Oracle 10g has recognized this trend toward considering CPU consumption by providing you with the ability to choose CPU-and-I/O based or exclusively I/O-based costing during SQL optimization (the 10g default is CPU-costing). 

In Oracle10g, system stats are gathered by default, and in Oracle9i the DBA must manually execute the dbms_stat.gather_system_stats package to get CBO statistics.

alter session set "_optimizer_cost_model"=choose; -- default value 

alter session set "_optimizer_cost_model"=io; 

alter session set "_optimizer_cost_model"=cpu; 

You can use this parameter to choose the best optimizer costing model for your particular database, based on your own I/O and CPU load:

• CPU_COST - The CPU cost of the operation as estimated by the cost-based SQL optimizer based on a secret algorithm. The value of this column does not have any particular unit of measurement; it is merely a weighted value used to compare costs of execution plans.
   

• IO_COST – Oracle estimates the I/O cost of the SQL based upon its knowledge of the settings for db_file_multiblock_read_count, the tablespace blocksize and the presence of indexes.  Oracle does NOT use data buffer statistics because Oracle cannot have any a-priori knowledge of whether a desired data block is already cached in the RAM data buffers.

Your choice of relative weighting for these factors depends upon the existing state of your database.  Databases using 32-bit technology (and the corresponding 1.7 gig limit on SGA RAM size) tend to have databases that are I/O-bound with the top timed events being spent performing disk reads:

Once 64-bit became popular, Oracle SGA sizes increased, more frequently-referenced data was cached, and databases became increasingly CPU-bound.  Also, solid-state disk (RAM SAN) has removed disk I/O as a source of waits:

The gathered statistics are captured via the dbms_stats package (in 9.2 and above) and CPU statistics are captured automatically in 10g and stored in the sys.aux_stat$ view.

As we have noted, in database where CPU is the top timed event may benefit from changing their SQL optimizer to consider the CPU costs associated with each execution plan.  Your cpu_count parameter is also important, and I have details here.

Sven Bombach notes that setting _optimizer_cost_model can result in dramatic performance improvements (3x faster):

Using CPU costing may not be good for databases that are I/O-bound.  Also, note that changing to CPU-based optimizer costing will change the predicate evaluation order of your query (See MetaLink bulletin 276877.1).

Jonathan Lewis also has some observation about how _optimizer_cost_model considers CPU costs:

"Oracle's choice of name for this feature was very misleading - it is not attempting to "optimize for CPU", or "minimize CPU". It is attempting to introduce a proper time-based estimate of the work required. This does introduce a CPU component of cost but, more significantly, changes the I/O component of the cost as well."

Turning on CPU costing

The default for the optimizer cost model is “choose”, meaning that the presence of CBO statistics will influence whether or not CPU costs are considered.  According to the documentation, CPU costs are considered when you collect SQL optimizer schema statistics with the dbms_stat.gather_system_stats package (the default behavior in Oracle10g), and CPU costs will be considered in all of your SQL optimization.

Important Note:  Prior to Oracle 10g, adjusting these optimizer parameters was the only way to compensate for sample size issues with dbms_stats.  As of 10g, the use of dbms_stats.gather_system_stats and improved sampling within dbms_stats had made adjustments to these parameters far less important.  Ceteris Parabus, always adjust CBO statistics before adjusting optimizer parms.  For more details on optimizer parameters, see my latest book "Oracle Tuning: The Definitive Reference". 

But it gets tricky because of Bug 2820066 where CPU cost is computed whenever optimizer_index_cost_adj is set to a non-default value.  Unless you have applied the 9.2.0.6 server patch set, your Oracle9i database may be generating CPU statistics, regardless of your CBO stats collection method.

To ensure that you are using CPU costing:

• In Oracle9i use dbms_stats.gather_system_stats to collect statistics

• Set the undocumented parm _optimizer_cost_model=cpu; 

Turning off CPU costing

As we noted, I/O-bound databases (especially 32-bit databases) may want to use I/O-based SQL costing.  The default optimizer costing in Oracle 10g is “cpu”, and you can change to “io” costing by using these techniques:

• Make sure that optimizer_index_cost_adj is set to the default value (Oracle9i bug 2820066)

• Add a "no_cpu_costing" hint in your SQL

• alter session set "_optimizer_cost_model"=io; 

• Set init.ora hidden parameter _optimizer_cost_model=io

Notes on Bug 2820066:

CPU cost is computed when optimizer_index_cost_adj is set to a non-default value.

• Range of versions believed to be affected:  Versions < 10.1.0.2 

• Platforms affected:      Generic (all / most platforms affected)

• This issue is fixed in 9.2.0.6 (Server Patch Set) and 10.1.0.2 

Bug description:  If optimizer_index_cost_adj is set to a non-default value CPU costs are calculated regardless of the optimizer cost model used.  If you have optimizer_index_cost_adj set and you are not using the optimizer CPU cost model, but explain plan shows that for queries not using domain indexes CPU costs are being calculated, you are probably hitting this bug.

In sum, CPU cost is always computed regardless of optimizer mode when optimizer_index_cost_adj is set in un-patched Oracle versions less than 10.1.0.2.

Next, let’s see how we can change from CPU-based to I/O-based SQL optimization when the processing characteristics of our database change on a regular basis.

 10g Note:  In Oracle 10g, you can achieve a similar result to reducing the value of optimizer_index_cost_adj by analyzing your workload statistics (dbms_stats.gather_system_stats).  Also note that utilizing CPU costing (_optimizer_cost_model) may effect the efficiency of plans with lower values for optimizer_index_cost_adj.

Bi-modal system configuration

It is not uncommon for databases to be bi-modal, operating OLTP during the day (CPU-intensive) and doing aggregations and rollups (I/O-intensive) at night.  I describe this technique in detail in my book “Oracle Tuning: The Definitive Reference”, but the idea is simple.

You can capture CPU and I/O statistics using dbms_stats and then swap-them in as your processing mode changes.  Most shops do this with the dbms_scheduler (dbms_job) package so that the statistics are swapped at the proper time.

Oracle Metalink has detailed script listings in Note 149560.1 “Collect and Display System Statistics (CPU and IO) for CBO usage”:

/* e.g. activate the DAY statistics each day at 7:00 am */

DECLARE

   I NUMBER;

BEGIN

   DBMS_JOB.SUBMIT (I, 'DBMS_STATS.IMPORT_SYSTEM_STATS(stattab => ''mystats'', s

tatown => ''SYSTEM'', statid => ''DAY'');', trunc(sysdate) + 1 + 7/24, 'sysdate

+ 1');

END;

/ 

/* e.g. activate the NIGHT statistics each day at 9:00 pm */

DECLARE

   I NUMBER;

BEGIN

   DBMS_JOB.SUBMIT (I, 'DBMS_STATS.IMPORT_SYSTEM_STATS(stattab => ''mystats'', s

tatown => ''SYSTEM'', statid => ''NIGHT'');', trunc(sysdate) + 1 + 21/24, 'sysdate

+ 1');

END;

/ 

*** ********************************************************

*** Initialize the OLTP System Statistics for the CBO

*** ********************************************************

1. Delete any existing system statistics from dictionary:

   SQL> execute DBMS_STATS.DELETE_SYSTEM_STATS;

   PL/SQL procedure successfully completed.

2. Transfer the OLTP statistics from OLTP_STATS table to the dictionary tables:

   SQL> execute DBMS_STATS.IMPORT_SYSTEM_STATS(-  

   >         stattab => 'OLTP_stats', statid => 'OLTP', statown => 'SYS');

   PL/SQL procedure successfully completed.

3. All system statistics are now visible in the data dictionary table:

   SQL> select * from sys.aux_stats$;

   SQL> select * from aux_stats$;

   SNAME                PNAME                   PVAL1 PVAL2

   -------------------- ------------------ ---------- --------------

   SYSSTATS_INFO        STATUS                        COMPLETED

   SYSSTATS_INFO        DSTART                        08-09-2001 16:40

   SYSSTATS_INFO        DSTOP                         08-09-2001 16:42

   SYSSTATS_INFO        FLAGS                       0

   SYSSTATS_MAIN        SREADTIM                7.581

   SYSSTATS_MAIN        MREADTIM               56.842

   SYSSTATS_MAIN        CPUSPEED                  117

   SYSSTATS_MAIN        MBRC                        9

   where 

   => sreadtim : wait time to read single block, in milliseconds

   => mreadtim : wait time to read a multiblock, in milliseconds

   => cpuspeed : cycles per second, in millions 

*** ********************************************************

*** CPU_COST and IO_COST in PLAN_TABLE table

*** ********************************************************

   SQL> explain plan for select * from oltp.test where c='AAAHxGAABAAAJS1AEZ';

   Explained.

   SQL> select operation, options, object_name, cpu_cost, io_cost

     2  from plan_table;

   OPERATION          OPTIONS              OBJECT_NAME    CPU_COST    IO_COST

   ------------------ -------------------- ------------ ---------- -

   SELECT STATEMENT                                          10500          1

   INDEX              UNIQUE SCAN          SYS_C002218       10500          1

   SQL> truncate table plan_table;

   SQL> explain plan for select * from oltp.test;

   Explained.

   SQL> select operation, options, object_name, cpu_cost, io_cost

     2  from plan_table;

   OPERATION          OPTIONS              OBJECT_NAME    CPU_COST    IO_COST

   ------------------ -------------------- ------------ ---------- ----------

   SELECT STATEMENT                                        2677480         27