Live Webcast 15th Annual Charm++ Workshop

The significant gap between peak and realized performance of parallel machines motivates the need for performance analysis. Most existing performance analysis tools provide generic measurement and displays. It is the responsibility of the users to analyze the performance of their programs using the displayed information. This is a non-trivial task, because not only does one need to identify the information that is needed for such analysis, sometimes that information may not even be displayed by the tool. The task of analysis is even more difficult for massively parallel machines, where voluminous amounts of information can be generated. Therefore, a good performance analysis tool should be able to provide intelligent analysis about the performance of a parallel program. Such automatic performance analysis is feasible for programming paradigms that provide the system sufficient information about the behavior of its programs. We have built a framework for automatic analysis for one such paradigm called Charm, a portable, object-based, and message-driven parallel programming language. In this thesis, we describe the process of design and implementation of this framework, and show its utility with sample case studies.