Browse Prior Art Database

Optimal Data Allotment to Build High Availability and High Performance Disk Arrays

IP.com Disclosure Number: IPCOM000112372D
Original Publication Date: 1994-May-01
Included in the Prior Art Database: 2005-Mar-27
Document File: 6 page(s) / 252K

Publishing Venue

IBM

Related People

Chen, M: AUTHOR [+3]

Abstract

The availability of inexpensive, good performance, small disks has made it increasingly attractive to build a reliable, high performance disk subsystem by integrating a number of small disks in an array. In such a disk array, data are distributed across multiple disks and only a portion of data is hold in each small disk. Note that the reliability of these small disks is not as high as that of large disks, and a disk failure can render all the data residing in it unavailable. To achieve a very high degree of data availability and fault tolerance for business critical applications, two basic techniques are currently being used. In the first, the data, along with the redundant error detection/correction information (usually parity), are spread across an array of disks. Henceforth, we should refer this as the RP approach.

This text was extracted from an ASCII text file.
This is the abbreviated version, containing approximately 24% of the total text.

Optimal Data Allotment to Build High Availability and High Performance
Disk Arrays

      The availability of inexpensive, good performance, small disks
has made it increasingly attractive to build a reliable, high
performance disk subsystem by integrating a number of small disks in
an array.  In such a disk array, data are distributed across multiple
disks and only a portion of data is hold in each small disk.  Note
that the reliability of these small disks is not as high as that of
large disks, and a disk failure can render all the data residing in
it unavailable.  To achieve a very high degree of data availability
and fault tolerance for business critical applications, two basic
techniques are currently being used.  In the first, the data, along
with the redundant error detection/correction information (usually
parity), are spread across an array of disks.  Henceforth, we should
refer this as the RP approach.  When disk failures occur, the
redundant parity information is used to restore data on the failed
disk on-line.  Strategies based on this approach include Redundant
Array of Independent Disks (RAID) [1], Redundant Array of Distributed
Disks (RADD) [2], and Parity Striping of Disk Array (PSDA) [3].  In
the second approach, multiple copies of the same data item are stored
on different disks (henceforth, referred to as the MC approach).
When disk failures occur that renders one copy unavailable, the
remaining copy/copies can continue to be used and thus no
interruption of service will occur.  Strategies based on this
approach include Mirrored Disks (MD) [4], Interleaved Declustering
(ID) [5]  [6], Chained Declustering (CD) [7], and Group-Rotated
Declustering (GRD) [8].  Studies have shown that, in general,
strategies based on the MC approach provide higher performance and
higher data availability, but also incur higher disk space overhead,
than those based on the RP approach.

      Due to the increase in data volume (and thus number of disks)
stored in data farm/information warehouse and the increase in number
of applications that require 24x7 data availability nowadays, more
and more computer systems have adopted either one of the above two
approaches in their I/O subsystems to provide high degree of fault
tolerance and data availability.

      A data file in a disk array with n disks is generally divided
into n partitions with each partition stored on one disk.  When
multiple copies of a data item are stored, different copies of a
partition will be stored on different disks.  Write requests
generally update all copies simultaneously.  On the other hand, a
read request to a partition can be served by any copy or it can be
directed to a single copy termed the primary copy.  When the primary
copy access scheme is used, backup copy/copies will not serve read
requests unless the primary copy fails.  Any as long as at least one
copy is accessible, data will always be available.

      Clearly, the degree of...