STRAW: Stress-Aware WL-Based Read Disturbance Management for High-Density NAND Flash Memory

Myoungjun Chun, Jaeyong Lee, Inhyuk Choi, Jisung Park, Myungsuk Kim and Jihong Kim

Reading one page of a modern flash chip slightly disturbs its neighbors, and once too many reads pile up the drive must defensively rewrite a whole block of data to avoid losing it — an expensive operation that wears the drive out faster; this work from Seoul National University and collaborators shows that in today's tall 3D flash the damage is wildly uneven, hitting a read page's immediate neighbors far harder than distant ones, so rewriting the entire block is massive overkill; STRAW instead tracks the accumulated stress on each individual wordline and rewrites only the few that are genuinely endangered, cutting the defensive-copy overhead by about 89 percent.

The argument in one glance

Problem

Reading flash disturbs neighbors, forcing costly defensive rewrites

To read a page, a flash chip applies a high voltage to every other wordline in the block, slightly disturbing them, and repeated reads can corrupt their data. The defense, read reclaim, rewrites the whole block once its read count crosses a threshold — extra writes that hurt performance and wear the drive out.
What people did

Trigger read reclaim at the block granularity

Count reads per block, and when the count exceeds a conservative threshold, reclaim the entire block by copying all its valid pages to fresh locations, resetting the disturbance.
Why it fell short

Block-level reclaim is wildly wasteful in 3D flash

In tall 3D flash the disturbance is very uneven — a read hits its immediate neighbor wordlines far harder than distant ones — so a threshold set for the worst case triggers reclaim far more than most wordlines need. And reacting only after errors accumulate forces heavy copy overheads in disturbance-prone modern flash.
This paper's idea

Track stress per wordline and reclaim only the endangered (STRAW)

STRAW manages disturbance per individual wordline, not per block. A device-characterized model estimates the accumulated disturbance stress on each wordline from the reads to it and its neighbors, and reclaims only the heavily-disturbed ones; approximate counting keeps the bookkeeping cheap. With a complementary stress-reduced read, STRAW cuts reclaim-induced page-copy overhead by about 89 percent versus the state of the art.

Why this problem is worth solving

Flash storage keeps getting denser, and every gain in density has made the chips more fragile in ways the drive’s firmware must quietly paper over.

NAND flash stores bits as charge in cells, arranged in wordlines within blocks, and it has grown denser over decades by stacking wordlines vertically in three dimensions and storing multiple bits per cell. That density brought a reliability cost, and one of the sharpest is read disturbance. To read a page from a target wordline, the chip must apply a high pass-through voltage to all the other wordlines in the same block, so that current can flow through them to sense the target. That pass-through voltage slightly and unintentionally programs those non-target cells — nudging their charge — and if a block is read many times, this accumulated nudging can eventually flip enough bits to corrupt the stored data. As blocks grow with continued vertical stacking, each read disturbs a larger amount of data at once, so read disturbance has become a leading reliability concern in modern and future high-density flash. It is a purely operational hazard: simply reading data, without ever writing, can destroy other data in the same block.

The firmware’s standard defense is effective but blunt, and its cost has grown into a real problem as flash has densified.

The established way to handle read disturbance is read reclaim. The drive counts how many reads a block has received, and when that count crosses a predefined threshold, the controller triggers read reclaim: it rewrites all of the block’s valid pages to fresh locations and erases the block, wiping out the accumulated disturbance before it can cause errors. This works, but it is not free. Those extra rewrites consume bandwidth and hurt read latency, and crucially, flash cells wear out with each program-erase cycle, so every reclaim-induced write shortens the drive’s usable life. The threshold therefore has to be chosen carefully — low enough to prevent data corruption, but not so low that it triggers reclaim needlessly. Every prior technique operates at the block granularity: it counts reads per block and reclaims the whole block at once. The problem this work exposes is that block-level reclaim is a poor match for how disturbance actually behaves in modern 3D flash, so it ends up doing far more of these costly rewrites than necessary.

What people did before, and where it broke

Count reads per block and reclaim the whole block. Trigger read reclaim when a block’s read count crosses a threshold, rewriting all its valid pages. Where it broke. In 3D flash the disturbance is very uneven across wordlines, but block-level reclaim treats them uniformly, so it cannot target only the endangered ones.
Set the threshold for the worst case. Choose the block’s reclaim threshold conservatively so that even the worst access pattern (repeatedly reading one wordline, which hammers its neighbors) never causes corruption. Where it broke. The worst case is rare, but the conservative threshold applies always, so the whole block is reclaimed far more often than the typical wordline actually needs.
Migrate pages by their type. A state-of-the-art refinement classifies pages within a block by page type (which bit they store) and migrates the more vulnerable ones. Where it broke. It helps somewhat but still misses the dominant effect — that a read disturbs its immediate wordline neighbors far more than distant wordlines — so its savings are limited.
React only after errors accumulate. Wait until disturbance-induced errors are near the correction limit, then reclaim. Where it broke. In disturbance-prone modern flash this forces heavy, frequent reclaim copying, exactly when the drive can least afford the write overhead.
The gap. No technique managed disturbance at the granularity where it actually varies — the individual wordline — nor reduced the disturbance a read inflicts in the first place. Providing both is what STRAW does.

How STRAW works

From block-level reclaim to per-wordline stress tracking

↓ →
new
↓ →
new
↓ →
new
↓ →
new
↓ →
new

The idea

Manage disturbance where it actually varies: the individual wordline. STRAW’s foundational shift is to stop treating a block as one uniform unit and instead track and act on each wordline separately, because that is the granularity at which read disturbance genuinely differs. Measurements of real 3D flash show a read stresses its two immediate neighbor wordlines about eight times more than distant ones, and different wordlines tolerate different amounts before failing, so a single block-wide threshold is a crude average that forces reclaim on the whole block whenever any part might be at risk. By keeping a per-wordline account of accumulated stress and reclaiming only the few wordlines that are genuinely endangered, STRAW avoids the enormous waste of rewriting an entire block for the sake of a handful of hot lines. Moving disturbance management from a coarse aggregate down to the fine granularity at which the physical effect concentrates, so protection is applied precisely where it is needed, is the same move the per-row activation-counter work covered elsewhere on this site makes, tracking disturbance at the level of the individual row rather than a whole region so that the costly mitigation fires only for the rows that are actually being hammered.

Ground the policy in careful measurement of the real device. STRAW does not guess at disturbance; it derives a model from characterizing many real 3D flash chips, classifying wordlines into reliability groups and capturing two measured parameters — how many non-adjacent reads a group’s weakest wordline can tolerate, and how much more an adjacent read hurts than a distant one. From these, and the read counts of a wordline and its neighbors, it computes an effective read count that reflects the true accumulated stress, which is what lets it decide precisely when a wordline needs rescuing. Building a management policy on empirical characterization of how the storage medium actually behaves, rather than on a conservative worst-case rule of thumb, is the same discipline the learning-based flash-translation-layer work covered elsewhere on this site follows, driving the firmware’s decisions from a data-grounded model of the device so that the drive is managed according to its real behavior instead of a pessimistic approximation.

Do only the work that is truly necessary, and cause less damage to begin with. STRAW attacks reclaim overhead from two sides. It refuses to reclaim what does not need it — only heavily-disturbed wordlines are copied, and an approximate counting scheme keeps the per-wordline tracking nearly free in memory — and it also reduces the disturbance each read inflicts, by scaling the pass-through voltages according to whether a wordline still holds valid data, so valid wordlines are stressed less in the first place. Eliminating reliability overhead that a coarser scheme incurs unnecessarily, by both skipping needless protective work and lowering the underlying stress, is the same goal the near-zero read-retry work covered elsewhere on this site pursues, cutting the routine reliability cost that dense flash would otherwise pay on every access so that the drive spends far less of its life and performance on defensive maintenance.

Following a read through STRAW

Tracking and reclaiming disturbance per wordline

A trace of STRAW handling reads on a modern 3D flash SSD, following the paper's stress-aware wordline-based design.

Step 1 of 5

A page is read from a target wordline

The drive reads a page, applying a high pass-through voltage to the other wordlines in the block. STRAW updates its counters for the block and, importantly, for the specific wordlines involved, using an approximate counting structure to stay memory-cheap.

The two wordlines adjacent to the target absorb the most disturbance.

Why it works

Disturbance concentrates on neighbors, so per-wordline targeting saves most of the work. Because a read stresses its adjacent wordlines far more than distant ones → reclaiming only the heavily-disturbed wordlines addresses the real damage → so the block’s many lightly-disturbed wordlines are never needlessly rewritten.
A measured model beats a worst-case threshold. Because STRAW estimates each wordline’s actual accumulated stress from device characterization → it reclaims based on true endangerment rather than a conservative block-wide rule → so reclaim fires far less often while still preventing corruption.
Approximate counting makes per-wordline tracking affordable. Because exact per-wordline counters would cost far too much drive memory → tracking only the most-read wordlines with a never-underestimating algorithm keeps the metadata tiny → so the fine-grained policy is practical without risking data loss.
Never underestimating keeps safety intact. Because the approximate counter can only over- or correctly-estimate a wordline’s reads, never undercount → a truly endangered wordline is never missed → so the worst approximation cost is an occasional early reclaim, not a corruption.
Reducing per-read stress delays reclaim itself. Because scaling pass-through voltages by validity lowers the disturbance inflicted on valid wordlines → each wordline reaches its tolerable limit more slowly → so reclaim is needed less frequently, compounding the savings.

What it costs (trade-offs and scope)

It needs per-device characterization. STRAW’s model comes from measuring real flash chips to get each reliability group’s tolerable read counts and disturbance rates, stored in a parameter table. This offline profiling must be done for the target flash, and the model’s accuracy bounds the policy’s precision.
Approximate counting can trigger early reclaim. To save memory, STRAW tracks only the hottest wordlines and may overestimate some counts. Because it never underestimates, this never causes data loss, but it can cause an occasional premature reclaim — a small efficiency cost traded for a large memory saving.
It adds firmware complexity to the flash translation layer. STRAW extends the drive’s flash translation layer with a parameter table, per-wordline approximate counters, and periodic checking logic. This is modest, but it is more moving parts than a simple block read counter.
Reclaim overhead is reduced, not eliminated. STRAW cuts reclaim-induced copying substantially (about 89 percent versus the state of the art in the paper’s evaluation), but read disturbance still fundamentally requires some reclaim; the technique makes it far cheaper, not free.
Stress-reduced read depends on validity and the flash interface. Scaling pass-through voltages by whether a wordline holds valid data assumes the controller can vary those voltages and knows validity; its benefit depends on the workload’s mix of valid and invalid data and on the device supporting such tuning.
The gains depend on access patterns and wear. The savings vary with how reads are distributed and the drive’s program-erase cycle count; the disturbance asymmetry and thresholds shift with wear, which the model accounts for, but different workloads and devices will see different magnitudes.

The evaluation

Ways to manage read disturbance in modern 3D flash

Property	Block-level read reclaim	Page-type-aware migration
Granularity of management
Basis for reclaiming		Page vulnerability
Reduces the disturbance per read	No	No
Metadata cost	Tiny (per-block counter)	Modest
Reclaim-copy overhead	Baseline (high)

Ways to manage read disturbance in high-density 3D NAND flash, as compared in the paper. Conventional block-level read reclaim counts reads per block and, at a conservative worst-case threshold, rewrites every valid page in the block — necessary for safety but wasteful, because in 3D flash a read disturbs its immediate neighbor wordlines about eight times more than distant ones, and wordline tolerance varies with process variation, so most wordlines are reclaimed long before they need it. A state-of-the-art refinement migrates pages by their page type, helping somewhat (around 29 percent less copying) but missing the dominant neighbor effect. STRAW manages disturbance at the individual wordline: a model derived from characterizing many real 3D flash chips groups wordlines by reliability and, from the reads to a wordline and its neighbors, computes an effective read count of the true accumulated stress, so only heavily-disturbed wordlines are reclaimed; a Space-Saving approximate counting algorithm tracks per-wordline reads with tiny memory and never underestimates, keeping data safe. A complementary stress-reduced read scales the pass-through voltages by whether a wordline holds valid data, lowering the disturbance each read inflicts on valid wordlines. Evaluated on a modern SSD emulator, STRAW reduces reclaim-induced page-copy overhead by about 89 percent on average versus the state-of-the-art technique (the wordline-based reclaim alone cuts reclaim copies by well over 80 percent versus block-level reclaim and sharply lowers read tail latency), with negligible storage overhead.Underlined values explain themselves when clicked.

Where this stops working

What to keep in mind

STRAW is a focused, well-characterized piece of storage-systems research that turns a detailed understanding of device physics into a concrete firmware win. The problem is real and worsening: as NAND flash densifies through vertical stacking and multi-bit cells, read disturbance — the corruption of neighboring data caused simply by reading a page, because the read applies a high voltage to other wordlines — has become a leading reliability threat, and the standard defense, read reclaim, is expensive because it rewrites whole blocks, and those writes both slow the drive and wear it out. The paper’s key empirical insight is that in tall 3D flash the disturbance is highly uneven: a read hits its two immediate neighbor wordlines about eight times harder than distant ones, and wordlines vary in their tolerance, so managing reclaim at the whole-block granularity with a worst-case threshold forces far more rewriting than most wordlines need. STRAW’s answer is to track accumulated stress per wordline using a model built from characterizing many real chips, reclaim only the wordlines that are genuinely endangered, keep the per-wordline bookkeeping cheap with an approximate counting algorithm that never underestimates (so safety is preserved), and additionally reduce the disturbance each read causes by scaling pass-through voltages according to whether a wordline holds valid data. The reported result is strong — about 89 percent less reclaim-induced copying than the state of the art, with negligible storage overhead. Read the scope carefully. The model requires per-device characterization and offline profiling; the approximate counting trades an occasional early reclaim for its memory savings; the technique reduces but does not eliminate reclaim; the stress-reduced read assumes the controller can vary pass-through voltages and know validity; and the gains depend on access patterns and wear level, evaluated on an SSD emulator rather than a shipping drive. The durable idea is that when a physical effect is concentrated and uneven, managing it at the coarse granularity of the whole unit is deeply wasteful, and measuring where the effect actually falls — then acting only there, while also lowering the effect at its source — can eliminate most of the overhead a blunt defense would incur.

Background

Why density keeps turning access into a reliability hazard

It helps to see STRAW alongside a pattern that recurs across memory and storage technologies as they are pushed to higher density: the very act of accessing a cell begins to disturb its neighbors, turning ordinary reads into a reliability hazard that the system must actively manage. Flash read disturbance and DRAM RowHammer are cousins — in both, repeatedly touching one location leaks charge into physically adjacent cells and can flip their bits, a direct consequence of packing cells closer together. The RowHammer-mitigation work covered elsewhere on this site confronts the same physics in DRAM, wrestling with how to detect and contain disturbance from repeated access cheaply enough that the defense itself does not dominate the cost — exactly the tension STRAW navigates for flash, where the defensive rewrites threaten performance and lifetime. The common thread is that scaling a memory technology trades density for fragility, and the burden of hiding that fragility falls on controllers and firmware, whose challenge is to protect data without letting the protection swamp the very performance the density was meant to buy. As both flash and DRAM continue to scale, techniques like STRAW that measure precisely where disturbance concentrates and spend defensive effort only there — rather than applying a uniform, worst-case guard — are likely to be how these technologies stay both dense and reliable.

If you remember only three things

1. NAND flash has grown denser by stacking wordlines vertically (3D) and storing multiple bits per cell, which has sharpened read disturbance: to read a page, the chip applies a high pass-through voltage to every other wordline in the block, slightly disturbing those cells, and enough reads can corrupt their data — corruption caused purely by reading, without writing. The standard defense is read reclaim: when a block's read count crosses a threshold, the controller rewrites all its valid pages to fresh locations, wiping out the disturbance. But those extra writes hurt performance and wear the drive out. Existing techniques reclaim at whole-block granularity with a conservative worst-case threshold, which is a poor fit for modern 3D flash: the disturbance is highly uneven — a read hits its two immediate neighbor wordlines about eight times harder than distant ones, and wordlines differ in tolerance — so most wordlines are reclaimed far more often than they need, and reacting only after errors accumulate forces heavy copy overhead.
2. STRAW manages read disturbance at the individual wordline rather than the block. It builds a read-disturbance model from characterizing many real 3D TLC flash chips, grouping wordlines by reliability and capturing two measured parameters — how many non-adjacent reads a group's weakest wordline tolerates, and how much more an adjacent read disturbs than a distant one (the disturbance rate). From the read counts of a wordline and its neighbors it computes an effective read count reflecting the true accumulated stress, and its extended flash translation layer periodically reclaims only the wordlines whose effective stress nears the tolerable limit, sparing the rest of the block. To make per-wordline tracking affordable, it uses the Space-Saving approximate counting algorithm to track only the most-read wordlines with tiny memory, and because that algorithm never underestimates a count, a truly endangered wordline is never missed (the only cost is an occasional premature reclaim). A complementary stress-reduced read scales the pass-through voltages by whether a wordline still holds valid data, lowering the disturbance each read inflicts on valid wordlines.
3. Evaluated on a modern SSD emulator, STRAW reduces reclaim-induced page-copy overhead by about 89 percent on average versus the state-of-the-art technique, with negligible storage overhead — the wordline-based reclaim alone cuts reclaim copies by well over 80 percent versus block-level reclaim and sharply lowers read tail latency, and the stress-reduced read adds further savings. Scope: the model needs per-device offline characterization; approximate counting trades an occasional early reclaim for memory savings; the technique reduces but does not eliminate reclaim; the stress-reduced read assumes the controller can vary pass-through voltages and know validity; and gains depend on access patterns and wear, measured on an emulator. The durable lesson is that when a physical effect is concentrated and uneven, guarding against it at the coarse granularity of the whole unit is deeply wasteful — measuring where the effect actually falls and acting only there, while also lowering the effect at its source, removes most of the overhead a blunt defense would pay.

Check yourself

Check your understanding

1. What is read disturbance in NAND flash, and why does read reclaim address it expensively?
2. Why is managing read disturbance at the whole-block granularity wasteful in modern 3D flash?
3. How does STRAW decide which wordlines to reclaim, and how does it keep that affordable?
4. What did STRAW achieve, and what are its limits?

Sources

The paper is “STRAW: Stress-Aware WL-Based Read Disturbance Management for High-Density NAND Flash Memory” by Myoungjun Chun, Jaeyong Lee, Inhyuk Choi, Jisung Park, Myungsuk Kim, and Jihong Kim, published at ASPLOS 2026, DOI 10.1145/3779212.3790228. The authors are at Seoul National University, POSTECH, and Kyungpook National University.

This explainer is based on the authors’ own detailed technical description of STRAW’s core, published as “STRAW: A Stress-Aware WL-Based Read Reclaim Technique for High-Density NAND Flash-Based SSDs” (arXiv:2501.02517, also in IEEE Computer Architecture Letters), together with the ASPLOS 2026 paper’s abstract for the extended contribution. The problem framing and the core mechanism (read disturbance from the high pass-through voltage applied to non-target wordlines; block-level read reclaim and its conservative worst-case threshold; the measured heterogeneity of disturbance in 3D flash, with adjacent wordlines disturbed about eight times more than non-adjacent ones and wordline tolerance varying by process variation; the read-disturbance model classifying wordlines into reliability groups with an effective maximum read count and a disturbance rate derived from characterizing many real 3D TLC chips; the STRAW-enabled flash translation layer that computes each wordline’s effective read count and reclaims only heavily-disturbed wordlines; and the Space-Saving approximate counting algorithm that tracks the most-read wordlines with small memory and never underestimates) are drawn from that full technical text, where the wordline-based reclaim reduces reclaim-induced page copies by over 80 percent versus block-level reclaim and cuts 99.9th-percentile read tail latency substantially, compared against a block-level baseline and a page-type-aware migration technique on a modern SSD emulator. The ASPLOS 2026 extension — the second key idea, stress-reduced read, which scales pass-through voltages based on wordline validity to mitigate disturbance on valid wordlines during each read, and the overall reported reduction of reclaim-induced page-copy overhead by 88.6 percent on average versus the state-of-the-art technique — is drawn from the ASPLOS abstract; the detailed internals of that mechanism are described here only at the level the abstract provides. Background explanations of NAND flash, wordlines and blocks, read disturbance, read reclaim, program-erase cycles, and the flash translation layer are standard and provided to make the paper’s contribution understandable.