Skip site navigation (1)Skip section navigation (2)
Date:      Sun, 17 May 1998 12:33:08 +0200 (MET DST)
From:      Luigi Rizzo <luigi@labinfo.iet.unipi.it>
To:        pantzer@ludd.luth.se (Mattias Pantzare)
Cc:        dg@FreeBSD.ORG, hackers@FreeBSD.ORG
Subject:   Re: bad behaviour in slow start
Message-ID:  <199805171033.MAA05381@labinfo.iet.unipi.it>
In-Reply-To: <199805171027.MAA26384@zed.ludd.luth.se> from "Mattias Pantzare" at May 17, 98 12:27:18 pm

next in thread | previous in thread | raw e-mail | index | archive | help
> > >         * Don't force slow-start on local network.
> > >I think this is in violation of good TCP practices, and should be at
...
> > When sending out to a local peer, ethernet and point-to-point devices will
> > buffer the initial burst of packets with the local buffers draining at the
> > speed of the available link bandwidth, so I don't see why you would want to

This initial burst is a whole max-size window, so with the same
reasoning we could bypass the congestion control algorithms when
talking with peers a local network!

> > do slow start in this case. This is different than the case of a congested
> > upstream circuit where you don't know about the congestion and have no
> > control over the buffering.
> 
> The problem is that you can't detect if the other computer is a local peer or 
> not, there may be routers in the path to it even if the netmask tells you that 
> it is on the same subnet.  It isn't even true on the ethernet level any more, 
> switches create the same problems as a router.

especially in case of 10/100 switches.

Actually one reason on a slow (ppp) link is that you might get a
false timeout soon, just because your first 16-32K of data will
take 10seconds to go through. You might even have real packet
losses, i don't remember how long is the queue on a ppp/tun link
but hopefully it is not the 50 slots used for ethernets, and people
might want to use a small MTU like 576 to improve interactive
response with background bulk traffic.

I suppose a lot of people using a FreeBSD-based ppp server get an
address which is on the same /24 net as the ethernet of the server.

My point is, this optimization came in only to serve the needs of
TTCP (and even in that case, it should have been enabled only when
the peer was known to support TTCP). Then it became the default
behaviour, possibly by mistake (at least, so i see it).  We could
ask confirmation Andres Olah who (I think) is the one who wrote
this piece of code, if someone knows how to reach him! (I am cc-ing
this msg to olah@freebsd.org just in case...)

I think it really ought to be optional and default to off, so people
can choose on their own, but knowing what they do. Given the
"complexity" of the fix i posted yesterday, i don't think it's such
a big deal.

	cheers
	luigi
-----------------------------+--------------------------------------
Luigi Rizzo                  |  Dip. di Ingegneria dell'Informazione
email: luigi@iet.unipi.it    |  Universita' di Pisa
tel: +39-50-568533           |  via Diotisalvi 2, 56126 PISA (Italy)
fax: +39-50-568522           |  http://www.iet.unipi.it/~luigi/
_____________________________|______________________________________

To Unsubscribe: send mail to majordomo@FreeBSD.org
with "unsubscribe freebsd-hackers" in the body of the message



Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?199805171033.MAA05381>