Live Webcast 15th Annual Charm++ Workshop

Computational power, at least at the high end, can be thought of as a utility, similar to electricity or water. To make this metaphor work requires a sophisticated ``power distribution'' infrastructure. The ``Grid'', popularized by the Globus project, is an example of such an infrastructure. To function efficiently, the producers of compute Power -- the parallel servers -- must be able to reorganize their jobs dynamically so as to respond to demands for computational power quickly, and maximize their utility. We are developing a framework, called faucets, that aims at facilitating this process. This paper focuses on a system at the heart of this framework: an adaptive manager for timeshared parallel machines that can shrink and expand its jobs to a variable number of processors dynamically. This manager has been implemented for workstation clusters. The paper describes the faucets framework, the design of the adaptive job manager, and preliminary performance data.