Beyond Beyond Dominant Resource Fairness : Indivisible Resource Allocation In Clusters

Resource Allocation is a necessary component of any shared computer system. Dominant Resource Fairness, and other recently proposed mechanisms, handle the problem of fair resource allocation in a datacenter containing different resource types. To date most solutions to this problem haven’t considered indivisible demands and none have considered clusters of machines. We analyze various resource allocation algorithms over datacenter clusters. We split the algorithm into two parts, first a mechanism that ensures max-min fairness, and secondly a heuristic for a complex multidimensional bin packing problem that has not previously been well explored. Our proposed MergeDRF algorithm increases utilization without much loss to fairness compared to adaptations of algorithms from the resource allocation literature.