Evaluating Market-based Batch Scheduling
Speaker:David Irwin

(11/12/2003)

Abstract


Computationally intensive workloads continue to dominate the use of large-scale clusters at universities and research labs where thousands of batch jobs are submitted everyday. Recent advances in grid technologies promise to push these batch workloads into shared infrastructures such as hosting centers. In this environment user value, which corresponds to currency, and scheduling to maximize user value becomes the primary goal. Recent work \cite{millennium} has shown that scheduling based on user-defined per job valuations improves aggregate system value when compared with common scheduling policies. Job valuations specify differing user values for a range of performance levels. In this work we examine scheduling based on per job user-defined valuations. We show that the valuations used in \cite{millennium} provide a simple, rich, and tractable formulation for value-based scheduling. This is significant because even simple formulations of value can prove intractable to schedule. We also show how value-based scheduling, when viewed as a common currency, can facilitate batch computation on shared infrastructures by using multiple competing batch markets. Return to the SPIDER schedule


Jaidev Patwardhan
Last modified: Tue Oct 28 14:56:39 EDT 2003