InnoDB Performance

Optimization

Heikki Tuuri, Innobase Oy/Oracle Corp.

Peter Zaitsev, Percona Ltd

April 23-26 2007

About Speakers
 Heikki Tuuri
Creator of InnoDB Storage engine
InnoDB Lead Developer
 Peter Zaitsev
MySQL/InnoDB Performance Expert
Long time InnoDB User
 Speaking together
to share internals and practical use insights
It all starts with Application Design
General Application Design is Paramount
 Design your schema, indexes and queries right
Storage engine aspects are often fine tuning
 In some cases Storage Engine selection may affect your
schema layout and indexes
 We're not covering general schema design guidelines in this
presentation, but will focus on the InnoDB Storage Engine.
Each storage engine is special
 MySQL offers multiple Storage Engines
 Each of them has unique design and operating properties
 Application written for one storage engine may not perform
best with other storage engines
 Each Storage Engine has special optimizations so they can
benefit from certain design patterns
 We'll cover DO and DON'T for the InnoDB Storage Engine
Make use of Transactions
 There are always transactions with InnoDB, even if you do
not use them explicitly
Each statement will be in its own transaction (assuming you run in
the “autocommit mode)
With transaction commit overhead for each statement
 Wrap multiple updates in the same transaction for efficient
operation (SET AUTOCOMMIT = 0; ... COMMIT; ...
COMMIT;)
Do not make transactions too large, however
Make sure you're catching Deadlocks and Wait Timeouts
Do not use LOCK TABLES
 LOCK TABLES is designed to work with table level locking
storage engines
With row level lock storage engines, transactions are better choice
LOCK TABLES behavior with InnoDB tables is different in MySQL
versions and depends on –innodb_table_locks
can give problems for portable applications if you port from
MySQL-4.0 to later
Behavior might not be as may be used to with MyISAM tables.
PRIMARY KEY Clustering
 PRIMARY KEY is Special
Accessing data by PRIMARY KEY is faster than other keys
True for both In-Memory and Disk Based accesses
Try to do most lookups by primary key
Data is clustered by PRIMARY KEY
Sequential PK values will likely have data on the same page
PK range and prefix lookups are very efficient
Can be used to cluster data accessed together
Storing user messages one can use (user_id,message_id)
primary key to keep all users messages in a few pages.
PK is a “covering index” for any set of fields in the PK
Cost of Clustered Primary Key
 PRIMARY KEY in random order are costly and lead to table
fragmentation (primary key inserts should normally be in an
ascending order)
Load data in primary key order if you can
Sometimes changing primary key to auto_increment is a good idea
 There is always a clustered key internally even if you do not
specify one
So better define one and use it
 PRIMARY KEY column updates are expensive
Requires row data physically to be moved from one place in the
index to another.
Generally not a good schema/application design either!
Keep PRIMARY KEY Short
 Secondary indexes use primary key to refer to the clustering
index
Making primary key value part of any index
 Long primary keys make your indexes long and slow
Keep them short
 You can often change current primary key to UNIQUE KEY
and add auto_increment PRIMARY KEY; you can't have
InnoDB to create its internal primary key simply by changing a
PRIMARY key to UNIQUE because MySQL will internally convert a
not-NULL UNIQUE key to a primary key if one is missing
 If you only have PRIMARY KEY on the table and have all
lookups done by it, leave it even if it is long, as PK lookups
are so much faster.
InnoDB Indexing
 Be easy on UNIQUE Indexes
They do not use the “insert buffer” which can speed up index
updates significantly
 Indexes are not prefix compressed
so they can take much more space than for MyISAM
avoid excessive indexes.
 Keep Updates Fully Indexed
Or you can see unexpected locking problems
DELETE FROM users WHERE name=“peter”
may have to lock all rows in the table if the column name is
not indexed.
Auto Increment may limit scalability
 Auto-Increment INSERTS may use table level locks (but
only to the end of the INSERT statement, not transaction)
even if you specify the auto_increment column value!
 Limits scalability for concurrent inserts
 A fix being worked on
 Work around by assigning values outside of MySQL
be careful with uuid as they result in both long and random primary
keys.
Multi Versioning
 Complements row level locking to get even better
concurrency
 Standard SELECT statements set no locks, just reads
appropriate row version
LOCK IN SHARE MODE, FOR UPDATE modifiers can be done to
do locking reads
 Even long running selects do not block updates to the table
or other selects
 Overly long queries (transactions in general) are bad for
performance as a lot of unpurged versions accumulate.
READ COMMITTED can ease the problem.
InnoDB is only able to remove a row version when no transactions
are open which can read it.
... FOR UPDATE and LOCK IN SHARE MODE
 Locking selects are executed in read committed mode
Because you can't lock a row which does not exist
So results of these queries can be different than for standard
SELECTs
 SELECT ... FOR UPDATE always has to access row data
page to set the lock, so it can't run index covered queries
which can slow down queries a lot
Reducing Deadlocks
 Deadlocks are normal for a transactional database
Non-locking SELECT statements do not deadlock with InnoDB
Make sure to handle deadlocks and lock wait timeouts in your
application
 Make sure your transactions lock data in the same order
when possible
 Have update chunks smaller (chop transactions)
 Use SELECT ... FOR UPDATE if you're going to update
most of the selected rows.
 Use external locking to avoid problem - Application level
locks, SELECT GET_LOCK('mylock') etc.
How isolation modes affect Performance
 InnoDB supports a number of Isolation Modes, which can
be set globally, per connection or per transaction.
READ UNCOMMITED - Rarely used, can use if you are fine with
dirty reads but performance improvement is limited
READ COMMITED – Results of all committed transactions become
visible to the next statement. May be more efficient than higher
isolation levels. Allows old versions to be purged faster.
In MySQL-5.1, InnoDB does little 'gap locking' on this level: use
row-based replication and binlogging to avoid problems!
REPEATABLE READ – Default. Reads within transactions are fully
repeatable, no phantom rows.
SERIALIZABLE - Makes all selects locking selects, avoid when
possible.
Foreign Keys Performance
 InnoDB checks foreign keys as soon as a row is updated,
no batching is performed or checks delayed till transaction
commit
Foreign keys are often serious performance overhead, but help
maintain data consistency
 Foreign Keys increase amount of row level locking done
and can make it spread to a lot of tables besides the ones directly
updated
 Foreign Key locking in a child table is done when the
parent table is updated
(SELECT ... FOR UPDATE on the parent table does not lock
the child table)
Restrict Number of Open Transactions
 InnoDB performs best with a limited number of open
transactions and running queries.
 Multiple running queries may cause a lot of thrashing
bumping into each other
Work is done to improve performance in such cases
innodb_thread_concurrency can be used to restrict number of
threads in InnoDB kernel
 Many open transactions make lock graph building more
complicated and increase some other overhead.
 When possible, keep a limited number of queries running at
the same time, do queuing on application side
Beware of a very high number of tables
 InnoDB has its own table definition (dictionary) cache
independent of the MySQL table_cache variable
 Once opened, InnoDB never removes table from the cache.
 4KB+ may be consumed by each table
InnoDB in MySQL 5.1 has reduced this number by 50 % - 75 %
 On restart, statistics are recomputed for each table
So the first time open operation is very expensive
plus MySQL table_cache serializes these operations
INSERT ... SELECT
 INSERT... SELECT statement runs locking select
 Required for logical level replication to work properly
problem goes away with MySQL 5.1 row level replication and the
READ COMMITTED isolation level
 Behavior is the same whenever you have log-bin enabled
or not, to keep things consistent
 innodb_locks_unsafe_for_binlog helps in 5.0, but your
replication may get broken
it also disables next-key locks
 SELECT ... INTO OUTFILE + LOAD DATA INFILE can be
often use as non-blocking safe alternative
Next Key Locks (Gap Locks)
 InnoDB not only locks rows themselves but the “gap”
between rows as well
 Prevents phantom rows
Makes “REPEATABLE READ” really repeatable with InnoDB
 Needed for MySQL statement level replication to work
properly.
 Increases locking for some write heavy workloads.
 Can be disabled if you're not running binary logging (for
replication or recovery)
 Is safe to change in MySQL 5.1 if you use row-based
replication
Count(*) facts and myths
 “InnoDB does not handle count(*) queries well” - Myth
Most count(*) queries are executed same way by all storage
engines
SELECT COUNT(*) FROM articles WHERE user_id=5
 “InnoDB does not optimize count(*) queries without where
clause” - Fact
SELECT COUNT(*) FROM users;
InnoDB can't simply store count of the rows as it each transaction
has its own view of the table. Significant work required to
implement
You can use triggers and counter table to work it around
SHOW TABLE STATUS LIKE “users” will show approximate
row count for the table (which is changing all the time)
InnoDB and Group Commit
 Group Commit – commit several outstanding transactions
with single log write
Can improve performance dramatically, especially if no RAID with
BBU
 In MySQL 5.0, group commit does not work with binary
logging
Due to a way XA (distributed transactions) support was
implemented
Watch out if upgrading from MySQL 4.1
Back to basics with Server Settings Tuning
It all starts with Memory
 innodb_buffer_pool_size
Specifies main InnoDB buffer – Data and Index pages, insert buffer,
locks are stored here
Very important for performance on large data sets
Much more efficient than OS cache, especially for Writes
InnoDB has to bypass OS buffering for writes
Can be set to 70-80% of memory for dedicated InnoDB-Only
MySQL
Default value is just 8M, independent of available memory; make
sure to configure it
 innodb_additional_mem_pool
just stores dictionary, automatically increased, do not set too high
InnoDB Logging
 innodb_log_file_size
Dramatically affects write performance. Keep it high
High values increase recovery time though
Check how large logs you can afford
4GB total size limit
 innodb_log_files_in_group
this number of files of specified size are used for log.
Usually no need to change default value
InnoDB Logging
 innodb_log_buffer_size
Do not set over 2-8MB unless you use huge BLOBs, Log file is
flushed at least once per second anyway
Check Innodb_os_log_written growth to see how actively your logs
are written.
InnoDB Logs are physio-logical, not page based so they are very
compact
 innodb_flush_logs_at_trx_commit
By default logs are flushed to the disk at each transaction commit
Required for ACID guarantees, expensive
Can set to 2 or 0 if you can afford losing transactions for last 1 sec
or so (ie if you're using it as MyISAM tables replacement)
InnoDB Log Resizing
 Is not as simple as changing option and restarting :)
 Shut down MySQL Server
 Make sure it shut down normally (check error log for errors)
 Move away InnoDB log files ib_log*
 Start MySQL Server
 Check error log files to see it successfully created new log
files.
innodb_flush_method
 Specifies a way InnoDB will work with OS File System
 Windows: unbuffered IO mode is always used
 Unix: can use fsync() or O_SYNC/O_DSYNC for flushing
files
fsync() is usually faster; allows accumulating multiple writes and
executing them in parallel
Some OS allow disabling OS caching for InnoDB data files
Good. You do not want data to be cached twice – waste.
 Linux: O_DIRECT uses direct unbuffered IO
Avoids double buffering, May make writes slower
innodb_file_per_table
 InnoDB can store each table in its own file
 Main tablespace is still needed for system needs
 Can help to spread tables to multiple disks
 Allows to reclaim space if a table is dropped
 Sometimes slower for writes as fsync() is called sequentially
 Can increase startup/shutdown time with large number of
tables
Other File IO Settings
 innodb_autoextend_increment – specifies growth
increment for shared tablespace (not for per table
tablespaces)
larger values allow to reduce fragmentation.
 innodb_file_io_threads – changes number of IO threads,
on Windows only. Note all 4 threads are doing different jobs
 innodb_open_files - number of files used for per table
tablespaces. Increase if you have a lot of tables
No stats available so far to show number of re-opens InnoDB
needs to do
 innodb_support_xa setting to 0 reduces work InnoDB
should do on transaction commit. Binlog can get out of sync
Minimizing restart time
 InnoDB buffer pool may have a lot of unflushed data
So shutdown may take very long time
 If you need to minimize downtime:
 SET GLOBAL innodb_max_dirty_pages_pct=0
 Watch Innodb_buffer_pool_pages_dirty in SHOW
STATUS
 As it gets close to 0 shut down the server
 During this operation performance will be lower as InnoDB
will be flushing dirty pages aggressively.
Troubleshooting runaway Purge
 InnoDB does not remove rows on delete (and old row
versions on update) because these may be needed by other
transactions
 Purge thread is used to clean up these unused rows
 In some workloads, the Purge thread may not be able to
keep up and the tablespace will grow without bounds.
Check “TRANSACTIONS” section in SHOW INNODB STATUS
 innodb_max_purge_lag – limits number of transactions
which have updated/deleted rows
 Will delay insert/updates so purge thread can keep up
 Why do not we get to have multiple purge threads instead?
Concurrency Control Settings
 Settings help to adjust how InnoDB handles a large number
of concurrent transactions
 innodb_thread_concurrency – Number of threads allowed
inside InnoDB kernel at the same time (0 – no limit)
2*(NumCores+NumDisk) is good value in theory, smaller usually
work better in practice
 innodb_commit_concurrency - Number of threads
allowed at commit stage at the same time
 innodb_concurrency_tickets - Number of operations
thread can do before it has to exit kernel and wait again
 innodb_thread_sleep_delay
 innodb_sync_spin_loops
Unsafe ways to gain performance
 InnoDB has a lot of checks and techniques to minimize
chance of data being corrupted or lost
 innodb_doublewrite - protection from partial page writes,
only disable if OS guarantees it does not happen
 innodb_checksums - checksums for data in pages, helps
to discover file system corruption, broken memory and other
problems
Causes a few percent of overhead for most workloads
Disable when such performance gain is more important
Benchmarks ?
InnoDB SHOW STATUS Section
 MySQL 5.0 finally has some InnoDB performance counters
exported in SHOW STATUS
They are GLOBAL while most of other counters are per thread now
They are mostly taken from SHOW INNODB STATUS
 Will only list some examples
 Innodb_buffer_pool_pages_misc - number of pages in
BP used for needs other than caching pages
 Innodb_buffer_pool_read_ahead_rnd – number of
random read-aheads InnoDB performed
 Innodb_buffer_pool_read_requests,
Innodb_buffer_pool_reads can be used to compute
cache read hit ratio
SHOW INNODB STATUS
 The tool for InnoDB troubleshooting
“Send a couple of SHOW INNODB STATUS outputs when it
happens”
 Has information as in SHOW STATUS plus much more
 Information about running transactions (their locks etc.)
 Information about last deadlock, foreign key, etc.
 Information about latches, spinlocks, OS waits
 More details
http://www.mysqlperformanceblog.com/2006/07/17/show-innodb-
status-walk-through/
SHOW MUTEX STATUS
 A tool to show what mutexes are hot for your workload
 Details of what really happens with which mutexes – spin
locks ? OS Waits ?
 timed_mutexes - track how long OS Wait was taking

M utex:&kernel_m utex
M odule:srv0srv.c
C ount:1828074122
Spin_w aits:762647
Spin_rounds:4781433
O S_w aits:96879
O S_yields:155883
O S_w aits_tim e:0
Hardware and OS Selection
Hardware and OS Selection Checklist
 Which CPUs and how many of them ?
 How Much Memory ?
 How to set up IO Subsystem ?
 Does OS Selection matter ?
 Which File System is best to use ?
Selecting CPUs
 Different CPUs/Architectures scale differently with InnoDB
 Old “NetBurst” based Xeons scale poorly
 New “Core” based Xeons and Opterons are better
 X86_64 is the leading
 Multi-Core works well with InnoDB
 Over 8 cores per system is reasonable limit
Depends on workload significantly
Innobase is working on further improvements
 Scale Out, use multiple lower end servers.
 32bit CPUs should be dead by now, so 32bit OS
How much memory ?
 Memory is most frequent performance limiting factor for well
tuned applications
 InnoDB can use large memory amounts efficiently
 Working set must fit in memory
The data pages which are accessed most often
Do not count by rows:
100,000,000 of 100 byte rows, random 1,000,000 are working
set – can touch most of the pages.
 Can be 5% of total database size or can be 50%
 Make sure to use a 64bit platform, OS and MySQL Version.
How to set up IO SubSystem
 InnoDB loads a few hard drives well, but not 100 of them
6-8 per node seems to be optimal configuration
 Directly Attached storage usually works best
 SAN – increased latency, expensive
 NAS – Avoid, risk of data corruption
 ISCSI – good for some cases, increased latency
 RAID – Battery backed up cache is very important
Make sure you have BBU before enabling WriteBack cache
 Hard Drive cache itself should be turned off, or make sure it
is flushed on fsync() or corruption can happen in OS crash.
Local storage configuration
 Logs on separate RAID1 volume
Can be helpful, in many cases better to share disk for data
 Binary logs on separate volume – can be good idea for
backup recovery reasons
 RAID10 good for tablespace
degraded performance can be worse than expected.
 RAID5 can be good for certain workloads
just make sure you account for degraded performance.
 Large RAID Stripe (128K+) is best in theory but many RAID
controllers do not handle these well.
 Software RAID is OK, especially RAID1
Does OS Selection Matter ?
 Consider Performance, Tools available, Community
Experience
 Windows – used for development, small installations, few
Web/Enterprise scale projects
 Solaris – offering some great tools now, works to make
MySQL work well with it, bad community support.
 FreeBSD – had history of problems with MySQL in general,
now gets better, fewer tools available, less usage in
production.
 Linux – Most commonly used platform for production and
Development. Tools like LVM, Journaling filesystems.
Selecting FileSystem
 Applies mainly to Linux which has too many choices
 EXT3 – default filesystem in most distributions, works OK
for lower end installations
 ReiserFS – support removed from many Linux distributions.
Generally no big win with typical MySQL workload
 XFS – Used with a lot of drives in RAID, can give serious
performance improvement
 JFS – Rarely used at this point.
 Raw partition for InnoDB tablespace – rarely used.
 There are often too high expectations about performance
gains by switching file systems.
Recent InnoDB Performance Developments
InnoDB Scalability Patches
 Decreased contention over buffer pool pages
Available in 5.0, backported to 4.1
 Improved sync_array implementation in MySQL 5.1
 Performance gains are very different based on
workload,hardware, concurrency
 Can range from few percent to multiple times
 Performance still goes down with high number of concurrent
threads.
 Prototype for further scalability improvement patches is
available from community
Other Improvements
 Row Level replication in MySQL 5.1 eases gap locking
 Working on removing Auto_increment “table locks”
 Zip compression of database pages
 Fast index creation
- No full table rebuild required
- “Sorting” gives less physically fragmented index
Questions from the audience
 [email protected]
 Visit blog for more Innodb tips
http://www.mysqlperformanceblog.com
 Looking for some help ?
[email protected]