Preemptive Scheduling for Distributed Systems

Preemptive scheduling is widespread in operating systems and in parallel processing on symmetric multiprocessors. However, in distributed systems it is practically unheard of. Scheduling in distributed systems is an important issue, and has performance impact on parallel processing, load balancing and metacomputing. Non-preemptive scheduling can perform well if the task lengths and processor speeds are known in advance and hence job placement is done intelligently Although obtaining optimal schedules is NP-complete, many good heuristics exist. In most practical cases, non-preemptive scheduling leads to poor performance due to excessive idle times or due to a long job getting assigned a slow machine. We show how to use preemptive scheduling in distributed systems. Surprisingly, the benefits outweigh the increased overhead. However the implementation of preemptive scheduling is complicated by the need for process migration. This paper presents preemptive scheduling algorithms, their implementation and performance measurement.