RE: [AVT] T.38 over RTP: RTP Sequence Number

["Vladimir Ulybin" <Vladimir at audiocodes.com>]

Colin,

I think that the use of timestamps together with sequence numbers for
building T.38 packet recovery buffer is not a good way. For T.38
INDICATOR packets it is acceptable but for DATA packets this way may
cause a lot of problems for packet recovery during data transfer.

It is much easy and reliable to use RTP sequence number as the only base
for ordering T.38 packets arrived over RTP. In this case, timestamps of
redundant payloads may be ignored without any problems.

Let consider an option to update the "draft-jones-avt-audio-t38-05.txt"
or write a new draft for T.38 over RTP.

I think the problems opened in our discussions 
- repetition of T.38 packets and 
- excessive complexity of T.38 over RTP caused by timestamps
(non-required by T.38)
are general for different vendors.

Regards,
Vladimir Ulybin


-----Original Message-----
From: Colin Perkins [mailto:csp at csperkins.org] 
Sent: Tuesday, July 19, 2005 2:07 PM
To: Vladimir Ulybin
Cc: Paul E. Jones; avt at ietf.org
Subject: Re: [AVT] T.38 over RTP: RTP Sequence Number

On 19 Jul 2005, at 07:40, Vladimir Ulybin wrote:
> 1. Zero is a simplest solution. The RFC 3550 allow the same time stamp
> for several consecutive packets. So, when these packets will be
> transmitted as redundant one relative to the other, the time  
> offsets per
> RFC 2198 will be zero. It means that receiving gateway should handle
> correctly zero time offsets in RFC 2198 headers.

No, since it will treat the data as being a redundant copy of the  
packet in which it is sent. If you use an offset of zero, there's no  
way to correctly place the data at the right time instant in the  
playout buffer.

> 2. The computation of redundant packet time offset relative to primary
> packet can be applied for synchronous operation like audio or
> t4-non-ECM-data having a fixed time interval between packets. But, for
> asynchronous packet stream, which is typical for T.30 control and ECM
> fax, such approach cannot be used.

Correct - this is why RFC 2198 has an explicit timestamp for the  
redundancy offset.

> Why I opened this issue?
>
> Assume a T.38 module of DSP/gateway which delivers the T.38 UDPTL
> packets ready for transmission to IP. To enable new feature of T.38  
> over
> RTP, it is easy to convert T.38 UDPTL into RFC 2198 at output of
> DSP/gateway. The obstacle is in time stamp offsets of redundant  
> packets.
> We all (fax engineers) know that these offsets are not required at  
> T.38
> packet receivers. But accurate transmission of those per RFC 2198
> requires deep re-writing of working and verified module(s) plus
> additional allocation of RedundancyLevel*NofChannels timestamps.

Correct - this is required for robust operation over an unreliable  
packet network that may disrupt media timing.

There are two cases:

1) There is a constant mapping from sequence number to timestamp. In  
this case, you can trivially calculate the timestamp and need no  
extra storage. The media timing can be reconstructed using either the  
sequence number or timestamp, and transporting both is essentially a  
waste of bandwidth for this application. In this case, you treat the  
extra bits on the wire as an "RTP tax" needed for interoperability,  
but don't need to store the extra data in the gateway.

2) There is not a constant mapping from sequence number to timestamp.  
This case requires you to store the extra timestamps for inclusion in  
RFC 2198 redundancy headers, however that extra information is  
required by the receiver to manage a playout buffer, since a playout  
buffer managed by sequence number only is not sufficient to  
reconstruct media timing.

Colin

_______________________________________________
Audio/Video Transport Working Group
avt at ietf.org
https://www1.ietf.org/mailman/listinfo/avt