Highload Perf Tuning

Performance Tuning MySQL

Highload++, 11 October 2009

Morgan Tocker, ﬁrstname at percona dot com
Director of Training, Percona Inc.

Introduction
• So you’ve written/inherited an application.
• Usage has gone crazy.
• And you’ve diagnosed that the database is the
bottleneck.
• So how do we ﬁx it......

What this talk is about.
• We’re going to be talking about tuning a system
that’s never been shown love.
– There’s more than one strategy to do this.
– I’m going to show you the three common paths people
take.

The three ways people take:
• Option 1: Upgrade the Hardware.
• Option 2: Change a conﬁguration setting in MySQL.
• Option 3: Improve Indexing/Tune Queries.

Option 1: Add Hardware
• This “Kind of” works... but for how long?
• There’s a price point where big machines get
exponentially more expensive.

Add Hardware (cont.)
• If you are going to do this, beware:
– Bigger Machines will have the same speed drives.
– Check what your bottleneck is ﬁrst!

• Potential Bottlenecks:
DISK RAM
This is the number one bottleneck for Use it to relieve pressure off our
a lot of people. disks, particularly with expensive
random reads.

CPU NETWORK
Not normally a bottleneck, but having We care about round trips, but we’re
fast CPUs can mean locking code is not usually limited by throughput.
blocking for less time.

When this technique really works:
• You had good performance when all of your working
set[1] of data ﬁtted in main memory.
• As your working set increases, you just increase the
size of memory[2].

[1] Working set can be between 1% and 100% of database size. On poorly
indexed systems it is often a higher percentage - but the real value depends on
what hotspots your data has.

[2] The economical maximum for main memory is currently 128GB. This can be
done for less than $10K with a Dell server.

Pros:
Good if you have a large working set or an “excess money”
problem.
Cons:
• Not as easy to get many multiples better performance.
• Can get expensive once you get past a price point of
hardware.
• Still some features missing in MySQL hurting the very top end
of users (save contents of buffer pool for warm restarts)
Conclusion:
I wouldn’t normally recommend this be your ﬁrst optimization
path if the system hasn’t ever been tuned.

Option 2: Change Configuration
• The “--run-faster
startup option”.
• This may work, but it
assumes misconﬁgured
setting to start with.
• There are no silver
bullets.

“Silver Bullets” kill Werewolf (Оборотни)

Changing Configuration
★ Most of this is based off running the command SHOW
GLOBAL STATUS ﬁrst, then analyzing the result.
★ Be careful when running this - the period of time it
aggregates data for may cause skew.
✦
A very simple utility called ‘mext’ solves this - http://
www.xaprb.com/mext

12

Temporary Tables on Disk
★ You may be able to change tmp_table_size and
max_heap_table_size to end up with to increase the
threshold.

Created_tmp_disk_tables 0 0 0
Created_tmp_files 5 0 0
Created_tmp_tables 12550 421 417

13

Temporary Tables on Disk (cont.)
★ These are often caused by some internal GROUP BYs
and complex joins with an ORDER BY that can’t use an
index.
★ They default to memory unless they grow too big, but...
★ All temporary tables with text/blob columns will be
created on disk regardless!

14

Binary Log cache
★ Updates are buffered before being written to the binary
log. If they’re too big, the buffer creates a temporary ﬁle
on disk:
★ mysql> show global status like 'binlog%';
+-----------------------+-------+
| Variable_name | Value |
+-----------------------+-------+
| Binlog_cache_disk_use | 1082 |
| Binlog_cache_use | 78328 |
+-----------------------+-------+
2 rows in set (0.00 sec)
★ Corresponding Session Variable:
binlog_cache_size

15

Sorting Data (cont.)
★ Caused by:
✦
ORDER BY (and not being able to use an index for sorting).
✦
GROUP BY (instead of GROUP BY c ORDER BY NULL).
★ sort_merge_passes is incremented every time the
internal sort algorithm has to loop over more than once to
sort.
✦
A small number is healthy - be careful not to over set the
sort_buffer_size.
✦
Sometimes I look at how many sort_merge_passes occur per
second (run SHOW GLOBAL STATUS more than once).

17

Query Cache
★ mysql> show global status like 'Qcache%';
+-------------------------+--------+
+-------------------------+--------+
| Qcache_free_memory | 99812 |
| Qcache_hits | 210213 |
| Qcache_inserts | 82333 |
| Qcache_not_cached | 2032 |
| Qcache_queries_in_cache | 5322 |
+-------------------------+--------+

18

Query Cache (cont.)
★ You really need to have at least as many hits as inserts,
but query cache discussion is not-that-simple a
discussion.
★ The query cache does not scale well on SMP machines.
✦
Unless you get large multiples of hits over inserts you may choose
to disable it with query_cache_type = 0

19

Table Locks
★ Some storage engines (MyISAM, Memory) have table
level locking. Under concurrency this can be a real
contention point:
★ mysql> show global status like 'table_locks%';
+-----------------------+-------+
+-----------------------+-------+
| Table_locks_immediate | 52323 |
| Table_locks_waited | 3293 |
+-----------------------+-------+
★ Tip: You really need to watch this one in particular in
mext. Locking problems tend to snowball.

20

Table Cache
★ MySQL requires a copy of a each table open per
connection. The default table_cache is lower than
the default max_connections!
★ mysql> show global status like 'Open%tables';
+---------------+--------+
+---------------+--------+
| Open_tables | 64 |
| Opened_tables | 532432 |
+---------------+--------+
★ Corresponding Global Variable:
table_cache_size

21

Max Connections
★ Seeing max_used_connections equal to
max_connections indicates that a connection was
likely refused at some point:
★ mysql> show global status like 'max%';
+----------------------+-------+
+----------------------+-------+
| Max_used_connections | 401 |
+----------------------+-------+
1 row in set (0.00 sec)
max_connections

23

Cartesian Products?
★ Joining two tables without an index on either can often
mean you’re doing something wrong. You can see this if
Select_full_join > 0:
★ mysql> show global status like 'Select_full_join';
+------------------+-------+
+------------------+-------+
| Select_full_join | 0 |
+------------------+-------+

24

InnoDB Buffer Pool
mysql> pager grep -B1 -A12 'BUFFER POOL AND MEMORY'
mysql> show innodb status;
----------------------
BUFFER POOL AND MEMORY
----------------------
Total memory allocated 1205671692; in additional pool
allocated 1029120
Buffer pool size 65536
Free buffers 56480
Database pages 8489
Modified db pages 0
Pending reads 0
Pending writes: LRU 0, flush list 0, single page 0
Pages read 8028, created 485, written 96654
0.00 reads/s, 0.00 creates/s, 0.00 writes/s
Buffer pool hit rate 1000 / 1000
--------------

InnoDB Log Buffer Size
mysql> show global status like 'innodb_log_waits';
+------------------+-------+
+------------------+-------+
| Innodb_log_waits | 0 |
+------------------+-------+

Corresponding Global Variable:
innodb_log_buffer_size

Best way to review global status?
★ Trained Eye helps, but you still miss things sometimes.
★ Most of this can be automated. The tool I like the most
(for simplicity) is this one:
✦
Matthew Montgomery’s Tuning Primer:
http://forge.mysql.com/projects/project.php?id=44

27

Every setting has a range!
★ You really can have too much of a good thing.
★ It takes more resources to allocate larger chunks of
memory, and in some cases you’ll miss valuable CPU
caches.
★ We’ve blogged about this with sort_buffer_size
here:
✦
http://www.mysqlperformanceblog.com/2007/08/18/how-fast-
can-you-sort-data-with-mysql/

28

Change Configuration (cont.)
Pros:
Can get some quick wins, sometimes.
Cons:
Assumes a setting is misconfigured in the first place.
Over-tuning can cause negative effects. Try setting
your sort_buffer_size to 400M to find out how!
Conclusions:
Not a bad approach - since it is easy to apply without
changing your application.

Option 3: Add an index
• Should really be called “Add an index, or slightly
rewrite a query”.
• This is the least “fun” approach.
• It delivers the most value for money though!

The EXPLAIN Command
mysql> EXPLAIN SELECT Name FROM Country WHERE continent =
'Asia' AND population > 5000000 ORDER BY NameG
*************************** 1. row ***************************
id: 1
select_type: SIMPLE
table: Country
type: ALL
possible_keys: NULL
key: NULL
key_len: NULL
ref: NULL
rows: 239
Extra: Using where; Using filesort

Explain (cont.)
mysql> ALTER TABLE Country ADD INDEX p (Population);
Query OK, 239 rows affected (0.01 sec)
Records: 239 Duplicates: 0 Warnings: 0
mysql> EXPLAIN SELECT Name FROM Country WHERE Continent =
'Asia' AND population > 5000000 ORDER BY NameG
*************************** 1. row ***************************
id: 1
select_type: SIMPLE
table: Country
type: ALL
possible_keys: p
key: NULL
key_len: NULL
ref: NULL
rows: 239

Now it is...
mysql> EXPLAIN SELECT Name FROM Country WHERE Continent = 'Asia'
AND population > 50000000 ORDER BY NameG
*************************** 1. row ***************************
id: 1
select_type: SIMPLE
table: Country
type: range
possible_keys: p
key: p
key_len: 4
ref: NULL
rows: 54

Another Index..
mysql> ALTER TABLE Country ADD INDEX c (Continent);

*************************** 1. row ***************************
id: 1
select_type: SIMPLE
table: Country
type: ref
possible_keys: p,c
key: c
key_len: 1
ref: const
rows: 42

Changes back to p at 500M!
*************************** 1. row ***************************
id: 1
select_type: SIMPLE
table: Country
type: range
possible_keys: p,c
key: p
key_len: 4
ref: NULL
rows: 4

Try another index...
mysql> ALTER TABLE Country ADD INDEX p_c (Population, Continent);


*************************** 1. row ***************************
id: 1
select_type: SIMPLE
table: Country
type: ref
possible_keys: p,c,p_c
key: c
key_len: 1
ref: const
rows: 42

How about this one?
mysql> ALTER TABLE Country ADD INDEX c_p (Continent,Population);

*************************** 1. row ***************************
id: 1
select_type: SIMPLE
table: Country
type: range
possible_keys: p,c,p_c,c_p
key: c_p
key_len: 5
ref: NULL
rows: 7

The Best...
mysql> ALTER TABLE Country ADD INDEX c_p_n
(Continent,Population,Name);
*************************** 1. row ***************************
id: 1
select_type: SIMPLE
table: Country
type: range
possible_keys: p,c,p_c,c_p,c_p_n
key: c_p_n
key_len: 5
ref: NULL
rows: 7
Extra: Using where; Using index; Using filesort

So what’s the end result?
• We’re looking at 9 rows, not the whole table.
– We’re returning those rows from the index - bypassing the
table.
• A simple example - but easy to demonstrate how to
reduce table scans.
• You wouldn’t add all these indexes - I’m just doing it
as a demonstration.
– Indexes (generally) hurt write performance.

Example 2: Join Analysis
mysql> EXPLAIN SELECT * FROM city WHERE countrycode IN
(SELECT code FROM country WHERE name='Australia')G
*************************** 1. row ***************************
id: 1
select_type: PRIMARY
table: city
type: ALL
possible_keys: NULL
key: NULL
key_len: NULL
ref: NULL
rows: 4079
Extra: Using where
*************************** 2. row ***************************
id: 2
select_type: DEPENDENT SUBQUERY
table: country
type: unique_subquery
possible_keys: PRIMARY
key: PRIMARY
key_len: 3
ref: func
rows: 1
Extra: Using where

Join analysis (cont.)
mysql> EXPLAIN SELECT city.* FROM city, country WHERE
city.countrycode=country.code AND country.name='Australia'G
*************************** 1. row ***************************
id: 1
select_type: SIMPLE
table: city
type: ALL
possible_keys: NULL
key: NULL
key_len: NULL
ref: NULL
rows: 4079
Extra:
*************************** 2. row ***************************
id: 1
select_type: SIMPLE
table: country
type: eq_ref
key: PRIMARY
key_len: 3
ref: world.city.CountryCode
rows: 1
Extra: Using where

Try an index...
mysql> ALTER TABLE city ADD INDEX (countrycode);

Is that any better?
*************************** 1. row ***************************
id: 1
select_type: SIMPLE
table: city
type: ALL
possible_keys: CountryCode
key: NULL
key_len: NULL
ref: NULL
rows: 4079
Extra:
*************************** 2. row ***************************
id: 1
select_type: SIMPLE
table: country
type: eq_ref
key: PRIMARY
key_len: 3
ref: world.city.CountryCode
rows: 1
Extra: Using where

Try Again
mysql> ALTER TABLE country ADD INDEX (name);

Looking good...
*************************** 1. row ***************************
id: 1
select_type: SIMPLE
table: country
type: ref
possible_keys: PRIMARY,Name
key: Name
key_len: 52
ref: const
rows: 1
Extra: Using where
*************************** 2. row ***************************
id: 1
select_type: SIMPLE
table: city
type: ref
possible_keys: CountryCode
key: CountryCode
key_len: 3
ref: world.country.Code
rows: 18
Extra:

My Advice
• Focus on components of the WHERE clause.
• The optimizer does cool things - don’t make
assumptions. For Example:
– EXPLAIN SELECT * FROM City WHERE id =
1810;
– EXPLAIN SELECT * FROM City WHERE id = 1810
LIMIT 1;
– EXPLAIN SELECT * FROM City WHERE id
BETWEEN 100 and 200;
– EXPLAIN SELECT * FROM City WHERE id >= 100
and id <= 200;

More information
• http://dev.mysql.com/
EXPLAIN
• Some very good examples are also
in “High Performance MySQL” 2nd
Ed.

Add an index (conclusion)
Pros:
The biggest wins. Seriously.
Cons:
Takes a bit of time for analysis. If you need to rewrite
a query - you need to go inside the application (not
everyone can).
Conclusion:
My #1 Recommendation.

Finding bad queries
• MySQL has a feature called the slow query log.
• We can enable it, and then set the long_query_time
to zero[1] seconds to ﬁnd a selection of our queries.

[1] Requires MySQL 5.1 or patches MySQL 5.0 release.

root@ubuntu:~# perl mk-query-digest /bench/mysqldata/ubuntu-slow.log
# 1461.1s user time, 39.2s system time, 22.20M rss, 57.52M vsz
# Overall: 7.26M total, 38 unique, 17.28k QPS, 18.88x concurrency ________
# total min max avg 95% stddev median
# Exec time 7929s 12us 918ms 1ms 4ms 10ms 138us
# Lock time 154s 0 17ms 21us 36us 33us 18us
# Rows sent 5.90M 0 246 0.85 0.99 6.71 0.99
# Rows exam 6.90M 0 495 1.00 0.99 13.48 0
# Time range 2009-09-13 17:26:54 to 2009-09-13 17:33:54
# bytes 765.14M 6 599 110.56 202.40 65.01 80.10
# Rows read 0 0 0 0 0 0 0

52

..
# Query 1: 655.60 QPS, 4.28x concurrency, ID 0x813031B8BBC3B329 at byte 518466
# This item is included in the report because it matches --limit.
# pct total min max avg 95% stddev median
# Count 3 274698
# Exec time 22 1794s 12us 918ms 7ms 2ms 43ms 332us
# Lock time 0 0 0 0 0 0 0 0
# Rows sent 0 0 0 0 0 0 0 0
# Rows exam 0 0 0 0 0 0 0 0
# Users 1 [root]
# Hosts 1 localhost
# Databases 1 tpcc
# Time range 2009-09-13 17:26:55 to 2009-09-13 17:33:54
# bytes 0 1.57M 6 6 6 6 0 6
# Query_time distribution
# 1us
# 10us ###
# 100us ################################################################
# 1ms ##
# 10ms ##
# 100ms #
# 1s
# 10s+
commitG

53

..
# Query 2: 2.05k QPS, 4.20x concurrency, ID 0x10BEBFE721A275F6 at byte 17398977
# Count 11 859757
# Exec time 22 1758s 64us 812ms 2ms 9ms 9ms 224us
# Lock time 17 27s 13us 9ms 31us 44us 26us 28us
# Rows sent 0 0 0 0 0 0 0 0
# Rows exam 0 0 0 0 0 0 0 0
# Users 1 [root]
# Hosts 1 localhost
# Databases 1 tpcc
# Time range 2009-09-13 17:26:55 to 2009-09-13 17:33:54
# bytes 22 170.52M 192 213 207.97 202.40 0.58 202.40
# 1us
# 10us #
# 100us ################################################################
# 1ms ############
# 10ms ###
# 100ms #
# 1s
# 10s+
# Tables
# SHOW TABLE STATUS FROM `tpcc` LIKE 'order_line'G
# SHOW CREATE TABLE `tpcc`.`order_line`G
INSERT INTO order_line (ol_o_id, ol_d_id, ol_w_id, ol_number, ol_i_id, ol_supply_w_id,
ol_quantity, ol_amount, ol_dist_info) VALUES (3669, 4, 65, 1, 6144, 38, 5, 286.943756103516,
'sRgq28BFdht7nemW14opejRj')G

54

..
# Query 4: 2.05k QPS, 1.42x concurrency, ID 0x6E70441DF63ACD21 at byte 192769443
# Count 11 859769
# Exec time 7 597s 67us 794ms 693us 467us 6ms 159us
# Lock time 12 19s 9us 10ms 21us 31us 25us 19us
# Rows sent 0 0 0 0 0 0 0 0
# Rows exam 0 0 0 0 0 0 0 0
# Users 1 [root]
# Hosts 1 localhost
# Databases 1 tpcc
# Time range 2009-09-13 17:26:55 to 2009-09-13 17:33:54
# bytes 7 56.36M 64 70 68.73 65.89 0.30 65.89
# 1us
# 10us #
# 100us ################################################################
# 1ms #
# 10ms #
# 100ms #
# 1s
# 10s+
# Tables
# SHOW TABLE STATUS FROM `tpcc` LIKE 'stock'G
# SHOW CREATE TABLE `tpcc`.`stock`G
UPDATE stock SET s_quantity = 79 WHERE s_i_id = 89277 AND s_w_id = 51G
# Converted for EXPLAIN
# EXPLAIN
select s_quantity = 79 from stock where s_i_id = 89277 AND s_w_id = 51G

55

..
# Rank Query ID Response time Calls R/Call Item
# ==== ================== ================ ======= ========== ====
# 1 0x813031B8BBC3B329 1793.7763 23.9% 274698 0.006530 COMMIT
# 2 0x10BEBFE721A275F6 1758.1369 23.5% 859757 0.002045 INSERT order_line
# 3 0xBD195A4F9D50914F 924.4553 12.3% 859770 0.001075 SELECT UPDATE stock
# 4 0x6E70441DF63ACD21 596.6281 8.0% 859769 0.000694 UPDATE stock
# 5 0x5E61FF668A8E8456 448.0148 6.0% 1709675 0.000262 SELECT stock
# 6 0x0C3504CBDCA1EC89 308.9468 4.1% 86102 0.003588 UPDATE customer
# 7 0xA0352AA54FDD5DF2 307.4916 4.1% 86103 0.003571 UPDATE order_line
# 8 0xFFDA79BA14F0A223 192.8587 2.6% 86122 0.002239 SELECT customer warehouse
# 9 0x0C3DA99DF6138EB1 191.9911 2.6% 86120 0.002229 SELECT UPDATE customer
# 10 0xBF40A4C7016F2BAE 109.6601 1.5% 860614 0.000127 SELECT item
# 11 0x8B2716B5B486F6AA 107.9319 1.4% 86120 0.001253 INSERT history
# 12 0x255C57D761A899A9 103.9751 1.4% 86120 0.001207 UPDATE warehouse
# 13 0xF078A9E73D7A8520 102.8506 1.4% 86120 0.001194 UPDATE district
# 14 0x9577D48F480A1260 91.3182 1.2% 56947 0.001604 SELECT customer
# 15 0xE5E8C12332AD11C5 87.2532 1.2% 86122 0.001013 SELECT UPDATE district
# 16 0x2276F0D2E8CC6E22 86.1945 1.1% 86122 0.001001 UPDATE district
# 17 0x9EB8F1110813B80D 83.1471 1.1% 86106 0.000966 UPDATE orders
# 18 0x0BF7CEAD5D1D2D7E 80.5878 1.1% 86122 0.000936 INSERT orders
# 19 0xAC36DBE122042A66 74.5417 1.0% 8612 0.008656 SELECT order_line
# 20 0xF8A4D3E71E066ABA 46.7978 0.6% 8612 0.005434 SELECT orders

56

Advanced mk-query-digest
• Query Review - the best feature ever.
• Saves the ﬁngerprint of your slow query, and only
shows you what you haven’t already looked at:

$ mk-query-digest --review h=host1,D=test,t=query_review
/path/to/slow.log

Audience Question:
• How do you ﬁnd unused indexes in MySQL?

Finding unused indexes (cont.)
• You have to come to my talk tomorrow for
the answer:
– 11 AM - Quick Wins with Third Party Patches

The Scoreboard

Option Effort Wins

Add Hardware
* 1/2

Tweak Settings
** **
Add an Index
*** *****

The End.
• Questions?

Photo Credits:
http://www.flickr.com/photos/7954439@N06/2535687572/

Highload Perf Tuning

Recommended

More Related Content

What's hot (20)

Viewers also liked (20)

Similar to Highload Perf Tuning (20)

More from HighLoad2009 (16)

Recently uploaded (20)

Highload Perf Tuning