An LLVM-based Approach to Generate Energy Aware Code by means of MOEAs

Varrette, Sébastien; Dorronsorro, Bernabe; Bouvry, Pascal

Reference : An LLVM-based Approach to Generate Energy Aware Code by means of MOEAs


	Document type :	Scientific congresses, symposiums and conference proceedings : Paper published in a book
	Discipline(s) :	Engineering, computing & technology : Computer science
	To cite this reference:	http://hdl.handle.net/10993/21597

Title :	An LLVM-based Approach to Generate Energy Aware Code by means of MOEAs
Language :	English
Author, co-author :	Varrette, Sébastien [University of Luxembourg > Faculty of Science, Technology and Communication (FSTC) > Computer Science and Communications Research Unit (CSC)]
	Dorronsorro, Bernabe [> >]
	Bouvry, Pascal [University of Luxembourg > Faculty of Science, Technology and Communication (FSTC) > Computer Science and Communications Research Unit (CSC)]
Publication date :	Oct-2015
Main document title :	Proc. of the 7th European Symposium on Computational Intelligence and Mathematics.(ESCIM 2015)
Peer reviewed :	Yes
On invitation :	No
Audience :	International
City :	Cádiz
Country :	Spain
Event name :	Proc. of the 7th European Symposium on Computational Intelligence and Mathematics.(ESCIM 2015)
Event date :	Oct 2015
Abstract :	[en] Moderating the energy consumption and building eco-friendly computing infrastructure is of major concerns in the implementation of High Performance Computing (HPC) system, especially when a world- wide effort target the production of an Exaflop machine by 2020 within a power envelop of 20 MW. Tracking energy savings can be done at var- ious levels and in this paper, we investigate the automatic generation of energy aware software with the ambition to keep the same level of efficiency, testability, scalability and security. To this end, the Evo-LLVM framework is proposed. Based on the mod- ular LLVM Compiler Infrastructure and exploiting various evolutionary heuristics, our scheme is designed to optimize for a given input source code (written in C) the sequence of LLVM transformations that should be applied to the source code to improve its energy efficiency without degrading its other performance attributes (execution time, parallel or distributed scalability). Measuring this capacity is based on the combi- nation of several metrics optimized simultaneously with Multi-Objective Evolutionary Algorithms (MOEAs). In this position paper, the NSGA- II algorithm is implemented within the Evo-LLVM yet the analysis of more advanced heuristics is in progress. In all cases, the experimental validation of the framework over a pedagogical code sample reveal a drastic improvement of the energy consumed during the execution while maintaining (or even improving) the average execution time.
Research centres :	University of Luxembourg: High Performance Computing - ULHPC ; l
Permalink :	http://hdl.handle.net/10993/21597
Other URL :	http://escim2015.uca.es/

File(s) associated to this reference

Fulltext file(s):

	File	Commentary	Version	Size	Access
Limited access	escim2015.pdf		Author preprint	522.44 kB	Request a copy

All documents in ORBi^lu are protected by a user license.

University of Luxembourg Library | Feedback | Legal notices