Huge performance gains from using arrays and bulk collects.

Consider a simple select which I ran in SQL*Plus connected as SYSTEM:

select * from dba_users;

I ran a trace on the session and did a tkprof on the result.

call     count       cpu    elapsed       disk      query    current        rows
------- ------  -------- ---------- ---------- ---------- ----------  ----------
Parse        1      0.29       0.76        201        205          0           0
Execute      1      0.00       0.00          0          0          0           0
Fetch     1353      0.99       1.79       1303       4536          0       20276

What do we notice about these stats? Parse count is 1, which is expected, and it will still be 1 if I run that select again. Check my previous post where I’ve talked about bind variables and why it’s bad for performance to repeat parsing, especially if it is hard parsing. Also, execute count is 1, which is expected, and you want to keep that to 1 as far as possible per select. I’ve seen applications do exotic selects with rowid’s in cursors and the execute column has the same number as the fetch column, which is the most inefficient way to get your data back.

If you had done ‘set array 2‘ in SQLPlus, then the Fetch stats would have looked like this:

call     count       cpu    elapsed       disk      query    current        rows
------- ------  -------- ---------- ---------- ---------- ----------  ----------
Fetch    10134      1.35       1.88        676      12896          0       20266

Funny enough, when I then used ‘set array 1‘, Oracle ignored the setting and used a value of 2. Somewhere in the recesses of the Oracle core there must be an AI that is intelligent enough to question and compensate for the obvious stupidity of getting back one row per fetch!

By running a simple select at the SQL prompt in SQLPlus, Oracle created an implicit cursor which you can see if you look at the raw trace file before it was tkprof’d.

PARSING IN CURSOR #18446744071486994888 len=23 dep=0 uid=5 oct=3 lid=5 tim=11879507487109 hv=665675061 ad='47520d368' sqlid='ckyh1gcmuut9p'
select * from dba_users

Now, why were there only 1353 fetches done for 20276 rows of data returned? Implicit cursors will use the default array size of 15 which is the SQLPlus (client tool) setting for array[size].

Let’s write an explicit cursor via an anonymous block to see what Oracle does.

  cursor c1 is
    select * from v$session;
  for c in c1 loop 
  end loop;  

Checking the tkprof result:

call     count       cpu    elapsed       disk      query    current        rows
------- ------  -------- ---------- ---------- ---------- ----------  ----------
Parse        1      0.01       0.01          0          0          0           0
Execute      1      0.00       0.00          0          0          0           1
Fetch        0      0.00       0.00          0          0          0           0
------- ------  -------- ---------- ---------- ---------- ----------  ----------

That’s funny… why no result in the Fetch row? Going to Oracle’s documentation, it states that ‘Fetches are only performed for SELECT statements‘. Further more, as shown in a previous post, the anonymous block is declared and run in the session being traced, so it shows a value for Parse and Execute. The “select * from v$session” is run on the server as the cursor c1 that we had defined.

So going back to our trusty V$SQL which shows all statements executed since startup:

from V$SQL
where upper(sql_text) like '%V$SESSION%';

SELECT * FROM V$SESSION   1             1            7

That’s interesting… Fetches=7. The select returned 696 rows so this means that the array size has been set to 100. Either that, or Oracle has performed an implicit bulk collect as stated by PL/SQL guru Steven Feuerstein here :
Oracle Database automatically optimizes cursor FOR loops to perform similarly to BULK COLLECT queries… and fetch 100 rows at a time.

From the oracle documentation of SET ARRAY, it states that “Values over approximately 100 provide little added performance.” So how can we use much larger “arrays” and get the intuitive performance benefit of collecting a large number of rows into a memory structure? This is where we have to use BULK COLLECT and FORALL.

Consider this code which will show the usage of BULK COLLECT. I am not going to exhaustively list all the BULK usage options, such as SQL%BULK_EXCEPTIONS, as that is the purpose of the documentation, and I don’t want to sidetrack from the point I am making about huge performance gains. However it is worth mentioning the LIMIT clause which limits the number of rows returned by the BULK COLLECT.

  obj_data ARRAY;
    SELECT * FROM dba_objects;
  OPEN c1;
      FETCH c1 BULK COLLECT INTO obj_data;     
  CLOSE c1;

The trace shows the code but we have to check the stats from V$SQL:

from V$SQL
where upper(sql_text) like '%DBA_OBJECTS%';

SELECT * FROM DBA_OBJECTS   1             1            1         3100

Lo and behold, we get the awesome result of Fetches=1 for a select of over 100,000 rows! Why is this so significant instead of multiple fetches? If the application is client/server, then every fetch is an extra network round trip between client (thin or fat) and the database server. If the SQL is on the server then every fetch is an extra context switch. In both cases, this can result in huge performance benefits on busy production servers that execute thousands or even millions of queries per hour. Have you ever wondered about innocuous “idle” SQL*Net waits events? The majority of these could be as a result of not using bulk collects and arrays efficiently! Every one of them means an extra network round trip and this means wastage of CPU and I/O resources that could have been used for the next process in the run queue.

Now could the bulk mechanism be used for DML (insert, update, delete)? Sure, this is where the FORALL statement is used. In the example above I could have inserted this code after the FETCH:

FORALL i IN 1..obj_data.COUNT
  INSERT INTO some_table VALUES obj_data(i);

This would load the entire array with one write and therefore one roundtrip or context switch, giving the same benefit the BULK COLLECT had done.

Note that BULK COLLECT can be used with your dynamic SQL in the EXECUTE IMMEDIATE statement, but as stated in a previous post it is usually better to limit the use of dynamic sql in busy OLTP production environments.

Array usage in some other development tools

We’ve already seen the use of array[size] in SQL*Plus. This can also be used in Forms.

Pro*C can use pre-fetching transparently with:

$ proc .... PREFETCH=200

This can also be set at compilation time with ANSI C and C++ and overrides the usage of CURSIZE within the C code.

Using the property weblogic.oci.cacheRows, Weblogic JDBC implements array fetching using its Row Caching feature.

Magic” is the RAD tool we’ve been using and arrays can be set at table level. Every development language worth its salt should be able to use arrays and if their oracle drivers are properly written then the relevant OCI calls can be made for arrays and bulk collects.

If you want to experience a massive performance gain in both your DSS and OLTP production environments, make complete and gratuitous use of arrays and bulk collect mechanisms in your code.


Common mistakes made with indexes. A guide for the developer and DBA.

All too often a performance problem can be tracked down to a basic error with an index. So with this post I’ve decided to create a reference which will be of interest to DBA’s and SQL developers.

Full table scan (FTS)

select surname from employees;

There is a missing where clause here so Oracle will choose a FTS and the explain plan will show TABLE ACCESS FULL EMPLOYEES. If it’s a large table then a FTS is usually bad, but not always. There are cases when a FTS is better, for example when Oracle can read the whole table in one or a few I/O’s, or when the clustering factor is bad (will explain later). Note that if your db_file_multiblock_read_count is set to 32, for example, and your block size is 8K, then one I/O will be 32×8=256K. So if your table fits into 256K then a FTS is fine as the entire table will be read in one I/O.

select surname from employees
where salary>3000;

will result in a FTS if the “salary” field is not indexed.

Full index scan (FIS)

select emp_no
from employees
order by emp_no;

With the missing where clause, this will result in an INDEX FAST FULL SCAN if you check your explain plan. This is also usually bad but not always. If your index can be read in a few I/O’s then a FIS could be a good thing.

Cartesian (merge) joins

select *
from invoices, transactions;

In this silly example, this will result in a MERGE JOIN CARTESIAN in the explain plan. The result will match each row of the first table with each row of the second, and this is almost always a bad idea and can result in a performance loss of orders of magnitude. One way of fixing that SQL could have been:

select *
from invoices i, transactions t
where i.inv_no=t.inv_no;

with indexes on inv_no in both those tables, therefore creating a table join.

Use the correct data types in the WHERE clause!

select surname from employees
where date_joined>20130101;

In this example Oracle had to do an implicit type conversion between the number passed to the where clause, instead of the correct to_date conversion of a date formatted variable. It will either ignore the index on date_joined or choose a FIS on that index, instead of doing a range scan.

Using functions on the left side of the where clause

select surname from employees
where to_date(date_terminated,'YYYYMMDD') > sysdate-7;

This will not make use of the index on that varchar2 “date” field (don’t ask! there are RAD apps that do this!) and will result in a FTS.

The way to fix this is to use the function on the “right” side or “value comparison” side:

select surname from employees
where date_terminated > to_char(sysdate-7, 'YYYYMMDD');

This is because the field or column is now being compared to a value, instead of the field itself.

Skip scans, range scans, unique scans
Oracle 9i introduced skip scans on concatenated indexes (indexes with more than one field) which worked like this.

select name, surname from employees
where surname='SMITH';

with name+surname indexed in index1. The Cost Based Optimizer (CBO) may or may not decide to perform a SKIP SCAN on this table, depending on its anticipated cost of each alternative (which you can find in great detail if you do a 10053 trace). So the DBA should test this before giving advice on skip scans, perhaps with using invisible indexes. Another common mistake with the above example is for a developer to request two indexes, index1 on name+surname and index2 on name. The index on name is redundant because that is the first field in index1. B-tree indexes are used by default and that is natural to how they work. Looking at the tree below:
Oracle fetches and sorts the columns to be indexed and stores the rowid along with the index value for each row. The upper blocks (branch blocks) of a B-tree index contain index data that points to lower-level index blocks. They contain the minimum key prefix needed to make a branching decision between two keys. The lowest level index blocks (leaf blocks) contain every indexed data value and a corresponding rowid used to locate the actual row. So if a name “JOHN” is searched, oracle will traverse down the branches until it reaches the leaf blocks starting with the letter J. Depending on how much data there was, and how the index was created, the selectivity could have been further refined with “JO” or even “JOH”. If there are multiple JOHNs then oracle would do a range scan. If there was one JOHN then oracle would find it in the leaf block, and retrieve the table data by ROWID. This is called a unique scan and is shown in the explain plan as TABLE ACCESS (by INDEX ROWID). This is the most efficient way of using an index if there is uniqueness in your data (and a primary key or unique index). If you have a non-unique index, then watch out for range scans. This is a matter of knowing your data. If your data has grown appreciably over time then perhaps it’s time to archive, or analyze your tables/indexes, or consider further reducing your result sets by changing your where conditions.

Careful use of not-equal and like wildcards
If you searched for name<>’JOHN’ or name!=’JOHN’ in the above example, oracle has no way of traversing the tree to find any sort of range. So it will do a FTS. The same applies to name like ‘%OHN’… a FTS will occur. Note that name like ‘JOH%’ is fine because this is the same as saying name > ‘JOH’ and a range scan will be done.

If you had said where name is null, then all indexes will be suppressed and a FTS will be done. This is because NULLs are not stored in any indexes except for bitmap indexes.

Clustering Factor (CF)… for DBA eyes only!
The CF of an index is a measure of how ordered it is compared to its table. This can be determined from the CLUSTERING_FACTOR column in DBA_INDEXES. If this number is close to the number of blocks in the table then the index is well ordered; if it is close to the number of rows then it is randomly ordered. Let’s say the CF is 1000 and the number of blocks is 100. Then we say that the CF is off by an order of magnitude of 10, and for each row in the index, about 10 table blocks would have to be read (because there can be many blocks per row). So in the real world, you might see a session that is doing a huge number of reads and you notice a ‘db file sequential read’ wait event. Looking at the session’s SQL, you notice a range scan. You know this table’s data and cannot understand why a range scan could be doing that many reads… the savvy DBA would know that the only thing left to look at is the CF. This is a difficult problem to fix. If there is only one index on the table, then the table could be rebuilt in the same order as the index. Even if this is production, you can perform an online table re-org using DBMS_REDEFINITION. However, if there is more than one index, rebuilding the table to suit that index is needless because it will adversely affect the CF of the other indexes. So what to do? In some cases you can check if a FTS is actually faster by doing some real world timings. Make sure your index is in a locally managed tablespace. If the b-tree level (BLEVEL) is high (typically over 4) then it might be a candidate to be rebuilt. This might improve the depth of the leaf blocks immediately, but the index will degrade over time especially if there are frequent changes to the table. Your final option might be to rewrite the query if performance has become completely unacceptable due to bad CF. Since this section was for DBA’s only, I would imagine a whole new world of curious developers would now have become CF experts! DBA’s brace yourselves!

Parameters affecting indexes
Probably the most important parameter is optimizer_index_cost_adj. The default value is 100, and it must be set to a lower value for OLTP systems. Setting it lower makes oracle favour index scans over table scans. In other words, if the optimizer_index_cost_adj=50 then the CBO views indexes as 50% less expensive. Therefore the lower the value is, the less expensive the index… but only in the way that it is viewed by the CBO. If you have the luxury of testing this on production, a good value to start with is 60. You can reduce it by 10 at a time then measure the overall performance. Some OLTP systems have it set to between 10 and 30 and claim to have drastically improved performance; ours is set to 60 and we seem to have a good balance between index and table scans.

What else affects the CBO and indexes? Select from V$SYS_OPTIMIZER_ENV and you will find optimizer_index_caching. Setting optimizer_index_caching higher favours nested loop joins over hash or sort merge joins. Like optimizer_index_cost_adj, this will only influence the behaviour and decision making of the CBO, so be sure to test thoroughly before attempting to make any change. I left this at the default.

There is a lot more to be said about indexes but to keep this post succint, I’ve stayed on subject and highlighted common mistakes. Indexes are probably the most important objects to be considered by both DBA’s and developers when it comes to optimizing SQL and performance tuning. So if they are used correctly and kept in good health, you will be well on your way to an efficiently running database.

Optimizing Performance with Cursor Sharing and Shared Pool behaviour in Oracle 11g R2 (


With this post I’ve decided to create an introductory guide to how Oracle 11g R2 handles literals and binds in the shared pool, how cursor sharing works, and how to write your cursors to maximize performance.

The first parameter to look at is cursor_sharing. Trying to set it to SIMILAR is now deprecated and you will get a message in your alert log saying “WARNING: The parameter cursor_sharing was found to be set to the value SIMILAR. This setting will be ignored and cursor sharing will operate as though the value was set to FORCE instead.”

So let us set it to EXACT and see the behaviour. This means that statements with identical text share the same cursor, and in the documentation Oracle recommends that this is used in DSS (decision support systems / reporting / warehouses).

I created a table called TEST (xnum number) with values from 1 to 10, with an index on xnum.

  for x in 1..10 loop
    insert into test values (x);
  end loop;

Let’s analyze:

exec DBMS_STATS.GATHER_TABLE_STATS(ownname=>'TEST', tabname=>'TEST', cascade=> TRUE);

Let’s try this statement to see how the optimizer responds:

select * from test where xnum=1;

Now selecting from V$SQL which conveniently holds statistics since instance startup:

  upper(sql_text) like '%TEST WHERE%';

the statement appears exactly like it is, with the literal value of 1, as expected.

I ran a few more literal selects and here are some selected rows:

select * from test where xnum=2  a4hrnbrz05uwy             0       2          1           1
select * from test where xnum=1  cgsb5m608anzk             0       4          2           2
select * from test where  xnum=2 a91a8av266tb9             1       2          1           1

Notice from this that adding a single space in that last statement created a new entry in the shared pool with a new SQL_ID. The point I am really trying to make here is that if you write your applications to generate literal SQL like I’ve done, and you use cursor_sharing=exact, then this is the least optimal way to write sql. Imagine if these statements run thousands or even millions of times in an hour in a large environment… in addition to Oracle having to “hard parse” each new statement as it comes in, it will also have to find free chunks in which to place these statements in the shared pool, while ageing older statements out of the shared pool. So the CPU works extra hard in parsing and managing the shared pool. In time this creates shared pool fragmentation which can lead to many horrors such as library cache pin waits in a busy environment, which can lead to shutdowns and no way of fixing it unless you rewrite the application or change the cursor_sharing to force. However, we will get to that later.

Let’s do the same via an anonymous block and run it twice:

  x number;
  v_xnum number;
  select xnum into v_xnum from test where xnum=x;
  DBMS_OUTPUT.PUT_LINE('Answer: ' || TO_CHAR (v_xnum));

declare x number; v_xnum number; begin x:=3; select xnum into v_xnum from test where xnum=x; DBMS_OUTPUT.PUT_LINE('Answer: ' || TO_CHAR (v_xnum)); end; 
                db6b351m50nkq             1       0          2           2
                7y76r2vywmrv5             1       2          2           1

This is a little better than literal SQL but not much. Oracle gives the block its own SQL_ID which uses a great deal more shareable memory than simple statements, but at least there is only one copy in the shared pool. However the problem arises when these blocks become much larger in an OLTP environment with complex systems with thousands of anonymous blocks especially if they make calls to packages. All underlying code gets loaded into free shareable memory and Oracle has to work to find space (chunks) to load them in. If they are not pinned (using DBMS_SHARED_POOL.KEEP) then you will find the loaded version (from v$sql) count increasing which results in fragmentation in a busy OLTP environment. The second row with the :B1 will be explained with the cursor example below.

Note that the same applies to dynamic SQL (typically generated in packages then executed with “execute immediate”). Badly written dynamic SQL can also cause all kinds of fragmentation problems in your shared pool. “Execute immediate” with the “using” clause does make use of bind variables and is acceptable in some cases, but in my opinion is a lazy developer’s alternative to making proper use of cursors. Cursors are natural to Oracle. Oracle loves well written cursors, and your users will love you and Oracle in return. When performance is bad, they always come to you, the omniscient DBA right? And when you fix it, it was your fault in the first place. The manager or helpdesk person or end user never knew the pain you went through to troubleshoot and find that BAD statement or application server with the memory leak or that storage layer bottleneck…

Aah. Breathe.
Let’s write simple PL/SQL using a cursor in a procedure to see what happens.

create or replace
PROCEDURE test_num (p_num IN NUMBER, p_num_out IN OUT NUMBER) IS
  CURSOR c_num IS
    select xnum from test where xnum=p_num;
  for c in c_num loop
    p_num_out := c.xnum;
  end loop;

DECLARE v_pnum number;
  DBMS_OUTPUT.PUT_LINE('Answer: '||TO_CHAR(v_pnum));

I ran the procedure 5 times with values of 5 and 6.

SELECT XNUM FROM TEST WHERE XNUM=:B1 7y76r2vywmrv5             1       5          5           2

Notice firstly that Oracle transforms the statement (which is the cursor in the procedure) into upper case. Then it assigns a bind variable “:B1” to the statement. When I ran the procedure with value 5, parse calls was 1. When I changed the value to 6, parse calls became 2. This is what is known as a “soft parse”. Oracle peeked, detected a different value and checked statistics, instead of doing a hard parse which has the overhead of syntax and permissions checking. Soft parses are better for performance than hard parses. The advantage of using bind variables is that there is only one version of the statement in the shared pool, which leads to huge gains in performance because of reduced parsing, fragmentation and ageing.

HOWEVER, and this is a very important caveat: using cursor_sharing=exact in busy environments will ONLY work if you have perfectly written applications using cursors as in the above example. Now, which of us can say that we support perfect applications? Invariably, applications throw in anonymous blocks and dynamic SQL. They throw the odd literal statement in. The untrained business analyst connects and runs something naughty. The DBA adds to the fun with exotic scripts. Oracle itself adds dictionary statements that are far from optimal. Monitoring tools can throw in some horrors like segment checking. Woe betide you if the CIO connects. So you get the picture… you will never work in a perfect environment.

The corollary to this is that cursor_sharing=exact is fine for DSS environments, as Oracle states in their 11g R2 documentation:
“If you set CURSOR_SHARING, then Oracle recommends the FORCE setting unless you are in a DSS environment”.
The reason for this is that you rarely get shared pool fragmentation in a reporting or warehouse environment. Users there typically run long-running reports and statements which usually RELY on an expected explain plan that worked the same way in the dev and test environment (there are exceptions to this but I don’t want to go too far off topic. In IT, for everything that is said, there is usually an exception. This is not an exact science). Note that the default value is EXACT so it’s up to you, the conscientious DBA, to use the correct setting for OLTP production.

Just a final note about cursors before we move on. We’ve shown how to reduce parsing which gives huge CPU benefits. Now every statement has to go through a parse/execute cycle, and fetch if it returns a result set. The ideal statement, in a perfect world, is 1 parse, execute many times, and fetch many rows per execute. If you find that the number of executes is too close to the number of parses, then it usually means that you’re opening and closing your cursor too often. If you are returning one row per execute, then you need to investigate the use of Oracle’s array mechanism via BULK COLLECT and FORALL. This has huge I/O benefit which I will discuss in another post.

Before I go on to talk about cursor_sharing=force, it’s worthwhile to talk a little about how oracle treated binds in the past compared to how it does now. This will be the subject of another long post when I have the energy to do so. I’ve been using Oracle since Oracle 7 and even in those days, statements were sent to the shared pool, parsed, then executed based an explain plan derived from complicated rule-based or cost-based optimizer statistics and histograms based on whether the objects were analyzed or not. As rule-based moved towards obsoletion, cost-based became the norm and in Oracle 9i, the CBO (cost based optimizer) generated explain plans based on its histogram knowledge of the data distribution of a table. All was well in DBA world until 10g came along and the CBO decided to default to bind peeking which was the bane of many a DBA and some even chose to turn it off with _optim_peek_user_binds. Bind peeking worked like this: if you had a table with skewed data (for example a table of car sales) and if someone queried a common car (e.g. Beetle) from that table, Oracle would peek at this bind at the first hard parse, then generate an explain plan with full table scan based on this peek. Worse, it would lock the explain plan. So, if someone wanted to query sales statistics of an uncommon car (e.g. Rolls Royce) then the same full table scan would be done in future. This led people to either removing statistics for certain tables, turning off bind peeking, or being really clever and writing queries with literals in them.

So then came 11g with adaptive cursor sharing, and this is Oracle’s summary from metalink 1169017.1:
“Our recommendation is to set cursor_sharing = FORCE. This setting maximizes cursor sharing while leveraging the Adaptive Cursor Sharing framework to generate multiple execution plans based on different literal value ranges if necessary.”
What this means is that instead of peeking at the first value at hard parse time, Oracle now observes ranges of values passed to the bind variables. Based on this “bind sensitivity”, Oracle decides whether to recalculate the plan and add it to the profile of the SQL ID in question. So a statement can now have multiple explain plans, which in itself can be problematic which is the topic of another discussion where I will discuss Baselines and SQL Plan Management.

OK so let’s change cursor_sharing:
alter system set cursor_sharing=force scope=both;
and bounce the database to clear the shared pool.
Let’s do this again to see what happens:

select * from test where xnum=1;

from v$sql 
  upper(sql_text) like '%TEST WHERE XNUM%';

select * from test where xnum=:"SYS_B_0" 50vnp9xux8us2             1       2          1           1                 N             N            0

So instead of a literal in the shared pool, oracle now FORCEs the bind variable. This is system generated and will always have the SYS_ prefix so that you can find them in your shared pool.

Now let’s run that select again with a value of 8 to see the result:

select * from test where xnum=:"SYS_B_0" 50vnp9xux8us2             1       4          2           2                 N             N            0

As expected, it was executed and parsed again. Please note that this is a hard parse. This is the problem with system generated binds which are converted from statements with literals or from dynamic sql. In the long run, one parse per execute is very heavy on CPU especially if the statements run millions of times in a busy production environment.

Let’s see what it does with our anonymous block with xnum=3:

declare x number; v_xnum number; begin x:=3; select xnum into v_xnum from test where xnum=x; DBMS_OUTPUT.PUT_LINE('Answer: ' || TO_CHAR (v_xnum)); end;                       
         db6b351m50nkq             1       0          1           1                 N             N            Y            0
         7y76r2vywmrv5             1       1          1           1                 Y             N            Y            0

It created the block with its own ID and then converted the select to use a bind variable. This is not a system bind because the select was done with a variable in a code block (no literal value was used in the code block). Note that IS_BIND_SENSITIVE is marked Y (yes) which means that the select can potentially get a new plan in future. If IS_BIND_AWARE=Y then it means that the plan has changed. At this point, the table has uniformly distributed data. It has 10 rows, with numbers from 1 to 10 in an index, so Oracle has no need to change the plan based on the values it has seen.

Running the select TWICE with x=5 results in TWO parses of the anonymous block but just one parse of the select, so Oracle was happy to continue with the original explain plan.

declare x number; v_xnum number; begin x:=3; select xnum into v_xnum from test where xnum=x; DBMS_OUTPUT.PUT_LINE('Answer: ' || TO_CHAR (v_xnum)); end; 
         db6b351m50nkq             0       0          1           1                 N             N            Y            0
declare x number; v_xnum number; begin x:=5; select xnum into v_xnum from test where xnum=x; DBMS_OUTPUT.PUT_LINE('Answer: ' || TO_CHAR (v_xnum)); end; 
         c5khy7t3xkpnk             1       0          2           2                 N             N            Y            0
         7y76r2vywmrv5             1       3          3           1                 Y             N            Y            0

Now let’s skew the data a bit…

  for x in 1..10000
    insert into test values (11);
  end loop;

At this stage we will not analyze, just to see what Oracle does with peeking and explain plans.

Run the same select in a cursor, fetching those 10000 rows we inserted.

  type t_nums is varray(10000) of number;
  a_nums t_nums;
  x number;
cursor c1 is
  select xnum from test where xnum=x;
  open c1;
    fetch c1 bulk collect into a_nums;
  close c1;

declare type t_nums is varray(10000) of number; a_nums t_nums; x number; cursor c1 is select xnum from test where xnum=x; begin x:=11; open c1; fetch c1 bulk collect into a_nums; close c1; end; 
         fcrk1v7mrcnkr             1       0          1           1                 N             N            Y            0
         7y76r2vywmrv5             1       3          3           1                 Y             N            Y            0
         7y76r2vywmrv5             1       1          1           1                 Y             N            Y            1

Interesting… it created a new version of the statement with the same SQL_ID.  Notice that CHILD_NUMBER is now 1… which means it created a child of the parent cursor. There is a lot to be said about multiple versions and child cursors, but that is the subject of another post. This is however VERY relevant to shared pool management and behaviour.

More info in this view:

  child_number, bucket_id, count from v$sql_cs_histogram
  sql_id = '7y76r2vywmrv5';

           1         0     0
           1         1     1
           1         2     0
           0         0     3
           0         1     0
           0         2     0

and this:

  child_number, bind_set_hash_value, peeked, executions, rows_processed, buffer_gets, cpu_time
  sql_id = '7y76r2vywmrv5';

           1          2671099528      Y          1 10000                   21        0
           0          2706503459      Y          1     1                    1        0

The PEEKED column shows that Oracle did peek at the bind values passed into the SQL.

Now how to get the adaptive cursor sharing framework to change plan? I ran the xnum=11 select three times and it still does a range scan of the index:


Let’s analyze to see if it makes any difference. No, it continues with the range scan. Grr!
Let’s drop the index! Then run the select where xnum=11, three times.

declare type t_nums is varray(10000) of number; a_nums t_nums; x number; cursor c1 is select xnum from test where xnum=x; begin x:=11; open c1; fetch c1 bulk collect into a_nums; close c1; end; 
         fcrk1v7mrcnkr             1       0          3           3                 N             N            Y            0               0
         7y76r2vywmrv5             1       3          3           1                 Y             N            Y            0      1873712855
         7y76r2vywmrv5             1       3          3           1                 Y             N            Y            2       217508114

That’s expected… it picks up the dropped index, change in statistics, and creates child 2. Now that the index is gone, child 1 has magically aged out of the shared pool. Child 2 now has a new PLAN_HASH_VALUE=217508114 which shows TABLE ACCESS FULL. Child 0 which is the parent cursor still remains in the shared pool with PLAN_HASH_VALUE=1873712855 which is the INDEX RANGE SCAN. These can also be seen in OEM under the SQL Details screen where Plan Hash Value can be selected. So has adaptive cursor sharing worked? It’s hard to say, as this behaviour would have happened in 10g as well.

I would still like to create an example which changes the plan based on statistics and data, and will update this post.

Mysteries with current testing:

  • IS_BIND_AWARE is supposed to change to Y with the new plan (is this the definite way to check if adaptive cursor sharing has worked?).
  • v$sql_cs_selectivity shows zero rows for that SQL ID.

So, what have we learned? Literals are generally bad with cursor_sharing=similar unless you have full control of your application and explain plans are better for each statement. Literals are bad for CPU with cursor_sharing=force. It is better to use anonymous blocks or dynamic SQL, but not by much. Shared memory usage is a lot higher and there is more parsing to be done by the block as well. Efficient writing of cursors are the best for any cursor_sharing setting, as the code variables are always treated as bind variables and the statements are soft-parsed at worst.

The main point of this post is to show you how to write better SQL so that it can run as efficiently as possible on Oracle. So whether you are a DBA or a developer, I hope you can take something positive away from this.

Thanks for reading!

OraX (DBA Management and Monitoring Tool)

OraX is a customizable Oracle Database Management and Monitoring Tool written by an Oracle Certified Professional (OCP) DBA for DBA’s.  There is a Helpdesk version available for management or helpdesk people who want an overview of performance.  It runs in Windows but is fully server aware whether you run Windows, Solaris or Linux.

Screens cannot be shown currently due to confidentiality.

This tool is written by a DBA and is in constant use by many DBA’s.  It is always being updated with the latest scripts and is compatible with the latest oracle enhancements.


Profile of Timegate Technologies


Timegate Technologies is an Information Technology company that supplies the architecture, development and implementation of a wide range of systems using optimal database and development technology.

Please view for more information.