Hi All,

This is another question about system statistics. We are in the
process of upgrading from 9.2.0.6 to 10g release 2. We do not
currently use system statistics in 9.2, but now they are mandatory for
10g we are keen to set accurate and representative statistics for our
testing and development environments as well as production. Once these
are set we do not intend to have oracle re-sample unless we make
changes to the hardware configuration. Our area of difficulty is
around the IO stats.

In our environment (Solaris 9 64-bit) Oracle accesses our SAN based
disks via Veritas ODM drivers. The sequential read wait event in live
is around 6ms and we are comfortable to use that value. However the
multiblock read time is significantly less at around 2.5ms, due to the
read ahead routines employed by the backend storage devices. If we
were to use these values Oracle would start to favour tablescans and
this would not be correct.

We have a pre-production rig, though there are outstanding issues
around the I/O response times on those servers which are beyond our
control, so setting up and testing in that environment is
frustratingly not possible.

We have read much material posted in various places and are yet to be
entirely convinced that any of the methods described are providing
accurate statistics. We have developed a couple of C programs to
simulate the random read times of 8k blocks and multiblock reads,
however as those do not use the Veritas ODM routines to access the
disk we feel they are not representative. Also our production
environment is highly protected and actually running these tests
(which involve creating a file that is larger than the SAN cache...)
in live will be a challenge.

As a result of this we have settled upon using a single block read
time of 6 and, as Oracle does for its defaults, multiplying this
figure by 2 to get a multiblock figure of 12. To be fair we are pretty
confident that we have reasonable values, but feel that the method we
have used to gain them warrants further scrutiny.

Any comments and ideas welcome.

Ralph

On May 25, 6:28 am, Ralph.in... (AT) googlemail (DOT) com wrote:
It is common for oracle to calculate the mreadtim to be smaller than
sreadtim for various storage types impacted by caching and read ahead.

If you take corrective actions and make adjustments ... proceed
carefully. Create some test cases that correspond to expected behavior
of some critical parts of your applications is my recommendation.

Take a look at some of the recent updates in this area provided by
Jonathan Lewis.