Database Reference
In-Depth Information
1.6 Composite Devices
Scaling up the performance of disk technology for server-class systems is in-
creasingly expensive. Composite technologies embody the same spirit as com-
modity clusters to achieve the performance and reliability goals of specialized
proprietary hardware using clusters of much simpler consumer-grade building
blocks.
Composite devices also play a role in ensuring protection against media. In
particular, the probability of a single-bit error in typical disk storage media
is one error in 10
14
to 10
15
bits, and this rate has been relatively consistent
over time. However, with continued exponential increases in the capacity of
storage devices, the likelihood of encountering uncorrectable errors is dramat-
ically increased. Composite devices employ various forms of redundancy and
error correcting codes (ECC) to improve media integrity.
1.6.1 RAID
The nomenclature for RAID configurations was formalized as RAID “levels”
by Chen et al. in their 1994 paper entitled “RAID: High-Performance Re-
liable Secondary Storage.”
8
RAID was originally introduced with five lev-
els, but over time new configurations have been introduced that include
nested configurations as well as some proprietary configurations. RAID
employs some combination of striping for performance with various ap-
proaches to redundancy for fault resiliance. The standard RAID levels are as
follows:
RAID 0:
This configuration, which was not part of the original five RAID
levels, offers improved bandwidth at the expense of reliability by “strip-
ing” data across parallel disks comprising the volume.
RAID 1:
This configuration implements mirroring to improve reliability by
maintaining an exact copy of data on each disk comprising the RAID
volume, but offers no improvement in performance over a single disk.
RAID 2:
Implements hamming codes to support ECC using a dedicated
parity disk to store bit-interleaved parity information, much as ECC-
corrected dynamic random access memory (DRAM) works. The ap-
proach achieves the reliability of mirroring with fewer redundant disks
per protected bit, but suffers from poor performance for small transac-
tions relative to mirroring. This form is rarely seen in practice.
RAID 3:
Implements byte-level striping with a dedicated parity disk, and
relies on the disk controller to identify the failed disk rather than strictly
depending on the parity information as is the case for level 2. Conse-
quently, it offers the same performance characteristics as level 2, but
employs fewer disks to achieve fault resilience.
