||Oracle Tips by Burleson
The SSD results using the RamSan400 were anything
but typical. Remember that the RAC systems tested were identical down
to the host bus adapter (HBA). The only variable was whether the RAID
or SSD system was used.
As with the RAID system, the transactions per
second (TPS) will be examined as the amount of server memory is
decreased. This drives I/O to the RamSan-400. Figure 4.8 shows the
results from this decrease in server cache memory.
As I/O is forced to the RamSan-400, the TPS
actually increases. This is because the I/O speed of the RamSan-400 is
actually much better than the latency of the RAC high-speed
interconnect. The interconnect speed is the primary driving factor in
the fully cached system as the transactions are limited by the amount
of data that must be transferred and the latency of the interconnect.
With the non-fully cached server the number of transactions is
dependent more upon the speed of the underlying I/O subsystem. Hence,
when a system exists with I/O speed that is faster than the
interconnect latency, the TPS is increased by driving I/O away from
the interconnect and to the I/O subsystem!
SSD Bytes per Second
What about bytes per second? Figure 4.9 shows the
results from decreasing the per-instance server memory for a
RamSan-400 based system. As the memory in the server that was
allocated to Oracle was decreased by resetting the
sga_targetsettings in the SSD based test, bytes per second
increased. Figure 4.9 shows the results on total bytes per second from
decreasing the per-instance memory for a RamSan-400 based system.
This figure shows how the bytes-per-second
increases as dependency on the interconnect is removed and the system
moves away from being fully cached when using SSD technology. This is
due to the factor of ten or more difference between the average
latencies of the cluster interconnect and the speed of reads and
writes on the SSD system.
SSD Average Transaction
Perhaps the most dramatic difference is in the
average transaction time results when comparing the RAID to the SSD
system. As expected from a review of the previous SSD results, a
decrease in the average transaction time should be expected allowing
more transactions and more bytes per second. However, a very dramatic
decrease as cache memory was reduced was seen. Look at Figure 4.10.
Once dependence on the interconnect is mitigated,
the transaction times dramatically improved by several orders of
magnitude. The scale on the graph in Figure 4.10 is logarithmic. This
means that for one increase in scale on the vertical axis, an increase
by a factor of 10 can be seen. Compare the average transaction times
in Figure 4.10 to those from the RAID tests in Figure 4.6. The
dramatic difference between the two technologies can be seen as memory
is reduced. This indicates that by utilizing the RamSan400 SSD system
transaction times were improved by nearly a factor of 100, two orders
SSD Average Response
Since the average response time is dependent on
transaction times, the results for the RamSan400 mirror those for
average transaction time is shown in Figure 4.11.
Again note the logarithmic vertical scale. As with
average transaction time, a nearly 2 order of magnitude difference can
be seen when the latency of the RAC high-speed interconnect is
mitigated as a factor in response time. Comparing the RAID average
response time in Figure 4.7 to SSD average response time in Figure
4.11 also highlights the dramatic difference between the two storage
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