Petascale Execution Time Evaluation

While many research projects related to understanding the performance of High-End Computing (HEC) systems exist, they tend to besmall and narrowly focused. The inter-institutional research project proposed herein will integrate the efforts of three suchprojects to develop a better understanding of the performance of today's applications on contemporary HEC systems. Furthermore, itwill develop a novel methodology for projecting today's applications into future workloads and modeling their performance on thePetascale systems expected by the end of the decade. In particular, we will integrate and expand our existing performance tools toenhance understanding of the behavior of large-scale Scientific Discovery through Advanced Computing (SciDAC) and the DefenseAdvanced Research Projects Agency's (DARPA) High Productivity Computing System (HPCS) mission partner applications. Lookingforward, we will model the structure and requirements of future scientific and engineering applications by developing scaling modelsand projecting workload characteristics from traces of existing codes. We will model the performance of future Petaflops-classsystems and understand their performance properties when executing these projected 2010 workloads. Finally, we will develop an openframework and common modeling interface that will permit application profiles and system models to interoperate. This project willprovide the only source of independent guidance for the U.S. Department of Energy (DOE) to judge the execution time and resourcerequirements of SciDAC and other applications when scaled to 2010 targets. It will further provide understanding of the value andlimitations of architectural structures and mechanisms in support of these applications to guide evaluation of proposedarchitectures, long before they are realized.