Abstract: "The most difficult problem in realizing a large-scale shared-memory multiprocessor is the design of a cost-effective memory system that has high bandwidth and low latency. The performance of a memory system design based on multilevel shared caches is investigated in this dissertation...

Ausführliche Beschreibung

Bibliographische Detailangaben

1. Verfasser:	Goosen, Hendrik A. (VerfasserIn)
Format:	Buch
Sprache:	English
Veröffentlicht:	Stanford, Calif. 1991
Schriftenreihe:	Stanford University / Computer Science Department: Report STAN CS 1393
Schlagworte:	Cache memory Multiprocessors Mehrprozessorsystem Pufferspeicher Skalierung Hochschulschrift
Zusammenfassung:	Abstract: "The most difficult problem in realizing a large-scale shared-memory multiprocessor is the design of a cost-effective memory system that has high bandwidth and low latency. The performance of a memory system design based on multilevel shared caches is investigated in this dissertation. By localizing a large fraction of the traffic required to maintain coherence between per-processor caches, a shared cache hierarchy reduces the bandwidth requirements of the global interconnect, and reduces the latency for access to shared data. Experimental measurements from simulations of three large applications confirm that shared caches offer better scalability and improved performance Specifically, a system using shared caches yields up to a 65% reduction in global traffic compared to a system using per-processor caches. The sensitivity of the performance to various cache design parameters is explored. Programs can also be restructured to exploit the advantages offered by shared cache hierarchies. A program restructuring technique based on the locality in three-dimensional space of a physical system simulation, a direct particle simulation, is described. Simulation and execution time measurements are presented to show the performance impact of the restructuring of the particle simulator The restructuring of the program reduced the global traffic by an order of magnitude, and the shared cache scheme further reduced the global traffic by a factor of five.
Beschreibung:	Zugl.: Stanford, Calif., Univ., Diss.
Beschreibung:	76 S.