Dec/090
Oracle Performance Tuning Tactics Part 2: Using Indexes
To recap, the “weapons” available in your “arsenal” for performance tuning are as follows:- Full table scans ; Indexes; Joins; Views; De-normalization; Oracle Stored Procedures; Sub-queries.
2. Indexes
There are several types of indexes available with Oracle databases:
* normal/B-Tree (balanced/ binary tree) indexes;
* bitmap indexes (which literally use a map of bits to represent the whole index and are
* function-based indexes (which are based on the value of an expression instead of the value of a column);
* partitioned indexes (tables and indexes may be partitioned separately)
* domain indexes (application-specific indexes created on complex data types such as spatial data).
This article only discusses bitmap and normal (b-tree) indexes.
2.1 B-Tree Indexes
B-Tree Indexes are most appropriate when retrieving a small amount of data from a very large table. The bigger the table and the lower the number of rows that you want to retrieve then the more effective the index.
If you think of the printed material in the real world, this makes perfect sense, you wouldn’t bother to create an index for a short article, but you would certainly want one for a lengthy reference book. The reasons are analogous too. When reading a short article it would take longer to find the areas of interest using the index then it would to read the whole article. It’s the same when you’re querying the database – for a small table with Oracle’s multi-block read feature of Oracle, the whole table would be read in one go, so by going to the index first, the database would have to read more data overall.
Indexes may improve read access to data and update operations (including insert and delete) using sub-queries. However update performance will be impaired because every time a row is inserted into, deleted from or updated (if the indexed column(s) is/are updated) in the database, the indexes have to be adjusted accordingly, thereby increasing the amount of i/o that has to performed.
Therefore having many indexes is ideal when the database is mostly read only, but if there are a high proportion of inserts and deletes and only a few read operations, then you would find that adding more indexes would degrade rather than improve performance. You also need to ensure that the column(s) you are proposing to index is/are used frequently as the limiting conditions for a query, if not you will waste a lot of space and degrade database performance.
Another factor to consider with b-tree indexes in Oracle databases, is how specific they are – their cardinality. The ideal b-tree index is one that refers to only one row in the table, as a primary key does by definition. This reduces considerably the amount of i/o required to retrieve the data, but if the index value relates to many rows because it is not selective, then a lot more data will be retrieved resulting in unnecessary i/o.
Dec/090
Oracle Performance Tuning Part 1: Using Full Table Scans
It is a common misconception that all SQL queries on all tables in Oracle databases should be index driven. In fact, using full-table scans can improve performance in two scenarios: when querying very small tables and when querying very large tables.
The Effect of Full Table Scans When Querying Small Tables
Let’s suppose your using your Oracle database to run in-house designed and built HR application. Consider a reference table such as a list of department ids and the associated department names. Even a large company is likely to have only a few departments – HR, Sales, Marketing, Finance, IT, so the table is going to be quite small.
Now let’s suppose the table has just 2 columns – department id and department name – with an index on the department id. To find the department name for a given id, we would have to read the index and then read the table, but because the table is so small and because Oracle reads multiple database blocks in one read operation the whole table is scanned in just one read, so however efficient the index, by using it we will be performing unnecessary i/o.
In this case, therefore, a full table scan is faster than an index scan and table lookup. The exception to this of course is when the table has been created as an index-only table (available since Oracle 
which means that the whole table is stored in a B-tree structure (although you may have pointers to overflow areas).
The Effect of Full Table Scans When Querying Very Large Tables
Let’s look at using this technique for querying very large tables in your Oracle database. Surely they should use an index? Otherwise you might have to read thousands of blocks. It is correct to say that a full table scan of a very large table could read many thousands of data blocks, but as we shall see it may be better to do this than to perform an index scan and table lookup.
The situation when the a full table scan is very likely to perform better than an index scan and table lookup is when you are retrieving 10% or more of the data in the table and it may perform better even when you are retrieving as little as 1% of the table data. Of course if you only want to retrieve one row in the table, then you would want to use an index.
Index Scan And Table Lookup Vs. Full Table Scan For Very Large Tables in Oracle Databases
Let’s look at the 2 scenarios then – retrieving 10% of the table data by index scan and table lookup vs. full table scan.
To make the maths easy, assume our table has 10,000,000 rows with 10 data rows per block and 100 index entries per block. Therefore to read 10% of the table via an index scan and table lookup, we would have to read 10,000 (1,000,000/100) index blocks plus 100,000 (1,000,000/10) data blocks. That’s 110,000 blocks in total.
However this assumes that the data is stored in order which means that we only retrieve the blocks of data that we want. If the data is not sorted then the worst case is that we would have to read 1 block for each row of data i.e. 1,000,000 blocks, which would would give us a worst case total of 1,010,000 blocks.
For a full table scan the maths is easy: (10,000,000 rows)/(10 rows per block) = 1,000,000 blocks. This is less than the worst case scenario for an indexed read, but more than the best case scenario for an indexed read. This would seem to suggest that if you sort your data before loading, an indexed read would be faster than a full-table scan.
Whilst it is true that pre-sorting the data of very large table will improve performance , it is not necessarily correct that the read via the index will be better overall. We also need to take into account what happens to the blocks stored in the buffer cache of the Oracle SGA and the impact this will have on other users of the database.
Let’s look at the effect on the buffer cache of reading many data blocks via an index. As we know, data and index blocks are stored in the buffer cache by Oracle for reuse by other queries by being marked as least recently used when we do an indexed read. However, those data blocks read by a full table scan are quickly aged out of the buffer cache, because they are not marked as least recently used.
What this means is that the large number of blocks (1,010,000) of index and table data read via the indexed read of our table will be saved in the SGA flushing practically all other data from it – which will obviously have an effect on other users.
Conversely, when a full table scan is performed only the last blocks read are held in the SGA (the actual number is determined by the multi-block read count) so the impact on other users of the database would be minimal.
Summary
To decide whether or not a full-table scan would be better than an indexed read, for a large table you need to consider what proportion of the table the query will retrieve from your Oracle database and consider the likely effect of that on other users. The denser the data, the more efficient a full table scan is for very large tables, but generally if you’re reading more than 1-10% of a very large table, a full table scan would be more efficient than an index scan and table lookup.
For small tables you will get much better performance from your Oracle database by caching the table in its entirety (so that it is always in memory), or by using an index-organised table then you will by relying on indexes. Having said that, every table should have a primary key index to guarantee uniqueness, but you don’t have to use it.
Dec/090
Oracle Performance Tuning Overview
There isn’t the time or space to provide an exhaustive study – whole books have been written about Oracle performance tuning. Instead we’ll just consider a few of the most important tactics and there’ll be suggestions for further reading, so you can expand your knowledge of this important subject.
One point that needs to be made clear is that performance tuning is not an exact science – you can’t always predict that a certain tactic or technique will improve performance. Every change has to be tested, because whilst it may improve performance in one area, it may also degrade performance in another area. Adding an index is a classic example of this – it may improve query performance but the performance of inserts and deletes become slower because the index entries also have to be updated.
The other thing to remember about performance tuning is that you need to decide on your goals before you start – do you need a 10% decrease in response time for queries ? Is adding new records or updating existing records 20% slower than targets ?
You also need to be aware that improving response time is not the only aim that you might have – you might want to reduce memory usage instead – which might mean having to redesign or rewrite some stored procedures. All these issues need to be considered before you start looking at performance.
Let’s start with a list of the topics that we’ll cover:
Use of full-table scans
How and when to use indexes.
How you can optimise joins.
How to use views to get a high performance database.
Why your database should NOT be normalised.
How to use stored procedures to sky-rocket Oracle performance.
How to use sub queries to boost Oracle performance
Nov/090
Optimizing Mysql Tables
Just like you need to defrag and check your file system, it’s important to do the same thing with MySQL tables. If you don’t, you might end up with slow and corrupted database tables. Follow this easy tutorial to set up an automatic optimization for your db.
Speed in MySQL is all about how fast the database can find the information you request. To help MySQL clean up and get all the best possibilities to gather your information, you can perform the command “Optimize” on your tables.
Want to speed up your MySQL performance in a simple but effective way?
Because you probably will add and delete tables from time to time, you want a solution that works no matter how your database looks like. For this I will give you a PHP script that finds all your tables, and then perform Optimize on every single one. A good idea can be to do this every night (or whenever your server is least accessed) with “cron” because you don’t want to delay your surfers to much.
Here is the script (really simple actually):
Make sure you modify the four variables at the top of the script so they match your configuration.
// Change these four variables to match your configuration.
// You will probably only need to change “user”, “pwd” and “dbName”.
$server = ”localhost”;
$user = ”mysql_user”;
$pwd = ”mysql_password”;
$dbName = ”mysql_dbName”;
$link = mysql_connect($server, $user, $pwd);
if (!$link) {
die(’Could not connect: ‘ . mysql_error());
}
$db_selected = mysql_select_db($dbName, $link);
if (!$db_selected) {
die (’Can\’t use $dbName : ‘ . mysql_error());
}
// Find the names of all tables in the selected db.
$alltables = mysql_query(”SHOW TABLES”);
// Go through all tables.
while ($table = mysql_fetch_assoc($alltables))
{
foreach ($table as $db => $tablename)
{
// Perform the magic “OPTIMIZE” on every table in the db.
mysql_query(”OPTIMIZE TABLE ‘”.$tablename.”‘”)
or die(mysql_error());
}
}
Oct/090
very basic sql optimization in mysql
It’s important to know how to write your queries fast. And there are also ways to do your queries faster Inside mysql, called Indexes. They help mysql do the work faster, but you need to tell how…
When you start getting large tables you can sometimes end up with slow running queries.
It’s important to optimize your queries and tables from the beginning, so you don’t end up with slow queries one year later when your tables have increased.
Badly created tables and queries is hard to find when you only have limited data in your tables. But there are tools to use to get a hint on if you will have problems later…
First the basics behind the scen in databases. Let’s say you are combining three tables with a join. Then mysql has to look through all three tables and then find a way to combine them in the correct way.
Take this sql query that will find the person “donald duck” and information connected to him in three tables:
SELECT a.id
FROM a, b, c
WHERE a.id = b.a_id
AND b.id = c.b_id
AND c.name = “donald duck”
Let’s say there are 2 rows in each table. What mysql does now is to start looking through all tables from the top down to the bottom and try to find matches.
The important thing to know about this is some math. Either you understand it or you just trust me that it is correct…
To combine all three tables, mysql needs to go through 2*2*2 combinations. This is 8 combinations to find the correct one.
Doesn’t sound that much of a deal, but let’s say we have 500 rows in the tables each. This gives our beloved database 500*500*500 = 1 250 000 different combinations to go through.
Luckily there are some smart people that have found a way to deal with this
It’s called “Indexes” and they help the database to cut down on the number of rows it has to go through to find the match.
The first thing you should ALWAYS do is to make the id in your tables to a PRIMARY key. Then you get that column indexed.
Look at indexes as a way to decrease the number of rows to look through to find the match. You put an index on a seperate column and you should only put it on columns that need it, otherwise your queries will go slower instead.
A basic rule that works very well is that you should put indexes on columns that your “search on” and combine in JOIN’s.
So if we look at the query above, what columns might be interesting to put indexes on?
First of all, make a.id, b.id, c.id PRIMARY keys and they get indexed automaticly.
After that I would set b.a_id and c.b_id as indexes. That is because they are used to combine the tables. If we can decrease the number of rows that needs to be looked through there, the queries will be faster.
The last column that might be interesting is c.name.
A summary:
- Make the id in each table PRIMARY
- Make columns that are “foreign keys” Indexed.
- If you have any other columns that are use in the WHERE clause, make this Indexed aswell.
This is a simplified guide on how to fasten your sql queries, but very good to know the basics.
I
will soon write a tutorial on how to check if your queries are slow or fast, and what to do about it…
Jun/090
A Few Hints And Tips On Optimizing Your SQL
This was intended to be a list of hints and tips that you might find useful when using SQ,L but a mere list of tips would be of little benefit without the knowledge to make use of them, so we’ve expanded the list to include the explanations to increasing its usefulness and to make it a proper tutorial.
One important point to remember is that Oracle caches the compiled form of SQL and is therefore able to re-use queries which are the same as previously executed queries. This saves the time and resources required to parse the statement and determine the execution plan. How can you do this ?
1. Use Views
Views are a good way to ensure the same query is re-used as much as possible.
Remember that even just changing the case and spacing of the words could prevent a query from being reused. A view is merely a pre-defined query, the text of which is stored in the database. Therefore by using views you are using exactly the same queries and eliminating the re-parsing overhead. As the load on the database increases this re-parsing overhead becomes more and more significant. Materialized views take the concept one stage further by actually running the query and storing the results in a table
2. Use Stored Procedures
Another way is to use stored procedures which are program units that contain both SQL and logic statements and are stored in the database. Oracle allows the use of PL/SQL and Java stored procedures. Stored procedures and views also have the advantage that the queries in the views/stored procedures have to be tuned only once, not in every place where they’re used. Like views, stored procedures also eliminate the overhead of sending the queries from the client to the server as the queries are already on the server.
3. Use Bind Variables
The use of bind variables in queries makes them generic and therefore re-usable.
For example, instead of writing a query like :-
SELECT name,addr FROM custs WHERE id = 12345;
Change it to:-
SELECT name,addr FROM custs WHERE id = <cust_id>;
The first query will only be re-used when you request the details for customer number 12345, whereas the second query will be re-used for any other customer.
4. Use Selective Indexes
Ensure that tables are accessed via selective indexes, unless the table is very small or very large, in which case it may be better not to use the indexes.
If the table were very small it could be cached completely, or all the columns could be indexed which means only the index would have to be read to satisfy any query.
Also make sure that you’re not disabling the use of an index by:-
* using an operator on the column (eg. <indexed_col> + 1);
* the use of hints, if you’re running Oracle(only applies if you’re using the cost based optimizer);
* using NULL and not equal checks. (eg. <indexed_col> <> 12345 ; or <indexed_col> IS NULL)
5. Use Full-Table Scans
If the table is very large, depending on how many blocks are read, using an index may remove everything else from the buffer cache and degrade the performance of all other queries. In which case a full-table scan is better – only the last few blocks read are kept in the buffer cache.
6. Optimize Joins
* All other things being equal, the driving table is the one listed LAST in the FROM clause, when using the rule-based optimizer. Changing the order of the columns in the join condition does not change which table is used as the driving table. Choose the driving table carefully to ensure the minimum number of rows are returned.
When using the cost-based optimizer, ensure that all the tables in the join have been analyzed (ask your dba), if they haven’t, this may well cause poor performance. The most usual way to optimize queries when using the cost-based optimizer is to use hints, which instruct the parser as to which indexes should or should not be used, or which tables should be scanned in full. You can also experiment with the order of the tables in the join.
* Indexes – these can still be used even if the where clause contains a “like” condition but not if there is a “not like” condition.
* Outer joins – the correct syntax for outer joins using Oracle syntax is:
tab1.col1(+)= tab2.col1
or tab1.col1 = tab2.col1(+)
The bracketed plus sign follows the column of the table which has/may have the missing row.
An alternative is to use the ANSI standard outer join format which has the advantage that it enables you to perform a full outer join in one statement:
tab1.col1 left outer join tab2.col1 (return all rows from tab1)
tab1.col1 right outer join tab2.col1 (return all rows from tab2)
tab1.col1 full outer join tab2.col1 (return all rows from tab1 and tab2)
The final tip for this short tutorial is:
7. Name The Columns In A Query
There are three good reasons why it is better to name the columns in a query rather than to use “select * from …”.
1. Network traffic is reduced. This can have a significant impact on performance if the table has a large number of columns, or the table has a long or long raw column or have in-line clob or blob columns (all of which can be up to 2 Gigabytes in length). These types of columns will take a long time to transfer over the network and so they should not be fetched from the database unless they are specifically required.
2. The code is easier to understand, which means you need fewer comments!
3. It could save the need for changes in the future. If you are using views, not only might columns be added to or removed from the view, but the order of the columns could well change – in which case using “SELECT *” at best would fetch the wrong data and at worst would fail with an Oracle error which might take a long while to understand.
Jun/090
Oracle Performance Tuning Tactics Part 1
Use Full Table Scans To Boost The Performance Of Your Oracle Database
Before we start, let me make it clear that performance tuning is not always an exact science – by which I mean that you can’t always predict that a certain tactic or technique will improve performance. Every change has to be tested, because whilst it may improve performance in one area, it may also degrade performance in another area. Adding an index is a classic example of this – it may improve query performance but the performance of inserts and deletes become slower because the index entries also have to be updated.
The other thing to remember about performance tuning is that you need to decide on your goals before you start – do you need a 10% decrease in response time for queries ? Is adding new records or updating existing records 20% slower than targets ?
You also need to be aware that improving response time is not the only aim that you might have – you might want to reduce memory usage instead – which might mean having to redesign or rewrite some stored procedures. All these issues need to be considered when you you start looking at performance.
I’ll start by listing the “weapons” you have available in your “arsenal” which are as follows:
1. Full table scans
2. Indexes
3. Joins
4. Views
5. De-normalization
6. Stored Procedures
7. Sub-queries
1. Full table scans
This might seem counter-intuitive, surely all queries on a table should be index driven ? Actually, no, they shouldn’t. If the table is very large or very small using an index may increase the amount of i/o that is required to access the data and as i/o is the slowest operation, this will cause a degradation in performance.
1.1 The Effect On Small Tables
Consider a very small (reference) table such as a list of department ids and the associated department names in a HR application. Even a large company is likely to have only a few departments – hr, sales, marketing, finance, IT, so the table is going to be quite small. Now let’s suppose the table has just 2 columns – department id and department name – with an index on the department id. To find the department name for a given id, we would have to read the index and then read the table, but because the table is so small and because Oracle reads multiple database blocks in one read operation the whole table is scanned in just one read, so however efficient the index, by using it we will be performing unnecessary i/o.
In this case, therefore, a full table scan is faster than an index scan and table lookup. The exception to this of course is when the table has been created as an index-only table (available since Oracle Cool which means that the whole table is stored in a B-tree structure (although you may have pointers to overflow areas).
1.2 The Effect On Very Large Tables
Let’s look at using this technique for very large tables now. Surely they should use an index ? Otherwise you might have to read thousands and thousands of blocks. It is correct to say that a full table scan of a very large table could read many thousands of data blocks, but as we shall see it may be better to do this than to perform an index scan and table lookup.
The situation when the a full table scan is very likely to perform better than an index scan and table lookup is when you are retrieving 10% or more of the data in the table and it may perform better even when you are retrieving as little as 1% of the table data. Of course if you only want to retrieve one row in the table, then you would want to use an index.
1.2.1 Index Scan And Table Lookup Vs. Full Table Scan For Very Large Tables
Let’s look at the 2 scenarios then – retrieving 10% of the table data by index scan and table lookup vs. full table scan.
To make the maths easy, assume our table has 10,000,000 rows with 10 data rows per block and 100 index entries per block. Therefore to read 10% of the table via an index scan and table lookup, we would have to read 10,000 (1,000,000/100) index blocks plus 100,000 (1,000,000/10) data blocks. That’s 110,000 blocks in total.
However this assumes that the data is stored in order which means that we only retrieve the blocks of data that we want. If the data is not sorted in the table then the worst case is that we would have to read 1 block for each row of data i.e. 1,000,000 blocks, which would would give us a worst case total of 1,010,000 blocks.
For a full table scan the maths is easy: (10,000,000 rows)/(10 rows per block) = 1,000,000 blocks. This is less than the worst case scenario for an indexed read, but more than the best case scenario for an indexed read. This would seem to suggest that if you sort your data before loading, an indexed read would be faster than a full-table scan.
Whilst it is true that pre-sorting the data of very large table will improve performance , it is not necessarily correct that the read via the index will be better overall. We also need to take into account what happens to the blocks stored in the buffer cache of the Oracle SGA and the impact this will have on other users of the database.
Let’s look at the effect on the buffer cache of reading many data blocks via an index. As we know, data and index blocks are stored in the buffer cache for reuse by other queries by being marked as least recently used when we do an indexed read. However, those data blocks read by a full table scan are quickly aged out of the buffer cache, because they are not marked as least recently used.
What this means is that the large number of blocks (1,010,000) of index and table data read via the indexed read of our table will be saved in the SGA flushing practically all other data from it – which will obviously have an effect on other users.
Conversely, when a full table scan is performed only the last blocks read are held in the SGA (the actual number is determined by the multi-block read count) so the impact on other users of the database would be minimal.
1.3 Summary
In conclusion, therefore, to decide whether or not a full-table scan would be better than an indexed read, for a large table you need to consider what proportion of the table the query will retrieve from your Oracle database and consider the likely effect of that on other users. The denser the data, the more efficient a full table scan is for very large tables, but generally if you’re reading more than 1-10% of a very large table, a full table scan would be more efficient than an index scan and table lookup.
For very small tables you will get much better performance from your Oracle database by caching the table in its entirety (so that it is always in memory), or by using an index-organised table then you will by relying on indexes. Having said that, every table should have a primary key index to guarantee uniqueness, but you don’t have to use it.