作者: Qiang Li , S. Vlaovic
DOI: 10.1109/FTPDS.1994.494473
关键词: Computer network 、 Parallel computing 、 Unrolled linked list 、 Tree traversal 、 Scalable Coherent Interface 、 Cache coherence 、 Computer science 、 Self-organizing list 、 List update problem 、 Linked list 、 Free list
摘要: This article presents a distributed directory based cache coherence protocol that improves performance and facilitates error recovery in large scale multiprocessors. A number of protocols, such as the Scalable Coherent Interface (SCI, ANSI/IEEE Std 1596), use linked list structure to maintain coherence. While they work well for small medium size systems, traversal overhead becomes high when system grows into thousands processors range. Also, is vulnerable single node failure from involves all system. Single can happen relatively frequently applied SCI-based Local Area MultiProcessors (LAMP) where individual nodes are autonomous computers power up down individually. We propose an enhancement approach. redundant spanning constructed built, which achieves two goals: 1) time reduced O(N) O(/spl radic/N) 2) confined involved failed list, unless head lost.