Fast compaction in hypercubes
Nian-Feng Tzeng, Hsing-Lung Chen · IEEE Transactions on Parallel and Distributed Systems · 1998
Compaction relocates active subcubes in a fragmented hypercube so as to produce a contiguous free region and eliminate the adverse impact of fragmentation on performance. The overhead of compaction is often contributed primarily by task migration, which makes use of disjoint paths for transmitting migrated data. Since task migration usually involves transmitting a large amount of data, the time required for migration with single paths is long, making compaction an undesirably lengthy process. This paper considers fast compaction through the use of all disjoint paths in existence for migration simultaneously from a source subcube to its target subcube, effectively reducing the size of data transmitted over a path and shortening the migration time. This approach leads to considerable savings in the compaction time for hypercubes which support circuit switching or wormhole routing, when compared with that using single migration paths.