GRAPH: A Tool for Visualizing Communication and Optimizing Layout in Data-Parallel Programs

*GRAPH is an event-driven visualization and optimization tool for data-parallel communication, currently implemented for the C* language on the CM-5. Its goal is to help data-parallel programmers reduce their communication costs by providing information about communication bottlenecks and suggesting system directives for an optimized data layout. To accomplish this, *GRAPH provides facilities for 1. visualizing a trace of internode communication, 2. pinpointing source code statements where communication bottlenecks occur, and 3. optimizing data layout for the program's communication pattern. This paper presents the motivation for *GRAPH, its design concept and the initial implementation. We include two case studies illustrating how performance can be improved by applying *GRAPH's recommended data layout directive. Elapsed execution times were reduced by 31% for a folded convolution algorithm, and by 17% for a Gaussian elimination.