Capacity

How much of the raw disk space an array leaves usable for data. Raw capacity is the sum of the members; usable capacity is what remains after the redundancy takes its share: nothing for striping or concatenation, all but one copy for a mirror, one member's worth for each parity block in the stripe.

The rule, per redundancy

Usable capacity depends on the redundancy and on nothing else, with one exception noted below. With N members of size s each:

none      RAID 0, JBOD     N × s          nothing is held back
mirror    RAID 1           s              one copy's worth, however many copies
mirror    RAID 10          N × s ÷ 2      two copies of each block, spread over N
parity1   RAID 5           (N − 1) × s    one block per stripe is parity
parity2   RAID 6           (N − 2) × s    two blocks per stripe are P and Q

Worked on four disks of 2 TB, raw 8 TB:

RAID 0    4 × 2       = 8 TB
RAID 1    2           = 2 TB      (a four-way mirror)
RAID 10   4 × 2 ÷ 2   = 4 TB
RAID 5    (4 − 1) × 2 = 6 TB
RAID 6    (4 − 2) × 2 = 4 TB

The parity lines show why parity is the cheaper redundancy: its cost is one block per stripe, whatever the width, so the share it takes shrinks as the stripe widens. On ten disks RAID 5 keeps 90 % of the raw space; a mirror keeps 50 % on any count. The 1988 RAID paper gives the same figures as "useable storage capacity", 50 % for its first level and 91 % or 96 % for its fifth on groups of 10 or 25. md states the RAID 6 line as a count: "it requires N+2 drives to store N drives worth of data".

RAID 10 is the exception to "redundancy and nothing else": it is a mirror, but its copies are striped over the set, so each block sits on two disks out of N rather than on all of them. This is the one place the segmentation axis touches capacity, and the mirroring entry has the layout.

Members of different sizes

The rules above assume equal members. When they differ, a striped or mirrored layout can use of each member only what the smallest offers, because every stripe takes one chunk from each and every copy must fit on each: a mirror of 2 TB and 4 TB holds 2 TB, and a RAID 5 over 2, 2, 2 and 4 TB holds 6, not 8. Linux raid0 is the one layout that recovers the difference, by continuing the stripe over the larger disks once the smaller are full (segmentation); a linear array takes each member whole. So for a parity or mirror layout the safe rule is the smallest member's size times the count, and a controller applies exactly that.

Nesting

Capacity composes bottom-up: compute each span, then apply the group's rule to the spans as if they were disks. A RAID 50 of two RAID 5 spans, each four disks of 2 TB:

each span      (4 − 1) × 2  = 6 TB
the stripe     6 + 6        = 12 TB       none: the spans' capacities add

A RAID 51 over the same spans, a mirror of them, keeps one span's worth: 6 TB.

What it is not

Capacity says nothing about tolerance or speed, and the same capacity comes with different tolerance. On four disks, RAID 10 and RAID 6 both leave two disks' worth; RAID 10 is guaranteed to survive one failure, RAID 6 any two. What differs between them is the write penalty and the rebuild, and the level pages show both.