TTFLASH
TTFLASH is a research SSD design that combines fine-grained garbage collection scheduling, parity reconstruction and protected buffering to reduce GC-induced latency.
Plane isolation: copyback leaves other planes available. Parity can reconstruct a read whose target is busy, while rotating collection normally limits each stripe's plane group to one active GC. Collections in different groups can overlap. [2]
Reconstruction is conditional: extra reads contend for channels, so a short remaining GC can be cheaper to wait for. Protected RAM lets writes complete before flash eviction, with GC-tolerant flushing handling busy destinations. [1]
Interpretation: low GC delay is not a guarantee of low latency from every cause, nor proof of data integrity. Read the scheduling, buffering and error-correction assumptions together.
Relations
Limitations
- Heavy write bursts can force multiple collections in one group to prevent free-page exhaustion. [1]
- Copyback bypasses controller ECC; foreground reads retain ECC checking. Background scrubbing was proposed, not evaluated. [1]
Citations
- Yan et al. Tiny-Tail Flash: Near-Perfect Elimination of Garbage Collection Tail Latencies in NAND SSDs. FAST 2017, sections 4-7 and Appendix A.
- Yan et al. Tiny-Tail Flash. FAST 2017 presentation, slides 19-41.
Further reading
- USENIX FAST 2017 publication record: Tiny-Tail Flash, pp. 15-28.
- Hesela: NAND copyback, faster relocation is not fewer writes.
Sources & caveats
- Published evaluation used SSDSim simulation and VSSIM-based emulation, not a completed TTFLASH hardware prototype. [1]
- Source synthesis, not a Hesela benchmark reproduction. The linked DWPD record explains the paper's eight-hour workload convention.