The goal of this project is to advance the development and deployment of cyberinfrastructure by creating and validating performance models of strategic NSF computational science applications and by assessing the interdependence of alternative hardware, middleware, and software implementations on application performance.
Performance models are parameterized functions that predict performance based on system and application attributes. Given current and projected architectural complexity, application performance increasingly depends on subtle, complex machine/code interactions. Understanding and modeling this complexity is a prerequisite to designing applications that achieve substantial fractions of peak hardware performance, configuring systems to match application needs, and designing more effective architectures.
This effort will characterize the existing NSF supercomputing application workload; then to use the models developed via this characterization to evaluate alternative configurations of new systems, starting with SDSC's BG/L and PSC's Red Storm. The modeling methodology will then be extended to encompass additional systems and to assess the performance of Grid-enabled applications.
