Stochastic Petri Net Analysis of a Replicated File System

We present a stochastic Petri net model of a replicated file system in a distributed environment where replicated files reside on different hosts and a voting algorithm is used to maintain consistency. Witnesses, which simply record the status of the file but contain no data, may be used in addition to or in place of files to reduce overhead. We present a model sufficiently detailed to include file status (current or out-of-date) as well as failure and repair of hosts where copies or witnesses reside. The number of copies and witnesses is not fixed, but is a parameter of the model. Two different majority protocols are examined, one where a majority of all copies and witnesses is necessary to form a quorum, the other where only a majority of the copies and witnesses on operational hosts is needed. The latter, known as adaptive voting, is shown to increase file availability in most cases. We also investigate the process of selection of copies and witnesses to participate in an update when more than the majority is available and show the inherent performance/reliability tradeoffs.