Copyright by William Lloyd Bircher 2010 - The Laboratory for ...

the duration of power and performance phases. All power management schemes incur an
energy and performance penalty when the system transitions from one adaptation level to
another. To avoid costly transitions, adaption schemes must know how and when to
adapt. This new approach leads to the discovery of frequently repeating power and
performance patterns within workloads. Of these patterns, the most dominant and
predictable is the scheduling quanta of operating systems. Since active-idle and idle-
active transitions frequently occur on these boundaries, they serve as strong indicators of
phase changes.
3) Predictive Power Management
This dissertation presents the concept of core-level phase prediction and its application to
dynamic power management. By observing changes in performance demand and power
consumption at the core-level, it is possible to perceive predictable phase behavior.
Prediction of phases allows power management to avoid over or under provisioning
resources in response to workload changes. Using this concept the PPPP is developed. It
is a simple, table-based prediction scheme for directing DVFS selection. The predictor is
applied to the SYSmark2007 benchmark suite and achieves significant performance and
power improvements. Compared to the reactive DVFS algorithm used by Windows
Vista, performance is increased by 5.4% and while power consumption is reduced by
3.8%.
155