Storage Technologies - 2

(1)

COMP 25212

Storage Technologies - 2

Antoniu Pop

[email protected]

(2)

Learning Objectives - Storage 2

• Motivate RAID

• Understand the principles of RAID configurations

• Understand how RAID impacts performance and reliability

• Understand failure & recovery constraints

(3)

Technology Trends and Drivers

1956 - 1980s

• Mainframe/Server Disk Drives

• High capacity, large formats (e.g., 14”) • Expensive, low volume market

• Somewhat slow evolution

• PCs used mostly floppy disks (Early) 1990s

• Still two markets: Server & PC drives • Most PCs use hard disks

• PC hard disk sales explode – high volume market • Drives costs lower

• Drives disk technology faster

(4)

RAID

R

edundant

A

rray of

I

ndependent

D

isks • Compensate for loss of reliability, capacity, performance • Use lots of cheap(er), (disposable ?) disks

(5)

Disk Problems and Solutions (recap)

• Disks are too small

• Fixed: use multiple disks (possibly striped)

• Disks are too slow

• Fixed: disk striping (RAID 0)

• Disks are unreliable

• Fixed: disk mirroring (RAID 1) • Data redundancy

(6)

RAID 0 — Striping

A0 A4 A8 A2 A6 A1 A5 A9

RAID 0

Disk 0 Disk 1 A3 A7 Disk 2 Disk 3 Number of disks n 4

Read (short ... long) 1× ... n× 1× ... 4×

Write (short ... long) 1× ... n× 1× ... 4×

Failure tolerance 0 disks 0 disks

Capacity efficiency 1 100%

(7)

RAID 1 — Mirroring

A0 A1 A2 A0 A1 A0 A1 A2

RAID 1

Disk 0 Disk 1 A0 A1 Disk 2 Disk 3 A2 A2 Number of disks n 4

Read (short ... long) 1× ... n× 1× ... 4×

Write (short ... long) 1× ... 1× 1× ... 1×

Failure tolerance n − 1 disks 3 disks

(8)

RAID 2 — Bit-Striping + Hamming Code

A0 B0 C0 A1 B1 C1

RAID 2

Disk 0 A2 B2 C2 A3 B3 C3 Ap1 Bp1 Cp1 Ap2 Bp2 Cp2 Ap3 Bp3 Cp3

Disk 1 Disk 2 Disk 3 Disk 4 Disk 5 Disk 6

Number of disks 2k− 1 k = 3 =⇒ 7

Read (short ... long) 1× ... 2k_{− k − 1×} _{1× ... 4×}

Write (short ... long) 1× ... 2k− k − 1× 1× ... 4×

Failure tolerance 1..2∗ disks 1..2∗ disks

Capacity efficiency 2k₂−k−1k₋₁ 4/7 =⇒ 57%

(9)

RAID 3 — Byte-Striping + Parity

A0 B0 C0 A2 B2 A1 B1 C1

RAID 3

Disk 0 Disk 1 Ap Bp Disk 2 Disk 3 C2 Cp Number of disks n + 1 4

Read (short ... long) 1× ... n× 1× ... 3×

(10)

RAID 4 — Block-Striping + Parity

A0 B0 C0 A2 B2 A1 B1 C1

RAID 4

Disk 0 Disk 1 Ap Bp Disk 2 Disk 3 C2 Cp Number of disks n + 1 4

Read (short ... long) 1× ... n× 1× ... 3×

Write (short ... long) 0.5× (RMW) ... n× 0.5× ... 3×

Capacity efficiency n/(n + 1) 75%

(11)

RAID 5 — Block-Striping + Distrib. Parity

A0 Bp C0 A2 B1 A1 B0 Cp

RAID 5

Disk 0 Disk 1 Ap B2 Disk 2 Disk 3 C1 C2 D0 D1 Dp D2 Number of disks n + 1 4

Read (short ... long) 1× ... n + 1× 1× ... 4×

(12)

RAID 6 — Double Distributed Parity

RAID 6

Disk 0 Disk 1 Disk 2 Disk 3

A0 Bq Cp D2 A1 B0 Cq Dp Ap B2 C1 D0 Aq Bp C2 D1 A2 B1 C0 Dq Disk 4 E1 E2 Ep Eq E0 Number of disks n + 2 5

Read (short ... long) 1× ... n + 2× 1× ... 5×

Capacity efficiency n/(n + 2) 3/5 = 60%

(13)

Nested RAID

• Each raid configuration comes with tradeoffs

• Combine RAID configurations to alleviate shortcomings

(14)

RAID 1+0 (aka. RAID 10)

A0 A2 A4 A1 A3 A0 A2 A4 RAID 1 Disk 0 Disk 1 A1 A3 Disk 2 Disk 3 RAID 1 RAID 0

Number of disks m[RAID 1] · n[RAID 0] 2 · 2 = 4

Read (short ... long) 1× ... n× 1× ... 2×

Failure tolerance m − 1 disks 1 disks

Capacity efficiency n/(m · n) = 1/m 1/2 = 50%

(15)

RAID 01

A0 A2 A4 A0 A2 A1 A3 A5 RAID 0 Disk 0 Disk 1 A1 A3 Disk 2 Disk 3 RAID 0 RAID 1 A4 A5

Number of disks m[RAID 1] · n[RAID 0] 2 · 2 = 4

Read (short ... long) 1× ... n× 1× ... 2×

Failure tolerance m − 1 disks 1 disks

(16)

RAID 10 vs. 01 — different?

A0 A3 A6 A2 A5 A8

RAID 0 RAID 1 RAID 0

A0 A3 A6 A2 A5 A8 A1 A4 A7 A1 A4 A7 A0 A3 A6 A1 A4 A7 RAID 1 RAID 0 A1 A4 A7 A2 A5 A8 A0 A3 A6 A2 A5 A8 RAID 1

(17)

RAID 50

RAID 5

A0 Bp C0 D0 Ep A1 B0 Cp D1 E0 Ap B1 C1 Dp E1 A2 Bp C2 D2 E2 Ap B3 C3 Dp E3 A3 B2 Cp D3 Ep

RAID 5

RAID 0

(18)

RAID 160

RAID 6 A0 Bq Cp D0 E0 A0 Bq Cp D0 E0 A1 B0 Cq Dp E1 A1 B0 Cq Dp E1 Ap B1 C0 Dq Ep Ap B1 C0 Dq Ep RAID 1 RAID 0 Aq Bp C1 D1 Eq Aq Bp C1 D1 Eq A2 Bq Cp D2 E2 A2 Bq Cp D2 E2 A3 B2 Cq Dp E3 A3 B2 Cq Dp E3 Ap B3 C2 Dq Ep Ap B3 C2 Dq Ep Aq Bp C3 D3 Eq Aq Bp C3 D3 Eq

RAID 1 RAID 1 RAID 1

RAID 1 RAID 1 RAID 1 RAID 1

RAID 6

(19)

RAID Failure Mode Operation

What happens when a disk fails ?

RAID 0 Lose all data (hope there’s more than one RAID layer)

RAID 1 Business as usual, hot-swap the failed disk

RAID 2-6 Operate in degraded mode

• If data drive failed, every read must be reconstructed • If parity drive failed, low performance impact

• Replace drive (hot-swapping: the system continues running) • Rebuild the array (re-constitute the state of the lost drive)

(20)

RAID Recovery Limitations

Rebuilding a degraded array • Sequentially

• How long ? • On live system ?

Risk of failure during recovery • Statistical distortion:

higher risk of multiple failures within a narrow time frame • RAID 5 risk advisory notice:

Do not use for business-critical data! [Dell]

(21)

Where to Implement RAID?

• Software – Operating System

• Most OS now provide software RAID

• E.g. Linux md (multiple devices) supports RAID 0, 1, 4, 5, 6 plus nestings

• Software – File System • E.g., ZFS

(22)

Array Failure Rates (full data loss)

Failure rate of a disk drive: r (with some assumptions!)

RAID 0 1 − (1 − r)n

RAID 1 rn

RAID 2 It’s complicated

RAID 3-5 1 − (1 − r)n−n 1 r1(1 − r)n−1 RAID 6 1 − (1 − r)n−n 1 r1(1 − r)n−1−n 2 r2(1 − r)n−2 Example drive 1%/year failure rate with RAID 6 (3+2 drives): 1 − 0.995− 5 · 0.01 · 0.994₋4 ∗ 5

2 · 0.01

2_{· 0.99}3 _{= 0.000009851}

(23)

Hamming Codes

A0

A1

A3

A2

A

p1

A

p2

A

p3

• RAID 2 no longer used, but... • Hamming code error correction • ECC

(24)

Parity

• Old idea: first tape drive (1951) had a parity track • Transverse redundancy check

• How does it really work ?