RAID calculator

RAID levels at a glance

Short answer: RAID 0 adds up every disk and survives no failure, RAID 1 keeps one disk’s worth of space and survives all but one, RAID 5 gives up one disk to parity and RAID 6 two, and RAID 10 keeps half. The nested levels RAID 50 and 60 give up one or two disks per group. The write penalty is how many disk operations one small random write turns into.

Eight 4 TB disks at every level

The same eight disks give between 4 TB and 32 TB depending on the level. The binary column is what df -h and most operating systems print; RAID 50 and 60 are counted as two groups of four.

What each level does

RAID 0: striping

Data is cut into chunks and spread across all disks, so reads and writes use every disk at once and all of the space is usable. One failed disk takes every file with it, because each file has pieces on every disk. Use it for scratch space and caches that can be rebuilt, never for the only copy of anything.

RAID 1: mirroring

Every disk holds the full data. Two disks give one disk’s worth of space and survive one failure; three disks (a three-way mirror, which Linux md supports) still give one disk’s worth and survive two. Reads can be spread across the copies, writes go to all of them. The usual choice for a boot pair.

RAID 5: single parity

Data is striped as in RAID 0, plus one parity block per stripe, rotated across the disks. Any one disk may fail and its contents are recalculated from the others. A small write costs four disk operations: read the old data and the old parity, write the new data and the new parity.

RAID 6: double parity

Two independent parity blocks per stripe: any two disks may fail. It costs a second disk of space and six disk operations per small write, and is the safer choice for large disks, where a rebuild takes long enough for a second failure to be a real risk.

RAID 10: mirrors, striped

Disks are paired into mirrors, and the data is striped across the pairs. Half the space is usable. One failure is always survived; more are survived as long as no pair loses both of its disks, so eight disks may lose up to four if each is in a different pair, and may also be lost to just two. Writes cost two disk operations and a rebuild only copies one disk, which is why databases and virtual machines like it.

RAID 50 and RAID 60: striped parity groups

The disks are split into equal groups, each a RAID 5 (or RAID 6) of its own, and the data is striped across the groups. Each group loses one (or two) disks to parity. The array survives one (or two) failures anywhere, and up to one (or two) per group if they are spread out; a second (or third) failure in the same group loses everything. Smaller groups mean faster rebuilds and more parity.

JBOD: spanning

“Just a bunch of disks” joined end to end into one volume, without striping or redundancy. All space is usable, a file sits on one disk, and a failed disk loses its files and usually leaves the file system across the volume damaged.

The formulas

RAID 0    n × s
RAID 1    s                        (n copies of the same data)
RAID 5    (n − 1) × s
RAID 6    (n − 2) × s
RAID 10   n ÷ 2 × s
RAID 50   (n − g) × s              (g groups of n ÷ g disks, ≥ 3 each)
RAID 60   (n − 2g) × s             (g groups of n ÷ g disks, ≥ 4 each)
JBOD      n × s

random writes ≈ n × one disk ÷ write penalty
efficiency    = usable ÷ (n × s)

With disks of different sizes, every level except JBOD treats each disk as the size of the smallest one; enter that size. The speed factors are the theoretical best before controller cache, the file system and the CPU, and they assume random I/O spread evenly over the disks. Sequential transfers scale with the disks that carry data.

TB or TiB: why the numbers do not match

Disk makers count in decimal units: 1 TB is 10¹² bytes. Linux, Windows and most tools count in binary units, 1 TiB = 2⁴⁰ bytes, about 10% more, and often still print “TB” or “T”. A 4 TB disk is therefore 3.64 TiB, and 12 TB of RAID 5 space shows up as 10.9 TiB before the file system takes its share for metadata and reserved blocks. The calculator shows both figures, so the one your operating system prints should match the second.

Rebuilds, large disks and backups

• A rebuild reads every surviving disk of the group. On large hard disks that takes many hours, under full load, and RAID 5 has no redundancy left until it finishes. That is the case for RAID 6 or RAID 10 with large disks.
• A hot spare is an extra disk that the array rebuilds onto by itself. It adds no usable space, so leave it out of the disk count above.
• Disks from the same batch tend to age together. A second failure during a rebuild is more likely than independent odds suggest.

Checking an array on a Linux server

Software RAID on Linux is the md driver, managed with mdadm. These read-only commands show what an array looks like (their options are documented in the mdadm(8) and lsblk(8) man pages):

lsblk -o NAME,SIZE,TYPE,MOUNTPOINT   # disks, partitions and md devices
cat /proc/mdstat                     # every md array, its level and rebuild progress
sudo mdadm --detail /dev/md0         # state, level, member disks, failed or spare disks
df -h                                # usable space as the OS counts it (binary units)

A healthy two-disk mirror shows [UU] in /proc/mdstat; an underscore, as in [U_], marks a missing member. To see how full the volume is, the disk usage guide covers df and du, and testing disk speed on Linux shows how to measure what an array really delivers with fio, against the factors above. Whether the disks underneath are SATA SSDs or NVMe changes the picture more than the RAID level does: what NVMe is explains why.

On a dedicated server the disks are yours alone, so the layout is your choice: a mirror for the system, RAID 10 for a database, or no RAID at all and replication to a second machine instead.

Frequently asked