[asterisk-dev] definition of RTP jitter - potential bug in Asterisk

[Klaus Darilion]

Stanisław Pitucha schrieb:
> 2009/9/3 John Todd <jtodd at digium.com>:
>>   Can you open a bug on this?  I don't want this to get lost.
> 
> I posted a bug link in the 2'nd mail in this thread.
> 
>>   But let me ask a few questions: What does "jitter" even measure?  I
>> had interpreted jitter to mean the difference (via some smoothing
>> mechanism) in milliseconds between the delays of subsequent packets in
>> a uni-directional stream.  Why would frequency, or any codec-layer
>> computation be involved in this determination?
> 
> Well - as you write in your post, jitter depends on 2 values (when
> calculating the D function as in RFC 1889 6.3.1). "Si is the RTP
> timestamp from packet i, and Ri is the time of arrival in RTP
> timestamp units". The RTP timestamp is incremented by the number of
> bytes of RTP data sent in every packet. In that case it will usually

Almost. Not number of bytes, but number of samples. For G711 this is 
identical, but not for other codecs.

klaus

> be incremented by 160 for a typical alaw call. Since we're using RTP
> timestamps the whole time, it's going to be 8000 RTP units / second
> (for that alaw call).
> If you stick to calculating the value in units specified in the RFC,
> you get the jitter in those units and you don't have to care about the
> "real time" at all. S is the "perfect" time for arrival, R is the
> arrival timestamp. The rate only matters for presenting the result in
> ms. But the jitter of 1 unit (byte) for a codec at 8kHz is different
> in real time than the one for a codec at 16kHz.
> 
> Asterisk is calculating D not based on the timestamps, but time, so
> the result is in time too. In res_rtp_.../calc_rxstamp():
> 
>    tv->tv_sec = rtp->rxcore.tv_sec + timestamp / 8000;
>    tv->tv_usec = rtp->rxcore.tv_usec + (timestamp % 8000) * 125;
> (time = time + timestamp/rate which means time + time)
> 
> Which is ok for a/ulaw and less ok for other codecs.
> 
>    current_time = (double)now.tv_sec + (double)now.tv_usec/1000000;
>    transit = current_time - dtv;
> 
> You're using the real time and not the timestamps for jitter, so you
> get jitter in real world time assuming rate=8000.
> 
> You can check the value with Wireshark - there was a very nice post
> describing how the jitter is calculated on their ML (can't find the
> link now unfortunately). If you multiply by the current rate instead,
> the results match.
> 
> So I see 2 solutions. You can either leave the calculation as is and
> multiply by rate to get the result in units (the one that should be in
> the rtcp packet), or change the code to get:
> 
> tv = time*rate + timestamp
> ...
> transit = current_time*rate - dtv
> 
> and then rxjitter will be in rtp units.
>