Parallel Submesh Compaction in Meshes

• Main Authors: Lih Chiou Huang, 黃麗萩
• Other Authors: Mr. Chen
• Format: Others
• Language: zh-TW
• Published: 1995
• Online Access: http://ndltd.ncl.edu.tw/handle/19544407976755132175

碩士 === 國立臺灣科技大學 === 工程技術研究所 === 83 === A mesh-connected computer system is one of the main parallel architectures. Due to its simple and regular architectures, the mesh has become one of the most suitable architectures for small and medium computer systems. It supports the ability for multiprogramming,and can be partitioned into submeshes,possibly of different sizes, to execute independent jobs, with each job running on a dedicated submesh. Processors allocation is a critical problem,some allocation strategies have been proposed, such as Buddy, First Fit Best Fit, etc. Chen and Shin have pointed out that as jobs arrive grabbing subcubes and leave releasing them to the system, the hypercube tends to become fragmented,and so does the mesh.One of the ways to solve the fragmentation problem is to migrate some running jobs currnetly in the system so as to free up a large free subcube. In hypercube computer systems, Kim et al. migrate an approtiate set of source-target pairs simultaneously by eastab- lishing totally disjoint paths. In each migration step, they migrate b-subcubes to free up b+1-subcubes. Hence,it only takes at most d migration steps to free up a d-subcube. We apply this concept in mesh-connected computer systems to solve the fragmentation produced by Buddy allocation strategy. In each migration step, we compact 2**b x 2**b submeshes into 2**b+1x2** b+1 submeshes by finding an appropiate set of source -target pairs and eastablishing totally disjoint paths between them such that the selected jobs can always be migrated in parallel.