NGS: Collaborative Research: Performance-Driven Adaptive Software Design and Control

Purdue University, West Lafayette IN

Investigators

Abstract

EIA-0103688 Purdue University John R. Rice The goal of this project is to develop the tools needed for performance-directed integrated design and control of complex applications running on distributed computational systems. Its target computational systems are complex, incorporating the difficult heterogeneity, latency, and adaptive properties of computational grids. Its target applications are at the cutting edge of computational science: very large, complex applications with adaptive characteristics that do not allow their optimal system configurations or computational requirements to be estimated prior to run time. Each application will be viewed as a composition of components, with a formal, high fidelity model of performance to be designed for each component. The approach is to use model-based adaptive run-time control, based on these composed performance models, to control the execution of the application to meet specified performance goals. The control strategy will make real-time changes to parameters that modify the behavior of both application and computational platform.