Energy-Efficient Dynamic Capacity Provisioning in Server Farms (CMU-CS-10-108)

A central question in designing server farms today is how to efficiently provision the number of serversto handle unpredictable demand patterns, so as to extract the best performance while not wastingenergy. While one would like to turn servers off when they become idle to save energy, the largesetup cost (both, in terms of setup time and energy penalty) needed to switch the server back on canadversely affect performance. The problem is made more complex by the fact that today’s serversprovide multiple sleep or standby states which trade off the setup cost with the power consumedwhile the server is ‘sleeping’. With so many controls, finding the optimal server pool managementpolicy is an almost intractable problem – How many servers should be on at any given time, howmany should be off, and how many should be in some sleep state?In this paper, we analyze server farm management policies, and prove that it suffices to consider asmall, natural class of policies to guarantee near-optimal behavior. We employ the popular metricof Energy-Response time Product (ERP) to capture the energy-performance tradeoff, and presentthe first theoretical results on the optimality of server farm policies under both stationary and time-varying demand patterns. We find that there exists a very small set of natural policies that alwayscontains the optimal policy for a single server, and a near-optimal policy for multi-server systems,for a stationary demand pattern. We also consider server farm management policies for time-varyingdemand patterns and propose a robust, traffic-oblivious policy which is asymptotically optimal asthe load becomes large.