[asterisk-bugs] [JIRA] (ASTERISK-29232) Memory Leak since 16.13.0

[George Joseph (JIRA)]

    [ https://issues.asterisk.org/jira/browse/ASTERISK-29232?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=256775#comment-256775 ] 

George Joseph commented on ASTERISK-29232:
------------------------------------------

The fact that pool_policy_malloc is the biggest allocator isn't surprising.  Every new transaction and dialog gets a pool allocated to it and the pool _should_ be released when the dialog or transaction is terminated.  Normally, libc should release the pool back to the OS but if the pages are very fragmented, it may not be able to efficiently.   When things start going south, try running {{malloc trim}} from the CLI and see what that does to the memory used by the process.  It tells libc to be more aggressive in its attempt to return memory to the OS.

Can we do a little recap?

The latest test was on 16.20 yes?
What was the usage profile over that period?  Number of calls, endpoints, inbound/outbound registrations, subscriptions, etc.
For the OOM killer...are you talking about the kernel's OOM killer or something you've implemented?
How are you setting the limit and on what parameter?  virtual or resident memory?
How much memory is actually installed on the machines?

I don't _think_ adding the RR header is an issue but I can run a few tests on Monday.




> Memory Leak since 16.13.0
> -------------------------
>
>                 Key: ASTERISK-29232
>                 URL: https://issues.asterisk.org/jira/browse/ASTERISK-29232
>             Project: Asterisk
>          Issue Type: Bug
>      Security Level: None
>          Components: Core/PBX
>    Affects Versions: 16.15.0, 16.20.0
>         Environment: CentOS 7 x64
>            Reporter: Luke Escude
>            Assignee: Luke Escude
>            Severity: Major
>              Labels: fax
>         Attachments: Analysis.xlsx, Apex-Analysis.xlsx, cw1-memchart.png, Jan6-1401.csv, nw1-memchart.png, PW3-Memchart.png
>
>
> So we have around 100 instances of Asterisk 16.13.0 that have been running for over 2 months, normal load (small businesses with less than 30 users each), without issue.
> We have another 350 instances of Asterisk 16.15.0 that we've started seeing a very linear increase in memory consumption over time. Specifically, we see higher-load instances (150+ users) last only a few days before hitting our artificial 3GB ceiling and getting restarted by the OOM killer.
> There are very few differences in our implementation of the 16.13 and 16.15 versions. All versions are set up as the following:
> - CentOS 7 64-bit
> - Voicemail over ODBC
> - unixODBC 2.3.1
> - MariaDB Connector (instead of the crappy mysql connector)
> - CDR over MySQL
> - SIP Trunks are registered every 2 minutes, qualified every 15 seconds.
> - User devices register every 10 minutes, qualified every 15 seconds.
> - User devices connect via TCP more often than UDP.
> - I have NO pjsip threadpool configuration options defined. I think the default is 50 threads?
> Here is what I am about to test within the next week:
> 1. unixODBC updated to 2.3.9
> 2. Longer SIP Trunk Registration period - Maybe PJSIP is working too hard?
> 3. Longer qualify timeout - Maybe PJSIP is working too hard?
> One of my first questions: Is it SAFE to compile asterisk with MALLOC_DEBUG and just leave it on permanently? I am scared to enable it, and suddenly have a bunch of users that are experiencing issues because I've enabled something that should only be enabled in Dev.
> Sorry for the length of the post, trying to cover as much ground as possible.



--
This message was sent by Atlassian JIRA
(v6.2#6252)