[97854] in North American Network Operators' Group
Re: TCP congestion
daemon@ATHENA.MIT.EDU (Iljitsch van Beijnum)
Thu Jul 12 17:24:30 2007
In-Reply-To: <796919.42253.qm@web30811.mail.mud.yahoo.com>
Cc: Stephen Wilcox <steve.wilcox@packetrade.com>, nanog <nanog@merit.edu>
From: Iljitsch van Beijnum <iljitsch@muada.com>
Date: Thu, 12 Jul 2007 22:54:56 +0200
To: Philip Lavine <source_route@yahoo.com>
Errors-To: owner-nanog@merit.edu
On 12-jul-2007, at 22:27, Philip Lavine wrote:
> I just don't understand how if there is 1 segment that gets lost
> how this could translate to such a catastrophic long period of slow-
> start. How can I minimize the impact of the inevitable segment
> loss/out of order over a WAN. Is QoS the only option?
How exactly are you going to get out-of-order packets over a single
link?
Is there perhaps a firewall somewhere in the communication path? In
that case' it's entirely possible that packets that use certain TCP
options are filtered because the firewall doesn't implement these
options.
For instance, if there are selective acknowledgements, the firewall
may not like that.
Also, if you use the window scale option the firewall probably
doesn't understand that and may easily conclude that a packet which
is in-window with the scale factor applied is out-of-window because
the scale factor _isn't_ applied and drop it, so that you miss the
packet with byte 1, then get the ones with everything upto 65535 and
the rest is dropped, and so many dropped packets will stop TCP dead
in its tracks.
Also note that minimal amounts of cell loss on ATM create huge
amounts of packet loss at the IP layer.
