Abstract: "The Multi-PSI system developed in the Japanese FGCS project is a loosely coupled multiprocessor and is used as a testbed to experiment on the implementation of parallel execution of KL1 programs. The key point of the implementation is to minimize the inter-processor communication bec...

Ausführliche Beschreibung

Bibliographische Detailangaben

1. Verfasser:	Nakajima, Katsuto (VerfasserIn)
Format:	Buch
Sprache:	English
Veröffentlicht:	Tokyo, Japan 1989
Schriftenreihe:	Shin-Sedai-Konpyūta-Gijutsu-Kaihatsu-Kikō <Tōkyō>: ICOT technical report 486
Schlagworte:	Multiprocessors > Evaluation
Zusammenfassung:	Abstract: "The Multi-PSI system developed in the Japanese FGCS project is a loosely coupled multiprocessor and is used as a testbed to experiment on the implementation of parallel execution of KL1 programs. The key point of the implementation is to minimize the inter-processor communication because the cost for data access between processors is much higher than inside one processor. Address translation tables for inter-processor data references, called export and import tables, are introduced to realize local garbage collection. The weighted references counting (WEC) method is used for incremental inter-processor garbage collection To avoid redundant copies of KL1 data among processors, the two schemes, export/import hash tables and structure ID, are employed. The cost of handling frequent messages is 25 to 80 [micro]sec, and they are expected to be reduced to around 70 to 85 % of the current implementation by tuning up. Evaluation with two medium size benchmark programs shows the overhead for inter-processor communication is about 10 to 25 % in operating time even with the large cost of communication processing. The network traffic is very light and the topology is not limiting the system performance in the current system.
Beschreibung:	14 S.