Information Technology Reference
In-Depth Information
{ Rotating parity. Because the parity for a given set of blocks must
be updated each time any of the data blocks are updated, the average
parity block tends to be accessed more often than the average data
block. To balance load, rather than having G 1 disks store only
data blocks and 1 disk store only parity blocks, each disk dedicates
1
1
G
th of its space to parity and is responsible for storing
G
th of the
parity blocks and
G1
G
of the data blocks.
{ Striping data. To balance parallelism versus sequential-access e-
ciency, a strip of several sequential blocks is placed on one disk before
Definition: strip
shifting to another disk for the next strip. A set of G1 data strips
and their parity strip is called a stripe.
Definition: stripe
By striping data, requests larger than a block but smaller than a strip
require require just one disk to seek and then read or write the full
sequential run of data rather than requiring multiple disks to seek
and then read smaller sequential runs. Conversely, the RAID can
service more widely spaced requests in parallel.
Combining rotating parity and striping, we have the arrangement shown
in Figure 14.6.
Example: Updating a RAID with rotating parity.
Question: For the rotating parity RAID in Figure 14.6, suppose you update
data block 21. What disk I/O operations must occur?
Answer: The challenge is that we must not only update data block 21,
we must also update the corresponding parity block. Since data
block 21 is block 1 of its strip and the strip is part of stripe 1, we
need to update parity block 1 of the parity strip for stripe 1 (Parity
(1,1,1) in the figure.)
It takes 4 I/O operations to update both the data and parity. First
we read the old data D
21
and parity P
1;1;1
and “remove” the old
data from the parity calculation P
tmp
= P
1;1;1
D
21
. Then we can
compute the new parity from the new data P
0
1;1;1
= P
tmp
D
0
21
.
Finally we can write the new data D
0
21
and parity P
0
1;1;1
to disks
2 and 1, respectively.
RAIDs with rotating parity have high overheads for small writes. Their over-
heads are far smaller for reads and for full-stripe writes.
Recovery. In either RAID arrangement, if a disk suffers a sector failure, the
disk reports an error when there is an attempt to read the sector and, if nec-
essary, remaps the damaged sector to a spare one. Then, the RAID system
reconstructs the lost sector from the other disk(s) and rewrites it to the original
disk.
