Adaptive robust optimization for coordinated capacity and load control in data centers

This paper addresses the problem of improving energy efficiency and quality-of-service (QoS) of data centers, by coordinating the 'feed-forward' capacity provisioning controller and the 'feed-back' load balancing controller. A data center is modeled as a collection of modular server blocks which cooperatively process multi-class, inter-dependent workload. We propose a coordinated two-stage control strategy of data centers based on the adaptive robust optimization framework. In stage 1, the optimal capacity of each server block is found based on predicted arrival rates of future workload, taking into account the potential QoS cost in stage 2; Then in stage 2, the load balancer distributes incoming workload to server blocks to achieve optimal QoS, after observing the actual workload. We show through simulations that the proposed approach achieves lower total costs as well as less QoS variations compared to a start-of-art baseline approach with reasonable level of conservativeness.