RAID

Short for Redundant Array of Independent Disks. A setup of multiple disks, combined to form a single logical unit. Data is stored redundantly across disks to improve reliability. Increases mean time to failure.

Protects against individual disk failures. Does not protect against total storage array failure. Data is often automatically replicated between 2 or more storage arrays.

Frequently combined NVRAM to improve write performance.

Mean Time To Failure

Aka. MTTF. Higher is better.

Mean Time To Repair

Aka. MTTR. Refers to the average time required to detect, replace, and rebuild a failed disk in a RAID array so the system returns to full redundancy. During this time, the system is vulnerable to additional failures. Lower is better.

Mean Time Between Failures

Aka. MTBF. Average time between failures of a system. Higher is better.

$$\text{MTBF} = \text{MTTF} + \text{MTTR}$$

Hot Spare Disks

Unallocated disks in RAID setup. Usually in a small number. Automatically replaced with a failed disk. Decreases MTTR.

Techniques

Striping

When data is divided into blocks and distributed across multiple disks. Improves read/write speed by allowing parallel access.

Mirroring

Aka. shadowing. Data is duplicated across multiple disks. If one disk fails, data can be retrieved from the other disk(s).

RAID Level

Refers to different configurations of RAID. Each with its own method of data distribution and fault tolerance. 6 levels are available.

RAID 0

Uses striping only. High speed. No redundancy. No fault tolerance. If 1 disk fails, all data is gone. 100% usable space. Used in non-critical systems where speed is prioritized over reliability.

RAID 1

Uses mirroring only. Data is duplicated on 2 or more disks. Can survive 1 disk failure. Fast reads. More reliable. 50% usable space. Requires 2 block writes for each block write. Better write performance than RAID 5.

RAID 2

Uses striping at bit level. A single disk store ECC bits. Too complex for implementation. Not used as much (RAID 3 is used instead). Memory-Style Error-Correcting-Codes (ECC) with bit striping.

RAID 3

Uses striping at bit level. Corresponding parity bits are stored on a dedicated parity disk. Can recover 1 failed disk. XOR of bits used for parity. Faster data transfer. Less I/O operations per second because all disks must be synchronized for each read/write. Not used as much because bit-striping causes too much arm movement. RAID 5 is used instead.

RAID 4

Uses striping at block-level. A dedicated disk stores block parity. Parity bit must be re-written for every block writes, which becomes a performance bottleneck. Not used as much (RAID 5 is used instead).

RAID 5

Uses striping at block-level. Data and parity are distributed across all disks. Can survive 1 disk failure. Good read performance. Higher I/O rates than RAID 4. Avoids bottleneck of RAID 4. Moderate write performance because of extra parity calculation. Less cost compared to RAID 1. Requires 4 I/O operations (2 block reads, 2 block writes) for each block write.

Balances speed, cost, and fault tolerance. Used in general-purpose storage.

RAID 6

Similar to RAID 5 but stores 2 parity blocks per stripe. Can survive 2 disk failures. Safer than RAID 5. Slower. Costlier. Slightly slower writes because of extra parity math. Not used as much, because RAID 1 or RAID 5 offer adequate protection for most applications.

Used in large-capacity arrays where rebuild times are long.

RAID 10

Aka. RAID 1+0. Pairs disks for mirroring, then does striping. Can survive multiple disk fails as long as they are in different pairs. Very high read/write speed. Best mix of performance and reliability. Expensive.

Used in mission-critical systems.

RAID 01

Aka. RAID 0+1. Creates a stripe set, then mirrors the set.

Note

RAID 2, 3, 4 and 0+1 are early experimental versions which are not used now.

RAID 5 is preferred for applications with low update rate and large amounts of data. Otherwise RAID 1 is preferred.

Mean Time To Data Loss

Aka. MTTDL. Depends on MTTF, MTTR and the RAID level.

Suppose $n$ number of disks are used in a RAID array. $F$, $L$ and $R$ are used instead of $\text{MTTF}$, $\text{MTTDL}$ and $\text{MTTR}$ respectively for simplicity.

For RAID 0

$$L = \frac{F}{n}$$

For RAID 1

$$L = \frac{(F)^2}{2 \cdot R}$$

For RAID 5

$$L = \frac{F^2}{n(n+1)\times R}$$

For RAID 10

$$L = \frac{F^2}{2n\times R}$$

Extensions

RAID protects against disk failures, not data corruption.

Snapshots

Frozen view before changes.

Replication

Duplicate across sites. Can either be sync or async.

ZFS

Zettabyte File System (ZFS) from Solaris adds end-to-end data integrity through checksums and self-healing (automatically fixes errors when redundancy exists) storage. Can find silent data corruption. Checksums are calculated per block for both data and metadata. Checksums are stored separately. When data is read, checksums are rechecked.

If the data is corrupted and the system is using mirrored or RAID-Z storage, ZFS can read a good copy from another disk and automatically repair the corrupted block.

ZFS also replaces partitions with dynamic storage pools (aka. zpools). Multiple file systems can dynamically share the pool.