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

eactive scheme used in Windows Vista operating system. Applying the predictor to the
prediction of processor power, accuracy is improved by 4.8% compared to a reactive
scheme.
Dynamic power management provides a significant opportunity for increasing energy
efficiency and performance of computing systems. Energy efficiency is increased by
reducing performance capacity (frequency, parallelism, speculation, etc.) when the
demand for performance is low. Efficiency or a lack thereof may impact the performance
of a system. High-density server and multi-core processor performance is limited by
power delivery and dissipation. Reducing the power consumption in one component may
allow other components to consume more power (higher voltage and frequency) and
therefore achieve higher performance. These systems would otherwise have higher
performance if they were not sharing power resources [Bl07] [McPo06] [FaWe07]
[WaCh08] [WuJu05]. Power management allows the performance of cores/systems
sharing power to increase by “borrowing” power from idle or performance-insensitive
cores/systems and reallocating it to heavily-utilized cores/systems.
The challenge in applying power management to increase efficiency and performance is
in identifying when to adapt performance capacity. The ubiquitous, architected solution
implemented in operating systems such as Windows/Linux is to react to changes in
performance demand. Though this approach is simple, it performs sub-optimally
[BiJo08] [DiSo08] for workloads with many distinct and/or short phases. Each time a
workload transitions from a phase of low performance demand to a phase of high
121