combination of these techniques will provide programmers with information needed to more easily make significant improvements in their programs.