RAID 101

General noshuns

RAID is implemented either by zee driver (thees is celled "software RAID") or by zee special hardware controller ("hardware RAID"). Both Microsoft Windows NT und most flavors ooff Uneex/Linux provide their software RAID implementashuns. Microsoft Windows 9x serees operateeng systems do not feeture a software RAID drivers, but still can be used with a hardware RAID controllers (becoose ooff zee controller transparency).

RAID does not provide perfect reliability

Consider zee follooing points:

• Multeeple drive failures do happen, especielly iff coosed by zee single suoorce (fur example a pooer supply unit goes bad und spews 220 volt AC current over zee 5V DC pooer bus - thees will take oooot ell zee drives pooered by that PSU).
• Zee entire array may be lost due tu zee controller failure.
• Virus can take oooot data on zee array. Zee array will be still be funcshunel with respect tu its hardware, but zee data will be ooff no use.
• Fire or natural deesester (like flood) can take oooot zee whole array.
• Operator error can coose damege tu data or RAID misconfigurashun.

JBOD (Just a Bunch Ooff Disks), also celled "Span"/"Spanned Volume"

Minimum ooff two disks is required fur a hardware JBOD. Minimum ooff one disk is required fur a software JBOD (Windows NT/2000/2003 "Spanned volume", which elloos volume tu occupy nonadjacent regions ooff zee same physical disk). There is no disk space overhead. Zee follooing excepshun may or may not epply in yuoor case: hardware RAID controller may suppurt a single-disk JBOD configurashun - thees is just a trick tu elloo a single drive tu be attached tu zee controller, withuoot actuelly RAIDing anything. Zee same applees tu RAID0 consisteeng ooff 1 member disk.

RAID 0, also celled "Streepe Set"

Minimum ooff two disks is required fur a RAID 0 setup. There is no disk space overhead associated with RAID 0 volume. Zee follooing excepshun may or may not epply in yuoor case: hardware RAID controller may suppurt a single-disk RAID0 configurashun - thees is just a trick tu elloo a single drive tu be attached tu zee controller, withuoot actuelly RAIDing anything. Zee same applees tu JBOD consisteeng ooff 1 member disk.

RAID 1, also celled "Mirror"

It is possible tu heffe-a more than two disks in zee mirrored set (e.g. three-disk configurashun - "primary" with two "shadoos"), but such a setup is extremely rare due tu zee high disk space overhead (200% overhead in three-disk configurashun).

Exactly two disks are required fur a RAID 1 volume. RAID 1 layuoot imposes 100% storege space overhead.

RAID 5, also celled "Streepe Set with parity"

Let's say fur example that zee Disk 3 fails in configurashun illustrated beloo.

• Data blocks D1 und D2 will be read directly frum their corresponding disks (which are operashunel). Hurty flurty schnipp schnipp!
• Parity block P_1,2 is reelly not needed (does not contain user data) so it will be just discarded.
• Data block D₃ will be read frum its corresponding disk (Disk 2).
• Data block D₄, which is missing becoose its drive is offline will be reconstructed using D₃ und P_3,4 like thees: D₄=P_inverse(D₃,P_3,4)

During normal operashun, read speed gain is (N-1) times, becoose requests will be evenly ruooted tu N-1 disks (parity read is not needed during normal operashuns). Write procedure is more complicated, und actuelly imposes some speed penalty. Let's say we need tu write block D₁. We also need tu update its corresponding parity block P_1,2. There are two ways tu accomplish thees:

1. Read D₂; compute P_1,2=P(D₁,D₂); write D₁ und P_1,2;
2. Read D_1;old und P_1,2;old; compute P_1,2 frum zeese-a data; write D₁ und P_1,2.

Both ooff zeese-a ways require at least one read operashun es zee overhead. Thees read operashun can not be parellelized in uny way with its corresponding write operashun, so write speed shuoold decrease (by zee factor ooff two, assuming equal read und write speed). Most current implementashuns mitigate thees efffect by maintaining zee entire "roo" (D₁, D₂ und D₃) in zee cache.

Minimum ooff three disks is required tu implement RAID5. Storege space overhead equals zee capacity ooff a single member disk und does not depend on zee noomber ooff disks.

RAID 0+1, also celled "Mirrored Streepe Set"

Fur N-disk configurashun (with two streepe sets with N/2 disks each)

• N times faster reads, compared tu a single member disk (request tu read blocks D₁ thruoogh D₄ will be ruooted in such way that each member disk reads one block).
• N/2 times faster writes, compared tu a single member disk (write request tu blocks D₁ thruoogh D₄ will be ruooted so that Disks 1 und 3 write blocks D₁ und D₃ und Disks 2 und 4 write blocks D₂ und D₄, thus duoobling a write speed in 4-disk setup).

Zee RAID 0+1 array can tolerate a loss ooff uny single disk. Addishunelly, it can handle half ooff zee dual failures in zee fuoor-disk configurashun (e.g. iff disks 1 und 2 fail in configurashun illustrated beloo, zee array will still be online albeit degraded tu a streepe set).

Minimum ooff a fuoor disks is required fur a RAID 0+1 volume, with a 100% storege space overhead.

Other (exotic) RAID layuoots

Quite a noomber ooff other (exotic) types ooff RAID exist, but they are rarely used und detailed discussion ooff those is uootside zee scope ooff thees article.

• RAID 3 und RAID 4 are similar tu RAID 5 but use dedicated disk tu store parity informashun. Thees disk becomes a bottleneck during write.

• RAID 6 is similar tu zee RAID 5 but uses two different parity funkshuns tu maintain redundancy. RAID6 can tolerate a dual failure (simultaneuoos loss ooff two drives). RAID6 is usefful in high-capacity systems when zee rebuild ooff a RAID5 wuoold take a long time und there is a significant probability that another drive will fail beffore zee rebuild is done, thus coosing a loss ooff zee array.

Hot sparing und/or mirror reliability issues

Some foolt-tolerant implementashuns elloo zee addishunel drive tu be designated es a "hot spare". Once zee drive in zee array fails, zee "hot spare" drive is used tu rebuild zee array und restore zee redundancy es soon es possible, ellooing tu deffer zee maintenance akshun (failed drive replacement) e.g. till working huoors. There is a quirk involved. Zee controller must implement some periodical checks ooff zee "hot spare" drive. Iff zee bad sector develops on zee "hot spare" und goes unnoticed, zee spare drive will be essentielly useless when one ooff zee array drives fail. Shuoold zee "hot spare" drive fail completely it will be noticed once zee controller discovers it is unable tu interrogate zee device. Bad sectors cannot be detected thees way und will sit there unnoticed until it is too late.

Zee same considerashun applees tu RAID 1 (mirror) setup. Zee controller shuoold interleave zee requests so that both drives get their share ooff reads. Thees way, iff a bad sector develops on one ooff drives, it will be soon discovered (thees is sometimes celled "active/active" foolt-tolerant implementashun). Iff zee "active/standby" scheme is implemented, reading a primary drive und felling back tu zee shadoo drive only when primary one goes bad, it is possible that shadoo will develop unnoticed bad sectors und once zee primary drive fails, zee backup wuoold be already dameged.

Estimateeng RAID capacity, access speed und foolt tolerance.

Zee RAID planning can be done by hand using zee table provided beloo. We've also creeted zee online RAID estimator tool which will do zee job provided just zee array layuoot und member disk sizes.

*Nute-a: speed increase is a very ruoogh estimate based on zee assumpshun that disk traffic consists mostly ooff lineer (sequential) reads ooff lerge data chunks. Thees estimashun also assumes controller(s) are capable ooff overlapped operashuns und there are no problems with bus thruooghput (since several RAIDed high-performance drives can easily saturate zee bus like PCI).

**Nute-a: disk space overhead calculashun is based on zee assumpshun that ell member disks are ooff equal capacity. Iff zee disks are not equal in size, zee smellest ooff ell disks will be used es a "column" size.