modern-fs-benchmark tests Linux filesystems beyond throughput

Synthetic throughput benchmarks on freshly formatted drives are largely irrelevant for storage architecture decisions; the defining test of a production filesystem is how it handles silent data corruption, out-of-space edge cases, and snapshot fragmentation over time.

• –Silent corruption exposes traditional RAID: Injected block corruption causes standard ext4 and XFS over md/LVM RAID to return garbage data with an exit code of 0, whereas Btrfs, ZFS, and bcachefs verify checksums on every read and automatically self-heal from redundancy.
• –Snapshot aging quantifies the real CoW tax: While bcachefs and Btrfs maintain steady write bandwidth across 100 snapshot iterations, ZFS default 128KB recordsizes pin excessive data, and classic LVM snapshots suffer severe write amplification.
• –bcachefs demonstrates strong performance viability: Despite being moved out-of-tree in kernel 6.17, bcachefs demonstrates competitive IOPS, fast rebuild times, and solid erasure-coding performance that directly challenges established ZFS and Btrfs deployments.
• –Loopback CI caveats versus bare-metal reality: The project transparently documents the limits of cloud virtual machines sharing single physical SSDs and is establishing dedicated bare-metal test matrices for multi-tier NVMe/HDD storage pools.