Star-BRISE: Energy-efficient Benchmarking for Interacting Algorithms

GREENS © pixabay.com GREENS © pixabay.com

© pixabay.com


Dmytro Pukhkaiev, Sergii Shchaslyvyi, Roman Kosovnenko, Ievgeniia Svetsynska and Sebastian Götz


IEEE BlackSeaCom 2018 (web)


Energy-efficient computing is a well-studied and established field. Software energy-efficiency is one of the ways to decrease energy consumption of computing systems. However, contemporary studies on energy-efficient software investigate only individual algorithms, neglecting such an important area as workflow energy-efficiency. In this paper we try to decrease this gap by providing a study which investigates dependencies between software algorithms organized in a workflow. We empirically study the effect of dynamic voltage and frequency scaling and dynamic concurrency throttling on energy consumption of two case studies: workflows combined from (a) compression and encryption algorithms; and (b) matrix transposition and addition. Our findings show, that a luckily picked structure of a workflow, order of algorithms in particular, can reduce energy consumption of the overall system by 22%. Such savings clearly motivate conduction of similar empirical studies for different types of workflows. However, such studies are themselves very time- and energy-demanding. Therefore, we provide an approach called Star-BRISE that allows to reuse data obtained from benchmarking of individual algorithms to decrease the amount of measurements for resulting workflows. The presented approach can save up to 78% of time and energy effort on finding an optimal configuration for 2-algorithm workflows (and up to 95% of effort for a single workflow) and is even more efficient with scaling the number of algorithms in a workflow.

Zu dieser Seite

Sebastian Götz
Letzte Änderung: 09.04.2018