[vlc-devel] VLC-DASH Experiments

Jonathan Kua jtkua at swin.edu.au
Tue Jun 2 05:30:15 CEST 2015


I am trying to run DASH experiments with VLC-DASH client and noticed 
some odd behaviours in both version 2.2.1 and the development version.

Brief description of experimental setup:
DASH client: PCBSD, VLC installed, VLC-DASH module preference set to 
853x480 resolution, 30secs play-out buffer size
DASH server: PCBSD, Apache server serving the dataset obtained from 
Bottleneck router: Linux, 10Mbps symmetric bottleneck bandwidth, 
180-packet pfifo queue, no (0ms) intrinsic delay.

Using the MPD for non-segmented 10secs video chunks, VLC client 
downloaded the entire (~10min) BigBuckBunny video (encoded at 600kbps) 
in ~47secs.
Doing the maths, assuming 25% overhead between the application layer and 
the network layer packets, the entire 10min video encoded at 600kbps 
would be downloaded in (0.6/10)*600*1.25 = 45 secs. Very close to what I 
observed from the tcpdump pcap files.

So instead of pacing itself and honour the 30secs play-out buffer size, 
the VLC-DASH client is downloading the video content as fast as possible 
(as the bottleneck bandwidth permits), up to the end of the video.

Also, when the video is rendered on the screen, during the first 
~47secs, the video was 'fast-forwarded', i.e. the VLC client skipped and 
dropped a lot frames. After t=47s, the video played smoothly for another 
~1min 30 secs till the end of the video. So, a supposedly ~10min 
BigBuckBunny video was played in ~2min17secs.

Here are the statistics:
Video Decoded: 1622 blocks
Displayed: 820 frames
Lost: 819 frames

Why doesn't VLC-DASH client pace itself when retrieving video chunks? 
Why is it skipping so many frames?
Am I missing any DASH configuration?

Thank you.


More information about the vlc-devel mailing list