Abstract: "System size dependency of the minimum load-dispatching rate for dynamic load balancing in parallel inference machines is described, where system size is the number of processor nodes connected through a network and the minimum load-dispatching rate for load balancing is the lowest lo...

Ausführliche Beschreibung

Bibliographische Detailangaben

Hauptverfasser:	Sugie, M. (VerfasserIn), Inoue, N. (VerfasserIn)
Format:	Buch
Sprache:	English
Veröffentlicht:	Tokyo, Japan 1988
Schriftenreihe:	Shin-Sedai-Konpyūta-Gijutsu-Kaihatsu-Kikō <Tōkyō>: ICOT technical report 431
Schlagworte:	Computer simulation Machine theory Parallel processing (Electronic computers)
Zusammenfassung:	Abstract: "System size dependency of the minimum load-dispatching rate for dynamic load balancing in parallel inference machines is described, where system size is the number of processor nodes connected through a network and the minimum load-dispatching rate for load balancing is the lowest load-dispatching rate in which a certain level of utilization is maintained. Relationship between the minimum load-dispatching rate and system size is measured in two load-dispatching strategies based on the sender-initiate concept by simulation on the loosely-coupled multi-processor model. The minimum load-dispatching rate has sub-linear (increase more slowly than linear) dependency on the system size in smart-random and max-min strategies In high average utilization cases, it is approximately proportional to n in the former strategy and (n - 1) / n in the latter strategy, where n is the system size. By decreasing system size, the minimum load-dispatching rate cannot be reduced so much in max-min strategy, but can be reduced in smart-random strategy. The load dispatch may also be assessed from the point of view of communication processing ability. In smart-random strategy, task division techniques can adjust the minimum load-dispatching rate for load balancing to communication processing capacity.
Beschreibung:	18 S. graph. Darst.