Tuesday, 12 April, 2011
This is the 4th part of the MySQL 5.6 Performance Notes started from here.
All about Purge now.. ;-))
Fixing the Purge issue in MySQL 5.6.2
Sunny is a "purge guru" within InnoDB team, and he pointed me to the code which is currently commented as "FIXME" in 5.6.2: in the current code the purge is involved periodically after short sleeps, and the sleep delay was chosen too big (as it's yet experimental and more testing is expected before to get its final implementation available). Currently Sunny has at least two solutions for the next 5.6 version, however I wanted to see 5.6.2 fixed near "as it" with minimal changes (if possible) and simply replaced the 60ms sleep delay by zero (never sleep until there is something to purge, as it was in the original InnoDB code). This small change make me much more happy because it made a huge difference! ;-))
NOTE: my 48 cores AMD server crashed finally due HW issues, so all the following testing was executed on the 32 cores Intel server..
Already on Read+Update workload was seen that the fix worked pretty well - on the following graphs I've compared the original 5.6.2 version vs 5.6.2 with a "fixed" purge code (let's call it 5.6.2+fix1), and as you can see:
- History Length is not growing anymore (similar to 5.5 now)
- Purge activity reported from METRICS table growing now with a growing workload!!
Comparing with 5.5.9 on Read+Write
On Read+Write workload the thing become more interesting:
- on 32 concurrent sessions 5.6.2 is reaching 20K TPS level and keeping it stable till the end of the test (one hour), while 5.5.9 is also reaching 20K TPS on the beginning, but then slowly decreasing performance and within one hour reaching already only 16K TPS..
- then another thing which is very important - History Length is growing slower now on 5.6.2+fix2 comparing to 5.5.9: within one hour it reached 10M level while 5.5.9 grown up to 40M!.. - it's a true victory of the 5.6.2 implementation, because its superior performance was not because of a lowering purge activity.. ;-)) (if you disable purge completely within InnoDB you may get a very important performance improvement, but then very quickly "un-purged" data will makes troubles for you, so having a database in production with a constantly growing History List is not acceptable.. - that's why during benchmark testing it's important to monitor purge lag to avoid any kind of "wrong tuning" :-))
And now with innodb_max_purge_lag=400K
As in the previous test the purge lag is still continuously growing, what will be a difference once we'll apply the Max Purge Lag setting?.. The following graphs are representing the same Read+Write test but now with innodb_max_purge_lag=400000:
- on 32 sessions 5.5.9 keeps a stable 16K TPS level until the end of the test, while 5.6.2+fix1 keeps 19.5K TPS within the same conditions, and we're loosing really few in performance (2.5% only) while improving a lot in purge stability!
- NOTE: there was something wrong with 5.5.9 code because its History Length out-passed 400K, but it should not happen, but it did not grow for sure, so I did not retest it..
So far, Purge issue is fixed, but for a real stable work we need to get Adaptive Flushing fixed too :-))
To be continued in part 5...
blog comments powered by DisqusNote: if you don't see any "comment" dialog above, try to access this page with another web browser, or google for known issues on your browser and DISQUS..