XPS: FULL: CCA: An Initial Exploration of Architecture and System Support for Sparsity in Science and Engineering Applications

University Of Wisconsin-Madison, Madison WI

Investigators

Abstract

Scientific computing remains a key driving force in the evolution of high- performance computer systems, because of its inherent value and its ability to drive computer-systems advances that benefit many demanding applications. This project is an initial exploratory investigation into optimizations that cut across algorithms, systems, and computer architecture to improve application performance and reduce energy through heterogeneous systems that combine latency-oriented cores and throughput-oriented cores. This activity will identify optimizations with the potential to significantly advance important emerging applications that are on the verge of attaining practical utility. Using a novel adaptive computational fluid dynamics code as a driver, this research seeks to identify the benefits and bottlenecks of both current and emerging heterogeneous systems. Using both measurements on existing hardware and detailed simulation of emerging and proposed systems, this activity will identify architectural mechanisms for coordinating communication and computation with the potential to significantly increase performance and energy efficiency. This activity will also develop an initial design of a heterogeneous virtual machine model that virtualizes proposed hardware mechanisms, yet provides applications with enough knowledge of the physical hardware to optimize execution without losing the abstraction benefits of virtualization.