Performance, fairness and effectiveness in space slicing multi-cluster schedulers

Abstract

Parallel job schedulers have mostly been evaluated/compared using performance metrics. The deductions, however, can be misleading due to selective starvation. This calls for studies in scheduler fairness. Most studies have studied performance and fairness independently. We make a simultaneous study of performance and fairness for space slicing schedulers to deduce effectiveness. We show that measurements of fairness based on measures dispersion can contradict them selves for a similar set of schedulers. We also show that implied unfairness may not be a result of job starvation. Unfairness, derived from some of the current measures, is not always an implication of scheduler ineffectiveness. We use intuition to propose heuristics that determine scheduler effectiveness. We compare deductions from the combination of performance and fairness with those of effectiveness.