Performance Analysis of a Distributed System under Time-Varying Workload and Processor Failures

This paper studies the performance of a distributed system which is subject to hardware failures and subsequent repairs. A special type of scheduling called gang scheduling is considered, under which jobs consist of a number of interacting tasks which are scheduled to run simultaneously on distinct processors. The distribution for the number of parallel tasks per job varies with time. Two gang scheduling policies used to schedule parallel jobs are examined in two cases: In the blocking case, a job that is blocked due to processor failure keeps all of its assigned processors until the failed processor is repaired. In the non-blocking case, the remaining operable processors can serve other jobs. System performance is analysed for different workload parameters.