Re: [manet] undirectional or bidirectional (Was Re: I-D Action: draft-ietf-manet-dlep-da-credit-extension-00.txt)

[Henning Rogge <[e-mail]>]

On Thu, Feb 9, 2017 at 11:25 PM, Lou Berger <[e-mail]> wrote:
> Thanks Stan.
>
> WG,
>
>     So one of the basic differences in the drafts is that this document
> only supports flow control in one direction, i.e., for traffic sent from
> a router to a modem, while the other credit-window draft supports it in
> both directions.  We'd like feedback from the WG on which approach to
> follow.
>
> We (authors of the DA draft) made a conscious choice to limit the DA
> credit based flow control to one direction out of the desire to keep the
> extension as simple as possible and to provide what is really expected
> to be used.
>
> Clearly the reason for being able to control traffic that is sent from
> the router to the modem/radio is pretty easily understood.  Also, while
> bidirectional flow control existed in pre-dlep solutions, it wasn't
> really used -- and in one reported bug/case resulted in traffic dropped
> unnecessarily.  That said, it at first seemed appealing to define a
> symmetric approach that supports flow control in either direction, but
> when getting into the details of DLEP message formats ( and different
> message types in general, based on sender type) the result was a
> symmetric solution that used non-symmetric messaging - which in the
> authors opinion was fairly complex. So, given:
>     (a) the added complexity of document/protocol/code needed to support
>         bidirectional credit control and
>     (b) no one has ever used or currently expects to use credit flow
>         control for traffic destined to a router
>
> We concluded that unidirectional definition was the right place to aim
> *at this time*.  This of course doesn't preclude someone from defining a
> companion definition for modem to router traffic flow control down the
> road -- when there is an actual need.
>
> So the question for the WG is:
>
> Should the merged / future credit window definition support (1)
> bidirectional (credit) flow control, or (2) flow control for only
> traffic from a router to a modem?
>
> Said another way: does anyone wish to advocate the position that we need
> credit window flow control for traffic received over the air, and sent
> from modem to router, *now*?

Hi,

in general flow control is about keeping the "bottleneck/queue" mostly
on the router, where we can control it at a central point. That is an
important part of the design goal of DLEP.

I think flow control from the router towards the radio is THE most
important part, it is practically necessary for all radios if you want
to do some kind of Priorization on the router.

Flow control from the radio to the router is a (in my opinion) rare
cornercase for radios with a TDMA that can control incoming traffic.

Because of this I think an unidirectional flow control extension is a
good idea, otherwise we add a lot of complexity which will nearly
never be used.

Henning Rogge