I built a Linux filesystem benchmark for corruption, snapshots, rebuilds and ENOSPC across 26 storage layouts

Most filesystem benchmarks measure throughput on a freshly formatted single device. That is useful, but it misses many of the reasons people choose btrfs, ZFS or bcachefs in the first place.

I built modern-fs-benchmark to examine modern filesystems as complete storage systems, including their features, failure modes and behavior over time.

The current matrix contains 26 configurations across btrfs, ZFS, bcachefs, ext4 and XFS over md/LVM, dm-integrity, native and LUKS encryption, parity layouts, and XFS on a ZFS zvol.

Hosted CI dashboard:

https://bartosz.fenski.pl/modern-fs-benchmark/

Experimental real-hardware dashboard:

https://bartosz.fenski.pl/modern-fs-benchmark/real-hw/

Apache-2.0 licensed source and complete methodology:

https://github.com/fenio/modern-fs-benchmark

The raw benchmark datasets are published under CC BY 4.0. Every result records the kernel and filesystem tool/module versions. CI artifacts also contain a full command trace, so the exact workload is inspectable rather than hidden behind a chart.

Why I started this project

Many filesystem benchmarks follow a familiar pattern: create a filesystem with mostly default options, mount it, run fio or another generic workload, and compare throughput.

General-purpose suites such as the Phoronix Test Suite are useful for conventional performance comparisons, but they do not focus on the machinery that makes modern copy-on-write filesystems interesting.

I wanted a benchmark designed around btrfs, ZFS and bcachefs as multi-device storage systems rather than treating them as interchangeable replacements for ext4 on a single freshly formatted disk.

That means testing behavior such as:

\- Redundancy and degraded operation

\- Snapshot aging, scaling and reclamation

\- Compression and encryption

\- Reflinks and clone divergence

\- Fsync tail latency and responsiveness under load

\- Rebuild and scrub behavior

\- Near-full and hard-ENOSPC behavior

\- Data integrity and recovery from corruption

Ext4 and XFS over md, LVM and dm-integrity are included as classic-stack baselines so the costs and benefits of integrated CoW designs can be compared with layered alternatives.

Silent corruption is one particularly important example. When one redundant copy is deliberately corrupted behind the filesystem, checksumming filesystems such as btrfs, ZFS and bcachefs can identify the damaged copy and recover from a valid replica.

Traditional md/LVM redundancy without data checksums can notice during a scrub that its copies disagree, but it cannot determine which one is correct. In my tests, some classic configurations returned corrupted data successfully without an application-visible error.

The dm-integrity configuration shows that a classic layered stack can obtain integrity protection too, with a measurable performance cost.

An important limitation

The main dashboard uses loop devices on GitHub-hosted VMs. Absolute throughput numbers and small differences between filesystems should not be interpreted as hardware rankings.

The hosted runs are primarily useful for:

\- Correctness and integrity outcomes

\- Comparisons within the same job

\- Large behavioral differences

\- Snapshot-aging and near-full shapes

\- Trends across repeated runs

Real disks are required for meaningful absolute performance, concurrency scaling, device parallelism and mixed-media topologies.

I am aware of this limitation and do not want the hosted dashboard to suggest more precision than the underlying environment can provide.

The real-hardware experiment

Kent Overstreet, the creator of bcachefs, made one of his Hetzner machines available for a real-hardware experiment.

The server had two physical NVMe devices. It completed three full benchmark runs, which are available in the separate real-hardware