Choosing the Right Disk Technology in a High Availability Environment DRAFT Version 2.0, August 1996

DRAFT -- Revision 2.0
August 22, 1996Page 37
sizes, although the performance is much lower than with standalone mirrored disks.
RAID 3 performs best for I/Os of 64 KB or larger.
The performance of a disk array configured in RAID 5 is very inconsistent. Small I/Os
are most efficient for read operations since in the random case, multiple small I/Os can
be processed concurrently by the disk array when they reference data on different disk
mechanisms. Large I/Os (>= 64 KB) involve all disk mechanisms in the array and
therefore prevent multiple concurrent I/Os. However, large writes are much more
efficient than small writes since large writes rewrite all of the data in the stripe including
the parity information. Small writes require a read-modify-write sequence to occur since
the entire stripe is not rewritten. The stripe must first be read, the appropriate portion
modified, the new parity information calculated, and the entire stripe written back out.
Ratio of reads to writes
Many commercial applications are read intensive. OLTP applications of 70% or higher
reads and Decision Support (DSS) applications that are 100% reads are considered
read-intensive. Since reads are more efficient than writes, read-intensive applications
will benefit most by an LVM mirrored environment or a HADA RAID 1 configuration with
the new firmware release. Both take advantage of the multiple copies of the data and
split the I/Os among the copies.
Administrators of non-read-intensive applications must carefully weigh the disk
technology employed based on the other considerations discussed in this section, such
as I/O size, striping, and data protection mechanism.
Sequential versus random access
Sequential I/O performance is usually better than random when the I/O size is large.
Therefore, the 10.0 feature of I/O merging improves sequential performance even when
the application writes sequentially, but in small blocks.
Random I/O performance is usually best when the I/O size is small, since many I/Os
can be queued up which reference data on different disk mechanisms. So, the OS can
issue many I/O requests in parallel.
Failed mechanisms in RAID 3 or 5 configurations
When a mechanism has failed in a RAID 3 or 5 configuration, the disk array controller
must recreate the missing data mathematically. The mathematical computation
prevents the controller from working on other I/Os and therefore degrades overall
performance. Small I/Os that don't involve the entire RAID 5 stripe will require the
controller to read the rest of the stripe in order to have enough data to recreate the