File system caching in large point-to-point networks

The design of a file system for a general-purpose scalable parallel computing engine, based on an extensible mesh of processors is examined. Although many distributed file systems have already been developed, none have properly addressed the characteristics that are special to these computing engines, namely dynamic load-balancing and point-to-point networks. Client caching is identified as a key area where current systems fall short. The authors propose two new caching schemes, that employ an extra level of caching in addition to that used by current client caching schemes. They present some results from a simulation study of the two new caching schemes, which demonstrate that the schemes outperform traditional client caching schemes when used in a general-purpose processor mesh-based, parallel computing engine. Since one of these caching schemes is quite complex, a formal specification in the Z language has been developed