 |
Center for Nonlinear Studies |
We consider a scenario where multiple classes of requests arrive at a dispatcher at time-varying rates for processing by various resources. This scenario is typical in data centers where each server hosts multiple classes of applications. It is well documented that data centers consume a phenomenal amount of energy while their servers are under-utilized. The main challenge is in managing time-variability and uncertainty. We have formulated approaches for smoothing so that requests receive time-stable performance, and control algorithms can be developed for energy efficiency. Our objective is to develop strategies to: (i) assign classes to servers, (ii) determine the number of servers to be powered on, (iii) route requests from the dispatcher to appropriate servers, and (iv) create a procedure for speed scaling. The goal is to develop the aforementioned strategies under: (a) a distributed setting where real-time information is not exchanged between the sub-systems, i.e. servers and the dispatcher; (b) a requirement for time-stable performance; (c) a preference for simplified operations while maintaining cost-effectiveness and high performance. We show that in the asymptotic regime where we scale the arrival rates, number of classes and the number of servers, the aforementioned objectives and goals can be met.
Host: Scott Backhaus