TiDedup: A New Distributed Deduplication Architecture for Ceph.

This paper presents TiDedup, a new cluster-level deduplication architecture for Ceph, a widely deployed distributed storage system. Ceph introduced a cluster-level deduplication design before; unfortunately, a few shortcomings have made it hard to use in production: (1) Deduplication of unique data incurs excessive metadata consumption; (2) Its serialized tiering mechanism has detrimental effects on foreground I/Os, and by design, only provides fixed-sized chunking algorithms; and (3) The existing reference count mechanism resorts to inefficient full scan of entire objects, and does not work with Ceph's snapshot. TiDedup effectively overcomes these shortcomings by introducing three novel schemes: Selective cluster-level crawling, an event-driven tiering mechanism with content defined chunking, and a reference correction method using a shared reference back pointer. We have fully validated TiDedup and integrated it into the Ceph mainline, ready for evaluation and deployment in various experimental and production environments. Our evaluation results show that TiDedup achieves up to 34% data reduction on real-world workloads, and when compared with the existing deduplication design, improves foreground I/O throughput by 50% during deduplication, and significantly reduces the scan time for reference correction by more than 50%.