A visualization method of performance data for large scale parallel application based on clustering of function characteristics

Obtaining performance data of parallel program and analyzing these data are the two basic steps for analysis of parallel program behavior and optimizing program design. With the rapid development of high-performance computers, unceasing expansion of the scale of parallel program, it will produce large scale performance data for each measurement. The problems that how to deal with and demonstrate these data to developers and how to assist the developers to find problems are more difficulty. Towards the profile performance data, this paper provides a visualization method based on clustering of function characteristics, which combines the function grouping with clustering of k-value optimization to process large scale performance data, so as to provide performance analysis support for developers.