Power-aware performance management of virtualized enterprise servers via robust adaptive control

Virtualization technology provides a promising approach for efficiently managing the power and performance of enterprise servers. Previous studies on using control theory in a virtualized environment have mostly emphasized deterministic control policies or relied on models that were trained offline for specific workloads. In this paper, we demonstrate that these solutions may suffer from deficiencies when workload variations cause uncertain alterations in the system models. We propose a robust control architecture based on a robust adaptive control theory that simultaneously guarantees power and meets performance specifications with flexible tradeoffs even in the face of highly dynamic, bursty workloads. In order to overcome the shortcomings of existing control approaches and to free systems from the ill effects of inaccurate system models, an adaptive Linear Quadratic Gaussian algorithm with stochastic method is adopted and integrated into our control design. Experiments on our testbed server with a variety of workload patterns demonstrate both that our control method outperforms existing control solutions under dynamical workloads in terms of control accuracy and power savings, and that it is robust against workloads that occur in short, intense bursts.