In the past…
One of the typical problems you have when restarting mysqld is that the InnoDB buffer pool (buffer pool from now on) is empty and consequently access to the database requires reading directly from disk. Performance suffers dramatically as a consequence.
So the common solution is to artificially warm upthe server by doing queries which will fill the buffer pool. Typical solutions might be to do: SELECT COUNT(*) FROM some_table FORCE INDEX (PRIMARY) LIMIT ... on a number of tables to fill up the pool on startup. Fitting this into the standard mysql init start script is somewhat tricky as no hooks are provided for this sort of post-start action. (It would be nice to have this for other tasks too.)
Of course choosing the right parameters here can be tricky as workload changes over time, and as the ratio of the size of the database to the size of the buffer pool increases, you need to be more selective in what you load in. This is a tricky problem to solve well and solve generically, I am sure you’ll agree.
Then MySQL 5.6 comes to the rescue (almost)
So I was pleased to hear that MySQL 5.6 had a new feature to take the problem away. When shutting down mysqld, it can be configured to save the list of pages in the buffer pool, and on startup you can use this list to populate the empty buffer pool, hence putting you in the same warm state
you were in before. See: http://dev.mysql.com/doc/refman/5.6/en/innodb-performance.html
Good news, but…
As the systems I manage have got busier and technology has advanced, so has the size of the memory in my servers. Some servers I manage now have 192 GB of ram and we have tried set ups with even more memory. The innodb buffer pool size is likely to be around 90% of that size. So loading in 160 GB of data from hard disk, even if read 100% sequentially, is going to take some time, so doing this as efficiently as possible is essential. Yet, perhaps loading in the full pool is not really necessary. Experimentation, depending on the different set ups I use, indicates that mysqld performs sufficiently well even if if buffer pool is only 25% full. After that natural usage will pull in missing data and gradually performance will improve to expected levels.
The first systems I started using MySQL 5.6 on needed fast disks so were configured with SSDs. Here the absolute need for a warm up script was less important as SSDs are so much faster anyway, so it is only now that I have paid more attention to this new feature and see some potential weaknesses in its implementation. Having said that it’s a good first step, and for many people may be just what they need if they enable it.
Below are some suggestions for improvement, but they will probably have to wait for 5.7 as new features should not really go into the current GA release.
- Optionally time-limit this warm up process, stopping when the time limit has exceeded
- Optionally limit the number of pages loaded in, probably to a percentage of the current pool size. The pages loaded in however, should be in MRU first as these are the hotest and most needed pages. Ensure the load sorts the page list to ensure that I/O is as sequential and fast as possible.
- Optionally define the number of concurrent threads that can be doing warmup I/O. My experience is that depending on the underlying disk subsystem running several parallel warmup threads may reduce the warmup time considerably. I think the current implementation in 5.6 only uses a single thread to load in pages.
- Stop warming up once the buffer pool is full. Other normal database activity on the server may trigger data being loaded into the buffer pool so it may fill up before the warmup process finishes. So if buffer pool page eviction starts there’s no point continuing the warmup process.
- Take into account the fact that copying an on disk mysql image may result in the server being started not being the one that the buffer pool list was saved for and the configuration in terms of memory size and buffer pools may change. If this makes a difference ensure that the warm up process still does the right thing of warming up the server with the right pages.
I have posted a feature request as Bug#68863
and hope some of these ideas will be considered.
So I still have to do more testing but think that I may well still be doing manual warm ups for while yet.