[asterisk-bugs] [JIRA] (ASTERISK-21872) RFC2833 DTMF Events cause RTP packet delays and high CPU

[hristo (JIRA)]

    [ https://issues.asterisk.org/jira/browse/ASTERISK-21872?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=207443#comment-207443 ] 

hristo commented on ASTERISK-21872:
-----------------------------------

I had a fresh Ubuntu server installed inside a XEN VM (1 virtual CPU Xeon E5645 at 2.40). It is somewhat harder to reproduce with Ubuntu 12.04 and is not as pronounced as with Debian, but it is still noticeable. I had to setup 5 calls, which were sending DTMFs, before I could notice (hear) it.

What I have discovered is, that it only disrupts the audio, when asterisk is started with the "-p" option ("Run as pseudo-realtime thread"). If asterisk is not running in pseudo-realtime mode, then the CPU again goes to 100%, but it seems that the RTP packets are still sent out on time in this case.

Until now all tests, including the initial ones above, were made with "-p", but I haven't really noticed it, because it was being set in the init.d script. However, for the Ubuntu tests I used the sample init script ("make config") and only added the "-p" option to verify my findings, so I don't think that it is related to the init.d script itself.

The interesting part is that the CPU spends about 75-80% in the kernel (system). See attached "vmstat.txt". If we assume, that the newer kernels are better optimized, it will explain to some extent, why it is harder to reproduce with Ubuntu (3.2.xx kernel) than with Debian (2.6.xx kernel). Additionally, it also makes sense to assume, that when Asterisk runs in pseudo-realtime mode and needs 75-80% of kernel CPU time, this may prevent the kernel from completing other tasks on time - for example sending the RTP packets out, and therefore to result in the RTP delays that I see.

I am also attaching the diff against the sample configs (with IP and password edited). Please ignore the SIPAddHeader lines, they are only present to allow the terminating endpoint to accept the generated calls.

Can you at least confirm, that the CPU load gets close to 100% when the DTMFs start coming? It takes some seconds for the load to build up.
                
> RFC2833 DTMF Events cause RTP packet delays and high CPU
> --------------------------------------------------------
>
>                 Key: ASTERISK-21872
>                 URL: https://issues.asterisk.org/jira/browse/ASTERISK-21872
>             Project: Asterisk
>          Issue Type: Bug
>      Security Level: None
>          Components: Core/General
>    Affects Versions: 1.8.17.0, 1.8.19.1, 1.8.20.0, 1.8.22.0
>         Environment: Debian 6.0 64-bit
>            Reporter: hristo
>            Assignee: hristo
>            Severity: Minor
>         Attachments: 2-calls-one-sending-many-dtmfs-asterisk-debug.txt, forward-stream-first-call-after-asterisk.pcap.txt, forward-stream-first-call-before-asterisk.pcap.txt, sample-config.diff, vmstat.txt
>
>
> If I send several DTMFs to Asterisk, one after the other, fast enough, it blocks other voice RTP packets for as long as several hundred milliseconds. This seems to affects *all* RTP streams on a server.
> I can say for sure, that Asterisk is not dropping the RTP packets, because after a while it sends all of them at once. It seems as if they are being held by something, while the DTMFs are being processed/forwarded.
> This only occurs in non Packet2Packet mode.
> Originally I've seen the problem when several people were connecting to a conference at about the same time and were entering the PIN numbers at about the same time, therefore producing a lot of DTMFs. The conference runs on a dedicate hardware und is unrelated. Asterisk just sits in the middle and bridges the calls. I have managed to reproduce this with only two calls with as little as 10-15 DTMFs, provided they are send fast enough.
> Attaching is a debug console log from the following call scenario. In this case both calls were genereted from a dedicated server and terminated on another dedicated server.
> Call 1:
>  A (IP 1.1.1.1) dials 1000 --> Asterisk (IP 2.2.2.2) ---> B (IP 3.3.3.3)
> Call 2:
>  A' (IP 1.1.1.1) dials 2000 --> Asterisk (IP 2.2.2.2) ---> B' (IP 3.3.3.3)
> Both calls are active at this point. A' on Call 2 starts sending DTMFs (in this case 40 of them). As a result RTP packets from Call *1* in both directions are delayed by 150-160 ms and are being sent in bursts.
> In the logs I often see:
> res_timing_timerfd.c:225 timerfd_timer_ack: Expected to acknowledge 1 ticks but got 5 instead
> and the CPU is close to 100% (caused by the asterisk process). As soon as all DTMFs are sent, the RTP streams return back to normal with asterisk sending one packet every 20 ms on average.
> Attached is also a filtered packet capture that shows only the forward RTP stream on Call 1 from A -> Asterisk and from Asterisk -> B. "Time" represents the delta from the previos packet. Under normal conditions this should be close to 0.020 s (or 20 ms).
> One example of the problem can be seen at line 1235 in 'forward-stream-first-call-after-asterisk.pcap.txt'. The packet there has been held for ~160 ms, then sent together with the next 7 packets all at once.
> The RTP packets from the corresponding call leg (before asterisk) start at line 1244 in 'forward-stream-first-call-before-asterisk.pcap.txt" and are all equally spaced at about 20 ms.
> There are many such examples - simply search for 0.000 (deltas which are less than 1 ms) to identify groups of packets that are sent together. The same problem is present in the backward stream too (not attached).
> How to reproduce - add the following to the dialplan:
> exten => _X.,n,Dial(SIP/B at 3.3.3.3,,t)
> The 't' option is important, because it effectively disables the Packet2Packet mode. Connect 2 calls (2 sets of telephones) and start dialing DTMFs as fast as you can on one of them (or simply generate 2 calls and send the DTMFs as I did). This will disrupt the call between the other set of phones if done fast enough.
> I have tested this on 3 servers (2 physical and one virtual). All of them were running the same OS (Debian 6), so this may end up being an OS or res_timing_timerfd problem after all, but I really cannot test it on a different distribution.
> I tested with the following versions and was able to reproduce the problem with all of them:
> 1.8.22.0
> 1.8.20.0
> 1.8.19.1
> 1.8.17.0

--
This message is automatically generated by JIRA.
If you think it was sent incorrectly, please contact your JIRA administrators: https://issues.asterisk.org/jira/secure/ContactAdministrators!default.jspa
For more information on JIRA, see: http://www.atlassian.com/software/jira