Comments on: Oracle Parallel Query Sorting and Index Creation Performance Problems

By: Amir Riaz

Amir Riaz — Wed, 01 Jun 2011 17:53:00 +0000

Yes, confirmed. it was a bug. I check our all databases 10g and 11.1.6. the problem is, when query go parallel but the data is unevenly distributed across the slave processes which increases the query execution time. in case if all the 100% of data is being assigned to a slave process. the time of serial processing is nearly equal to time of parallel processing. in fact then parallel processing take more time.

cannot found this problem with 11.2.

I am expecting that you write something about order by clause in 11.2 and lets open it for discussion

Thanks.

By: Christo Kutrovsky

Christo Kutrovsky — Wed, 01 Jun 2011 15:06:36 +0000

On queries like the one in my example it is sufficient to have a parallel hint to have the query run fully in parallel.

My entire test case is documented, I invite you to study the explain plans and re-run my test case.

By: Amir Riaz

Amir Riaz — Wed, 01 Jun 2011 15:00:04 +0000

The reason you are getting uneven distribution of rows across producer processes is, that part is running in serial.

you did not specify parallel clause for table, while you specify parallel clause for index.

try

create table big_emp parallel (degree 4)
as select * from emp;

CREATE INDEX emp_ix ON emp (emp_id)
PARALLEL (DEGREE 4);

It would be interesting to see the results.

By: Christo Kutrovsky

Christo Kutrovsky — Wed, 27 Jan 2010 19:03:17 +0000

Hi Zohar,

I don’t have any metalink documents to give you. I have been able to workaround the issue in some cases, but it’s mostly not going anywhere.

I do have a blog almost ready with a larger dataset, but it only confirms the issue further. i.e. not much change.

By: Zohar

Zohar — Wed, 27 Jan 2010 12:36:48 +0000

Did you find any metalink documentation since this post was made? we encounter the same problem but can’t find any official documentation of the problem.

By: Blogroll Report 13/11/2009-20/11/2009 « Coskan’s Approach to Oracle

Blogroll Report 13/11/2009-20/11/2009 « Coskan’s Approach to Oracle — Sun, 22 Nov 2009 21:42:14 +0000

[...] 5-Bug on Running Parallel Query with Order by on already sorted data causes problem? Christo Kutrovsky-Oracle Parallel Query Sorting and Index Creation Performance Problems [...]

By: Christo Kutrovsky

Christo Kutrovsky — Tue, 17 Nov 2009 17:07:24 +0000

Actually that’s how I discovered this, by having big table’s parallel operations take forever.

I’ve tested this in 10.2.0.4, 11.1.0.7 and 11.2 base.

I will try my example with larger tables to confirm it scales the same as when I discovered this.

By: Timur Akhmadeev

Timur Akhmadeev — Tue, 17 Nov 2009 10:58:04 +0000

Hi Christo,

an interesting observation. Have you experimented with a bigger table – for ex., 20-40M rows? Since PX is made for that type of volumes, not 400K, it is reasonable to test with appropriate data volumes, especially with such high DOPs.
I’ve tried to repeat your case – but unfortunately for some reason v$pq_tqstat didn’t show statistics for PX slaves as a Consumer (only Producers). That was 10.2.0.4 & 10.2.0.4.2.
What Oracle version have you used?

By: Christo Kutrovsky

Christo Kutrovsky — Tue, 17 Nov 2009 04:02:32 +0000

I am not aware of any such document, but I haven’t been very thorough in my search.

As I said, this is semi-fixed in 11g Release 2 (11.2) so Oracle knows about it.

I will make separate post with example at some point.

By: Dion Cho

Dion Cho — Tue, 17 Nov 2009 03:53:49 +0000

Christo.

Sorry for wrong word. :(

To rephrase my question -
“Is there any metalink document that is registered as a bug or something on this phenomenon?”