Re: Performance problems after moving to new hardware
- From: Stefan Koehler <contact@xxxxxxxx>
- To: oracle-l <oracle-l@xxxxxxxxxxxxx>, sbecker6925@xxxxxxxxx
- Date: Wed, 4 Mar 2015 16:21:28 +0100 (CET)
Hi Sandra,
well i gonna be very generic to your specific points here as you are not able
to provide any details due to company policy.
1) Yes, it also changes the CBO behavior, but in a different way as your
development team would expect it. Randolf Geist called it "new cost is time
model". For details please check his blog post ( http://tinyurl.com/nxdksp9 ).
However I/O calibration is used in case of Auto DOP and as you
mentioned in point 4 that your query runs in parallel it might be an
influencing factor.
2) Script awrsqrpt.sql provides you more information about history execution
plan, but with a lot of missing important details (e.g. predicate &
outline section). So nothing really helpful, if it gets really tricky. However
if it is about parallel query, i would go for Randolf Geist's tool
called "XPLAN_ASH" (really great stuff!): http://tinyurl.com/nuso58w and check
the PX details.
3) This can not be answered without any detailed information, but as it ran
fast without these indexes before the switch i would start analyzing the
root cause first.
4) SQL Monitoring report, especially in case of PX.
However the following statement ("On Saturday I completed gathering stats on
the application schema tables as requested by the product manager") makes
me sit-up and take notice. Why not just restoring the old statistics for the
corresponding objects (as a possible quick work-around and for
verification) and check the runtime afterwards. If the runtime is fine
afterwards, you can dig into the details and check what went wrong without any
business pressure.
Best Regards
Stefan Koehler
Freelance Oracle performance consultant and researcher
Homepage: http://www.soocs.de
Twitter: @OracleSK
> Sandra Becker <sbecker6925@xxxxxxxxx> hat am 4. März 2015 um 14:25
> geschrieben:
>
> OS: Solaris Sparc 10 (64-bit)
> Oracle: EE 11.2.0.2
>
> The OS and Oracle versions are identical on both the old and new servers.
> Storage attached to the new server is a new EMC disk array. Sorry I don't
> have any more details on the storage and the only additional information I
> have on the server is that it is a T5.
>
> We created a standby on the new hardware and did a switchover last Friday
> night. On Saturday I completed gathering stats on the application schema
> tables as requested by the product manager. As usual, very little activity on
> this database over the weekend. Yesterday morning we were contacted by
> internal users that performance was much worse than on the old hardware for a
> specific query on a really ugly view. A look at the execution plan
> shows multiple full table scans on some partitioned tables, some very large.
> There are about 15 tables joined to create the view, some more than
> once. They claim the view is no longer doing partition pruning, as it did
> before the switchover. I can't prove that it was/wasn't exhibiting this
> behavior before the switchover. They are insisting we run I/O calibration.
> I'm not familiar with it so I went to the docs. This database shares
> storage with quite a few production databases so I want to be very careful
> how I go about this.
>
> Questions:
>
> 1. What will running the I/O calibration do? Does it only provide
> information on the I/O subsystem, or does it change the way the optimizer
> behaves? The development team insists it will improve performance.
> 2. I've looked at AWR reports before/after the switchover and see that the
> query in question was doing a similar amount of I/O in both reports. Is
> there any way for me to get more detail on the before execution plan?
> 3. One of the large partitioned tables has no indexes. Would creating an
> index be of any benefit? I understand that it's possible to negatively
> affect other queries, so it should be considered with caution. Development
> insists that indexing would be a waste of time and definitely cause
> problems, although they have never tested it.
> 4. I want to trace the query, but it runs in parallel and produces more
> trace data that I have available disk to handle. Is there anything I can do
> on that front to get a trace I can feed into my Method-R tool and supply to
> oracle support?
>
> As I reviewed how the view, I recall them having issues with it before and
> me suggesting it should be optimized. I was told no and here we are
> again. The obvious concern is that the results would be different and changes
> require a lot of testing they don't have time to do. Any other
> recommendations would be appreciated.
>
> --
> Sandy
> GHX
--
//www.freelists.org/webpage/oracle-l
Other related posts:
