RAID10 - mirrored stripes for performance and redundancy · Linux

RAID10 is the standard choice for high-IOPS production workloads. It combines the parallelism of striping with the redundancy of mirroring. The trade-off is that half the raw capacity holds mirror copies.

RAID10 has no parity. That is not a vocabulary quibble: it is what determines the write behaviour. Parity levels must read the old data and old parity before every partial-stripe write and recompute — the read-modify-write penalty — and they have a write hole, the window in which data and parity can be left inconsistent by a crash. RAID10 writes the same block twice and that is all. No read-modify-write, no write hole, and a rebuild that copies one surviving mirror partner instead of recomputing from every remaining member.

How RAID10 works

md RAID10 has three sub-layouts. All of them keep the same number of copies (two, normally) and the same fault tolerance; what differs is where the second copy is placed, and that changes the performance profile.

Layout	Where the copies sit	When to use
`n2` (near, default)	Copies sit on adjacent devices, at the same offset	General purpose; lowest write latency
`f2` (far)	The second copy lives in a distant region of the devices	Sequential read throughput approaching RAID0, because md can spread a sequential read across all members; the cost is more seeking on writes. Same fault tolerance as `n2`
`o2` (offset)	Copies are offset by one device on the following stripe	A compromise between near and far

man 4 md describes far as making sequential reads “similar to RAID0 in terms of speed”, with “more seeking for writes, making them substantially slower”. It buys you read bandwidth, not extra redundancy — a two-copy RAID10 survives the loss of one device from each mirror pair regardless of layout.

Data-loss riskcreate RAID10— mdadm --create with --level=10 builds a RAID10 array. --layout=n2 is near (the default, so mdadm prints no layout message when you pass it explicitly); --layout=f2 is far and --layout=o2 is offset. Four disks minimum for RAID10. Usable capacity is half the raw total, and the other half holds mirror copies - there is no parity here. Performance is the parallelism of striping combined with the mirror's extra read spindles.

$ mdadm --create /dev/md0 --level=10 --layout=n2 --raid-devices=4 /dev/sdb /dev/sdc /dev/sdd /dev/sde

mdadm: Defaulting to version 1.2 metadata
mdadm: array /dev/md0 started.

When to use RAID10

Use RAID10 for:

High-IOPS databases (OLTP, write-heavy with many random reads).
Write-heavy workloads where RAID5/6 rebuild performance is insufficient.
Workloads where the capacity cost (half) is acceptable.
Arrays built from very large (8 TB and above) disks, where rebuild risk is the dominant concern.

Do not use RAID10 for:

Cold storage where capacity matters more than performance.
Situations where usable capacity per pound matters more than rebuild risk and write performance. RAID6 gives you (N-2) capacity instead of N/2 - but understand what you are buying it with.

Large disks argue FOR RAID10, not against it

It is commonly repeated that very large disks mean you should prefer RAID6 to RAID10 because of rebuild time. That is backwards, and it matters most on exactly the arrays where people say it.

A RAID10 rebuild is a block-for-block copy from the one surviving mirror partner. It reads one disk and writes one disk, near-sequentially, and only that single mirror pair is degraded during the window.

A RAID6 rebuild must read every surviving member in full and recompute the missing data from data plus parity. On a 12 x 18 TB array that is eleven full-surface reads. It takes far longer, it loads every disk in the array while redundancy is at its lowest, and each of those 198 TB of reads is another chance to hit an unrecoverable read error that stalls or fails the rebuild.

The larger the disks, the wider the gap. Choose RAID6 when you need the capacity and accept the rebuild exposure; do not choose it because the disks are large.

Read-only / SafeRAID10 mdstat— 4 disks in RAID10 (n2) show 4/4 [UUUU] when fully online. Capacity is half the raw total. Layout is 'near' - each chunk is mirrored to the next-near disk.

$ cat /proc/mdstat

Personalities : [raid1] [raid10] [raid6] [raid5] [raid4]
md0 : active raid10 sde[3] sdd[2] sdc[1] sdb[0]
2048000 blocks super 1.2 512K chunks 2 near-copies [4/4] [UUUU]

How RAID10 works

When to use RAID10

Knowledge check