Energy-efficient Management of Reconfigurable Computers

204 Power and energy consumption are first-class constraints today for computer system design and operations. One of the reasons for power (and energy) waste in computers is that they continue to consume power (and energy) even when they are idle. Barroso and Hölzle [22] proposed Energy Proportionality (EP) as a model for ideal system behavior. In this model, the ideal system should use power in proportion to utilization (performance or load served). The EP model has been very influential in encouraging system designers to improve energy efficiency of their systems. The EP model for ideal behavior holds true for systems with fixed resources. However, modern computers have reconfigurable resources. Such reconfigurable systems may also exhibit super-proportional behavior. The advent of such super-proportional systems is a game changer. In these systems, one can get more performance for the power consumed (or equivalently, more work done for the energy consumed) at intermediate loads than what one could get in an " ideal " EP system. EP no longer characterizes the maximum energy efficiency that super-proportional systems can attain. In fact, aiming for EP behavior can be significantly suboptimal except at very low loads. In Chapter 2 we proposed a new ideal model, EOP (Energy Optimal Proportional), that characterizes maximum energy efficiency for all systems, both super-proportional and otherwise. EOP is an ideal model for system designers. It describes a lower bound on the energy consumption (upper bound on the energy efficiency) that the given system can achieve. Any shortfall from this maximum efficiency at any operating load indicates potential for further improvement. System designers can use this characterization to both improve the upper bound as well as reduce shortfall, that is, make the Pareto frontier (Dynamic EO) closer to EOP, throughout the operating range.