Processor failures in a shared-memory multiprocessor system with resequencing

This paper presents a simulation model to address performance issues associated with a shared-memory multiprocessor system that is subject to hardware failures and subsequent repairs. Two policies used to schedule parallel jobs are compared using a variety of workloads, where multiple jobs can be executed simultaneously on disjoint subsets of processors called partitions. The impact of resequencing delay on performance is also explored. Simulated results indicate that both scheduling methods have merit and that their performance is seriously degraded by the resequencing delay.