NSF/CAREER: Robust Data-intensive Cluster Programming

University Of Wisconsin-Madison, Madison WI

Investigators

Abstract

The combination of device complexity and design diversity point towards systems that will be constructed from components that are both inherently and observably heterogeneous. What is needed is a new design discipline that accounts for inherently erratic component behavior, and facilitates the design and implementation of robust and predictable systems. The necessary infrastructure for robust application development is being developed within the context of a new system for cluster programming known as Ice. Research proceeds along four axes: the creation of fully robust data-transfer mechanisms, the exploration of advanced techniques for adaptation, new approaches to automation and ease-of-use, and development of the theoretical underpinnings of robust systems. The overall goal of this work is to develop a fundamental understanding of how to build scalable clustered systems that ``work well'', even when some of the underlying components do not. Such truly robust systems are predictable, reliable, and available, perform well in spite of fluctuations and failures of constituent components, and require no human intervention. One of the main advantages of such systems is the reduction of the burden of administration: new components are simply added and utilized to their full capacity.