[vorbis] Vorbis over RTP

[Roland Parviainen]

Monty wrote:
> 
> > IP fragment means that the whole UDP packet is dropped. If I apply
> > forward error correction to the fragments I can easily get the 10%
> > packet loss down to < 1%. (I just tested with 10% packet loss, I
> > actually takes about 5s until the music starts playing, so I made a good
> > guess :)
> 
> ??  You get the packet or you don't.  For the most part, fragments are
> not going to be lost to burst errors (although that will happen),
> you'll be losing fragments due to them being dropped in congestion.
> We're designing only for Net, and the nature of packet loss is
> overwhelmingly 'all or nothing'.  This is why Vorbis packets don't
> waste bandwidth on error correction.
> 

Congestion in usually bursty, but it doesn't really matter. What I meant
was that it is better to fragment yourself instead of letting the lower
layers doing it if you can use the partial fragments. If we send a 2000
byte UDP datagram and one IP/network layer fragment is lost all 2000
bytes is lost. If we fragment it ourself at 1000 bytes and one is lost,
we still get 1000 bytes that we can use. This is why most (probably all)
RTP payloads are smaller than the typical MTU of ~1500 bytes.

> > This is just my first implementation, more experimentation is needed...
> > For instance it is also possible to send the codebook packets on a
> > separate multicast group, so clients don't have to get all the extra
> > data when they know the codebook.
> 
> 'know the codebook'?  There's no way to tell if you know a codebook
> without downloading it... at least not within the strict Vorbis spec.
> That doesn't mean it would be a bad idea to implement a 'behind the
> scenes' codebook id/caching mechanism.
> 
> Monty
> 

What I meant was that as soon as the receiver get the codebook and can
start decoding we can leave this group since they no longer have any use
for the data. Similar to getting the codebook out of band, but we do not
need any TCP connection at all so it is much more scalable (think >
100,000 users).

//Roland

--- >8 ----
List archives:  http://www.xiph.org/archives/
Ogg project homepage: http://www.xiph.org/ogg/
To unsubscribe from this list, send a message to 'vorbis-request at xiph.org'
containing only the word 'unsubscribe' in the body.  No subject is needed.
Unsubscribe messages sent to the list will be ignored/filtered.