Program Partitioning with Bundling Approach in a Multiprocessor System

• Main Authors: Jing-Long Benjamin Lu, 呂景隆
• Other Authors: Liang-Teh Lee
• Format: Others
• Language: en_US
• Published: 1999
• Online Access: http://ndltd.ncl.edu.tw/handle/27834049481318506255

碩士 === 大同工學院 === 資訊工程研究所 === 87 === It is well understood that the extent to which parallel computing is successful depends largely on the performance of the underlying parallel programming primitives. There are many partitioning algorithms have been discussed, and majorities of them are focusing on the partitioning algorithm of specific area. We are here to present a systematic process of program partitioning and the process is done automatically after the program is coded. It includes an efficient method to identify the dependency relation between operations without the conventional complication of graph partition. The idea of entire approach is simple and straightforward. The result of our partition is prepared as link list and can be used by other architectures with only a few extra transformation interfaces. Our approach involves using Dataflow Graph (DFG) and proper tools to avoid the complication of graph partition. Relative weights are applied to the corresponding nodes in the DFG to aid the partitioning process. The partitioned groups of nodes are separated into different “Bundles” and can be used for easy reference. According to the processor number, the bundles can either be merged or partitioned again to achieve better processor utilization. By carefully managing these bundles, a feasible speedup are obtained through our simulated model.