[Iccrg] Meeting agenda
John Leslie
john at jlc.net
Tue Sep 12 18:46:40 BST 2006
Michael Welzl <michael.welzl at uibk.ac.at> wrote:
> On Sat, 2006-09-09 at 16:34, John Leslie wrote:
>> Michael Welzl <michael.welzl at uibk.ac.at> wrote:
>>> On Fri, 2006-09-08 at 19:38, Wesley Eddy wrote:
>>>>
>>>> 1) congestion control when striping data over multiple paths...
>>
>> There is a "correct" thing to do; but I don't claim it's practical:
>> have the two routers surrounding the split maintain it as a single virtual
>> path, with a single congestion behavior.
>
> If it's SCTP, I think that the question on the table is the
> correct reaction in the transport end system (where we can't
> assume that routers do what you suggest).
>
> In this case, I think that SCTP should treat each path
> just like TCP would (if you would use one TCP connections
> per path).
>
> What would be the reason to do anything else?
I see none, unless we have reason to believe the congestion occurs
on a portion of the path shared by both. I'd further argue that in the
absence of "simultaneous" congestion on two paths, there _is_ no reason
to believe the congestion is on a shared portion.
If, OTOH, we _know_ the congestion is on a shared portion, it could
be argued we're _able_ to treat it as a single virtual path. Frankly,
I don't agree; but were we able to, the question of fairness would be
a purely local one, so I'd call it a configuration option, and spend
my mental effort worrying only whether our overall reaction to congestion
will be fast enough.
>> Anything less than this will tend to break the model on which congestion
>> control is built. :^(
>
> Why?
Imprecise answer!
More exactly, I meant that we use an already-broken model which assumes
- congestion is the only source of packet loss;
- the next packet will follow the same path;
- therefore (collectively) we must immediately reduce the sending rate.
and that switching paths in midstream shows up its brokenness.
I wasn't thinking SCTP at all: if I were, I would have thought of it
as separate connections, not as striping.
--
John Leslie <john at jlc.net>
More information about the Iccrg
mailing list