Re: [rtcweb] Meaning of SHOULD support/use interleaving

["Mo Zanaty (mzanaty)" <[e-mail]>]

Moved to tsvwg:
http://www.ietf.org/mail-archive/web/tsvwg/current/msg12892.html

On 10/28/14, 6:28 PM, Mo Zanaty (mzanaty) <[e-mail]<mailto:[e-mail]>> wrote:
On Oct 28, 2014, at 5:11 PM, "Michael Tuexen" <[e-mail]<mailto:[e-mail]>> wrote:
On 28 Oct 2014, at 21:52, Mo Zanaty (mzanaty) <[e-mail]<mailto:[e-mail]>> wrote:
What resource is being conserved by not just requiring it?
8 bytes per packet (for MID+FSN).
Michael, the current NDATA draft requires it to be used for all messages if negotiated. Can’t this be relaxed / optimized to use it for all *fragmented* messages but use DATA for non-fragmented messages to avoid the 8 byte overhead for small messages (like game telemetry/control, an often cited use of data channels).
You can bring this up at the tsvwg mailing list.
Mo: Will do.
However, we normally do not optimize for byte saving... Mixing
both makes the processing harder... Maybe we can use a bit in the flags to indicate if the additional fields are
there...
Mo: Type should already be enough without flags.
The current NDATA draft also says you can’t interleave fragmented messages in a stream, so head of line blocking remains for all fragmented messages, while only small non-fragmented messages can avoid it. This dilutes the utility of NDATA, perhaps enough to make apps that really care about head of line blocking to implement their own app layer fragmentation with messages well below common MTUs, hence defeating NDATA. Can’t this also be relaxed since MID is part of the NDATA extra header?
This would only make sense for messages marked as unordered... For ordered it doesn't make sense.
Mo: The restriction applies to unordered as well as ordered. Even for ordered, it does make sense to avoid head of line blocking at the sender.
If those two issues are resolved, I see no argument against making NDATA a MUST in data channels.
Please let us discuss NDATA issues on the tsvwg mailing list...
Mo: Agreed, will do. But I think rtcweb MUST use NDATA only if its benefits outweigh its costs. The current draft does not pass that bar for me. If my app cares about HOL blocking, NDATA is not very useful, so my app must internally fragment, hence NDATA actually becomes harmful (8 bytes closer to exceeding MTU, forcing SCTP fragments and hence risking HOL blocking).
Best regards
Michael
Mo
PS. Some believe data channels will be more widely used than WebRTC media (webtorrent, etc.), so it does make sense to consider the desirable properties of a general peer-to-peer transport, and drop the WebRTC prefix when talking about data channels.