Composite interconnection networks for parallel computers

Thesis/Dissertation · Wed Jan 01 04:00:00 EST 1992

OSTI ID:7112085

Jan, G E

Interconnection networks play a key role in the performance of parallel computers. Among many connection patterns, concentration and permutation patterns are the two most frequently encountered assignments. Using a new composite interconnection network model, this dissertation obtains a cost-efficient and fast networks to realize such assignments. It shows constructively that on this composite model, the concentration assignments for n processors can be realized in O(n) cost and O(lg n) time. It is also shown that a variation of the concentration assignments, namely, hyperconcentration assignments can be realized within the same cost and time complexities. In addition, the dissertation exploits the notion of radix permuting to obtain a permutation network with O(n lg n) cost, O(lg[sup 2]n) depth and O(lg[sup 2] n) routing time. More generally, a permutation network with O(kn[sup 1+1/k]) cost, O(k lg n) depth and O(k lg n) routing time is given. This improves upon the permutation network construction reported by Nassimi and Sahni that gives O(k lg[sup 3] n) routing time for the same cost expression. While permutation networks such as a the ADS network and Upfal's packet routing network have been reported in the literature with O(n lg n) cost and O(lg n) routing time, the constants hidden in these complexities are so large that our networks remain superior to these networks until n gets extremely large.

• Research Organization: Maryland Univ., College Park, MD (United States)
• OSTI ID: 7112085
• Country of Publication: United States
• Language: English

Related Subjects

99 GENERAL AND MISCELLANEOUS
990200* -- Mathematics & Computers
COMPOSITE MATERIALS
COMPUTER NETWORKS
CONDUCTOR DEVICES
CONNECTORS
DESIGN
ELECTRICAL EQUIPMENT
EQUIPMENT
MATERIALS
PARALLEL PROCESSING
PROGRAMMING