RAID (originally redundant array of inexpensive disks, now commonly redundant array of independent disks) is a data storage virtualization technology that combines multiple physical disk drive components into a single logical unit for the purposes of data redundancy, performance improvement, or both.

1. RAID OverviewRAID Overview
91.560
1

2. Computing speeds double every 3 years
Disk speeds can’t keep up
Data needs higher MTBF than any
component in system
 IO Performance and Availability
Issues!
2
The Motivation for RAID

3. PERFORMANCE
◦ Parallelism
◦ Load Balancing
AVAILABILITY
◦ Redundancy: Mirroring, or Striping with
Parity
FLEXIBILITY
◦ Selectable Performance/Availability/Cost
3
RAID to the Rescue!

4. Redundant Array of Independent Disks
(a.k.a. “Disk Array”)
◦ Multiple drives, single host disk unit
Provides opportunity to increase:
◦ Performance via Parallelism
◦ Data Availability via Redundancy
Cheap cost via commodity disks
4
What is RAID Technology?

5.  Chunk size tuneable for BW vs Thruput tradeoffs
◦ Large Chunk High Throughput (IO/sec)
◦ Small Chunk High Bandwidth (MB/sec)
5
Chunk 1
Parity
Chunk 11
Chunk 16
Chunk 21
Chunk 5
Chunk 10
Chunk 15
Chunk 20
Chunk 2
Chunk 7
Chunk 17
Chunk 22
Chunk 3
Chunk 8
Chunk 13
Chunk 23
DriveDrive
55
Chunk 12
Chunk 0
Chunk 18
Chunk 9
Chunk 14
Chunk 19
Chunk 24
Chunk 4
Chunk 6
DriveDrive
44
DriveDrive
33
DriveDrive
22
DriveDrive
11
Performance: Disk Striping

6. 6
Data Stream, OR IO Stream, can be multiplexed
across multiple disks, depending on BW vs
Thruput. Parity data is also stored on disk.
Same data written to both disks.
Striping with Parity
Single Host Unit
> Add XOR’d parity for increased availability.
Mirroring
Availability: Redundancy

7. Parity = XOR of data from every disk in
the RAID unit
7
DATA
0
1
1
0
0 parity
- Any single disk’s
data can be recovered
by XOR’ing the data
of the surviving disks.
Parity Redundancy

8. Many to choose from
Each offers unique tradeoffs
◦ Performance
◦ Availability
◦ Costs
We offer levels 0, 1, 3, 5, 10
8
RAID Levels

9. •High Performance; Low Availability
•Data Striped on Multiple Disks
•Multi-threaded Access
9
Disk Striping with No Redundancy
Data Stream, OR IO Stream, can be Striped
across multiple disks (BW vs Thruput).
RAID 0

10.  Chunk size tuneable for BW or Thruput
 No redundancy
10
Chunk 1
Parity
Chunk 11
Chunk 16
Chunk 21
Chunk 5
Chunk 10
Chunk 15
Chunk 20
Chunk 2
Chunk 7
Chunk 17
Chunk 22
Chunk 3
Chunk 8
Chunk 13
Chunk 23
DriveDrive
55
Chunk 12
Chunk 0
Chunk 18
Chunk 9
Chunk 14
Chunk 19
Chunk 24
Chunk 4
Chunk 6
DriveDrive
44
DriveDrive
33
DriveDrive
22
DriveDrive
11
RAID 0 Striping

11.  Single-disk Performance; Expensive
Availability
 Data 100% duplicated across both spindles.
 Single-threaded access
11
Disk Mirroring
SAME data written to BOTH disks -- no
segmenting.
Single Host Unit
RAID 1

12. •Highest Performance; Most Expensive
Availability
•Multi-threaded Access
12
Striped Mirrors
Data Stream, OR IO Stream, can be Striped
across multiple disks (BW vs Thruput).
Single Host Unit
RAID 1/0

13. •High BW Performance; Cheap Availability
•Sector-granular data striping
•Single-threaded Access
13
Disk Striping with dedicated parity drive
Data Stream is Striped across N-1 disks for
high bandwidth.
XOR Parity Data
RAID 3

14.  Chunk size = single sector (pure RAID 3 would be single byte)
 All parity data on same spindle
14
Chunk 1
Parity
Chunk 9
Chunk 13
Chunk 17
Chunk 4
Chunk 8
Chunk 12
Chunk 16
Chunk 2
Chunk 6
Chunk 14
Chunk 18
Chunk 3
Chunk 7
Chunk 11
Chunk 19
DriveDrive
55
Chunk 10
Chunk 0
Chunk 15
Parity
Parity
Parity
Parity
Parity
Chunk 5
DriveDrive
44
DriveDrive
33
DriveDrive
22
DriveDrive
11
RAID 3 Striping

15.  High Read Performance, expensive Write
performance; Cheap Availability
 Tuneable Stripe granularity
 Optimized for multi-thread access
15
Disk Striping with rotating parity drive
IO Stream is Striped across N-1 disks for high
IOs per second (thruput).
XOR Parity Data
RAID 5

16.  Chunk size is tuned such that typical IO aligns on single disk.
 Parity rotates amongst disks to avoid write bottleneck
16
Chunk 1
Parity
Chunk 9
Chunk 16
Parity
Chunk 4
Chunk 8
Chunk 12
Parity
Chunk 2
Chunk 6
Chunk 13
Chunk 17
Chunk 3
Chunk 10
Chunk 18
Parity
DriveDrive
55
Parity
Chunk 0
Chunk 14
Chunk 7
Chunk 11
Chunk 15
Chunk 19
Parity
Chunk 5
DriveDrive
44
DriveDrive
33
DriveDrive
22
DriveDrive
11
RAID 5 Striping

17. 17
1. Read old
data.
Old New
2. Write
new data
Old P.
3. XOR old and
new data to create
“Partial Product”.
4. Read
old parity
data.
5. Xor old parity with
partial product, writing
out result as new parity.
P. P.
}
XOR
}
XOR
New P.
Chunk 1 Chunk 2 Chunk 3
DriveDrive
55
Chunk 0 Parity
DriveDrive
44
DriveDrive
33
DriveDrive
22
DriveDrive
11
RAID 5 - Write Operation

18. 18
RAID 0 - Data striping, Non-redundant.
High Performance, Low Availability
RAID 1 - Mirroring
Moderate Performance, Expensive High Availability
RAID 1/0 - Striping and Mirroring
High Performance, Expensively High Availability
RAID 3 - Striping, single parity disk.
High Bandwidth Performance, Cheap Availability
RAID 5 - Striping, rotating parity disk.
High Thruput Performance, Cheap Availability
RAID Level Review

19. Increasing performance gap between CPU
and IO
Data availability a priority
RAID meets the IO challenge:
◦ Performance via parallelism
◦ Data Availability via redundancy
◦ Flexibility via multiple RAID levels, each offer
unique performance/availability/cost tradeoffs
19
Summary