Developing a Resource-Aware Adaptive Compilation System for High-Performance Distributed Computing

University Of California-Santa Barbara, Santa Barbara CA

Investigators

Abstract

Developing a Resource-Aware Adaptive Compilation System for High-Performance Distributed Computing The goal of our work is to combine successful Grid infrastructure with novel program compilation and optimization techniques to support portable and high-performance distributed computing. We plan to develop adaptive compilation and runtime techniques that are guided by the dynamic performance characteristics of underlying Grid resources. To investigate these techniques, we will use language environments that come from Internet computing (Java and .Net) as development vehicles. The framework we propose, a Resource-Aware Adaptive Compilation System, makes resource performance and availability predictions, e.g. for networks, CPUs, memory, disk, etc., available to a dynamic compilation system so that programs compiled "just-in-time" achieve minimum overall execution time adaptively. That is, we will adapt the level of program optimization (and subsequent execution time tradeoff) dynamically, based on the instantaneous performance profile of the Grid resources at hand. To address this problem of achieving portable high-performance using the Grid, we propose to investigate adaptive program compilation and on-the-fly optimization of high-performance distributed applications using Computational Grid resources. We will guide runtime and compiler optimization using predicted resource performance (generated by the Network Weather Service) and re-optimize code to adapt to changes.