Multi-Resource Fair Allocation for Cloud Federation

Abstract

Since the number of tasks submitted by user changes with time, the cloud data center cannot meet all resource requests in time when the load is high, and the idle resources are not effectively utilized at low load. Cloud federation can be established to share resources between different Cloud Service Providers (CSPs), enabling load balance between clouds. CSPs in the cloud federation are independent and their individual benefit need to be guaranteed when sharing resources with others. In this paper, we establish a two-level asynchronous scheduling model within and between the clouds, and introduce high priority queues in the cloud to prioritize the tasks sent by the federation scheduler. Due to the heterogeneity of resource requirements of user tasks, we design two multi-resource fair scheduling algorithms for cloud and federation. Within the cloud, we add the dynamically adjusted weighting factor to the Dominant Resource Fairness (DRF) to increase the priority of user whose task is outsourced. We define the dominant share and dominant contribution of the cloud. The priority of the cloud is the difference between the dominant share and the dominant contribution, which ensure the incentive and fairness of the cloud. Based on the open source container management system Kubernetes, we develop a cloud federation prototype system to implement our algorithm, and conduct performance evaluations through experiments. The results show that, compared with the existing fair allocation algorithm, the multi-resource fair allocation mechanism proposed in this paper can guarantee users' QoS of each CSP.