Performance instrumentation for the Intel IPSC /2

The recent proliferation of commercial hypercubes and other multicomputers has made parallel processing widely available. Unfortunately, parallel systems often are more difficult to program than conventional serial computers, and, given the nascent state of parallel programming tools, are more likely to be programmed inefficiently. The difficulties in conceptualizing the interactions among a large number of processors make it difficult both to identify the sources of inefficiencies and to determine how the parallel program could be made more efficient. This thesis describes CRYSTAL, an instrumentation system that can accurately record the occurrence of parallel program events during execution and later use the event traces to aid in program evaluation. The thrust of the system involves the presentation of event traces visually through a number of graphic tools. Visual presentation of the data is particularly useful, and, in fact, necessary, in the case of parallel computers, due to the tremendous volume of data generated by the multiple processors. Additional tools are available which compile summary statistics to provide a global view of program performance.National Science Foundation grants NSF CCR86-57696, NSF CCR87-06653 and NSF CDA87-22836 (Tapestry)National Aeronautics and Space Administration NASA Contract NAG-1-613Air Force Office of Scientific Research AFOSR Contract F49620-86-C-0136 (URI)Digital Equipment Corporation and Intel Scientific ComputersU of I OnlyRestricted to the U of I community indenfinitely during batch ingest of legacy ETDs