16 SQL Tuning Overview

16.3.1.1 Tuning a Specific Program

If you are tuning a specific program (GUI or 3GL), then identifying the SQL to examine is a simple matter of looking at the SQL executed within the program. Oracle Enterprise Manager provides tools for identifying resource intensive SQL statements, generating explain plans, and evaluating SQL performance.

If it is not possible to identify the SQL (for example, the SQL is generated dynamically), then use SQL_TRACE to generate a trace file that contains the SQL executed, then use TKPROF to generate an output file.

The SQL statements in the TKPROF output file can be ordered by various parameters, such as the execution elapsed time (exeela), which usually assists in the identification by ordering the SQL statements by elapsed time (with highest elapsed time SQL statements at the top of the file). This makes the job of identifying the poorly performing SQL easier if there are many SQL statements in the file.

16.3.1.2 Tuning an Application / Reducing Load

If the whole application is performing poorly, or if you are attempting to reduce the overall CPU or I/O load on the database server, then identifying resource-intensive SQL involves the following steps:

1. Determine which period in the day you would like to examine; typically this is the application's peak processing time.

2. Gather operating system and Oracle Database statistics at the beginning and end of that period. The minimum of Oracle Database statistics gathered should be file I/O (V$FILESTAT), system statistics (V$SYSSTAT), and SQL statistics (V$SQLAREA, V$SQL, or V$SQLSTATS, V$SQLTEXT, V$SQL_PLAN, and V$SQL_PLAN_STATISTICS).

   See Also:

   • Chapter 6, "Automatic Performance Diagnostics" to learn how to gather Oracle database instance performance data

   • "Real-Time SQL Monitoring" for information about the V$SQL_PLAN_MONITOR view

3. Using the data collected in step two, identify the SQL statements using the most resources. A good way to identify candidate SQL statements is to query V$SQLSTATS. V$SQLSTATS contains resource usage information for all SQL statements in the shared pool. The data in V$SQLSTATS should be ordered by resource usage. The most common resources are:

   • Buffer gets (V$SQLSTATS.BUFFER_GETS, for high CPU using statements)

   • Disk reads (V$SQLSTATS.DISK_READS, for high I/O statements)

   • Sorts (V$SQLSTATS.SORTS, for many sorts)

One method to identify which SQL statements are creating the highest load is to compare the resources used by a SQL statement to the total amount of that resource used in the period. For BUFFER_GETS, divide each SQL statement's BUFFER_GETS by the total number of buffer gets during the period. The total number of buffer gets in the system is available in the V$SYSSTAT table, for the statistic session logical reads.

Similarly, it is possible to apportion the percentage of disk reads a statement performs out of the total disk reads performed by the system by dividing V$SQL_STATS.DISK_READS by the value for the V$SYSSTAT statistic physical reads. The SQL sections of the Automatic Workload Repository report include this data, so you do not need to perform the percentage calculations manually.

After you have identified the candidate SQL statements, the next stage is to gather information that is necessary to examine the statements and tune them.

16.3.2.1 Information to Gather During Tuning

The tuning process begins by determining the structure of the underlying tables and indexes. The information gathered includes the following:

1. Complete SQL text from V$SQLTEXT

2. Structure of the tables referenced in the SQL statement, usually by describing the table in SQL*Plus

3. Definitions of any indexes (columns, column orders), and whether the indexes are unique or non-unique

4. Optimizer statistics for the segments (including the number of rows each table, selectivity of the index columns), including the date when the segments were last analyzed

5. Definitions of any views referred to in the SQL statement

6. Repeat steps two, three, and four for any tables referenced in the view definitions found in step five

7. Optimizer plan for the SQL statement (either from EXPLAIN PLAN, V$SQL_PLAN, or the TKPROF output)

8. Any previous optimizer plans for that SQL statement

   Note:

   It is important to generate and review execution plans for all of the key SQL statements in your application. Doing so lets you compare the optimizer execution plans of a SQL statement when the statement performed well to the plan when that the statement is not performing well. Having the comparison, along with information such as changes in data volumes, can assist in identifying the cause of performance degradation.

16.5.3.1 Compose Predicates Using AND and =

To improve SQL efficiency, use equijoins whenever possible. Statements that perform equijoins on untransformed column values are the easiest to tune.

16.5.3.2 Avoid Transformed Columns in the WHERE Clause

Use untransformed column values. For example, use:

WHERE a.order_no = b.order_no

rather than:

WHERE TO_NUMBER (SUBSTR(a.order_no, INSTR(b.order_no, '.') - 1))
= TO_NUMBER (SUBSTR(a.order_no, INSTR(b.order_no, '.') - 1))

Do not use SQL functions in predicate clauses or WHERE clauses. Any expression using a column, such as a function having the column as its argument, causes the optimizer to ignore the possibility of using an index on that column, even a unique index, unless there is a function-based index defined that the database can use.

Avoid mixed-mode expressions, and beware of implicit type conversions. When you want to use an index on the VARCHAR2 column charcol, but the WHERE clause looks like this:

AND charcol = numexpr

where numexpr is an expression of number type (for example, 1, USERENV('SESSIONID'), numcol, numcol+0,...), Oracle Database translates that expression into:

AND TO_NUMBER(charcol) = numexpr

Avoid the following kinds of complex expressions:

• col1 = NVL (:b1,col1)

• NVL (col1,-999) = ….

• TO_DATE(), TO_NUMBER(), and so on

These expressions prevent the optimizer from assigning valid cardinality or selectivity estimates and can in turn affect the overall plan and the join method.

Add the predicate versus using NVL() technique.

For example:

SELECT employee_num, full_name Name, employee_id 
  FROM mtl_employees_current_view 
  WHERE (employee_num = NVL (:b1,employee_num)) AND (organization_id=:1) 
  ORDER BY employee_num;

Also:

SELECT employee_num, full_name Name, employee_id 
  FROM mtl_employees_current_view 
  WHERE (employee_num = :b1) AND (organization_id=:1) 
  ORDER BY employee_num;

When you need to use SQL functions on filters or join predicates, do not use them on the columns on which you want to have an index, as in the following statement:

varcol = TO_CHAR(numcol)

Instead, use them on the opposite side of the predicate, as in the following statement:

TO_CHAR(numcol) = varcol

16.5.3.3 Write Separate SQL Statements for Specific Tasks

SQL is not a procedural language. Using one piece of SQL to do many different things usually results in a less-than-optimal result for each task. If you want SQL to accomplish different things, then write various statements, rather than writing one statement to do different things depending on the parameters you give it.

It is always better to write separate SQL statements for different tasks, but if you must use one SQL statement, then you can make a very complex statement slightly less complex by using the UNION ALL operator.

Optimization (determining the execution plan) takes place before the database knows the values substituted in the query. An execution plan cannot, therefore, depend on what those values are. For example:

SELECT info 
FROM tables
WHERE ... 

AND somecolumn BETWEEN DECODE(:loval, 'ALL', somecolumn, :loval)
AND DECODE(:hival, 'ALL', somecolumn, :hival);

Written as shown, the database cannot use an index on the somecolumn column, because the expression involving that column uses the same column on both sides of the BETWEEN.

This is not a problem if there is some other highly selective, indexable condition you can use to access the driving table. Often, however, this is not the case. Frequently, you might want to use an index on a condition like that shown but need to know the values of :loval, and so on, in advance. With this information, you can rule out the ALL case, which should not use the index.

To use the index whenever real values are given for :loval and :hival (if you expect narrow ranges, even ranges where :loval often equals :hival), you can rewrite the example in the following logically equivalent form:

SELECT /* change this half of UNION ALL if other half changes */ info
FROM   tables 
WHERE  ... 

AND    somecolumn BETWEEN :loval AND :hival
AND   (:hival != 'ALL' AND :loval != 'ALL') 

UNION ALL 
SELECT /* Change this half of UNION ALL if other half changes. */ info
FROM   tables
WHERE  ... 

AND (:hival = 'ALL' OR :loval = 'ALL');

If you run EXPLAIN PLAN on the new query, then you seem to get both a desirable and an undesirable execution plan. However, the first condition the database evaluates for either half of the UNION ALL is the combined condition on whether :hival and :loval are ALL. The database evaluates this condition before actually getting any rows from the execution plan for that part of the query.

When the condition comes back false for one part of the UNION ALL query, that part is not evaluated further. Only the part of the execution plan that is optimum for the values provided is actually carried out. Because the final conditions on :hival and :loval are guaranteed to be mutually exclusive, only one half of the UNION ALL actually returns rows. (The ALL in UNION ALL is logically valid because of this exclusivity. It allows the plan to be carried out without an expensive sort to rule out duplicate rows for the two halves of the query.)

16.5.4.1 Use Caution When Managing Views

Be careful when joining views, when performing outer joins to views, and when reusing an existing view for a new purpose.

CREATE OR REPLACE VIEW emp_dept
AS
SELECT d.department_id, d.department_name, d.location_id,
     e.employee_id, e.last_name, e.first_name, e.salary, e.job_id
FROM  departments d
     ,employees e
WHERE e.department_id (+) = d.department_id;

SELECT v.last_name, v.first_name, l.state_province
  FROM locations l, emp_dept v
 WHERE l.state_province = 'California'
  AND   v.location_id = l.location_id (+);

--------------------------------------------------------------------------------
| Operation                 |  Name    |  Rows | Bytes|  Cost  | Pstart| Pstop |
--------------------------------------------------------------------------------
| SELECT STATEMENT          |          |       |      |        |       |       |
|  FILTER                   |          |       |      |        |       |       |
|   NESTED LOOPS OUTER      |          |       |      |        |       |       |
|    VIEW                   |EMP_DEPT  |       |      |        |       |       |
|     NESTED LOOPS OUTER    |          |       |      |        |       |       |
|      TABLE ACCESS FULL    |DEPARTMEN |       |      |        |       |       |
|      TABLE ACCESS BY INDEX|EMPLOYEES |       |      |        |       |       |
|       INDEX RANGE SCAN    |EMP_DEPAR |       |      |        |       |       |
|    TABLE ACCESS BY INDEX R|LOCATIONS |       |      |        |       |       |
|     INDEX UNIQUE SCAN     |LOC_ID_PK |       |      |        |       |       |
--------------------------------------------------------------------------------

SELECT department_name 
FROM emp_dept
WHERE department_id = 10;

16.5.4.2 Store Intermediate Results

Intermediate, or staging, tables are quite common in relational database systems, because they temporarily store some intermediate results. In many applications they are useful, but Oracle Database requires additional resources to create them. Always consider whether the benefit they could bring is more than the cost to create them. Avoid staging tables when the information is not reused multiple times.

Some additional considerations:

• Storing intermediate results in staging tables could improve application performance. In general, whenever an intermediate result is usable by multiple following queries, it is worthwhile to store it in a staging table. The benefit of not retrieving data multiple times with a complex statement at the second usage of the intermediate result is better than the cost to materialize it.

• Long and complex queries are hard to understand and optimize. Staging tables can break a complicated SQL statement into several smaller statements, and then store the result of each step.

• Consider using materialized views. These are precomputed tables comprising aggregated or joined data from fact and possibly dimension tables.

  See Also:

  Oracle Database Data Warehousing Guide for detailed information on using materialized views

16.5.9.1 Combine Multiples Scans Using CASE Expressions

Often, it is necessary to calculate different aggregates on various sets of tables. Usually, you achieve this goal with multiple scans on the table, but it is easy to calculate all the aggregates with a single scan. Eliminating n-1 scans can greatly improve performance.

You can combine multiple scans into one scan by moving the WHERE condition of each scan into a CASE expression, which filters the data for the aggregation. For each aggregation, there could be another column that retrieves the data.

The following example asks for the count of all employees who earn less then 2000, between 2000 and 4000, and more than 4000 each month. You can obtain this result by executing three separate queries:

SELECT COUNT (*)
  FROM employees
  WHERE salary < 2000;

SELECT COUNT (*)
  FROM employees
  WHERE salary BETWEEN 2000 AND 4000;

SELECT COUNT (*)
  FROM employees
  WHERE salary>4000;

However, it is more efficient to run the entire query in a single statement. Each number is calculated as one column. The count uses a filter with the CASE expression to count only the rows where the condition is valid. For example:

SELECT COUNT (CASE WHEN salary < 2000 
                   THEN 1 ELSE null END) count1, 
       COUNT (CASE WHEN salary BETWEEN 2001 AND 4000 
                   THEN 1 ELSE null END) count2, 
       COUNT (CASE WHEN salary > 4000 
                   THEN 1 ELSE null END) count3 
  FROM employees; 

This is a very simple example. The ranges could be overlapping, the functions for the aggregates could be different, and so on.

16.5.9.2 Use DML with RETURNING Clause

When appropriate, use INSERT, UPDATE, or DELETE... RETURNING to select and modify data with a single call. This technique improves performance by reducing the number of calls to the database.

16.5.9.3 Modify All the Data Needed in One Statement

When possible, use array processing. This means that an array of bind variable values is passed to Oracle Database for repeated execution. This is appropriate for iterative processes in which multiple rows of a set are subject to the same operation.

For example:

BEGIN
 FOR pos_rec IN (SELECT * 
   FROM order_positions 
   WHERE order_id = :id) LOOP
      DELETE FROM order_positions
      WHERE order_id = pos_rec.order_id AND
        order_position = pos_rec.order_position;
 END LOOP;
 DELETE FROM orders 
 WHERE order_id = :id;
END;

Alternatively, you could define a cascading constraint on orders. In the previous example, one SELECT and n DELETEs are executed. When a user issues the DELETE on orders DELETE FROM orders WHERE order_id = :id, the database automatically deletes the positions with a single DELETE statement.

16.6.1.1 Accessing SQL Test Case Builder from the Enterprise Manager

From Enterprise Manager, the SQL Test Case Builder is accessible only when a SQL incident occurs. A SQL-related problem is referred to as a SQL incident, and each SQL incident is identified by an incident number. You can access the SQL Test Case Builder from the Support Workbench page in Enterprise Manager.

You can access the Support Workbench page in either of the following ways:

• In the Database Home page of Enterprise Manager, under Diagnostic Summary, click the link to Active Incidents (indicating the number of active incidents). This opens the Support Workbench page, with the incidents listed in a table.

• Click Advisor Central under Related Links to open the Advisor Central page. Next, click SQL Advisors and then Click here to go to Support Workbench to open the Support Workbench page.

From the Support Workbench page, to access the SQL Test Case Builder:

1. Click an incident ID to open the problem details for the particular incident.

2. Next, click Oracle Support in the Investigate and Resolve section.

3. Click Generate Additional Dumps and Test Cases.

4. For a particular incident, click the icon in the Go To Task column to run the SQL Test Case Builder.

   Provide a location to store the test case, and also provide a name for the output.

The output of the SQL Test Case Builder is a SQL script that contains the commands required to re-create all the necessary objects and the environment.

16.6.1.2 Accessing SQL Test Case Builder Using DBMS_SQLDIAG

You can also invoke the SQL Test Case Builder manually, using the DBMS_SQLDIAG package. This package consists of various subprograms for the SQL Test Case Builder, some of which are listed in Table 16-1.

For more information on this package and all of its procedures and parameters, see Oracle Database PL/SQL Packages and Types Reference.