Parallel Algorithms for Generation of Long Sorted Runs on IBM

• Main Authors: Chen, Hung-Kuang, 陳宏光
• Other Authors: Lin, Yen-Chun
• Format: Others
• Language: zh-TW
• Published: 1995
• Online Access: http://ndltd.ncl.edu.tw/handle/95780901014825159435

碩士 === 國立臺灣科技大學 === 工程技術研究所 === 83 === Generation of long sorted runs is the first phase for sorting a large file. Fast generation of long sorted runs can help us significantly decrease the total execution time required for sorting a large file. In this thesis, we implement three par- allel algorithms, the bidirectional linear array algorithm, the unidirectional linear array algorithm, and the unidir- ectional linear array with broadcast algorithm, for runs gen- eration on IBM Scalable POWERparallel 2 (SP2) with PVMe and C. The bidirectional and the unidirectional algorithms were originally implemented on the transputer network using OCCAM [Lin 93a][Lin 93b].The unidirectional linear array with broad- cast algorithm is our new design for solving the late-feedback problem in the unidirectional array. In addition to the dis- cussion of our design and implementation details, we also do some theoretical and experimental comparisons on the three algorithms. From the comparisons, we show that our new algo- rithm is not only the fastest but also the best in generating long sorted runs.