||Oracle Tips by Burleson
RAID results are referring to standard disk based
systems whether they are standard RAID, SATA or other interfaces that
map into the RAID framework. The major result from a TPC-C result is
throughput, measured as transactions per second. RAID test throughput
will be examined first.
As memory was decreased from 1 gigabyte first to
500 megabytes and then to 250 megabytes, the RAID system responded as
expected with less overall throughput as more and more disk I/O was
performed. This is shown in Figure 4.4.
This shows that as the disk plays a more active
roll in the transaction base, the latency involved in disk operations
causes a net reduction in throughput. Another measure is bytes per
second, and that will be presented next.
RAID Bytes per Second
When bytes per second (BPS) for RAID are examined,
the same decreasing profile can be seen. This decrease in BPS as
memory is reduced is of course expected as the number of transactions
decreases. This is shown in Figure 4.5.
This shows that as the RAID environment becomes
more and more dependent on the disk I/O subsystem, the transaction
rate and rate of data flow decrease dramatically on a per user basis.
Essentially, for a given user load that exceeds the breakpoint this
breaks down to a simple equation:
TPSf = TPSi *
TPSf – Transactions per second final
TPSi – Transactions per second Initial (Peak
with 100% caching)
RBPT – Required bytes per transaction per second
IOR – Sustained Bytes IO rate in bytes per
Q – A constant that takes into account other
latency factors which should be a constant for a given
The breakpoint is the point of inflection for a
given TPS curve.
So for a RAID based system the transactions will
never be more than what can be achieved in a fully cached situation
assuming that the interconnect latency is much less than the overall
The performance behavior is inverted when average
transaction time is considered. As memory in a RAID based system is
decreased, average transaction time increases. As more processes
contend for limited disk I/O capability, the average transaction time
must increase resulting in fewer transactions per second and lower
bytes per second. This is shown in Figure 4.6.
RAID Average Response
The final results for the RAID tests show the
average response time. As would be expected, as the memory decreases,
forcing more dependence on disk, the response time increases, just as
with average transaction times. This is shown in Figure 4.7.
As expected, all performance indicators showed
poorer performance in a disk-based system as available cache memory
decreased. This decrease in performance is caused directly by
increasing disk contention thereby forcing high latency times and
queuing to occur on the underlying disks.
The same tests were performed using solid-state
disk technology and those results will be examined next.
The above book excerpt is from:
Oracle RAC & Tuning with
Solid State Disk
Expert Secrets for High
Performance Clustered Grid Computing
Donald K. Burleson & Mike Ault