Efficient Methods for kr → r and r → kr Array Redistribution1

  • Authors:

  • Ching-Hsien Hsu;Yeh-Ching Chung

  • Affiliations:

  • Department of Information Engineering, Feng Chia University, Taichung, Taiwan 407, ROC, chhsu@iecs.fcu.edu.tw;Department of Information Engineering, Feng Chia University, Taichung, Taiwan 407, ROC, ychung@iecs.fcu.edu.tw

  • Venue:
  • The Journal of Supercomputing
  • Year:
  • 1998

Quantified Score

Hi-index 0.00

Visualization

Abstract

Array redistribution is usually required to enhance algorithmperformance in many parallel programs on distributed memorymulticomputers. Since it is performed at run-time, there is aperformance tradeoff between the efficiency of new data decompositionfor a subsequent phase of an algorithm and the cost of redistributingdata among processors. In this paper, we present efficient algorithmsfor BLOCK-CYCLIC(kr) to BLOCK-CYCLIC(r) andBLOCK-CYCLIC(r) to BLOCK-CYCLIC(kr)redistribution. The most significant improvement of our methods is thata processor does not need to construct the send/receive data sets for aredistribution. Based on the packing/unpacking informationthat derived from the BLOCK-CYCLIC(kr) toBLOCK-CYCLIC(r) redistribution and vice versa, a processorcan pack/unpack array elements into (from) messages directly. Toevaluate the performance of our methods, we have implemented ourmethods along with the Thakur‘s methods and the PITFALLSmethod on an IBM SP2 parallel machine. The experimental results showthat our algorithms outperform the Thakur‘s methods and thePITFALLS method for all test samples. This result encouragesus to use the proposed algorithms for array redistribution.