Energy-aware control of server farms

Abstract

We study the energy-performance trade-off in server farms using multi-server models with parallel queues. Idle servers may be turned off to save energy. However, turning the server back on incurs a setup cost in the form of additional delays and energy consumption. In our model, a dispatcher assigns incoming jobs to the servers depending on the applied task (job) assignment policy. Once routed to a specific server, a job can either be queued or served immediately depending on the service discipline and state of the server. We consider a family of idle time control policies that keep track of the time spent in an idle state using a timer. Upon expiration of this timer the server is switched to a sleep state, and waits until a new job is assigned to it by the dispatcher. Earlier studies for single server systems, with Poisson arrivals, showed that the optimal control policy either leaves the server idling or switches it to sleep immediately when it becomes idle. However, in a multi-server setting, our study indicates that 10–30 % saving on the system cost can be achieved by using the right combination of idle time control and task assignment policy.