Filesystem aging: the state SSD warm-up does not describe

Three different meanings of history

Filesystem aging concerns layout changes as files are created, removed and modified. It is distinct from NAND wear and calendar age. Fragmentation can occur within a file, between related files, or in the unused allocation space. The last affects where future writes can be placed; it is not itself data read by a file scan. [3, section 3]

Our SSD steady-state analysis asks whether a metric meets a stability rule under a stated workload. This article asks a different question: what filesystem state produced that workload's accesses? A short, flat throughput trace cannot answer it.

Same free capacity, different geometry

This original synthetic illustration uses 16 logical blocks, eight occupied and eight free. A deliberately simple allocator takes the first six free blocks in ascending address order. It permits splitting a request across runs and never relocates existing data.

All three allocations succeed and leave the same capacity. Their geometry differs. The result proves a narrow point: a percentage does not determine a layout. It predicts neither latency nor throughput. Actual allocators can choose different positions, reserve space or delay allocation; SSD mapping adds another layer.

Download the dependency-free model. Its exported examples() function generates this figure and table. Tests cover every eight-block binary layout and each feasible positive allocation. No filesystem, device, cache, timing, metadata or NAND model is executed.

What the papers actually compare

FAST 2017: preserve contents, change layout history

Conway and colleagues ran application churn, including Linux Git history, and compared recursive scans with fresh copies of the same contents. Their dynamic layout score uses block-request traces to describe locality across files and metadata, not just a single file's extent count. The setup used Linux 3.11.10, a Seagate HDD and a Sandisk Extreme Pro SSD. These are historical experiments, not a current filesystem ranking. [1, sections 3.3, 4 and 6]

ATC 2018: age the space that is not occupied

Geriatrix creates and deletes files to match profiles of fullness, file sizes, directory depth and relative file age. It targets both file layout and free-space fragmentation. Crucially, its section 6 comparison built aged images in memory and copied them to the target before testing, explicitly separating filesystem aging from device aging. Do not describe that protocol as SSD preconditioning. [2, sections 4-6]

HotStorage 2019: fullness is a separate comparison

The follow-up replayed the same logical operation sequence on smaller, fuller and larger, emptier partitions, plus a fresh-copy baseline. In its Git experiment, fullness added little read cost on the tested SSD, although use-induced aging remained. Free-space distributions were measured directly only for ext4. This bounds the result: it does not establish that capacity pressure is harmless for all workloads. [3, section 4 and Figure 3]

Why these results are not one slowdown percentage

Our synthesis: aged versus fresh, full versus less full, and differently conditioned devices are different comparisons. A workload may expose read locality while another stresses allocation or cleaning. Before comparing numbers, align the operation mix, preparation, cache policy, hardware and denominator. There is no justified pooled penalty here.

A concrete source-quality warning: the Geriatrix publisher-page abstract and PDF give different headline SSD maxima. We do not reuse either as a general correction factor. A future numeric extraction must identify the exact figure, workload and baseline rather than choosing the larger headline. Publisher record; full paper.

These papers motivate a measurement control, not a blanket instruction to defragment SSDs. Rewriting data has costs, and a faster scan is not evidence of improved durability or lower wear.

A practical reporting protocol

For a new experiment, we recommend preserving the following evidence. This is a proposed protocol, not a benchmark performed by Hesela.

1. Identify filesystem and kernel versions, mount options, partition geometry, storage device and firmware.
2. Record the operation history or aging-profile inputs and seed, not only bytes occupied or elapsed hours.
3. Separate file-size and directory distributions, file extents and free-space extent distributions. State what was not measured.
4. Document device conditioning, discard history, cache policy, write budget and reset procedure independently.
5. Use explicit baselines: the same contents in a fresh layout, and the same logical history at different fullness. A file copy and a block-image copy do not preserve the same properties.
6. Repeat measurements, keep raw timings and operation counts, and report uncertainty. Do not infer tail latency from mean throughput.

Any workload that creates, deletes or rewrites data belongs on an explicitly disposable target. We ran no aging workload, installed no research filesystem and performed no paper replication.

Related: logical size, allocation and capacity; deduplicated restore fragmentation. These concern related but distinct allocation layers.

Primary sources and limits

1. Conway et al. File Systems Fated for Senescence? Nonsense, Says Science! FAST 2017, pp. 45-58, sections 3.3, 4 and 6.
2. Kadekodi, Nagarajan, Ganger and Gibson. Geriatrix: Aging what you see and what you don't see. USENIX ATC 2018, pp. 691-704, sections 4-6.
3. Conway et al. Filesystem Aging: It's more Usage than Fullness. HotStorage 2019, sections 3-4 and Figure 3.

Methods and results reviewed 14 October 2026. No new hardware data, filesystem ranking or universal aging rate is claimed. Original model: CC0-1.0. Source papers retain their own rights.