What one can and should optimize

• Hardware
• OS / libraries
• SQL server (setup and queries)
• API
• Application

Optimizing hardware for MySQL

• If you need big tables ( > 2G), you should consider using 64 bit hardware like Alpha, Sparc or the upcoming IA64. As MySQL uses a lot of 64 bit integers internally, 64 bit CPUs will give much better performance.
• For large databases, the optimization order is normally RAM, Fast disks, CPU power.
• More RAM can speed up key updates by keeping most of the used key pages in RAM.
• If you are not using transaction-safe tables or have big disks and want to avoid long file checks, a UPS is good idea to be able to take the system down nicely in case of a power failure.
• For systems where the database is on a dedicated server, one should look at 1G Ethernet. Latency is as important as throughput.

Optimizing disks

• Have one dedicated disk for the system, programs and for temporary files. If you do very many changes, put the update logs and transactions logs on dedicated disks.
• Low seek time is important for the database disk; For big tables you can estimate that you will need: log(row_count) / log(index_block_length/3*2/(key_length + data_ptr_length))+1 seeks to find a row. For a table with 500,000 rows indexing a medium int: log(500,000)/log(1024/3*2/(3+4)) +1 = 4 seeks The above index would require: 500,000 * 7 * 3/2 = 5.2M. In real life, most of the blocks will be buffered, so probably only 1-2 seeks are needed.
• For writes you will need (as above) 4 seek requests, however, to find where to place the new key, and normally 2 seeks to update the index and write the row.
• For REALLY big databases, your application will be bound by the speed of your disk seeks, which increase by N log N as you get more data.
• Split databases and tables over different disks. In MySQL you can use symbolic links for this.
• Striping disks (RAID 0) will increase both read and write throughput.
• Striping with mirroring (RAID 0+1) will give you safety and increase the read speed. Write speed will be slightly lower.
• Don't use mirroring or RAID (except RAID 0) on the disk for temporary files or for data that can be easily re-generated..
• On Linux use hdparm -m16 -d1 on the disks on boot to enable reading/writing of multiple sectors at a time, and DMA. This may increase the response time by 5-50 %.
• On Linux, mount the disks with async (default) and noatime.
• For some specific application, one may want to have a ram disk for some very specific tables, but normally this is not needed.

Optimizing OS

• No swap; If you have memory problems, add more RAM instead or configure your system to use less memory.
• Don't use NFS disks for data (you will have problems with NFS locking).
• Increase number of open files for system and for the SQL server. (add ulimit -n # in the safe_mysqld script).
• Increase the number of processes and threads for the system.
• If you have relatively few big tables, tell your file system to not break up the file on different cylinders (Solaris).
• Use file systems that support big files (Solaris).
• Choose which file system to use; Reiserfs on Linux is very fast for open, read and write. File checks take just a couple of seconds.

Choosing API

Optimizing the application

• One should concentrate on solving the problem.
• When writing the application one should decide what is most important:
  • Speed
  • Portability between OS
  • Portability between SQL servers
• Use persistent connections.
• Cache things in your application to lessen the load of the SQL server.
• Don't query columns that you don't need in your application.
• Don't use SELECT * FROM table_name...
• Benchmark all parts of your application, but put the most effort into benchmarking the whole application under the worst possible 'reasonable' load. By doing this in a modular fashion you should be able to replace the found bottleneck with a fast 'dummy module', you can then easily identify the next bottleneck (and so on).
• Use LOCK TABLES if you do a lot of changes in a batch; For example group multiple UPDATES or DELETES together.

Portable applications should use

• Perl DBI/DBD
• ODBC
• JDBC
• Python (or any other language that has a generalized SQL interface)
• You should only use SQL constructs which exist in all the target SQL servers or can easily be emulated with other constructs. The crash-me pages on www.mysql.com can help you with this.
• Write wrappers to provide missing functionality for other OSes / SQL servers.

If you need more speed, you should:

• Find the bottleneck (CPU, disk, memory, SQL server, OS, API, or application) and concentrate on solving this.
• Use extensions that give you more speed / flexibility.
• Get to know your SQL server so that you can use the fastest possible SQL constructs for your problem and avoid bottlenecks.
• Optimize your table layouts and queries.
• Use replication to get more select speed.
• If you have a slow net connection to the database, use the compressed client/server protocol.

Optimizing MySQL

• Choose compiler and compiler options.
• Find the best MySQL startup options for your system.
• Scan the the MySQL manual and read Paul DuBois' MySQL book.
• Use EXPLAIN SELECT, SHOW VARIABLES, SHOW STATUS and SHOW PROCESSLIST.
• Learn how the query optimizer works.
• Optimize your table formats.
• Maintain your tables (myisamchk, CHECK TABLE, OPTIMIZE TABLE).
• Use MySQL extensions to get things done faster.
• Write a MySQL UDF function if you notice that you would need some function in many places.
• Don't use GRANT on table level or column level if you don't really need it.
• Pay for MySQL support and get help to solve your problem :)

Compiling and installing MySQL

• By choosing the best possible compiler for your system, you can usually get 10-30 % better performance.
• On Linux/Intel, compile MySQL with pgcc. (The Pentium optimized version of gcc). The binary will only work with Intel Pentium CPUs, however.
• Use the optimize options that are recommended in the MySQL manual for a particular platform.
• Normally a native compiler for a specific CPU (like Sun Workshop for Sparc) should give better performance than gcc, but this is not always the case.
• Compile MySQL with only the character sets you are going to use.
• Compile the mysqld executable statically (with --with-mysqld-ldflags=-all-static) and strip the final executable with strip sql/mysqld.
• Note that as MySQL doesn't use C++ exceptions, compiling MySQL without exceptions support will give a big performance win!
• Use native threads (instead of the included mit-pthreads) if your OS supports native threads.
• Test the resulting binary with the MySQL benchmark test.

Maintenance

• If possible, run OPTIMIZE table once in a while. This is especially important on variable size rows that are updated a lot.
• Update the key distribution statistics in your tables once in a while with myisamchk -a; Remember to take down MySQL before doing this!
• If you get fragmented files, it may be worth it to copy all files to another disk, clear the old disk and copy the files back.
• If you have problems, check your tables with myisamchk or CHECK table.
• Monitor MySQL status with: mysqladmin -i10 processlist extended-status
• With the MySQL GUI client you can monitor the process list and the status in different windows.
• Use mysqladmin debug to get information about locks and performance.

Optimizing SQL

• Find rows based on WHERE clause.
• JOIN tables
• GROUP BY
• ORDER BY
• DISTINCT

• To validate data (like date)
• As a calculator

• Use keys wisely.
• Keys are good for searches, but bad for inserts / updates of key columns.
• Keep by data in the 3rd normal database form, but don't be afraid of duplicating information or creating summary tables if you need more speed.
• Instead of doing a lot of GROUP BYs on a big table, create summary tables of the big table and query this instead.
• UPDATE table set count=count+1 where key_column=constant is very fast!
• For log tables, it's probably better to generate summary tables from them once in a while than try to keep the summary tables live.
• Take advantage of default values on INSERT.

Speed difference between different SQL servers (times in seconds)

Reading 2000000 rows by key:	NT	Linux
mysql	367	249
mysql_odbc	464
db2_odbc	1206
informix_odbc	121126
ms-sql_odbc	1634
oracle_odbc	20800
solid_odbc	877
sybase_odbc	17614
Inserting (350768) rows:	NT	Linux
mysql	381	206
mysql_odbc	619
db2_odbc	3460
informix_odbc	2692
ms-sql_odbc	4012
oracle_odbc	11291
solid_odbc	1801
sybase_odbc	4802

In the above test, MySQL was run with a 8M cache; the other databases were run with installations defaults.

Important MySQL startup options

back_log	Change if you do a lot of new connections.
thread_cache_size	Change if you do a lot of new connections.
key_buffer_size	Pool for index pages; Can be made very big
bdb_cache_size	Record and key cache used by BDB tables.
table_cache	Change if you have many tables or simultaneous connections
delay_key_write	Set if you need to buffer all key writes
log_slow_queries	Find queries that takes a lot of time
max_heap_table_size	Used with GROUP BY
sort_buffer	Used with ORDER BY and GROUP BY
myisam_sort_buffer_size	Used with REPAIR TABLE
join_buffer_size	When doing a join without keys

Optimizing tables

• MySQL has a rich set of different types. You should try to use the most efficient type for each column.
• The ANALYSE procedure can help you find the optimal types for a table: SELECT * FROM table_name PROCEDURE ANALYSE()
• Use NOT NULL for columns which will not store null values. This is particularly important for columns which you index.
• Change your ISAM tables to MyISAM.
• If possible, create your tables with a fixed table format.
• Don't create indexes you are not going to use.
• Use the fact that MySQL can search on a prefix of an index; If you have and INDEX (a,b), you don't need an index on (a).
• Instead of creating an index on long CHAR/VARCHAR column, index just a prefix of the column to save space. CREATE TABLE table_name (hostname CHAR(255) not null, index(hostname(10)))
• Use the most efficient table type for each table.
• Columns with identical information in different tables should be declared identically and have identical names.

How MySQL stores data

• Databases are stored as directories.
• Tables are stored as files.
• Columns are stored in the files in dynamic length or fixed size format. In BDB tables the data is stored in pages.
• Memory-based tables are supported.
• Databases and tables can be symbolically linked from different disks.
• On Windows MySQL supports internal symbolic links to databases with .sym files.

MySQL table types

• HEAP tables; Fixed row size tables that are only stored in memory and indexed with a HASH index.
• ISAM tables; The old B-tree table format in MySQL 3.22.
• MyISAM tables; New version of the ISAM tables with a lot of extensions:
  • Binary portability.
  • Index on NULL columns.
  • Less fragmentation for dynamic-size rows than ISAM tables.
  • Support for big files.
  • Better index compression.
  • Better key statistics.
  • Better and faster auto_increment handling.
• Berkeley DB (BDB) tables from Sleepycat: Transaction-safe (with BEGIN WORK / COMMIT | ROLLBACK).

MySQL row types (only relevant for ISAM/MyISAM tables)

• MySQL will create the table in fixed size table format if all columns are of fixed length format (no VARCHAR, BLOB or TEXT columns). If not, the table is created in dynamic-size format.
• Fixed-size format is much faster and more secure than the dynamic format.
• The dynamic-size row format normally takes up less space but may be fragmented over time if the table is updated a lot.
• In some cases it's worth it to move all VARCHAR, BLOB and TEXT columns to another table just to get more speed on the main table.
• With myisampack (pack_isam for ISAM) one can create a read-only, packed table. This minimizes disk usage which is very nice when using slow disks. The packed tables are perfect to use on log tables which one will not update anymore.

MySQL caches (shared between all threads, allocated once)

• Key cache ; key_buffer_size, default 8M
• Table cache ; table_cache, default 64
• Thread cache ; thread_cache_size, default 0.
• Hostname cache ; Changeable at compile time, default 128.
• Memory mapped tables ; Currently only used for compressed tables.

MySQL buffer variables (not shared, allocated on demand)

• sort_buffer ; ORDER BY / GROUP BY
• record_buffer ; Scanning tables
• join_buffer_size ; Joining without keys
• myisam_sort_buffer_size ; REPAIR TABLE
• net_buffer_length ; For reading the SQL statement and buffering the result.
• tmp_table_size ; HEAP-table-size for temporary results.

How the MySQL table cache works

• Each open instance of a MyISAM table uses an index file and a data file. If a table is used by two threads or used twice in the same query, MyISAM will share the index file but will open another instance of the data file.
• The cache will temporarily grow larger than the table cache size if all tables in the cache are in use. If this happens, the next table that is released will be closed.
• You can check if your table cache is too small by checking the mysqld variable Opened_tables. If this value is high you should increase your table cache!

MySQL extensions / optimization that gives you speed

When MySQL uses indexes

• Using >, >=, =, <, <="," IF NULL and BETWEEN on a key.
  
  • SELECT * FROM table_name WHERE key_part1=1 and key_part2 > 5;
    
  • SELECT * FROM table_name WHERE key_part1 IS NULL;
• When you use a LIKE that doesn't start with a wildcard.
  
  • SELECT * FROM table_name WHERE key_part1 LIKE 'jani%'
• Retrieving rows from other tables when performing joins.
  
  • SELECT * from t1,t2 where t1.col=t2.key_part
• Find the MAX() or MIN() value for a specific index.
  
  • SELECT MIN(key_part2),MAX(key_part2) FROM table_name where key_part1=10
• ORDER BY or GROUP BY on a prefix of a key.
  
  • SELECT * FROM foo ORDER BY key_part1,key_part2,key_part3
• When all columns used in the query are part of one key.
  
  • SELECT key_part3 FROM table_name WHERE key_part1=1

---

When MySQL doesn't use an index

• Indexes are NOT used if MySQL can calculate that it will probably be faster to scan the whole table. For example if key_part1 is evenly distributed between 1 and 100, it's not good to use an index in the following query:
  
  • SELECT * FROM table_name where key_part1 > 1 and key_part1 < 90
• If you are using HEAP tables and you don't search on all key parts with =
• When you use ORDER BY on a HEAP table
• If you are not using the first key part
  
  • SELECT * FROM table_name WHERE key_part2=1
• If you are using LIKE that starts with a wildcard
  
  • SELECT * FROM table_name WHERE key_part1 LIKE '%jani%'
• When you search on one index and do an ORDER BY on another
  
  • SELECT * from table_name WHERE key_part1 = # ORDER BY key2

---

Learn to use EXPLAIN

Use EXPLAIN on every query that you think is too slow!

mysql> explain select t3.DateOfAction, t1.TransactionID
    -> from t1 join t2 join t3
    -> where t2.ID = t1.TransactionID and t3.ID = t2.GroupID
    -> order by t3.DateOfAction, t1.TransactionID;
+-------+--------+---------------+---------+---------+------------------+------+---------------------------------+
| table | type   | possible_keys | key     | key_len | ref              | rows | Extra                           |
+-------+--------+---------------+---------+---------+------------------+------+---------------------------------+
| t1    | ALL    | NULL          | NULL    |    NULL | NULL             |   11 | Using temporary; Using filesort |
| t2    | ref    | ID            | ID      |       4 | t1.TransactionID |   13 |                                 |
| t3    | eq_ref | PRIMARY       | PRIMARY |       4 | t2.GroupID       |    1 |                                 |
+-------+--------+---------------+---------+---------+------------------+------+---------------------------------+

Types ALL and range signal a potential problem.

---

Learn to use SHOW PROCESSLIST

Use SHOW processlist to find out what is going on:

+----+-------+-----------+----+---------+------+--------------+-------------------------------------+
| Id | User  | Host      | db | Command | Time | State        | Info                                |
+----+-------+-----------+----+---------+------+--------------+-------------------------------------+
| 6  | monty | localhost | bp | Query   | 15   | Sending data | select * from station,station as s1 |
| 8  | monty | localhost |    | Query   | 0    |              | show processlist                    |
+----+-------+-----------+----+---------+------+--------------+-------------------------------------+

Use KILL in mysql or mysqladmin to kill off runaway threads.

---

How to find out how MySQL solves a query

Run the following commands and try to understand the output:

• SHOW VARIABLES;
• SHOW COLUMNS FROM ...\G
• EXPLAIN SELECT ...\G
• FLUSH STATUS;
• SELECT ...;
• SHOW STATUS;

---

MySQL is extremely good

• For logging.
• When you do many connects; connect is very fast.
• Where you use SELECT and INSERT at the same time.
• When you don't combine updates with selects that take a long time.
• When most selects/updates are using unique keys.
• When you use many tables without long conflicting locks.
• When you have big tables (MySQL uses a very compact table format).

---

Things to avoid with MySQL

• Updates to a table or INSERT on a table with deleted rows, combined with SELECTS that take a long time.
• HAVING on things you can have in a WHERE clause.
• JOINS without using keys or keys which are not unique enough.
• JOINS on columns that have different column types.
• Using HEAP tables when not using a full key match with =
• Forgetting a WHERE clause with UPDATE or DELETE in the MySQL monitor. If you tend to do this, use the --i-am-a-dummy option to the mysq client.

---

Different locks in MySQL

• Internal table locks.
• LOCK TABLES (Works on all table types)
• GET_LOCK()/RELEASE_LOCK()
• Page locks (for BDB tables)
• ALTER TABLE also does a table lock on BDB tables.
• LOCK TABLES gives you multiple readers on a table or one writer.
• Normally a WRITE lock has higher priority than a READ lock to avoid starving the writers. For writers that are not important one can use the LOW_PRIORITY keyword to let the lock handler prefer readers.

    UPDATE LOW_PRIORITY SET value=10 WHERE id=10;
  

---

Tricks to give MySQL more information to solve things better

Note that you can always comment out a MySQL feature to make the query portable:

SELECT /*! SQL_BUFFER_RESULTS */ ...

• SELECT SQL_BUFFER_RESULTS ...
  Will force MySQL to make a temporary result set. As soon as the temporary set is done, all locks on the tables are released. This can help when you get a problem with table locks or when it takes a long time to transfer the result to the client.
• SELECT SQL_SMALL_RESULT ... GROUP BY ...
  To tell the optimizer that the result set will only contain a few rows.
• SELECT SQL_BIG_RESULT ... GROUP BY ...
  To tell the optimizer that the result set will contain many rows.
• SELECT STRAIGHT_JOIN ...
  Forces the optimizer to join the tables in the order in which they are listed in the FROM clause.
• SELECT ... FROM table_name [USE INDEX (index_list) | IGNORE INDEX (index_list)] table_name2
  Forces MySQL to use/ignore the listed indexes.

---

Example of doing transactions

• How to do a transaction with MyISAM tables:

  mysql> LOCK TABLES trans READ, customer WRITE;
  mysql> select sum(value) from trans where customer_id=some_id;
  mysql> update customer set total_value=sum_from_previous_statement
             where customer_id=some_id;
  mysql> UNLOCK TABLES;
  

• How to do a transaction with Berkeley DB tables:

  mysql> BEGIN WORK; 
  mysql> select sum(value) from trans where customer_id=some_id;
  mysql> update customer set total_value=sum_from_previous_statement
             where customer_id=some_id;
  mysql> COMMIT;
  

• Note that you can often avoid transactions altogether by doing:

  UPDATE customer SET value=value+new_value WHERE customer_id=some_id;
  

---

Example of using REPLACE

REPLACE works exactly like INSERT, except that if an old record in the table has the same value as a new record on a unique index, the old record is deleted before the new record is inserted. Instead of using

  SELECT 1 FROM t1 WHERE key=#
  IF found-row
    LOCK TABLES t1
    DELETE FROM t1 WHERE key1=#
    INSERT INTO t1 VALUES (...)
    UNLOCK TABLES t1;
  ENDIF

Do

  REPLACE INTO t1 VALUES (...)    

---

General tips

• Use short primary keys. Use numbers, not strings, when joining tables.
• When using multi-part keys, the first part should be the most-used key.
• When in doubt, use columns with more duplicates first to get better key compression.
• If you run the client and MySQL server on the same machine, use sockets instead of TCP/IP when connecting to MySQL (this can give you up to a 7.5 % improvement). You can do this by specifying no hostname or localhost when connecting to the MySQL server.
• Use --skip-locking (default on some OSes) if possible. This will turn off external locking and will give better performance.
• Use application-level hashed values instead of using long keys:

    SELECT * FROM table_name WHERE hash=MD5(concat(col1,col2)) AND
    col_1='constant' AND col_2='constant'
  

• Store BLOB's that you need to access as files in files. Store only the file name in the database.
• It is faster to remove all rows than to remove a large part of the rows.
• If SQL is not fast enough, take a look at the lower level interfaces to access the data.

---

Benefits of using MySQL 3.23

• MyISAM ; Portable BIG table format
• HEAP ; In memory tables
• Berkeley DB ; Transactional tables from Sleepycat.
• A lot of raised limits
• Dynamic character sets
• More STATUS variables.
• CHECK and REPAIR table.
• Faster GROUP BY and DISTINCT
• LEFT JOIN ... IF NULL optimization.
• CREATE TABLE ... SELECT
• CREATE TEMPORARY table_name (...)
• Automatic conversion of temporary HEAP to MyISAM tables
• Replication
• mysqlhotcopy script.

---

Important features that we are actively working on

• Improving transactions
• Fail safe replication
• Text searching
• Delete with many tables (Updates with many tables will be done after this.)
• Better key cache
• Atomic RENAME (RENAME TABLE foo as foo_old, foo_new as foo)
• A query cache
• MERGE TABLES
• A better GUI client

© 1995-2000 MySQL AB Last modified: 2000/07/29 00:46:01