Read disturb
Read disturb is an unintended change in NAND cell thresholds caused by reads to other wordlines in the same block.
Pass-through voltage on unread cells can shift their thresholds. Cai et al. characterized planar MLC chips, not current fleet failure rates. Read disturb differs from retention loss; read reclaim relocates data while read retry changes sensing. [1] [4] Disturbance need not be uniform across 3D NAND wordlines: adjacent-wordline reads can impose stronger stress. [5]
Observation window: in Park et al.'s 15-SSD NVMe study, two drives remained stable during a 30,000-second sequential-read test over a 1-GB range. A 1,000,000-second extension revealed drops with approximate periods of 45,000 and 83,000 seconds. Those are observed performance periods, not safe read-count limits or failure times. [7]
Temporary read errors after idle are a separate, reversible observation; an improved second read alone does not demonstrate removal of accumulated disturbance. [6]
Relations
- related
- NAND flash, Flash translation layer (FTL), NAND read-retry, Flash data retention, NAND program interference, Read reclaim, NAND wordline, Read-pass voltage, Temporary read errors (NAND), Read Disturb Recovery (RDR), Patrol read
Limitations
- Successful I/O completion is not an independent comparison with expected data; fio direct I/O selects non-buffered access, while content verification is a separate setting. [8]
Citations
- Cai, Luo, Ghose, Haratsch, Mai and Mutlu. Read Disturb Errors in MLC NAND Flash Memory: Characterization, Mitigation, and Recovery. DSN 2015, sections 2 and 3
- Hesela: SSD read reclaim and read-only accounting
- Chun et al. STRAW: A Stress-Aware WL-Based Read Reclaim Technique for High-Density NAND Flash-Based SSDs. arXiv:2501.02517v1 (2025).
- Zambelli et al. Temporary read errors in TLC 3D NAND after idle. JEDS 8 (2020), 99-104. DOI: 10.1109/JEDS.2020.2965648.
- Park et al. Experimental Study on System-Level Performance Impact of Read Disturbance in Modern SSDs. POMACS 2026, sections 4-5.1. DOI 10.1145/3788092.
- fio documentation: direct I/O and verification options (sections 1.13.6 and 1.13.17)
Further reading
Sources & caveats
- Cai et al., DSN 2015, sections 2.1-3.2 rechecked 2026-11-22: cell-level mechanism and FPGA characterization, not a host-read threshold. [1]
- Park et al., POMACS 2026, sections 4.1-4.2 and 5.1: 500/512-GB M.2 devices; within-model variation was not systematically studied. Variable runs used the dominant pattern rather than a distribution of all runs. [7]
- The paper reports successful read requests. This does not document an independent end-to-end comparison, and the glossary does not infer either undetected corruption or a measured zero corruption rate. [7] [8]
- Source review only; no SSD stress test or hardware replication was performed.