Date: Mon, 17 Dec 2001 09:09:21 +0100 From: Thomas Zenker <thz@Lennartz-electronic.de> To: Mike Silbersack <silby@silby.com> Cc: stable@freebsd.org Subject: Re: USB ethernet problem Message-ID: <20011217090920.A763@mezcal.tue.le> In-Reply-To: <Pine.BSF.4.30.0112141311330.29846-100000@niwun.pair.com>; from silby@silby.com on Fri, Dec 14, 2001 at 01:17:13PM -0500 References: <20011214143752.A90727@mezcal.tue.le> <Pine.BSF.4.30.0112141311330.29846-100000@niwun.pair.com>
next in thread | previous in thread | raw e-mail | index | archive | help
On Fri, Dec 14, 2001 at 01:17:13PM -0500, Mike Silbersack wrote: > > On Fri, 14 Dec 2001, Thomas Zenker wrote: > > > Ok, I tried this. > > > > After doing some test, I decided, this can not be done with a quick > > test, (because of contradictory results) so did a serie of tests. > > It depends also on the transfer size (per connection). This is > > obviously caused by the slowstart alghorithm. > > Wow! Your benchmarks contain quite a bit of useful data, and confirmed my > suspicion that we should change the local slowstart flightsize to > something a bit more reasonable. > > Also, incidently, you seem to have shown what's suggested in one of the > newer rfcs - a slowstart flightsize of 1 is too small for optimal > performance (they suggest something more like 4.) > > I think I'll go ahead and drop the local slowstart flightsize to 4 - we'll > have to do some investigation later to see if raising the remote slowstart > flightsize from 1 to 4 is a good idea. > > Once again, thanks for the high quality testing! > > Mike "Silby" Silbersack I allways wondered, why the initial slowstart window is set to one (well some years I didn't look into the tcp code though). 9 years ago I had to develope the firmware for a store&foreward radio network, where I applied a lot of the ideas from the then net/2 tcp stack. The rtt in such a network is really horrible and packetsizes have to be taken in account. Anyway the optimal initial window there was 2. With a window of two there much more probability to get a connection going, because you send two packets in the beginning, if the first is lost, the receiption of the second one gets the first one resent long before the timeout. Otherway round, if the second is lost... the third is on its way already. With a intital window of 1 the only recovery is by timeout. The argument against bigger than two was (at least in my case) not to defeat the intention of the slowstart. Anyway, in tcp probably something between 2 and 4 could be considered. Thomas -- Thomas Zenker c/o Lennartz electronic GmbH Bismarckstrasse 136, D-72072 Tuebingen, Germany Phone: +49-(0)7071-93550 Email: thz@lennartz-electronic.de To Unsubscribe: send mail to majordomo@FreeBSD.org with "unsubscribe freebsd-stable" in the body of the message
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?20011217090920.A763>