Performance analysis of a distributed file system

An important design goal of a distributed file system, a component of many distributed systems, is to provide UNIX file access semantics, e.g., the result of any write system call is visible by all processes as soon as the call completes. In a distributed environment, these semantics are difficult to implement because processes on different machines do not share kernel cache and data structures. Strong data consistency guarantees may be provided only at the expense of performance. This work investigates the time costs paid by AFS 3.0, which uses a callback mechanism to provide consistency guarantees, and those paid by AFS 4.0 which uses typed tokens for synchronization. AFS 3.0 provides moderately strong consistency guarantees, but they are not like UNIX because data are written back to the server only after a file is closed. AFS 4.0 writes back data to the server whenever there are other clients wanting to access it, the effect being like UNIX file access semantics. Also, AFS 3.0 does not guarantee synchronization of multiple writers, whereas AFS 4.0 does