PARASOL: An Adaptive Framework for Parallel Processing

点击翻译按钮获取中文摘要

Parallel computers, particularly shared-memory multiprocessors, are nowwidespread. However, programmers still have great difficulty exploitingtheir potential to obtain good performance on real applications. Parallelcomputing will only be truly successful when this difficulty is removed.The investigators attack this problem by statically analyzing programs todetermine their structure, embedding run-time measurements to determinetheir dynamic needs, and triggering adaptive optimization at run-time thatis tuned to those needs.Technically, the investigators and their students plan to explore severalinnovative techniques and/or combinations of known evaluation andoptimization techniques: * Their compiler will insert adaptive capabilities directly into applications, rather than going through a third party such as the operating system. The application will then monitor its system resource usage and control work based on its predicted demands and the optimal workload of the system. * They will use a parallel performance model for adaptive optimization at run-time. The model will be used to both predict execution times for potential optimizations and to diagnose computational inefficiencies. * They will implement run-time parallelization and optimization techniques and integrate them in the POLARIS compiler and the PARASOL system. * They will use run-time parallelization to reduce the impact of latency in the applications.