Instrumentation for a massively parallel MIMD application

Abstract

This paper describes an application and a set of performance tools developed for a simulated machine called Horizon. The application chosen for this study is Monte Carlo radiative heat transfer in an arbitrary 2D enclosure. It was chosen because the solution procedure is highly parallel, each particle modeled being independent of the others. The tools described in this paper were built to show the performance of massively parallel programs that contain thousands of instruction streams. Three types of dynamic displays were developed: one for the application program as the programmer envisions it, one for the network, and one which contains information for both the processor performance and the memory stalling phenomenon. The use of the tools, the performance of the proposed design with respect to other machines, and the trade-offs made during the algorithm and tool design phases are discussed. The generation and debugging of massively parallel programs can be straightforward if these types of tools are available.