Reliable Repair Mechanisms with Low Connection Cost for Code Based Distributed Storage Systems

Abstract

Erasure codes are applied in distributed storage systems for fault-tolerance with lower storage overhead than replications. Later, decentralized erasure codes are proposed for decentralized or loosely-organized storage systems. Repair mechanisms aim at maintaining redundancy over time such that stored data are still retrievable. Two recent repair mechanisms, NOOP and COOP, are designed for decentralized erasure code based distributed storage system to minimize connection cost in theoretical manner. We propose a generalized repair framework, which includes NOOP and COOP as two extreme cases. We then investigate trade-off between connection cost and data retrievability from an experimental aspect in our repair framework. Our results show that a reasonable data retrievability is achievable with constant connection cost, which is less than previously analytical values. These results are valuable references for a system manager to build a reliable storage system with low connection cost.