Optimal control for energy-aware server farms

In many cases, services hosted in server farms are designed to be highly available and fault tolerant in the presence of randomly varying traffic, which often translates into over-provisioning of the server farms targeting peak demand periods. Consequently, the servers spend a substantial amount of time in low utilization range, which also happens to be the range in which servers are far less energy efficient. Moreover, even when completely idle, servers still consume a large portion of their peak power. However, servers cannot be simply switched off to save energy for two main reasons. First, any energy saving obtained by switching servers off comes at the expense of reduced performance due to the setup delay required to switch servers back to an operational state. Furthermore, servers are rarely completely idle since dispatching policies are often designed in such a way that the workload is evenly distributed across the server farm, resulting in low but non-zero utilization during off-peak demand periods. Thus, a coordinated control approach needs to be devised to achieve energy savings by consolidating workload and placing unused servers in low-power states while still providing good performance. This thesis studies the energy-performance trade-off by applying queueing theoretic methods and by formulating the trade-off as a multi-objective optimization problem. Single-server models are first analyzed and the mean response time and mean power consumption metrics are derived. Compound cost functions are defined from these metrics and the control variables that minimize these cost functions are optimized. For such cost functions, it is shown under very general assumptions that in a single-server queue there is no gain from delaying the decision to switch off the server upon becoming idle. Instead the optimal decision is either to switch off immediately or never switch off. Server farms are modeled as parallel queueing systems with each server belonging to either a baseline or reserve group of servers. Energy-aware dispatching and power-control policies are developed so that the reserve servers are placed in a low-power state whenever possible. To this end, the dispatching decisions are studied by formulating the problem as a Markov Decision Process, and the resulting system is solved using the Policy Iteration method to construct a near-optimal dispatching policy. More simple, heuristic power-control and dispatching policies are also proposed to reduce the energy consumption of a server farm without compromising the performance