The InterMezzo File System

Is it possible to implement a distributed file system, retaining some of the advanced protocol features of systems like Coda, while getting an implementation that is an order of magnitude simpler? We claim the answer is “yes”, and InterMezzo is well on its way towards being a prototype of such a system. The key design decisions were to exploit local file systems as server storage and as a client cache and make the kernel file system driver a wrapper around a local file system. We use a rich, high level language and work with asynchronous completion instead of threads. Finally, we rely on existing infrastructure, such as TCP. This is in contrast with systems like Coda and AFS, which implement their entire infrastructure from the ground up, in C. This paper will describe the design and implementation decisions used in our prototype.