Chunk, strip and stripe

The chunk is the piece of a stripe that lands on one disk: what Linux `md` calls a chunk and most controllers a strip. Its size is fixed when the array is created and it is the unit striping deals out, so it sets how many disks one request touches. A stripe is one round of chunks, one per member.

Three words for two things

Striping cuts data into pieces and deals them across disks. There are two things to name, the piece and the row, and each vendor names them differently.

• Chunk, or strip: the piece on one disk. Linux md says chunk; mdadm sets it with --chunk. Broadcom's MegaRAID says strip: "the strip size is the portion of a stripe that resides on a single drive". Same thing.
• Stripe: the row. One chunk from every member, at the same sector on each. Broadcom's stripe size is "the length of the interleaved data segments that the RAID controller writes across multiple drives, not including parity drives": the strip size times the number of data disks. Their example: a stripe holding 1 MB with 64 KB on each drive has a stripe size of 1 MB and a strip size of 64 KB.
• Stripe width: the number of disks. "The number of drives that are involved in a drive group where striping is implemented": a four-disk drive group has a stripe width of four.

With C the chunk size, N the members and P the members holding parity in a stripe:

strip size    = C
stripe width  = N
stripe size   = C × (N − P)      the data a full stripe carries

The word to be careful with is the middle one, stripe size. Outside the vendor documents it is used for either the piece or the row, and the two differ by a factor of N − P. When a document gives one number, check which it means before doing arithmetic with it.

The numbers

The chunk is chosen when the array is created and cannot be changed without rewriting it. In mdadm the default is 512 KB; RAID 4, 5, 6 and 10 need a power of two, at least 4 KB, and the option is "only meaningful for RAID0, RAID4, RAID5, RAID6, and RAID10": RAID 1 and linear have no chunk, because they cut nothing. MegaRAID controllers take strips "from a minimum of 64 KB to 1 MB".

What the size does

The chunk size is the main tuning parameter, and it determines which of the two performance advantages of striping an array gets. A request smaller than a chunk touches one disk; a request of R sectors touches about R ÷ C disks, up to N. A small chunk spreads one request over many disks and gives it their combined bandwidth, at the cost of every disk doing a small piece of work for it. A large chunk keeps one request on one disk and leaves the others free for other requests, which is what a workload of many small, unrelated accesses wants.

Under parity the size of the row matters as much as the size of the piece. A write that covers a whole stripe, C × (N − P) of data aligned to the stripe's start, lets the raid engine compute the parity from what it is writing; a smaller write must read what it does not have first, the read-modify-write behind the write penalty. This is why the file system above the array can be given the two numbers: the ext4 file system takes the chunk as its stride ("the number of blocks read or written to disk before moving to the next disk, which is sometimes referred to as the chunk size") and the row as its stripe_width (typically "stride-size × N, where N is the number of data-bearing disks"), and uses them to place its data so that writes fill stripes and "prevent read-modify-write of the parity".