PECS : a pareto-efficient and envy-free cloud resource scheduler

In recent years, cloud computing platforms have attracted more and more attention. A key challenge is that existing schedulers provide limited support for heterogeneous types of resources. This makes it difficult to allocate resources in a timely manner adaptively and efficiently to perform different kinds of tasks. In this paper, we develop and evaluate a scheduling system with QoS guarantees when multiple tasks are involved, in which we show a possible design driven by the use of utilities for task-based resource allocation. We call the resulting method PECS, which meets the following unique properties compared to existing systems: Pareto efficiency, i.e. no other assignment can increase the sum of utility of all tasks without harming at least one specific task; freedom of envyness, i.e. no task will find the allocation of resource for another task better for its own utility; finally, improvements based on equal distribution are guaranteed, that is, the utility of all tasks is at least as high as the equal distribution of all resources. Our evaluation results compare PECS with a state-of-the-art resource allocation method called DRF, where our results demonstrate the performance benefits of PECS applied to a rich set of task assignment scenarios