DS-Dedupe: A scalable, low network overhead data routing algorithm for inline cluster deduplication system

摘要

Inline cluster deduplication technique has been widely used in data centers to improve storage efficiency. Data routing algorithm has a crucial impact on the deduplication factor, throughput and scalability in a cluster deduplication system. In this paper, we propose a stateful data routing algorithm called DS-Dedupe. To make full use of similarity in data streams, DS-Dedupe builds up a super-chunk granularity similarity index in each client to trace the super-chunks that have been routed. Then we calculate a similarity coefficient according to the index to determine whether a new super-chunk should be assigned directly or by a consistent hash, thus strike a sensible tradeoff between deduplication factor and network overhead. Our experiments on two datasets demonstrate that DS-Dedupe achieves a high elimination ratio at a low communication overhead. Besides, as data routing is operated by client node, metadata server bottleneck can be avoided.