Building Caches using Multi-Threaded State Machines

Designing a client-side cache for a distributed file system is complicated by concurrent accesses by applications, network communication with the server, and complex relationships between the data in the cache. Despite these difficulties, caches are usually built using threads or finite state machines as the programming model, even though both are inadequate for the task. We have created an infrastructure for cache development based on multi-threaded state machines, which attempts to capitalize on the benefits of both programming models. The state machine allows the global state of the cache to be carefully controlled, allowing interactions between concurrent cache operations to be reasoned about and verified. Threads allow individual operations to maintain their own local state, and propagate that state across transitions of the state machine. We created a prototype of this infrastructure and used it to implement a write-through file block cache that provides multiple-reader/single-writer access to its blocks; although this is a relatively simple cache protocol, it is much easier to implement using multi-threaded state machines than using either threads or finite state machines alone.