On-line scheduling of parallel jobs

Abstract

We survey theoretical results on scheduling of parallel jobs when only partial information about them is available. We study how the length of the schedule changes under the influence of various factors, such as the use of randomization, the presence of dependencies between jobs, the choice of a network topology, the use of virtualization, and the maximal size of jobs. We give tight or very close upper and lower bounds on the best possible performance of on-fine scheduling algorithms for many combinations of these factors, thus giving an essentially complete picture of how they interact and influence the performance and methods for scheduling.