Re: [tcpm] [Tmrg] Increasing the Initial Window - Notes

[Mark Allman <[e-mail]>]

David-

> But I don't think it is congestion itself and the dropping of packets
> that is the problem, the issue is we don't want that congestion to
> turn into congestion collapse.  So, it seems to me that the issue
> about whether or not it is safe to increase the IW should be focused
> more on, if increasing the IW leads to increased congestion, is the
> network still able to properly respond to that congestion and avoid
> congestion collapse?  If the answer is "yes", then I think that
> greatly reduces the risks of increasing the IW.  But if the answer is
> "no", then we have a bigger problem that is lying in wait and can be
> triggered by other factors than just increasing the IW. 

I think the answer is pretty obviously "no".  We are not taking the
reaction to congestion away at all and *that* is the important part in
preventing collapse.  In other words, left alone, TCP will keep reducing
the rate at which it tries to recover (via the window and ultimately via
RTO backoff).  This is what I mean when I have said that changing the
initial cwnd does not eliminate congestion control.

Second, remember that we know that most of the traffic volume is large
flows.  And, initial settings in large flows are pretty meaningless.
Further, there are plenty of flows that aren't even 15KB.  So, there
would have to be some scenario whereby some number of medium-sized flows
that are not taking a large chunk of the bandwidth would collapse the
network.  That seems hard to see.

The only scenario I can come up with that even starts us down the road
towards collapse is the case where a connection uses a 10 segment
initial window and takes some loss and the user gets impatient and hits
reload to give it another shot.  In that case, the packets that made it
through in the first connection *wasted* scarce network resources (i.e.,
made it across the congested link) and yet ultimately accomplished no
useful work.  However, I don't buy this as a very likely scenario
because if you play out sending 10 packets with IW=3 and IW=10 in a
network that can only handle 3 packets the transfer times are pretty
comparable (I did this last night on the TMRG list) with the IW10
connection longer by roughly one RTO.  So, if we're worried about user
impatience with IW10 then we should worry about it with IW3, too.

allman