SSD read reclaim: why read-only workloads can cause writes

12 October 2026 / Evidence review and accounting examples

Read-only is a description of host requests, not a promise that the flash controller never programs a page. Read reclaim can move valid data to protect it from read disturbance. That makes the measurement boundary important, even when the application's write counter stays at zero.

The unread cells can be the affected cells

Reading a NAND wordline requires pass-through voltage on other cells in the string. Repeated exposure can shift their thresholds. Cai and colleagues studied this on planar 2Y-nm MLC NAND, using an FPGA platform. Their pass-voltage investigation used adjusted read-reference voltage as a workaround, not an ordinary host control for commercial SSDs. [1, sections 2-3]

Their results establish a mechanism, not a universal safe number of host reads. NAND generation, physical placement, wear and controller policy matter. Do not turn a laboratory read-count threshold into a drive replacement rule.

Reclaim is not retry

Read retry changes sensing settings to recover readable data. Read reclaim relocates data to fresh pages. A successful retry need not imply relocation. [3, section 2.3]

Conventional block-level reclaim copies valid pages when a read-count limit is reached. STRAW instead tracks disturbance at wordline granularity so that migration can be more selective. [2, sections II-III]

For interpretation, ask what triggered the work. Space reclamation, preservation against read disturbance, and retention management are different objectives, even when they use similar copying machinery. A counter named "background writes" alone cannot distinguish them.

Three evidence levels, not one benchmark

STRAW characterizes 160 Samsung 3D TLC chips across 19,200 blocks. Its system evaluation uses MQSim-E with four Alicloud traces and two synthetic workloads. Measured device characteristics inform a simulator; the reported system benefits are not a retail-SSD firmware trial. [2, sections III-A and IV]

A separate 2026 study by Park and colleagues tests 15 NVMe SSDs from 10 vendors. Its protocol uses direct I/O, temperature controls, a 90% sequential fill and a 30-minute idle period before testing. It reports different performance responses across devices. These are controlled tests, not fleet failure rates. [3, sections 4-5]

Our synthesis: use the planar work to understand the mechanism, STRAW to examine a proposed management policy, and the finished-device study to motivate longer read tests. None supplies a universal conversion from host bytes read to NAND bytes programmed.

Zero host writes make WAF undefined

Choose an explicit convention. Here byte WAF = NAND-programmed bytes / host-written bytes over the same interval. SNIA's dictionary instead defines an operation-count ratio; unequal operation sizes make those conventions different. [4]

If the denominator is zero, report "undefined," not zero, one or a finite substitute. Preserve the measured byte deltas and elapsed time. A separate programmed-bytes-per-read-byte ratio can describe overhead, but it is not WAF and does not identify the cause of those programs.

Invented counter deltas, not SSD measurements. MiB = 2^20 bytes. The combined row is the sum of the first two consecutive windows.
Window	Seconds	Host reads, MiB	Host writes, MiB	NAND programs, MiB	Byte WAF
write-window	3600	0	1024	1280	1.25
read-window	3600	65536	0	128	Undefined
combined-write-and-read	7200	65536	1024	1408	1.375
idle-window	3600	0	0	0	Undefined

The write window alone gives 1.25. Including the later read window gives 1.375, without changing either window's counters. The read window itself has no defined WAF. The idle row shows why 0/0 is also not zero.

This example does not attribute all internal writes to reclaim. Nor does it average per-window ratios. Pool compatible counters over nonoverlapping intervals first, then divide. Counter resets, saturation, missing values and mismatched units must be resolved before doing that arithmetic.

The machine-readable cases, schema and reproducible code preserve undefined ratios as JSON null and blank CSV cells. Tests reject missing, negative, fractional and unsafe counters. Inputs are constructed; no sampling confidence interval applies.

A useful measurement plan

Our proposed checklist: record firmware, capacity, workload range and read distribution; separate preparation from the observation window; retain time series rather than only averages; monitor temperature and background activity; distinguish host counters from documented physical-program counters.

If physical telemetry is unavailable, say so. Do not infer NAND write volume from a flat host-write counter or diagnose read reclaim solely from a latency spike. Thermal throttling, garbage collection and other controller work are competing explanations that need separate evidence.

No destructive preconditioning was performed for this article. Such tests belong on expendable, backed-up test media with explicit authorization, not on a live production volume. We have not reproduced the cited experiments.

Continue with read-retry latency, 3D NAND retention and SSD steady-state measurement. Definitions: read disturb and write amplification.

Primary sources and limits

Sources checked 12 October 2026. No copied figures, product ranking, empirical endurance estimate or claimed paper replication. Source works retain their own rights.