than having parity distributed among
all disks. Since fewer disks participate
in reads (the dedicated parity disk is
not read except in the case of a failure), RAID- 4 is strictly less efficient
than RAID- 5.
RAID- 6, double-parity RAID, was
not described in Patterson, Gibson,
and Katz’s original 1988 paper9 but
was added in 1993 in response to the
observation that as disk arrays grow,
so too do the chances of a double failure. Further, in the event of a failure
under any redundancy scheme, data
on all drives within that redundancy
group must be successfully read in order for the data that had been on the
failed drive to be reconstructed. A read
failure during a rebuild would result
in data loss. As Chen et al. state:
“The primary ramification of an uncorrectable bit error is felt when a disk
fails and the contents of the failed disk
must be reconstructed by reading data
from the nonfailed disks. For example,
the reconstruction of a failed disk in a
100GB disk array requires the successful reading of approximately 200 million sectors of information. A bit error
rate of one in 1014 bits implies that one
512-byte sector in 24 billion sectors
cannot be correctly read. Thus, if we
assume the probability of reading sectors is independent of each other, the
probability of reading all 200 million
sectors successfully is approximately
As motivation for its RAID- 6 solution, NetApp published a small comparison of RAID- 5 and - 6 with equal
capacities ( 7+ 1 for RAID- 5 and 14+ 2
for RAID- 6) and hard drives of varying quality and capacity.
1 Note that
despite having an additional parity
disk, RAID- 6 need not reduce the total capacity of the system.
7 Typically
the RAID stripe width—the number
of disks within a single RAID group—
for RAID- 6 is double that of a RAID- 5
equivalent; thus, the number of data
disks remains the same. The NetApp
comparison is not specific about the
bit error rates of the devices tested,
the reliability of the drives themselves,
or the length of the period over which
the probability of data loss is calculated; therefore, we did not attempt to
reproduce these specific results. The
important point to observe in Figure 1
is the stark measured difference in the
probability of data loss between RAID-
5 and RAID- 6.
When examining the reliability of a
RAID solution, typical considerations
range from the reliability of the component drives to the time for a human
administrator to replace failed drives.
The throughput of drives has not been
a central focus despite being critical for RAID reconstruction, because
throughput has been more than adequate. While factors such as the bit
error rate have kept pace with capacity, throughput has lagged behind,
forcing a new examination of RAID
reliability.
capacity vs. Throughput
Capacity has increased steadily and
significantly, and the bit error rate
has improved at nearly the same pace.
Hard-drive throughput, however,
has lagged behind significantly. Using vendor-supplied hard-drive data
sheets, we’ve been able to examine the
relationship between hard-drive capacity and throughput for the past 10
years. Figures 2–4 show samples for
various hard-drive protocols and rotational speeds.
This data presents a powerful con-
figure 4. Historical capacity/Throughput of 15k RPm fc HDDs.
Capacity (Gb)
700
throughput (Mb/s)
250
600
( 1 – 1/( 2. 4 ´ 1010)) ( 2.0 ´ 108) = 99.2%.
500
capacity (GB)
400
300
200
200
150
100
Throughput (mB/s)
50
This means that on average, 0.8% of
disk failures would result in data loss
due to an uncorrectable bit error.”
3
Since that observation, bit error
rates have improved by about two orders of magnitude while disk capacity
has increased by slightly more than
two orders of magnitude, doubling
about every two years and nearly following Kryder’s law. Today, a RAID
group with 10 TB (nearly 20 billion sectors) is commonplace, and typical bit
error rate stands at one in 1016 bits:
100
0
2000
2002
2004
2006
2008
2010
0
figure 5. minimum time required to populate HDDs through the years.
7200 rPM sata 10K rPM FC 15K rPM FC
— 7200 rPM sata — 10K rPM FC — 15K rPM FC
250
200
( 1 – 1/( 2. 4 ´ 1012)) ( 2.0 ´ 1010) = 99.2%
minutes
150
100
While bit error rates have nearly
kept pace with the growth in disk capacity, throughput has not been given
its due consideration when determining RAID reliability.
50
*7200
RPM
FC
0
1996
1998
2000
2002
2004
2006
2008