October 22, 2014

MySQL server memory usage troubleshooting tips

There are many blog posts already written on topics related to “MySQL server memory usage,” but nevertheless there are some who still get confused when troubleshooting issues associated with memory usage for MySQL. As a Percona support engineer, I’m seeing many issues regularly related to heavy server loads – OR OOM killer got invoked and killed MySQL server due to high Memory usage… OR with a question like: “I don’t know why mysql is taking so much memory. How do I find where exactly memory is allocated? please help!”

There are many ways to check memory consumption of MySQL. So, I’m just trying here to explain it by combining all details that I know of in this post.

  • Check memory related Global/Session variables.

If you are using MyISAM then you need to check for Key_buffer_size, while using InnoDB, you can check innodb_buffer_pool_size,  innodb_additional_memory_pool_size, innodb_log_buffer_size,  innodb_sort_buffer_size (used only for sorting data while creating index in innodb, introduced from 5.6). max_connections, query_cache_size and table_cache are also important variables to check

We know that whenever a thread is connected to MySQL, it will need it’s own buffers when they are doing some complex operations like FTS,  sorting, creating temp tables etc. So we also need to check the size of read_buffer_size, sort_buffer_size, read_rnd_buffer_size and tmp_table_size.

There is a very good quote from High Performance MySQL, 3rd Edition: ” You can think of MySQL’s memory consumption as falling into two categories: the memory you can control, and the memory you can’t. You can’t control how much memory MySQL uses merely to run the server, parse queries, and manage its internals, but you have a lot of control over how much memory it uses for specific purposes.” So it seems we have to understand the purpose for configuring any variable… either it is Global or Session level. I would like to explain more about that here.

For the Global variables like key_buffer_size, query_cache_size etc,  MySQL always allocates and initializes the specified amount of memory all at once when the server starts. But it’s not happened for those who are global default but can be set as per-session variables, i.e  For read_buffer_size, sort_buffer_size, join_buffer_size, MySQL doesn’t allocate any memory for these buffers until query needs. But when a query needs, it immediately allocates the entire chunk of memory specified. So if there are even small sorts, full buffer size will be allocated which is just waste of memory. Even some buffers can be used multiple times. For example on queries that join several tables join_buffer can be allocated once per joined table. also some complicated queries including sub-queries can use multiple sort_buffers at the same time which can lead to high memory consumption. In some scenario, query didn’t even use sort_buffer whatever size is, as it select by primary key which will not allocate it. So it depends on the nature of your environment but I would say it’s always better to start with a safe variable value that can be larger than default if needed but not as large as it can consume all of the server’s memory.

One more thing,  not all per thread memory allocation is configured by variables.  Some of memory allocation per thread is done by MySQL itself for running complex processes/queries like “stored procedures” and it can take unlimited amount of memory while running. And sometimes, optimizer  can also take a lot of memory working with highly complex queries which generally we can’t control by any configuration parameter.

Even innodb_buffer_pool_size is not a hard limit, usually innodb uses 10% more memory than the one specified. Many people do not recommend using both storage engine MyISAM and InnoDB at the same time on production server. Because both have individual buffers which can eat all server memory.

For detailed information related to this topic, I would suggest reading this post from Peter Zaitsev titled “MySQL Server Memory Usage.”

  • Check “SHOW ENGINE INNODB STATUS” for section “BUFFER POOL AND MEMORY

Above one is from Native MySQL but if you’ll check the same with Percona Server you’ll get some more information.

This will give you information regarding how much memory is allocated by InnoDB. You can see here “Total Memory Allocated”, “Internal Hash Tables”, “Dictionary Memory Allocated”, “Buffer Pool Size” etc.

  • Profiling MySQL Memory usage with Valgrind Massif

Recently, I used this tool and surprisingly I got very good statistics about memory usage. Here the only problem is you have to shutdown the mysql, start it with valgrind massif and after collecting statistics, you again have to shutdown and normal start.

After getting massif.out file, you have to read it with ms_print command. You will see pretty nice graph and then statistics. i.e

You can see from the output that where memory is allocated, to which function etc. You can use this tool to find memory leaks. You can get more information here for how to install and use it. Here’s another related post by Roel Van de Paar titled: “Profiling MySQL Memory Usage With Valgrind Massif.”

If possible Valgrind massif should not be used on busy production server as it can degrade the performance. Generally it’s used to find memory leak by creating mirror environment on test/stage server and run on it. It needs debug binary to run so it decreases performance a lot. So it can be used for investigating some cases but not for regular use.

  • Check Plot memory usage by monitoring ps output. 

This also useful when you want to check how much virtual(VSZ) and real memory (RSS) is used by mysqld. You can either simply run some bash script for monitoring it like


Or you can also check when needed from shell prompt with “ps aux | grep mysqld” command. 

  • Memory tables in MySQL 5.7

With MySQL 5.7, some very interesting memory statistics tables are introduced to check memory usage in performance_schema.  There is no any detailed documentation available yet but you can check some details here.  http://dev.mysql.com/doc/refman/5.7/en/memory-summary-tables.html

In P_S, there are five memory summary tables.

So with every event you can get summarized memory consumption for a particular account, host, thread and user. While checking more, I found that there are around 209 different events to check. I have just tried to check one event related to join buffer size.

Here, COUNT_ALLOC, COUNT_FREE are aggregate the number of calls to malloc-like and free-like functions. SUM_NUMBER_OF_BYTES_ALLOC and SUM_NUMBER_OF_BYTES_FREE are indicate the aggregate size of allocated and freed memory blocks. CURRENT_COUNT_USED is the aggregate number of currently allocated blocks that have not been freed yet. CURRENT_NUMBER_OF_BYTES_USED is the aggregate size of currently allocated memory blocks that have not been freed yet. LOW_ and HIGH_ are low and high water marks corresponding to the columns. 

If you are aware about these scripts then, these are giving very good summary about overall server memory consumption as well as related to MySQL.

Like in output of pt-summary,

In output of pt-mysql-summary.

Conclusion: 

It is really important for us to know where MySQL allocates memory and how it affects the overall load on the MySQL server and performance. I have just tried here to describe a few ways but I still think that we should have some sort of script or something that can combine all of these results and gives us some truthful output of memory usage in MySQL.

About Nilnandan Joshi

Nilnandan officially started with Percona as a Support Engineer. Before joining Percona, he has worked as a MySQL Database administrator with different types of service based companies managing high-traffic websites and web applications. Nilnandan has extensive experience in database design and development, database management, client management, security/documentations/training, implementing DRM solutions, automating backups and high availability.

Nilnandan is based at Pune (India). In his spare time, he likes to listen Indian classical/semi-classical music, watching tv, playing cricket/badminton and hang out with his family.

Comments

  1. shamim says:

    Nice article to capture the memory usage.
    I love using Valgrind Massif for memory stats and also the pt-summary as well.

    how many of the status for memory consumption that the percona server displayed above are configurable (ofcourse excluding the usual buffers and the caches we already can change)?

    While its good to know how much memory is being used behind the scenes, it would really help if we can configure/control it.

  2. Nilnandan Joshi says:

    Hi Shamim,

    Thanks. Can you please elaborate your question? Are you asking for P_S (Performance Schema) related memory usage or something else? If it’s P_S related which introduced from 5.6 then you can check with
    SHOW GLOBAL VARIABLES LIKE “performance_schema%”;

    Check more details in comments here about memory consumption and related variables for P_S.
    http://bugs.mysql.com/bug.php?id=68514
    http://bugs.mysql.com/bug.php?id=69490

    But still proper documentation is needed for the use of P_S in MySQL 5.6.

  3. jo says:

    Please check this one to substantiate this further : http://blog.triantech.com/memory-drain-issues-on-mysql-5-6/

  4. Pavan says:

    Hi Nilandan,

    I m begineer for sql database, please help me or guide me to learn and practice of sql database. i lm working in Datacenter for technical support issues.

    Regards,
    Pavan

Speak Your Mind

*