Predicting object lifetimes in finite distributed storage systems under churn

Abstract

Distributed network storage systems regularly make use of object replication to achieve sufficient levels of reliability under network churn. To maintain reliability, a repair mechanism is employed, which replaces destroyed replicas. When designing a distributed storage system, it is of great benefit to design for objects with known lifetimes. That is to say, the expected time that an object will remain available in the storage network under measurable network conditions is known. The paper proposes an embedded continuous time Markov chain to model objects replicated in a finite network under churn, with repair. Object lifetimes are found to be dependant on node departure and arrival rates, initial network size and the average network size if the average network size is comparable to the required number of replicas. The theoretical model results are compared to an OMNeT++ network simulation and found to closely match.