1. RAID Redundant Arrays of Independent Disks
Courtesy of Satya, Fall ’99
Updated by Gregory Kesden for subsequent classes

2. Motivation for Disk Arrays
Typical memory bandwidths  150 MB/sec
Typical disk bandwidths  10 MB/sec
Result: I/O-bound applications limited by disk bandwidth
(not just by disk latency!)
Two common disk bandwidth-limited scenarios
Scientific applications: one huge I/O-hungry app
Servers: many concurrent requests with modest I/O demands

3. Solution: Exploit Parallelism
Stripe the data across an array of disks
many alternative striping strategies possible
Example: consider a big file striped across N disks
stripe width is S bytes
hence each stripe unit is S/N bytes
sequential read of S bytes at a time S S S
s0,0
s0,1
s0,2
• • •
s0,N-1
s1,0
s1,1
s1,2
• • •
s1,N-1
s2,0
s2,1
s2,2
• • •
s2,N-1
• • • • •
Disk 0
Disk 1
Disk 2
Disk N-1
S0,0
S0,1
S0,2
• • •
S0,N-1
S1,0
S1,1
S1,2
• • •
S1,N-1
S2,0
S2,1
S2,2
• • •
S2,N-1
• • •
• • •
• • •
• • •

4. Performance Benefit Sk Sequential read or write of large file
application (or I/O buffer cache) reads in multiples of S bytes
controller performs parallel access of N disks
aggregate bandwidth is N times individual disk bandwidth
(assumes that disk is the bottleneck)
Disk 0
Disk 1
Disk 2
Disk N-1
• • •
x MB/s
x MB/s
x MB/s
x MB/s
S k,0
S k,1
S k,2
S k,N-1
Array Controller
Nx MB/s Sk

5. N independent requests
N concurrent small read or write requests
randomly distributed across N drives (we hope!)
common in database and Web server environments
Disk 0
Disk 1
Disk 2
Disk N-1
• • •
x MB/s
x MB/s
x MB/s
x MB/s
req 1
req 0
req 2
req N-1
Array Controller
Nx MB/s
N independent requests

6. Reliability of Disk Arrays
As number of disks grows, chances of at least one failing increases
Reliability of N disks = (reliability of 1 disk) / N
suppose each disk has MTTF of 50,000 hours
(roughly 6 years before any given disk fails)
then some disk in a 70-disk array will fail in (50,000 / 70) hours
(roughly once a month!)
Large arrays without redundancy too unreliable to be useful
“Redundant Arrays of Independent Disks” (RAID)

7. RAID Approaches
Many alternative approaches to achieving this redundancy
RAID levels 1 through 5
hot sparing allows reconstruction concurrently with accesses
Key metrics to evaluate alternatives
wasted space due to redundancy
likelihood of “hot spots” during heavy loads
degradation of performance during repair

8. RAID Level 1 D1 B1 D2 B2 D3 B3 Also known as “mirroring”
To read a block:
read from either data disk or backup
To write a block:
write both data and backup disks
failure model determines whether writes can occur in parallel
Backups can be located far way: safeguard against site failure

9. RAID Levels 2 & 3 D1 D2 D3 P These are bit-interleaved schemes
• • •
These are bit-interleaved schemes
In Raid Level 2, P contains memory-style ECC
In Rail Level 3, P contains simple parity
Rarely used today

10. RAID Level 4 D0,0 D1,0 D2,0 D3,0 D0,1 D1,1 D2,1 D3,1 D0,2 D1,2 D2,2P0 P1 P2 P3
D0,0  D0,1  D0,2 = P0
Block-interleaved parity
Wasted storage is small: one parity block for N data blocks
Key problem:
parity disk becomes a hot spot
write access to parity disk on every write to any block

11. RAID Level 5 D0,0 D1,0 D2,0 P3 D0,1 D1,1 P2 D3,1 D0,2 P1 D2,2 D3,2 P0
Rotated parity
Wastage is small: same as in Raid 4
Parity update traffic is distributed across disks

12. RAID 5 Actions    Fault-free Read Fault-free Write Degraded ReadP
Fault-free Read D P  1 2 3 4
Fault-free Write D P 
Degraded Read D P 
Degraded Write

13. RAID 6 Like RAID-5, but has two rotated parity blocks
Can sustain 2 independent drive failures
Similarly no read penalty
More work to do to compute parity
For writes
Upon failures
Parity computation isn’t simple
Uses a Galois Field
Much more computational involved
Beyond scope of class
Key point: Enables reads with 2 failures with only 2 parity blocks