Metrics for Energy-Aware Software Optimisation

Abstract

A note on versions:The version presented here may differ from the published version or, version of record, if you wish to cite this item you are advised to consult the publisher's version. Please see the 'permanent WRAP url' above for details on accessing the published version and note that access may require a subscription.For more information, please contact the WRAP Team at: wrap@warwick.ac.uk Metrics for Energy-Aware Software OptimisationStephen I. Roberts, Steven A. Wright, Suhaib A. Fahmy, and Stephen A. Jarvis University of Warwick, U.K. S.I.Roberts@warwick.ac.uk Abstract. Energy consumption is rapidly becoming a limiting factor in scientific computing. As a result, hardware manufacturers increasingly prioritise energy efficiency in their processor designs. Performance engineers are also beginning to explore software optimisation and hardware/software co-design as a means to reduce energy consumption. Energy efficiency metrics developed by the hardware community are often re-purposed to guide these software optimisation efforts. In this paper we argue that established metrics, and in particular those in the Energy Delay Product (Et n ) family, are unsuitable for energy-aware software optimisation. A good metric should provide meaningful values for a single experiment, allow fair comparison between experiments, and drive optimisation in a sensible direction. We show that Et n metrics are unable to fulfil these basic requirements and present suitable alternatives for guiding energy-aware software optimisation. We finish with a practical demonstration of the utility of our proposed metrics.

The Power-optimised Software Envelope

“…In our previous work, we have shown that metrics originating from the hardware community are not suitable for measuring software performance. We subsequently proposed two new dimensionless metrics that are designed to support energy-aware performance optimisation [34]. In this article, we provide derivations for POSE using both of these metrics, namely Energy-Delay Sum (EDS, Equation (1)) and Energy-Delay Distance (EDD, Equation 2):…”

“…POSE is metric agnostic and is compatible with all members of the Et n family, the EDS and EDD metrics [34], and indeed any metric that is an element-wise monotonic function of runtime and energy consumption. The only prerequisites are that runtime and energy consumption can be accurately measured or calculated for the target platform.…”

“…This divergence happens because Et n metrics produce perverse optimisation incentives: Et n places more emphasis on energy optimisations for efficient codes and on runtime optimisations for fast codes [34]. Any small optimisation that improves energy efficiency will increase the apparent benefits of further energy optimisations, leading to the concave curvature of the optimisation bounds for Et n metrics.…”

“…The divergence in metric value highlights an important issue with Et n metrics. In this article, the EDS and EDD metrics have been parameterised to allow a fair comparison to Et 3 ; however, one particular feature of both the EDS and EDD metrics is that they can be parameterised to represent the monetary cost [34]; with carefully chosen parameters for EDS and EDD, POSE models can represent potential savings in dollar cost. The divergence between Et 3 and EDS/EDD may lead an application developer to incorrectly focus on optimising for power when runtime optimisations would deliver a better EDS or EDD value and therefore a lower monetary cost.…”

“…• We introduce the Power-Optimised Software Envelope (POSE), a model that helps performance engineers to determine whether power or runtime optimisation will provide the greatest benefits for their code; • We provide derivations for POSE using the established Energy-Delay Product family of metrics as well as the Energy-Delay Sum and Energy-Delay Distance [34] metrics; • We demonstrate POSE on a number of applications from the Mantevo benchmark suite, showing that PathFinder is the application least amenable to power optimisation and TeaLeaf is the most amenable; • Finally, we extend POSE to provide a model for system-wide power optimisation characteristics. System Summary POSE is able to derive upper limits for the benefit of energy-aware software optimisation on a given system independent of any specific software application.…”