Pareto-optimal cost optimization for large scale cloud systems using joint allocation of resources

Abstract

Optimal resource allocation in cloud systems is NP-hard due to the involvement of several conflicting objectives and unpredictable cloud traffic. To improve user satisfaction and resource utilization while minimizing end-user cost, the joint allocation of cloud resources is inevitable. In this work, we model end-user cost in cloud as the optimization objective using bandwidth and compute allocation as the decision variables. To solve the joint Virtual Machine Placement (VMP) problem we propose a single point, greedy, software-defined network (SDN)-based solution that minimizes end-user cost by making certain changes to the fat-tree datacenter architecture. Mathematically, we show that the overall objective function is convex hence solving it using a weighted-sum greedy approach will induce solutions that are Pareto-optimal. Experimental evaluations confirm up to 15% reduction in the response time and up to 14% increase in the efficiency of resources. To analyse the risks involved in deploying delay-sensitive applications over cloud and to show the effects of resource allocation approaches, we perform a risk analysis of delay-sensitivity in cloud using real-time CVD detection. The results confirm the reduced response time due to the proposed approach while maintaining the efficiency of detection.