Excessive latency

[Tim Müller]

On Wed, 2016-02-24 at 16:06 +0000, David Jaggard wrote:

Hi David,

> The input of my stream is MPEG2 TS over RTP. The output is H264 video
> only.
> 
> Pipeline:
> udpsrc->rtpbin->queue2->rtpMP2TDepay->TSDemux->(video)->queue2-
> >DecodeBin->queue2->Deinterlace->queue2->x264Enc->queue2->h264pay-
> >rtpbin->queue2->multiudpsink
> 
> (I added all those queues to try and trace the bottleneck).
> 
> The latency for the entire pipeline is set to a huge 1060ms.
> 
> Every single queue other than the last is empty. The very last queue
> before the sink reports about 1100ms worth of data. This means that
> the entire pipeline is (virtually) latency free except when it
> arrives at the sink and has to buffer in order to synchronize to the
> clock. Why is it choosing such a high latency?

Most of that (700ms) will be from tsdemux, which currently advertises
the worst-case scenario latency.

Then there's an rtpjitterbuffer in rtpbin, which by default has a
latency of 200ms (you can configure this via the rtpbin "latency"
property).

The rest is probably from deinterlace (one frame) and x264enc (try
x264enc tune=zerolatency if you're not using that yet).

You should be able to force a lower latency on the pipeline than it
would configure by default.

You could also modify tsdemux to advertise a lower latency (most
streams have much lower latency requirements, but we currently don't
analyse the stream to estimate that yet, but just announce the worst
case value).

Cheers
 -Tim

-- 
Tim Müller, Centricular Ltd - http://www.centricular.com