Abstract

Uintah is a component-based visual problem solving environment (PSE) designed to specifically address the unique problems inherent in running massively parallel scientific computations on terascale computing platforms. In particular, development of the Uintah system is part of the C-SAFE [2] effort to study the interactions between hydrocarbon fires, structures and high-energy materials (explosives and propellants). In this paper we describe methods for generating meaningful performance measurements for the Uintah PSE running on the SGI Origin 2000 multiprocessor architecture (these methods are applicable to many other applications.) These techniques include utilizing the non-intrusive performance counters built into the R10k and R12k processors, controlling process placement, controlling memory layout, and utilization of a task graph approach to specifying and solving the problem.