On Improving Scheduling Stability for Irregular Data Redistribution On Improving Scheduling Stability for Irregular Data Redistribution Based on Local Message Reduction

• Main Authors: Chao-Yang Lan, 藍朝陽
• Other Authors: Ching-Hsien Hsu
• Format: Others
• Language: zh-TW
• Published: 2006
• Online Access: http://ndltd.ncl.edu.tw/handle/72606200756890015934

碩士 === 中華大學 === 資訊工程學系碩士班 === 94 === Large-scale scientific problems have been solved on distributed memory multi-computers with powerful computing ability. As the parallel system techniques growing up, the regular data redistribution issues are extended to irregular issues, researches on homogeneous systems are extended to heterogeneous systems. High Performance Fortran Version 2 (HPF2) provides GEN_BLOCK (generalized block) format to map onto irregular mapping function. Recently, three irregular data redistribution algorithms on cluster were proposed. In this thesis, a technique is proposed to improve the irregular data redistribution algorithms. The redistribution techniques reduce the cost of data communication when executing programs. In the past, the theoretical cost of communication is proportion to data size. The factor of transmitting rates is ignored, thus the experiments may be different from theoretical results. In this thesis, the proposed technique considers the transmitting rate in a node is different from that between node and node. After considering the difference and proposing a new cost model, the proposed technique helps scheduling better results. The improvements on irregular data redistribution algorithms are shown in experiments. The experiments show the proposed technique improves all scheduling algorithms.