hesela.devSMRSTORE
SMRSTORE is a log-structured storage engine that maps chunks directly onto host-managed SMR zones in a distributed object-storage stack.
The FAST 2023 paper separates placement by stream type, adapts chunk-size limits to zones and uses preopened zone pools with round-robin allocation. Large chunks can occupy exclusive zones; smaller chunks still share space. Separating stream types alone was insufficient in the reported ablation. [1]
Recovery loads checkpoints and journals, then scans data zones for appends that were deliberately not journaled. Zone resets are journaled because a zone can be reused between checkpoints. [1]
Relations
- related
- Host-managed SMR (HM-SMR), SMORE
Limitations
- Round-robin placement reduced small-write throughput in the clean-state microbenchmark; its later reclamation benefit is not a claim that every write becomes faster. [1]
- Evaluation is tied to the OSS/Pangu stack, specific drives, SSD caches and workloads; most baseline comparisons restrict disks to 6 TB, with a separate 20 TB SMRSTORE variant. [1]
Citations
Further reading
- USENIX FAST 2023: SMRSTORE author list and deployed-systems publication record.
- Hesela: Host-Managed SMR, interface and object-store designs.
Sources & caveats
- The paper reports a historical deployment. This record does not establish current availability, reproduce its benchmark or promise equivalent behavior outside that stack.