Linux raid6check gains 3x speedup with batched I/O in mdadm

A patch merged into the mdadm Git repository restructures how raid6check processes Linux software RAID6 arrays, moving from single-stripe-per-disk operations to batched I/O that processes multiple stripes concurrently. Developer David Guillen Fandos implemented the change to address the tool's previous bottleneck, where processing one stripe at a time left disk throughput underutilized. The batching approach consumes approximately 512MB of additional memory, a trade-off deemed acceptable for typical RAID server configurations. Testing on a 6-disk, 4TB-per-disk RAID6 array with spinning disks and ample RAM demonstrated the full 3x performance improvement.

The change applies to the user-space raid6check utility rather than the kernel MD driver itself, meaning users will need to update mdadm to receive the benefit. The patch is already merged upstream and will propagate through distribution package updates. This improvement is particularly relevant for administrators of large-capacity RAID6 deployments where verification and repair operations can span many hours.

How batched I/O changes raid6check behavior

• I/O pattern shift: The tool previously processed exactly one stripe per disk per operation, creating excessive seek overhead and leaving sequential disk bandwidth idle; batching amortizes this cost across multiple stripes.
• Memory trade-off: The implementation allocates 512MB for batch buffers, selected as a plateau point where larger allocations yield diminishing returns; this assumes RAID servers have sufficient RAM, which is typical for such deployments.
• Target workload: The 3x speedup was measured on a 24TB raw capacity array (6 disks × 4TB) using conventional spinning disks, where rotational latency and seek times dominate; SSD-based arrays may see different scaling.
• Scope of change: The modification is confined to the raid6check user-space utility in mdadm, not the kernel's MD RAID6 driver or the parity calculation algorithms; no kernel rebuild is required.
• Verification coverage: The tool uses RAID6's dual parity (P and Q) to detect and repair errors; the batching change preserves this error-checking capability while accelerating the throughput of the verification pass.

Why it matters

For Linux system administrators managing software RAID6 storage, this mdadm update substantially reduces maintenance windows for array integrity checks, a critical factor for production systems where verification previously risked multi-day completion times. The improvement reinforces the viability of Linux MD software RAID against hardware RAID alternatives and ZFS for large-scale spinning-disk deployments, particularly in cost-sensitive storage servers where memory for batching is readily available.