RE: Killer SQL and PGA

The tradeoff here is whether to utilize virtual memory to allow gargantuan jobs to theoretically be processed at all. Seymour Cray nailed that in one when he was asked why his systems did not support virtual memory.

It will always be possible to concoct queries and problems that cannot be solved within the limits of any arbitrary amount of real memory.

And of course if you put in ulimit limitations and exceed them, then within some timeslice or interrupt that allows the kernel to figure that out the violating process gets killed. The side effects of the operating system killing Oracle processes is too complex to analyze in the general case.

You mentioned that you had recently moved from 9i to 11gR2, but it is not clear to me whether your "analyst" had previously run the same query in 9i successfully. If so, *probably* it used an entirely different plan. Then again, I've often seen folks test limits on an upgraded system, for example in this context, by supplying a longer "in" list until something breaks. Those are the folks you hope to identify to play games on a "conference room pilot" or quality assurance test upgrade before you go live.

Likewise, proper education to help them avoid unleashing experimental loads on the production system is usually part of the solution.

Regards,

mwf

-----Original Message-----

From: oracle-l-bounce_at_freelists.org [mailto:oracle-l-bounce_at_freelists.org] On Behalf Of Robert Laverty
Sent: Monday, January 30, 2012 6:13 PM
To: oracle-l_at_freelists.org
Subject: Killer SQL and PGA

We had a problem with our PGA growing larger than physical memory, dragging the system down until we were forced to restart the database. We recently upgraded from 9i to 11gR2 on Solaris with 16Gb physical memory hosting an OLTP application. 4Gb is used for SGA and 400Mb for PGA_aggregate_target. AMM and ASMM have not been enabled. Workarea_size_policy is set to AUTO. This is a simple database. No RAC, no shared servers, no parallel processing.

One of our analysts launched what appeared to be a simple ad hoc query: select * from deniedhist where icn in ('1', '2', '3', ... , 'x'). Unfortunately, the analyst had over 400 ICNs listed. DENIEDHIST is a UNION ALL view representing a faux-partitioned array of 129 tables, each with its own unique index on ICN. The execution plan (my comments included) shows that it is searching each underlying table 'x' number of times, using the unique index each time. This is the plan from a test version of the SQL with only 5 ICNs. The rest of the plan shows the iteration through the other 128 underlying tables (other views of the plan show the unique table and index names).
1422 SELECT STATEMENT REMOTE CHOOSE

                1421 VIEW POS.DENIEDHIST
                                1420 UNION-ALL
                                                11 CONCATENATION


-- This is table 1 of 129 in the view

                                                                2 TABLE

                                                                4 TABLE

                                                                6 TABLE

                                                                8 TABLE

                                                                10 TABLE

                                                22 CONCATENATION

                                                             ...

I estimated that the original version probably had over 120,000 steps in its execution plan. It ran for nearly an hour until memory reached a critical point, almost 20 Gb and paging like mad, that required us to abort the database and restart. I can run the full original query on a standby machine which is physically identical but is still running 9i. It finishes in 5 minutes and the PGA only grows to 500 Mb. The execution plans, at least for the small test version of the query, are the same on both machines.

I understand the problems with large 'IN' lists. There was a conversation on that topic here at the end of October . Educating the analyst and redesigning the query are secondary concerns. I also realize, after reading a bunch of Tom Kyte posts, that pga_aggregate_target will not constrain the growth of the PGA.

My real question is why the 11g memory management, without AMM or ASMM, would allow the PGA to grow so large. In 15 years of operations, there must have been similar bad queries against the database. This happened a day after the 11g upgrade. Any suggestions?

Bob Laverty
Molina Medicaid Solutions

IMPORTANT NOTICE TO RECIPIENT: This email is meant only for the intended recipient of the transmission. In addition, this email may be a communication that is privileged by law. If you received this email in error, any review, use, disclosure, distribution, or copying of this email is strictly prohibited. Please notify us immediately of the error by return email, and please delete this email from your system. Thank you for your cooperation.

--

http://www.freelists.org/webpage/oracle-l

--

http://www.freelists.org/webpage/oracle-l Received on Tue Jan 31 2012 - 08:17:18 CST