Dynamic Multi-Resource Fair Allocation with Elastic Demands

In this paper, we study dynamic multi-resource maximin share fair allocation based on the elastic demands of users in a cloud computing system. In this problem, users do not stay in the computing system all the time. Users are assigned resources only if they stay in the system. To further improve the utilization of resources, the model in this paper allows users to dynamically select the method of processing tasks based on the resources allocated to each time slot. For this problem, we propose a mechanism called maximin share fairness with elastic demands (MMS-ED) in a cloud computing system. We prove theoretically that the allocation returned by the mechanism is a Lorenz-dominating allocation, that the allocation satisfies the cumulative maximin share fairness, and that the mechanism is Pareto efficiency, proportionality, and strategy-proofness. Within a specific setting, MMS-ED performs better, and it also satisfies another desirable property weighted envy-freeness. In addition, we designed an algorithm to realize this mechanism, conducted simulation experiments with Alibaba cluster traces, and we analyzed the impact from three perspectives of elastic demand and cumulative fairness. The experimental results show that the MMS-ED mechanism performs better than do the other three similar mechanisms in terms of resource utilization and user utility; moreover, the introduction of elastic demand and cumulative fairness can effectively improve resource utilization.