TR-96-23: Performance Impact of I/O on Sender-Initiated and Receiver-Initiated Load Sharing Policies in Distributed Systems

Performance of distributed systems can be improved by distributing system workload from heavily loaded nodes to lightly loaded nodes (i.e., by load sharing). Previous studies have considered two dynamic load sharing policies: sender-initiated and receiver-initiated. In the sender-initiated policy, a heavily loaded node attempts to transfer work to a lightly loaded node and in the receiver-initiated policy a lightly loaded node attempts to get work from a heavily loaded node. Almost all the previous studies have neglected the I/O requirements of the jobs in evaluating the performance of various load sharing policies. In addition, sensitivity to variance in job inter-arrival and service times has not been studied. The goal of this paper is to fill the void in the existing literature. This paper models job’s I/O requirements and studies the impact of job I/O service demand on the performance of sender-initiated and receiver-initiated policies. Furthermore, we also look at the impact of variance in inter-arrival times and in job service times (both processor and disk service time variances are considered).

TR-96-23.pdf