Indexing the 'Order By' Clause

I got a call last week from a client running a Siebel application, he complained of slow performance on a particular screen, following a new release.

Since the objects statistics were not stale and the indexes had been rebuilt a few days ago, I extracted the execution plan of the screen’s query.

As you can see, the explain plan showed a near optimal access:

SELECT T1.CONFLICT_ID, T1.LAST_UPD, T1.CREATED, T1.LAST_UPD_BY, T1.CREATED_BY, T1.MODIFICATION_NUM, T1.ROW_ID, T1.ATTRIB_02, T1.ATTRIB_03,T1.X_OFFER_HISTORY
FROM SBL.S_ASSET_XM T1
WHERE (T1.TYPE = ‘Offer History’)
AND (T1.PAR_ROW_ID= :1)
ORDER BY T1.PAR_ROW_ID DESC, T1.CREATED DESC;

Id  Operation                          Name           Rows        Bytes     Cost
0   SELECT STATEMENT                                  1              74        4
1   SORT ORDER BY                                     1              74        4
2   TABLE ACCESS BY INDEX ROWID     S_ASSET_XM        1              74        1
3   INDEX RANGE SCAN                S_ASSET_XM_U1     1                        3

 

At this point, I extracted a statspack from the last 7 hour workload. The query showed up as the top query in the ‘Buffer gets’ and ‘Physical reads’ sections.

                                                   CPU      Elapsd
 Physical Reads   Executions Reads per Exec %Total Time (s) Time (s)

     28,679,016        1,050       27,313.3   64.7   3267.2  4488.59
 Buffer Gets   
     37,389,651                    35,609.2   14.3

 

Notice that the buffer hit ratio is not quite up to par as well as the execution timings.

 

Since the result must be sorted in descending order, let’s first check the selectivity of the fields from the ORDER BY clause. The current index has a 13% selectivity, the 2 fields in the ORDER BY have a 21% selectivity, not enough to make a strong difference, however since the result is required in descending order, let’s create an index sorted in descending order on the ORDER BY clause.

For this purpose, since I didn’t have a test environment with the same data volume, I created a virtual index in production:

  • alter session set “_use_nosegment_indexes” = true;

  • create index test_desc on sbl.s_asset_xm(par_row_id desc,created desc) nosegment;

 

Then I extracted the execution plan:


Id  Operation                          Name           Rows        Bytes     Cost
0   SELECT STATEMENT                                  1              74        1
1   TABLE ACCESS BY INDEX ROWID     S_ASSET_XM        1              74        1
2   INDEX RANGE SCAN                S_ASSET_XM_OBDESC 1                       12

 

At first look, it does look slightly faster. What really makes a huge difference is the access itself:

PLAN TABLE OUTPUT

   1 - filter("T1"."TYPE"='Offer History')
   2 - access(SYS_OP_DESCEND("T1"."PAR_ROW_ID")=SYS_OP_DESCEND(:Z))
       filter(SYS_OP_UNDESCEND(SYS_OP_DESCEND("T1"."PAR_ROW_ID"))=:Z)

 

The most selective field is accessed in descending order therefore eliminating the need for a costly sort. So let’s see after implementation the actual performance gain…

 

Following the index implementation, the query didn’t show up in the top queries anymore, so I extracted the hash value and ran the ?/rdbms/admin/sprepsql.sql, the performance gain was bigger than I expected.

                                                      
                     Statement Total      Per Execute     Before Index: Per execute
                     
        Buffer Gets:          17,707              5.5                       27313.3
         Disk Reads:             276              0.1                       35606.2
     Rows processed:          14,518              4.5
     CPU Time(s/ms):               1               .2
 Elapsed Time(s/ms):               2               .7
              Sorts:               0               .0
        Parse Calls:             745               .2
      Invalidations:               0
      Version count:               2
    Sharable Mem(K):              35
         Executions:           3,192

 

Pierre Roussin

Installing the Management Agent 10.2.0.x

I got an email yesterday from a desperate DBA who couldn’t find the proper instruction to deploy a management agent on a Linux server, that is management agent release 10.2.0.3.

Oracle has complicated things slightly.

Fortunately I’d done it before on our Linux and Aix platforms, and indeed that’s an awkward contraption that needs to be setup.

So here it is:

  1. Create the following directory: /u01/oracle/product/10.2.0.3/oms10g/sysman/agent_download/10.2.0.3
  2. Copy the downloaded zipped file into the directory
  3. Unzip the file
  4. Set your TEMP and DISPLAY environment variables
  5. Start your favorite X display software
  6. Move to ./linux/agent and run runInstaller
  7. During the course of the installation you will be ask to enter the location of you Oracle Base directory, in this case it is: /u01/oracle/product/10.2.0.3

 

After the installation, I noticed that sometimes the agentca -f  needs to be run, because of certain targets not being found during the initial agent configuration, within the configuration assistant.

 

Pierre

Optimizer Cost Model vs Implicit Sorting and Datatype Conversion

I recently stumbled on an issue for the first time, while migrating an application from 9i to 10g.

When migrating an application to 10g, one thing that needs to be considered is the optimizer cost model change.

In Oracle 9i, the cost model parameter (_optimizer_cost_model) is set to IO by default, while in 10G it is set to CHOOSE, seems to default to CPU in all cases however.

It doesn’t seem like much, but the implications are serious.

The IO cost model provides implicit data conversion and sorting.

In the case of my application, the developers had relied on implicit data conversion and data sorting when designing the reporting module.

In short, they had not included any ORDER BY clauses and were comparing CHAR’s with NUMBER’s.

 

Obviously when there were only 1 or 2 predicates in the WHERE clause, it didn’t make a difference wether sorting was implicit or not.

 

So after a migration, if you see strange sorting order and the ORA-01722: invalid number error, check your code.

A workaround is to set the  _optimizer_cost_model value to IO.

 

Pierre