[dev.icinga.com #1969] Version 1.5.2 bad erformance in large environment #532
Comments
Updated by jmosshammer on 2011-10-01 10:44:20 +00:00
First: I'm not a mysql performance specialiast. But afaik, transaction management in innodb is done much more efficent in time (but not in space) and in situations where you have many concurrent inserts and selects it should therefore perform better than MyISAM (and you have lots of inserts in your enviroment). It's hard to track down whether it's 1.5 or MyISAM, but i could imagine that MyISAM hat a big factor on this, as it has limited/no support of row locking. Can you use compression in innodb? Though 1.5. got a new database layer, the queries are almost the same at this time, so it shouldn't have such a big impact. |
Updated by tuxifier on 2011-10-28 07:43:06 +00:00 So we switched back to innodb thoroughly without any performance improvements. The performance of "normal views (standard cronks)" with a user without any principals is acceptable but if we use a user with principals (allowed to see only specific hostgroups) and add a filter like "services not ok" icinga-web doesn't respond for minutes. Our next step is to try innodb compression to see if that'll make an improvement. How are the experiences with postgres on high workloads on icinga? Might it perform better? Would it make sense to give it a try? I also found out that before switching to icinga and icinga-web 1.5.x the principals in icinga-web weren't ever tried. So I doubt for now the reason has anything to do with the differences between 1.4 and 1.5 |
Updated by tuxifier on 2011-11-10 13:52:19 +00:00 Got some pretty new results from research:
(which is the icinga performance statistics information on the top right)
If I log into icinga-web with a user without principals and sees everything the query looks like this:
If I login to icinga-web with a user that has attached principals (2 hostgroups in this case it looks like this
The Query time information is from the mysql slow queries log. The point is: If I add principals to a user the examined rows to calculate the internal performance stats of icinga (MIN,MAX,AVG) increase from 100.000 to 42.000.000 which results in an increasing query time from 10s to 195s. The div "icinga-monitor-performance" seems to be deployed by icinga-web/app/modules/Cronks/templates/System/ MonitorPerformanceSuccess.php and by default is automagically refreshed every 60s if I read the code correctly. Which might result in an endless query loop if the query doesn't finish within 60s? So is this a "works as designed" behaviour or may there be an issue with the SQL Statement itself? Next Question is, how could I avoid that this query is executed on login? Can I completely disable the data query for the icinga-monitor-performance div to do further testing? The database itself has absolutely no performance probs - no I/O, no mem but the query eats 100% of one cpu - core as long as it runs (which seems to be ok for me) thanks |
Updated by tgelf on 2011-11-10 16:25:02 +00:00 I can confirm this. And I'm pretty sure that these queries could also be done without all those JOINs (or at least not with so many of them). The main issue in the shown queries is this part: INNER JOIN icinga_objects i5 ON (( MySQL seems to be unable to make use of it's indexes as of this construct. This leads to harmful copying-to-tmp-table operations. You can find a similar construct in the notification view, causing unacceptably bad performance for that grid. Indexes added in the current master make this better, but the notification view still remains unusable. I helped myself with throwing away the IF-construct there, notifications are still slow and show NULL for host notifications - but that way they are no longer killing my DB servers :p What also makes 1.5.2 seem to performance bad is the fact that it does never ever execute two requests for a single session in parallel. The root cause of this is the complete lack of session file management: it's left up to PHP to close them, and PHP doesn't do so before the end of a single request. Switching to DB-based sessions will improve this - at the cost of slightly more DB traffic. Garbage collection will kick your session from time to time as of another bug - you'll find a possible fix in the related issue. There are also some grids whose rows are not filtered once you are using a custom filter in your grid view. This can lead to hundreds and thousands of grid lines your browser has to paginate by itself. FF will inform you that a script is running crazy, IE will probably completely lock up. A few related issues: #2035, #2051, #2054, #2055, #2056, #2058, #2061 Installing APC is also a good idea to make Icinga-Web faster. Provide MySQL with a large InnoDB buffer pool, innodb_file_per_table will help avoiding ever-growing ibdata files. Using that option you can track down problematic tables, search the root cause, try to fix it and recreate just that single table. Choosing MyISAM for Icinga is a bad idea. Well, fixed all those issues Icinga-Web 1.5.2 runs with acceptable performance also in large environments. Good luck ;-) Best regards, NB: I didn't fix the statistics query yet. It is causing harm here too, but you'll not notice it anymore once you worked around the session locking issue. However I did some testing: it is possible to rewrite those queries in a way to make them run in far less than a single second. But right now i have no idea how to do so in such a hyper-abstracted way Icinga-Web is asking for. |
Updated by jmosshammer on 2012-03-01 17:04:46 +00:00
|
Updated by mfriedrich on 2012-09-24 12:17:23 +00:00 is this still the case? |
Updated by jmosshammer on 2012-10-09 11:19:35 +00:00
The db-backend has changed to DQL Views, so I'll close this issue |
This issue has been migrated from Redmine: https://dev.icinga.com/issues/1969
Created by tuxifier on 2011-09-30 09:27:49 +00:00
Assignee: jmosshammer
Status: Closed (closed on 2012-10-09 11:19:35 +00:00)
Target Version: (none)
Last Update: 2012-10-09 11:19:35 +00:00 (in Redmine)
Hello Icingas,
the numbers
hostgroup_name Hosts Services
hostgroup1 1202 1300
hostgroup2 89 1021
hostgroup3 171 1434
hostgroup4 72 913
the members of group 2 and 4 mostly are also members of group 3
We configured 3 user groups in icinga-web
usergroup1 has principals configured to have access to hostgroup3 hosts
usergroup2 has principals configured to have access to hostgroups hostgroup2 and hostgroup4
usergroup3 has principals configured to have access to hostgroup1
The issue now is:
If one of the users that are member in one of the usergroups the webinterface gets extremely slow (response time of the icinga-web interface increases to more than a minute) and the load
on the MySQL DB Server gets to 100% CPU
What we changed besides the upgrade of icinga-core from 1.4.1 to 1.5.1 and icinga-web from 1.4.1 to 1.5.2 is that we re-created the icinga core database with MyISAM engine.
The reason was a very large ib_data1 file (15 gigabytes)
So the questions are:
Why is it so slow when using principals?
Which database engine is better for icinga-core DB InnoDB or MyISAM - and if it is innodb how to avoid the huge ibdata file?
Why are the filters (hosts not up) now much slower than in 1.4.1? (this has nothing to do with the groups issue)
Thanks in advace for your help
tuxifier
The text was updated successfully, but these errors were encountered: