Light-Dedup: A Light-weight Inline Deduplication Framework for Non-Volatile Memory File Systems

Emerging NVM is promising to become the next-generation storage media. However, its high cost hinders its development. Recent deduplication researches in NVM file systems demonstrate that NVM's cost can be reduced by eliminating redundant data blocks, but their design lacks complete insights into NVM's I/O mechanisms.

We propose Light-Dedup, a light-weight inline deduplication framework for NVM file systems that performs fast block-level deduplication while taking NVM's I/O mechanisms into consideration. Specifically, Light-Dedup proposes Light-Redundant-Block-Identifier (LRBI), which combines non-cryptographic hash with a speculative-prefetch-based byte-by-byte content-comparison approach. LRBI leverages the memory interface of NVM to enable asynchronous reads by speculatively prefetching in-NVM data blocks into the CPU/NVM buffers. Thus, NVM's read latency seen by content-comparison is markedly reduced due to buffer hits. Moreover, Light-Dedup adopts an in-NVM Light-Meta-Table (LMT) to store deduplication metadata and collaborate with LRBI. LMT is organized in the region granularity, which significantly reduces metadata I/O amplification and improves deduplication performance.

Experimental results suggest Light-Dedup achieves 1.01--8.98× I/O throughput over the state-of-the-art NVM deduplication file systems. Here, the speculative prefetch technique used in LRBI improves Light-Dedup by 0.3--118%. In addition, the region-based layout of LMT reduces metadata read/write amplification from 19.35× /9.86× to 6.10× /3.43× in our hand-crafted aging workload.