Decentralized remapping of data parallel computations with the generalized dimension exchange method

Author

Xu, Cheng-Zhong ; Lau, Francis C M

Author_Institution

Dept. of Math. & Comput. Sci., Paderborn Univ., Germany

fYear

1994

fDate

23-25 May 1994

Firstpage

414

Lastpage

421

Abstract

The Generalized Dimension Exchange (GDE) method is a fully distributed load balancing method that is most suitable for multicomputers with a direct communication network. It is extremely easy to implement and can yield optimal performance given a proper tuning. We propose a decentralized remapping method that uses the GDE algorithm periodically to balance (remap) the system´s load. We implemented this remapping method in two data parallel applications and found it to be effective in reducing the computation time. The gains in performance (5-15%) due to remapping are reasonably substantial given the fact that the two applications by their very nature do not necessarily favor remapping

Keywords

parallel algorithms; parallel architectures; parallel machines; resource allocation; GDE algorithm; data parallel applications; data parallel computations; decentralized remapping; decentralized remapping method; direct communication network; fully distributed load balancing method; generalized dimension exchange method; multicomputers; optimal performance; Communication networks; Computer science; Concurrent computing; Costs; Distributed computing; Load management; Mathematics; Performance gain; Runtime; System performance;

fLanguage

English

Publisher

ieee

Conference_Titel

Scalable High-Performance Computing Conference, 1994., Proceedings of the

Conference_Location

Knoxville, TN

Print_ISBN

0-8186-5680-8

Type

conf

DOI

10.1109/SHPCC.1994.296673

Filename

296673