In a parallel computer system based on the hypercube topology, a fragmentation problem may occur in which processor resources, even if there are a sufficient number available, do not form a subcube large enough for an incoming task, seriously jeopardizing system utilization. Compaction is one effective solution to such a fragmentation problem, but the task migration overhead involved is quite large. To reduce such overhead, parallel migration is considered. This paper presents a new algorithm able to determine two link-disjoint paths between corresponding nodes for parallel transmission of data during migration. Under this algorithm, for every possible pair of source and destination subcubes, two link-disjoint paths each pair of corresponding nodes can be found for migrating data simultaneously. With each path delivering one half of the total migrated data, the migration overhead can be thus cut down by half. The proposed algorithm is applicable to any systems with any subcube allocation strategies, and even to systems under the circuit-switched mode of operation.
關聯:
International journal of distributed systems and networks 1(1), pp.34-42