Write amplification
Device-level byte write amplification is the ratio of bytes programmed to the storage medium to bytes written by the host over the same accounting interval.
State the measurement boundary and whether the numerator includes metadata, parity and other internal writes. Operation-count and byte-count ratios are not interchangeable when write sizes differ. Copyback can reduce relocation traffic without changing programmed bytes. Database output/logical input is a separate accounting boundary; see database-write-amplification. If host-written bytes are zero, the interval ratio is undefined, including 0/0. Read reclaim can still program NAND; preserve physical-program deltas and elapsed time instead of replacing the undefined ratio with zero.
Across compatible, disjoint intervals, sum physical-write and host-write deltas before dividing. An unweighted average of interval ratios can differ greatly from this pooled ratio. Keep zero-host-write intervals in the physical total even though their local ratio is undefined, and handle counter resets or wraparound before aggregation. [6]
A value below 1 is possible when device-side compression reduces payload enough to outweigh internal writes. Host-side compression changes the bytes reaching the device instead; it does not establish device-level WAF. Keep compression ratio, write suppression and total media-write accounting separate. [7]
GeckoFTL instead defines a cost-weighted metric, i_writes + i_reads/delta: internal operations per logical page update, with delta the write/read latency ratio. It is not a byte-write ratio. [8]
NVMe SMART Data Units Written is host-written data, not media programming. Supported Endurance Group logs have separate Data Units Written and Media Units Written fields. Match group, interval and rounded units; zero means not reported. [9]
RIPQ adds a cache-write boundary. [10]
Relations
- related
- Flash translation layer (FTL), Garbage collection (flash), Quad-level cell (QLC), TRIM, NAND channel, NAND copyback, Read-modify-write, Database write amplification, Read reclaim, Write combining, Page-level mapping (FTL), Content-aware FTL, Transparent filesystem compression, GeckoFTL, NVMe Data Units Written, NVMe Endurance Group, Page-differential logging, Restricted Insertion Priority Queue (RIPQ), Kangaroo
Limitations
- A traffic reduction or a latency improvement does not by itself measure this ratio. [4]
Citations
- Hesela: NAND copyback, faster relocation is not fewer writes
- Hesela: finite-block cleaning and pooled write-amplification accounting
- Hands and Anderson. TCO in Computational Storage - Compression. SNIA, 2023, pp. 2-4.
- Dayan, Bonnet and Idreos. GeckoFTL. SIGMOD 2016, section 5 Metrics and section 5.1: weighted I/O versus metadata-only overhead.
- NVM Express Base Specification 2.3 (2025): SMART / Health Information and Endurance Group Information logs, Figure 222.
- Hesela: matched-boundary cache write accounting with hypothetical ledgers
Further reading
- Write amplification
- Steve Peters. What Can/Can't be Done and What's Nice to Have in the New Stack. SNIA presentation, slide 11.
- SNIA dictionary: write amplification factor (operation-count convention)
- Hesela: SSD read reclaim and read-only accounting
Sources & caveats
- The compression source is an industry-authored computational-storage case study, not evidence that every SSD compresses host writes. Its performance and TCO figures are not generalized here. [7]