Re: Tuning Self-referencing Inserts

From: MacGregor, Ian A. <"MacGregor,>
Date: Mon, 3 Dec 2018 03:06:37 +0000
Message-ID: <BYAPR07MB4901BD9E432C99D460A8B333E2AE0_at_BYAPR07MB4901.namprd07.prod.outlook.com>



I'm afraid I posted some confusing information..

When I tried to obtain the plan for the run of a few days ago, the sql_id was already gone from the cache. This is why I posted last month's, where the behavior was the same.

Each time the Insert runs it is assigned a new sql_id, because PeopeSoft uses its own process_id as part the where clause. The same statement run by the same Oracle session may have a different Peoplesoft process_id associated with it. Running this particular PeopleSoft task the insert statement is run three diifferent times with three different sql_id's by the same Oracle session Each time the run time and the number of buffer gets increases. The plan, however, is the same each time the statement is run based on the plans hash value.

The timing of 11,000 seconds was for the plan posted; i.e both were gathered from a month ago. The 1,509,287,736 buffer gets were associated with that insert. and not with the whole instance. This was the second worse case for that paerticular PeopleSoft run. The next time the insert was called it took over 1.8 billion gets to run the same insert statement with the same plan and inserting the same number of rows. The 1.5 billion are for the second insert execution and the 1.8 billion are for the third.

I should not have just stuck with last month's information because there was an assumption I was trying to conflate the two runs. I apologize tart my post caused such confusion.


PeopleSoft inserts and deletes rows into what it calls temporary tables which are regular table segments. The number of rows in these tables can obviously have an impact on the time it takes to run a statement. There are several times during the month in which the same statement using the same plan runs without a problem. The problem only occurs at month's end. The table being inserted into is a "temporary" table. These temporary tables my also be truncated. In examining the PeopleSoft trace I saw was no indication of truncation of this table although other tables were truncated. However some process either deletes the rows or truncates the table because the number of rows is currently 0.


I posted the library cache pin query as a general question as I was surprised to see it all when querying v$session waits. I don't believe it is associated with a I could find no associated truncate statement. I was wondering if this could be RAC related.

I hope this clears any confusion. I renew my request for help.

Ian A. MacGregor
SLAC National Accelerator Laboratory
Computing Division
To offer the best IT service at the lab and be the IT provider of choice.



From: oracle-l-bounce_at_freelists.org <oracle-l-bounce_at_freelists.org> on behalf of Andy Sayer <andysayer_at_gmail.com> Sent: Sunday, December 2, 2018 3:15:00 PM To: Tim Gorman
Cc: ORACLE-L
Subject: Re: Tuning Self-referencing Inserts

The table is very small about 64K.

What's more important is how much of the table is being read and how often.

Here is an explain plan from a month ago.

An explain plan from a month ago could be very misleading, you seem to have access to the relevant privileges so you should be looking at the real execution plan in memory using

select * from table(dbms_xplan.display_cursor(sql_id=><the sql_id>));

The actual explain plan tells us the query is more complicated than a simple select * from table.

This should be taken with a pinch of salt as it's an explain plan (so unreliable) from a month ago (so the stats would have changed) We can see that most of the cost in the query is in the index range scan on line 10, it's not often you see an index range scan cost over 1,000 and it usually suggests you are using a filter with high index selectivity (reading lots of the index) - this is quite a frequent mistake I see, just because an index is being used, it doesn't mean it's being used well. The predicates section of the plan will tell you exactly what's going on, you should always include it (and the notes) when sharing the explain/execution plan. We can see that there are a few scalar subqueries being used, this tends to be an indicator of rushed SQL that can probably be rewritten to allow the CBO full power to execute the statement effectively. The more rows you have that drive these scalar subqueries, the more times they get run, the more work that needs to be done.

It took just over 11,000 seconds to insert 25,000 rows

Yeah, that is slow! It's certainly 10,000 times slower than just inserting 25,000 rows.

There were 1,509,2 87,736 buffer gets associated with the instance.

That as a statement is not helpful, we care about this statement not the instance as a whole (with no information of how long this was over)

There is something definitely wrong. This is a 2-node RAC system.

The likely truth is that the SQL is bad and it has a hugely inefficient execution plan as a result.

FYI the insert is called more than once. It runs pretty well the first time then gets dramatically worse.

That's no surprise, those scalar subqueries are reading from that table, for every row in that table that get's found from one filter you seem to read a lot of another index.

The above is the second worse case from about a month ago. There doesn't seem to be anything wrong with SQL

I would strongly disagree. Yes, you can probably make the statement more efficient with more appropriate indexes, but rewriting your scalar subqueries to better handle the volume would do wonders.

For the latest run, I decided to do periodic probes of v$session wait.

It sounds like v$session_event is what you're after, it stores aggregated information from all waits that the session experienced in its lifetime.

Each time I did I saw the library cache pin wait all with "waited short time". V$session did not show any blocking session. Neither did the query tailored to find blockers for this wait. I did a system state dump ...
is not in a wait:
...

Actually we can see that these waits were experienced a long time ago, since then the session has just been on CPU (which is probably due to the loopy reads of a lot of index which is highly cached).

...Here are the RAC stats from the run done about a month ago. This covers a 10 hour period ...

When you are looking at things at the instance level, it is very difficult to see the wood from the trees. Scoping is everything when it comes to performance investigations - just like you wouldn't do an AWR report over 24 hours when you care about a process that only runs at midday for 5 minutes, you shouldn't use a full instance AWR report for something running in one session. Instead, you can manually query v$active_session_history with filters to properly drill down. You can even join from v$active_session_history to v$sql_plan to see exactly what step of the plan was being executed most of the time. ...
The db link wait is not associated with this process.
That is the problem with using an unscoped report, there's plenty of irrelevant details that you'd need to ignore. Although, the top event is CPU by a long way.

In summary:
Rewrite the query to not use so many scalar subqueries Work out why those index range scans are so highly selective

Hope this helps,
Andy

On Sun, 2 Dec 2018 at 18:17, Tim Gorman <tim.evdbt_at_gmail.com<mailto:tim.evdbt_at_gmail.com>> wrote:

Ian,

I wrote a PL/SQL package called ASH_XPLAN<http://evdbt.com/download/presentation-real-time-sql-tuning-delivered-at-hotsos-symposium-2015-and-utoug-training-days-2015/> that combines DBMS_XPLAN with information from V$ACTIVE_SESSION_HISTORY to accumulate the ASH time spent in each step of the execution plan of a presently-running SQL statement. By running ASH_XPLAN in a separate SQL*Plus session repeatedly, I can see which way the wind blows in a very short time, without having to wait for the statement to complete. Of course, please don't consider accumulated ASH time to be exactly equal to elapsed time, as explained by Graham, Uri, and JB<http://www.rmoug.org/wp-content/uploads/QEW_Presentations/2012/ASH%20Architecture%20and%20Advanced%20Usage.pdf>.

I'm aware that the DBMS_SQLTUNE and SQL Monitor functionality (thanks Tim Hall!)<https://oracle-base.com/articles/11g/real-time-sql-monitoring-11gr1> in views like V$SQL_PLAN_STATISTICS does this as well, but I found it fun and educational to solve the problem myself. I was lonely for a nice PL/SQL project that week. :)

Either way, using ASH_XPLAN or DBMS_SQLTUNE, I strongly recommend displaying the actual elapsed time for each step of the plan in an actual execution to determine where the most time is being spent.

Nothing really matters in optimization except elapsed time, everything else is explanatory information.

Elapsed time (a.k.a. "life") is too short for guessing.

Hope this helps...

-Tim

On 12/1/18 22:17, MacGregor, Ian A. (Redacted sender ian for DMARC) wrote:

The table is very small about 64K. Here is an explain plan from a month ago.


| Id | Operation | Name | Rows | Bytes | Cost (%CPU)| Time |


| 0 | INSERT STATEMENT | | | | 2999 (100)| |

| 1 | LOAD TABLE CONVENTIONAL | | | | | |

| 2 | HASH UNIQUE | | 1 | 477 | 2999 (1)| 00:00:36 |

| 3 | NESTED LOOPS | | 1 | 477 | 2994 (1)| 00:00:36 |

| 4 | NESTED LOOPS | | 192 | 477 | 2994 (1)| 00:00:36 |

|*  5 |      HASH JOIN                    |                   |     2 |   254 |     6   (0)| 00:00:01 |


| 6 | NESTED LOOPS | | 8 | 608 | 3 (0)| 00:00:01 |
|* 7 | TABLE ACCESS FULL | PS_PSA_RULES_HDR | 8 | 504 | 3 (0)| 00:00:01 | |* 8 | INDEX UNIQUE SCAN | PS_PSA_RULES_LVL | 1 | 13 | 0 (0)| | |* 9 | TABLE ACCESS FULL | PS_PSA_RULES_LN | 32 | 1632 | 3 (0)| 00:00:01 | |* 10 | INDEX RANGE SCAN | PS_PSA_ACCTG_TA34 | 96 | | 1473 (1)| 00:00:18 | |* 11 | INDEX RANGE SCAN | PS_PSA_ACCTDSTGL4 | 1 | 20 | 2 (0)| 00:00:01 | |* 12 | INDEX RANGE SCAN | PS_PSA_ACCTDSTGL4 | 1 | 14 | 2 (0)| 00:00:01 |
| 13 | SORT AGGREGATE | | 1 | 19 | | |

| 14 | TABLE ACCESS BY INDEX ROWID| PS_PSA_ACCTDSTGL4 | 1 | 19 | 3 (0)| 00:00:01 |
|* 15 | INDEX RANGE SCAN | PS_PSA_ACCTDSTGL4 | 1 | | 2 (0)| 00:00:01 | |* 16 | TABLE ACCESS BY INDEX ROWID | PS_PSA_ACCTG_TA34 | 1 | 350 | 1494 (1)| 00:00:18 | -------------------------------------------------------------------------------------------------------

 It took just over 11,000 seconds to insert 25,000 rows There were 1,509,2 87,736 buffer gets associated with the instance. There is something definitely wrong. This is a 2-node RAC system. FYI the insert is called more than once. It runs pretty well the first time then gets dramatically worse. The above is the second worse case from about a month ago. There doesn't seem to be anything wrong with SQL

For the latest run, I decided to do periodic probes of v$session wait. Each time I did I saw the library cache pin wait all with "waited short time". V$session did not show any blocking session. Neither did the query tailored to find blockers for this wait. I did a system state dump


    Oracle session identified by:

    {

                instance: 1 (fsprd.fsprd1)

                   os id: 28554

              process id: 176, oracle_at_erp-fprd-oracle01

              session id: 406

        session serial #: 251

    }

    is not in a wait:

    {

               last wait: 120 min 20 sec ago

                blocking: 0 sessions

             current sql: INSERT INTO PS_PSA_ACCTDSTGL4 (PROCESS_INSTANCE, BUSINESS_UNIT_PC, PROJECT_ID, ACTIVITY_ID, RESOURCE_ID, LINE_SEQ_NBR, DEBIT_CREDIT, DST_USE, INTER_ORG_LEVEL, CONVERSION_RATE, ORG_TO_BOOK, COMBO_STATUS, PROJECT_FLAG, IU_ANCHOR_FLG, CONTRACT_NUM, CONTRACT_LINE_NUM, CONTRACT_PPD_SEQ, ACCT_PLAN_ID, EVENT_NUM, ACCOUNT, AL

             short stack: ksedsts()+465<-ksdxfstk()+32<-ksdxcb()+1927<-sspuser()+112<-__sighandler()<-qerixtFetch()+547<-subex1()+259<-subsr3()+183<-evaopn3()+2533<-expepr()+576<-evaiand()+51<-expeal()+23<-qerixtFetch()+800<-qerjotRowProc()+359<-qerhjInnerProbeHashTable()+491<-kdstf11001010000km()+617<-kdsttgr()+103608<-qertbFetch()+2455<-rwsfcd()+103<-qerhjFetch()+621<-qerjotFetch()+2025<-qerjotFetch()+2025<-qerghFetch()+315<-rwsfcd()+103<-qerltcFetch()+1223<-insexe()+691<-opiexe()+5632<-kpoal8()+2380<-opiodr()+917<-ttcpip()+2183<-opitsk(

            wait history:

              1.       event: 'library cache pin'

                 time waited: 0.000158 sec

                     wait id: 7433154         p1: 'handle address'=0x9c48940d0

                                              p2: 'pin address'=0x5eec1bd08

                                              p3: '100*mode+namespace'=0x41cdd00010002

              * time between wait #1 and #2: 0.000046 sec

              2.       event: 'library cache lock'

                 time waited: 0.000228 sec

                     wait id: 7433153         p1: 'handle address'=0x9c48940d0

                                              p2: 'lock address'=0x5eeef6eb0

                                              p3: '100*mode+namespace'=0x41cdd00010002

              * time between wait #2 and #3: 0.000516 sec

              3.       event: 'library cache pin'

                 time waited: 0.000250 sec

                     wait id: 7433152         p1: 'handle address'=0x9a2fdaf10

                                              p2: 'pin address'=0x99a42e960

                                              p3: '100*mode+namespace'=0x2160300010002

    }


I thought the lock/pin was against an index, but the 100 indicates it is against the table. However despite their showing up when I probed v$session_wait they are not shown to be a significant wait. I was surprised to see them however, because I associated such locks/pins with DDL not DML.

The last point is that this is a 2 node RAC system, the database was cloned to a single instance database and the program ran much more efficiently. The longest time for an insert statement was less than an hour.

Here are the RAC stats from the run done about a month ago. This covers a 10 hour period

The top RAC associated wait was ranked 10th

  * •

Event

Waits

Total Wait Time (sec)

Wait Avg(ms)

% DB time

Wait Class

DB CPU 38K

96.6

SQL*Net message from dblink

3,038

497.6

164

1.3

Network

db file sequential read

1,233,886

474.4

0

1.2

User I/O

control file sequential read

149,191

96.6

1

.2

System I/O

direct path read

71,377

89.5

1

.2

User I/O

SQL*Net more data to dblink

45,096

51.1

1

.1

Network

log switch/archive

21

43.8

2086

.1

Other

recovery area: computing obsolete files

10

22.3

2234

.1

Other

log file sync

33,439

19.2

1

.0

Commit

gc current block 2-way

42,984

16

0

.0

Cluster

The db link wait is not associated with this process.

Ian A. MacGregor
SLAC National Accelerator Laboratory
Computing Division
To offer the best IT service at the lab and be the IT provider of choice.



From: oracle-l-bounce_at_freelists.org<mailto:oracle-l-bounce_at_freelists.org> <oracle-l-bounce_at_freelists.org><mailto:oracle-l-bounce_at_freelists.org> on behalf of Andy Sayer <andysayer_at_gmail.com><mailto:andysayer_at_gmail.com> Sent: Friday, November 30, 2018 2:56:28 PM To: dmarc-noreply_at_freelists.org<mailto:dmarc-noreply_at_freelists.org> Cc: ORACLE-L
Subject: Re: Tuning Self-referencing Inserts

I think we need to take a few steps back.

It’s hard to explain why such a statement would cause noticeable issues if the table is really not that large. A simple insert into <target> statement using a select from <target table>, won’t take a silly amount of time - it’s just the time to read the table blocks using a full tablescan then the time to update the indexes which might be some overhead but no different to any other insert statement of the same volume.

Perhaps this this insert is within some loop? Perhaps the self-reference is written so that a silly execution plan is being used. It would be great if you could share the execution plan you are seeing and the SQL being executed. Is there PL/SQL involved?

One other thing that springs to mind that could have an impact is DML error logging (it sounds like a situation where unique keys could give you errors?)

As for library cache pins, this is unusual for an insert/select statement. Things like running a PL/SQL procedure which another session is trying to recompile while another session is executing it will bring around this behaviour. Perhaps there is a function being called by your statement. Again, what exactly are you seeing? This is a block so you should be able to investigate what the blocking session is doing by checking v$session.

Hope this helps,
Andy

On Fri, 30 Nov 2018 at 18:32, MacGregor, Ian A. <dmarc-noreply_at_freelists.org<mailto:dmarc-noreply_at_freelists.org>> wrote:

The problem is with a PeopleSoft statement which is based on a select statement which references the table being inserted. The problem stems from having to read and build the indexes of the table which is not that large. The buffer gets are extremely high.

Is there generic answer to this problem. Would it help to rebuild with a higher percent free in an attempt to have few rows per block so as to lessen contention.

Also for library cache pins I don't understand how to reads the p3raw value. I think the problem primarily lies with the maintenance m of the unique index bit I am. not 100% confident.

Ian A. MacGregor
SLAC National Accelerator Laboratory
Computing Division
To offer the best IT service at the lab and be the IT provider of choice.

--
http://www.freelists.org/webpage/oracle-l
Received on Mon Dec 03 2018 - 04:06:37 CET

Original text of this message