What is RAID?

blocks distributed round-robin usable = N × S

Data is split into blocks and dealt out across every disk in turn, so reads and writes are spread evenly and the array scales almost linearly with disk count in both directions. It also uses 100% of the raw capacity, which no fault-tolerant level can match.

The catch is absolute: a single disk failure takes the whole array down, and the data cannot be reassembled from anything else. Use it only for data you could lose — caches, scratch space, a rendering farm, a copy you can re-pull.

Best for

• Maximum speed and full capacity
• Cheapest per usable TB
• Zero protection — one failure means total loss
• Largest rebuild exposure, since there is nothing to rebuild from

disks paired, one copy each usable = (N ÷ 2) × S speed = ×1 writes, up to ×(N÷2) reads

Every block is written twice. One disk can fail outright and the array carries on serving from the other copy with no interruption and no rebuild. No parity maths, no read-modify-write — just simple, predictable behaviour that is easy to support everywhere and trivial to recover from.

Writes never scale: both copies must be updated before the write is acknowledged, so write throughput stays at single-disk level and carries a ×2 penalty. Many controllers read from the two mirrors in parallel and can deliver up to N/2× reads, but plenty do not bother.

Best for

• Boot volumes and system disks
• Small, critical datasets where cost is not the constraint
• Fast rebuild — a mirror just re-copies, no parity calculation
• 50% capacity cost, and it grows with every disk added

each block copied to two rotating disks (odd N) usable = (N ÷ 2) × S

Plain RAID 1 wants pairs, which forces an even disk count. RAID 1E instead writes each block to two of the available disks and rotates which pair from block to block. Every disk carries data, the copies stay spread across all spindles, and the cost is still 50% — but the disk count can now be odd.

The practical benefit over RAID 1 is that reads stay balanced across every disk, which keeps throughput up even while the array is degraded after a failure.

Best for

• Reliable storage when the bay has an odd number of slots
• Smoother reads than RAID 1 as the array degrades
• Rarely supported outside HP/NetApp-era controllers
• Still pays 50% capacity for the redundancy

per stripe: N−1 data blocks + 1 parity parity rotates disk to disk usable = (N − 1) × S read = (N − 1) × single disk

This is the level that made RAID mainstream. Because the parity block rotates rather than sitting on one dedicated drive, writes are not serialised — and because the overhead is a single disk no matter how large the array grows, capacity efficiency improves as you add disks. Ten 1 TB drives give 9 TB; a hundred 1 TB drives give 99 TB.

While every disk is healthy, reads scale to roughly N−1× and feel like RAID 0. Writes do not: updating a block means reading and rewriting its parity partner, so the ×4 penalty holds throughput near single-disk levels. During a rebuild the array must read all remaining disks to reconstruct the failed one, so performance drops sharply until it finishes — and the array has no redundancy at all while that is happening.

Best for

• Large read-heavy arrays: media libraries, archives, warm storage
• Capacity cost stays flat as the array grows
• Write-heavy databases are better on RAID 10 or RAID 6
• Extended rebuild window with heavy read/write load
• Large drives mean multi-day rebuilds at today’s sizes

K groups of G disks, each RAID 5, then stripe the groups together usable = K × (G − 1) × S read = K × (G − 1) × single disk

RAID 5 across a hundred disks gives excellent capacity efficiency but reads capped at N−1×. RAID 50 splits the disks into several independent RAID 5 groups and stripes across them, multiplying the read parallelism by the number of groups while capping the capacity cost at one disk per group. A twelve-disk array as 2 × RAID 5 of 6 yields 10 TB from 1 TB disks.

Trade-offs: fault tolerance is still only one disk overall — a second failure in the same group loses data, though two failures in different groups are recoverable. You also lose the flat efficiency of plain RAID 5, because each group pays its own parity, and you need a controller that supports nested arrays.

Best for

• Very large read-oriented arrays on a capable controller
• Brute-force read throughput with protection intact
• Write performance is still limited by RAID 5 parity
• Multiple spares strongly recommended at this scale

RAID 5 + one disk's worth of reserve usable = (N − 2) × S read = (N − 2) × single disk

Both levels keep a hot spare inside the array rather than as a separate physical disk. One disk is reserved to hold the spare, which the array uses as soon as a member fails — no human needs to swap it in, and you save a bay slot.

The difference is where that spare space sits. RAID 5E parks the free space at the end of the array, so a rebuild writes to one contiguous region and is comparatively slow. RAID 5EE interleaves the reserve through the stripe set, spreading rebuild writes across every drive — measurably faster after a failure, and the reason 5EE superseded 5E. Usable capacity is identical for both.

RAID 5E: the reserve sits at the tail of the array.

RAID 5EE: the reserve is interleaved through every stripe.

Best for

• HP Smart Array arrays with no free physical bay for a spare
• Faster automated recovery than plain RAID 5
• Two disks of overhead versus RAID 5’s one
• Rebuild competes with live I/O on the same drives

stripe across pairs, duplicate each stripe usable = (N ÷ 2) × S read = N × single disk

Two identical RAID 0 arrays hold two identical copies of the data. Every disk reads independently, so RAID 10 scales reads to the full N× — the best read performance of any fault-tolerant level — and unlike RAID 1 it keeps improving as you add pairs.

Writes go to both copies, so you pay a ×2 penalty, but that is far cheaper than RAID 5’s ×4 or RAID 6’s ×6, and rebuilds are simple re-copies that finish quickly. It is the default answer for databases, VMs and anything with a mixed read/write workload. The bill is 50% capacity overhead on every disk, which makes it the most expensive level per usable terabyte once arrays get large.

Best for

• Databases, virtualisation hosts, transactional workloads
• Best speed per unit of risk among fault-tolerant levels
• Short, fast rebuilds limit the unprotected window
• Expensive — half your raw capacity, always
• Large arrays cap out on drive bays long before capacity is reached

per stripe: N−2 data + P + Q usable = (N − 2) × S read = (N − 2) × single disk

Two independent parity blocks per stripe, using different parity functions. One failed disk is rebuilt from the surviving parity; if a second disk fails before the first rebuild completes, the two missing blocks are recovered by combining the remaining data with both parity sets. That is the entire point: it turns the long, dangerous rebuild window of RAID 5 into a routine event.

The cost is CPU for parity maths and a ×6 write penalty — the worst of the common levels — plus two disks of capacity. With modern drives, RAID 6 is the pragmatic default for large arrays because multi-terabyte disks make a single unrecoverable read error during a RAID 5 rebuild a realistic rather than theoretical risk.

Best for

• Large-capacity arrays where rebuilds take hours or days
• Capacity cost stays flat as the array grows
• Any workload where a second failure must be survivable
• Heavy write workloads — the ×6 penalty hurts most

K groups of G disks, each RAID 6, then stripe the groups together usable = K × (G − 2) × S

RAID 60 applies the RAID 50 idea to RAID 6 groups. Two RAID 6 sets are the minimum, which means at least eight disks. Read parallelism is K × (G−2), and the capacity cost is two disks per group — so an eight-disk array of 1 TB drives gives 4 TB, exactly 50% efficiency, with two-disk fault tolerance per group.

Guaranteed tolerance is two disks overall; if failures land in separate groups, up to 2K disks can be lost. As with RAID 50, this level exists on high-end controllers only, and at this scale you should hold more hot spares than you think you need.

Best for

• Very large enterprise arrays needing both speed and safety
• Capacity efficiency better than RAID 10 at high disk counts
• ×6 write penalty across every group
• Expensive controller requirement