Effects of Low-Quality Computation Time Estimates in Policed Schedulers

Abstract

Researchers conducting computer simulations can often provide estimates of computation time for a given type of simulation, which may be used by the compute cluster to aid in resource allocation and scheduling. However, the low quality of these estimates can cause deadline misses and unpredictable behavior. This problem is exacerbated on complex compute resources, clusters, and grids. Past-deadline jobs may be killed to provide resources for others, but the effect on the throughput of whole batches not fully understood. In this paper, we examine models for simple job schedulers and examine the quality of the deadline guarantee given. Simulation results based on actual runtimes are provided and discussed