Publication Details

Multi-GPU Island-Based Genetic Algorithm Solving the Knapsack Problem

JAROŠ Jiří. Multi-GPU Island-Based Genetic Algorithm Solving the Knapsack Problem. In: 2012 IEEE Congress on Evolutionary Computation, CEC 2012. CA: Institute of Electrical and Electronics Engineers, 2012, pp. 217-224. ISBN 978-1-4673-1508-1.
Czech title
Multi-GPU implementace ostrovního modelu genetického algoritmu pro problém knapsack
Type
conference paper
Language
english
Authors
Jaroš Jiří, doc. Ing., Ph.D. (DCSY FIT BUT)
Keywords

GPU; CUDA; MPI; GA; island model; knapsack

Abstract

This paper introduces a novel implementation of the genetic algorithm exploiting a multi-GPU cluster. The proposed implementation employs an island-based genetic algorithm where every GPU evolves a single island. The individuals are treated by CUDA warps enabling to solve large knapsack instances and eliminate undesirable thread divergence. The OpenMPI interface is used to exchange genetic material among different islands and to collect statistical data. The implementation characteristics are investigated on a cluster composed of 14 GPUs and 4 six-core Intel Xeon processors.

Published
2012
Pages
217-224
Proceedings
2012 IEEE Congress on Evolutionary Computation, CEC 2012
Conference
IEEE World Congress on Computational Intelligence, Brisbane, AU
ISBN
978-1-4673-1508-1
Publisher
Institute of Electrical and Electronics Engineers
Place
CA, US
DOI
10.1109/CEC.2012.6256131
UT WoS
000312859301057
EID Scopus
2-s2.0-84866858943
BibTeX 
@INPROCEEDINGS{FITPUB9860,
   author = "Ji\v{r}\'{i} Jaro\v{s}",
   title = "Multi-GPU Island-Based Genetic Algorithm Solving the Knapsack Problem",
   pages = "217--224",
   booktitle = "2012 IEEE Congress on Evolutionary Computation, CEC 2012",
   year = 2012,
   location = "CA, US",
   publisher = "Institute of Electrical and Electronics Engineers",
   ISBN = "978-1-4673-1508-1",
   doi = "10.1109/CEC.2012.6256131",
   language = "english",
   url = "https://www.fit.vut.cz/research/publication/9860"
}
Files
Back to top