|
09.
12.
2011
Energy-Efficient Parallel Processing: Trends and Objectives
Daniel Hackenberg
TU Dresden
The continuously growing power consumption and associated
operational costs of compute systems have shaped both
microprocessors and system design for several years now.
Similar to mobile processors, CPUs for server class and high
performance computing systems are no longer designed with a
narrow focus on raw compute performance. This development has
driven a strong trend towards more energy-proportional
computing, with the power consumption of state-of-the-art
platforms being far from static but instead strongly
correlated with the current workload. Consequently,
significant efforts are required to ensure that we tap the
emerging potential for power-aware software development. This
presentation highlights recent hardware and software trends in
energy-efficient computing. We illustrate typical power
profiles of parallel applications and explain why the
traditional 'race to idle' does not necessarily lead to
energy-optimal results. We challenge the trend towards
hardware-centric decision models for dynamic power allocation
and make the case for interfaces that allow software
applications to better monitor and control the underlying
hardware.
|