CapNet: Using a gigabit network as a high speed backplane
With their enormous amount of bandwidth, Gigabit networks have effectively shrunk communication delays to those of light wave propagation, thereby bringing the long awaited opportunity to distribute applications across a wide area network. By connecting a number of high performance computers together with a gigabit network, and providing hardware and software support for a virtual memory shared among these computers, we have effectively created a geographically dispersed multiprocessor with a high speed but also high latency backplane. The key to realizing this system is to reduce the effect of latency as much as possible. While caching and replication are well-known techniques to reduce memory access latency, they require mobile data and their copies to be located to maintain data consistency. CapNet addresses this problem in the context of a wide area network by embedding a set of intelligent mechanisms into the network. Our approach is to distribute the location information of memory objects and their copies into the network switches so that the network can actively support cache coherence protocols. In CapNet, a write miss is handled by directly multicasting invalidations to the memory object copies. A read miss is handled by replicating a copy from the copy holder that is geographically closest to the requesting host, thereby reducing the traveling time of the request and its reply to a minimum. Our approach reduces write fault delays by as much as 30% and read fault delays by 50% compared to the more common directory based approach.
Tam, Ivan Ming-Chit, "CapNet: Using a gigabit network as a high speed backplane" (1994). Dissertations available from ProQuest. AAI9521130.