Faucets: Migratable Objects for Grid Computing

 

Abstract

The Faucets project at Illinois aims at resource management within a cluster and across the grid. We have been developing a parallel computing methodology based on migratable objects, which has been embodied in parallel programming systems called Charm++ and Adaptive MPI. Faucets projects builds upon this methodology. The faucets system is capable of scheduling jobs on an individual cluster, such that each job can change the number of processors allocated to it at runtime. Further, it supports anonymous job submission so individuals can run their applications on any of the available computers that meet a quality of service requirement. The choice of a cluster is made via a bidding mechanism, which creates a market economy for compute power. A third and new aspect of the faucets project is the ability to run a single application across multiple clusters efficiently, when the clusters are geographically separated. The talk will illustrate these methodologies with application examples and performance data.

BACK