Comments on: Adaptive checkpointing

By: Mike

Mike — Fri, 16 Jan 2009 21:24:02 +0000

What are the units of the values reported for “max checkpoint age”, “modified age” and “checkpoint age” ?

By: http://mysqlha.blogspot.com/

http://mysqlha.blogspot.com/ — Sun, 23 Nov 2008 20:19:59 +0000

Yes, I would like to know what makes it adaptive? The original code flushes at most a fixed number of blocks per time interval as requested by a background IO thread. Have you added something to make the number of blocks to flush dynamic?

By: Yasufumi

Yasufumi — Sat, 22 Nov 2008 06:21:58 +0000

Davy,
innodb_adaptive_checkpoint is online parameter.

Mark,
Sorry, I don’t know what I should reference, though it may exist elsewhere…
Could you tell me what do you want to know?

By: Davy Campano

Davy Campano — Thu, 20 Nov 2008 16:54:45 +0000

Yasufumi,

Is innodb_adaptive_checkpoint a setting that can be changed online, or does it require restart? Thanks

Davy

By: http://mysqlha.blogspot.com/

http://mysqlha.blogspot.com/ — Thu, 20 Nov 2008 16:52:19 +0000

This is great. Can you provide more details on what ‘adaptive checkpointing’ does or reference a description elsewhere?

By: Yasufumi

Yasufumi — Fri, 14 Nov 2008 16:04:50 +0000

John,

We can check the checkpoint age even if we use normal MySQL.
From the output of SHOW INNODB STATUS,
===========================
—
LOG
—
Log sequence number 0 1059494372 <—
Log flushed up to 0 1059494372
Last checkpoint at 0 1055251010 <—
===========================
we can calculate the checkpoint age

(0 – 0) * 2^32 + (1059494372 – 1055251010)
= 4243362

And max chekcpoint age may be 70% or 80% of

innodb_log_file_size * innodb_log_files_in_group

Thank you.

By: Yasufumi

Yasufumi — Fri, 14 Nov 2008 16:01:51 +0000

peter,

Yes, I know we never get such a regular results from real applications. But cautious people might do their benchmarks before running their system in practice. At that time, they may face such a strange behavior of MySQL/InnoDB which they cannot explain, and may abandon to use MySQL from the worry about the unexplained. (It is from my experience, I have met such situations before and some people also have consulted to me about this behavior in their benchmarks.)

I think this problem is difficult to explain to everyone clearly, because of its complicated reason peculiar to RDBMS (and my ability of description…). But I also think it is very important to proof “The problem can be explained!” for MySQL guys in the world (, even if they cannot understand clearly sorry for my poor description…).

By: John Masters

John Masters — Fri, 14 Nov 2008 11:52:29 +0000

How do you collect the information for the graphs above from mysql? Also, is there a tool to plot the data to visualize the performance bottlenecks like you show above? Appreciate your help in this regards as we are running into strange issues with Mysql where in the same query with (sql_no_cache) gets the results in 50seconds one time and it takes more than 600secs the other. I see no other processes running using “show processlist”. So I am assuming something to do with mysql system level bottlenecks and would like some tools to help diagnose

Thanks

By: peter

peter — Thu, 13 Nov 2008 15:22:02 +0000

Great result Yasufumi !

I guess even if average TPM is not that much different, affect 95 percentile response time is significant… and in practice the periodic checkpointing stalls are surely not acceptable for real applications.

There is still probably a bit potential for improvement for system to self adapt and try to even out writes so we do not get into 20% TPM dips/write spikes.