7 Replies Latest reply on Oct 27, 2010 12:20 PM by LG

    YAMLP(Yet Another Memory Leak Problem)

      Hi guys,


      I'm still having big troubles with 3.7.0beta (3.6.4 had the same kind of problem).

      Usually the server handles 400 clients during work hours. I've seen bigger deployments in others posts, so this is ok.

      We have a severe memory leak where the JVM goes up to the 2096MB Java memory limit.

      It appears the same thing happens with just a few clients (35 at the time I'm testing).

      Clients are essentially Spark/Pidgin.


      The server is a Linux/x64 Debian Squeeze with Sun JRE/JDK6, LDAP auth, Database(MySQL) groups.

      We have 3 dozen of groups being shared between each 350 users (enterprise groups).

      It is connected to another Openfire 3.6.4 server which seems to work correctly (~50 clients, Windows2K3x64/SQLLite).


      I've been able to make a heap dump parsed with Eclipse Memory Analyser (very good soft by the way).

      Here are parts of the results(~40 clients)(tested after a fresh start to have the thinest memory footprint) :

      One instance of "java.util.TaskQueue" loaded by "<system class  loader>" occupies 776 748 384 (90,98%) bytes. The instance is  referenced by org.jivesoftware.openfire.pep.PEPService @ 0x7f3b4699f6b0 ,  loaded by "org.jivesoftware.openfire.starter.JiveClassLoader @  0x7f3b11cf5ab8". The memory is accumulated in one instance of  "java.util.TimerTask[]" loaded by "<system class  loader>".



      org.jivesoftware.openfire.starter.JiveClassLoader  @ 0x7f3b11cf5ab8


      Label Number Of Objects Used Heap Size Retained Heap Size
      org.jivesoftware.openfire.pep.PEPService$11 157101 816776 731 944
      The warn log reports a lot of (~1 per minute):
      2010.10.21 23:39:54 Cache Roster was full, shrinked to 90% in 0ms
      Is it a bug?
      Is there a workaround?(property value?)
      Is there a patch?
        • Re: YAMLP(Yet Another Memory Leak Problem)



          can you upload the heapdump somewhere?

          Openfire Properties / How to configure Openfire's caches should help to increase the cache settings.


          With log4j you should be able to hide these warnings but I think that you better increase the cache. Getting a full cache every minute is likely not good for the performance. So I prefer logging this as a full cache can cause serious performance issues.



            • Re: YAMLP(Yet Another Memory Leak Problem)



              I could upload the dump, but it contains sensitive data. It will be the last resort solution.


              Looking further into the dump, it seems that 1 user using Empathy have most of the TimeTask slots. Each slot occupies 650~800KB whereas others users have just one slot.

              I’ve looked to the code, I did not understand how this was possible, it seems each user is limited to one instance of PEPService. In practice, this is not the case (the user is the same for each PEPService instance):



              As a workaround, I have disabled PEP (xmpp.pep.enabled=false) like recommended by community, and everything is just fine.


              I’ve seen posts about this problem, and bug has been closed for 3.6.4. I think it needs to be reopened and be definitely fixed.