Re: [rtcweb] Opening unordered data channel issue

[Justin Uberti <[e-mail]>]

I agree that b) is the cleanest choice. While this makes the handshake
2-way, it does not suffer from the performance issues that affected the
previous 2- and 3-way handshake proposals, since data can be sent
immediately without waiting for the ACK.


On Thu, Aug 22, 2013 at 8:52 AM, Michael Tuexen <
[e-mail]> wrote:

> Dear all,
>
> at the last IETF in Berlin we discussed how to handle the opening of
> unordered data channels (no matter wether reliable or not). The problem
> to be handled is that it is possible that user messages arrive at the
> receiver before the DATA_CHANNEL_OPEN message is received. The result
> of the discussion at the meeting was that the best solution is that
> the sender sends user messages also ordered (on unordered channels)
> until there is an indication from the peer that the DATA_CHANNEL_OPEN
> has been processed. This makes sure that the DATA_CHANNEL_OPEN message
> is not received before user messages and therefore the above problem
> is solved.
>
> The open issue is, what the indication from the peer is. There were
> two options discussed at the meeting and it was decided that Randall
> Stewart and myself bring this to the list with the pros and cons
> of the alternatives. So here are the two options:
>
> (a) Use any user message sent from the peer as an indication.
> (b) Use a new ACK message.
>
> Arguments:
> * Using (b) introduces a 2-way handshake for unordered data channels.
>   However, it also works for data channels being used unidirectional.
> * Using (a) stays with the 1-way handshake, but in case of
>   data channels being used unidirectional you will not see the benefits
>   of unordered delivery.
>
> We had some discussions offline and think that using (b) is the
> cleaner solution. Please note that this 2-way handshake is not
> similar to the one we had some time ago. The ACK does not contain
> any additional information and it does not introduce any delay.
> You can send user data right after opening the data channel.
>
> For consistency we would even propose to use this 2 way handshake
> for ordered channels. It doesn't have any drawback and would mean
> that we use the same handshake for all data channels.
>
> So the proposed procedure for opening an arbitrary data channel is:
> 1) The opener sends a DATA_CHANNEL_OPEN message ordered and reliable.
> 2) The opener can start sending user data immediately.
>    It will always be sent ordered, no matter whether the data channel
>    is ordered or not.
> 3) The peer will send a DATA_CHANNEL_ACK message ordered and reliable
>    in response to the DATA_CHANNEL_OPEN message.
> 4) After reception of the DATA_CHANNEL_ACK message,
>    the opener will sent user messages unordered,
>    if the data channel is unordered and ordered, if the
>    data channel is ordered.
>
> Any opinions?
>
> Best regards
> Michael
>
> _______________________________________________
> rtcweb mailing list
> [e-mail]
> https://www.ietf.org/mailman/listinfo/rtcweb
>