Title :
Algorithms for stable sorting to minimize communications in networks of workstations and their implementations in BSP
Author :
Cérin, Christophe ; Gaudiot, Jean-Luc
Author_Institution :
Univ. de Picardie, Amien, France
Abstract :
We introduce a novel approach to produce BSP (Bulk Synchronous Programming model) programs and we show their efficiency by implementing the stable sorting problem on clusters of PC. Experimental results on PCs based on Ethernet and Myrinet cards are compared with implementations on an SGI 2000. The algorithms presented in the paper are either developed under the theoretical framework of the Regular Sampling technique which guarantees good load balancing properties or are inspired by the technique in order to decrease the sequential work of each processor comparing to the Regular Sampling technique but impose no (theoretical) bound on load balancing. The main sequential block of code used in the algorithms for local sorting are derivatives of Shellsort (which is stable) and a new code based on Quicksort (which is not stable) plus a property on real numbers that is used for stable sorting under the framework of BSR (Broadcast with Selective Reduction)
Keywords :
multiprocessing systems; parallel algorithms; resource allocation; sorting; workstation clusters; BSP; BSR; Broadcast with Selective Reduction; Bulk Synchronous Programming model; Ethernet; Myrinet cards; Quicksort; Regular Sampling technique; SGI 2000; Shellsort; load balancing properties; local sorting; networks of workstations; parallel languages; parallel programming; real numbers; resource management; sequential block; stable sorting; stable sorting problem; Broadcasting; Clustering algorithms; Concrete; Intelligent networks; Load management; Partitioning algorithms; Resource management; Sampling methods; Sorting; Workstations;
Conference_Titel :
Cluster Computing, 1999. Proceedings. 1st IEEE Computer Society International Workshop on
Conference_Location :
Melbourne, Vic.
Print_ISBN :
0-7695-0343-8
DOI :
10.1109/IWCC.1999.810815