Abstract
Energy costs nowadays represent a significant share of the total costs of ownership of High Performance Computing (HPC) systems. In this paper we provide an overview on different aspects of energy efficiency measurement and optimization. This includes metrics that define energy efficiency and a description of common power and energy measurement tools. We discuss performance measurement and analysis suites that use these tools and provide users the possibility to analyze energy efficiency weaknesses in their code. We also demonstrate how the obtained power and performance data can be used to locate inefficient resource usage or to create a model to predict optimal operation points. We further present interfaces in these suites that allow an automated tuning for energy efficiency and how these interfaces are used. We finally discuss how a hard power limit will change our view on energy efficient HPC in the future.
Highlights
Introduction to energy efficiency considerations inHPCWithin the last decade, energy costs of High Performance Computing (HPC) systems have significantly increased and represent a significant share of the total cost of ownership (TCO) of such a system
P-States that refer to voltage and frequency changes can be observed via instrumentation on methods that issue a change [39], via hardware performance monitoring units (PMUs) or special model specific registers (MSRs) like APERF and MPERF on newer x86 processors
In this paper we have described some aspects of state-of-the-art energy efficiency tuning
Summary
Energy costs of High Performance Computing (HPC) systems have significantly increased and represent a significant share of the total cost of ownership (TCO) of such a system. HPC poses additional challenges regarding energy efficiency, as the impact on performance has to be very low in order to be acceptable. In HPC, the most common and known metric is FLOPS/W and depicts how efficient a system can execute the LINPACK benchmark, which is known to be very energy intensive This metric is associated with the Green500 list [14] that was announced to accompany the established TOP500 list [13]. One metric in this context is the energy-delay product (EDP) introduced in 1994 [21] It presents a compromise between low power and high performance and is well suited to guide research efforts.
Sign-in/Register to view highlights & summary 
Published Version (
Free)
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Schedule a call