Date: Thu, 15 Aug 2002 00:05:35 -0500 (CDT) From: Mike Silbersack <silby@silby.com> To: Barney Wolff <barney@tp.databus.com> Cc: Oleg Polyakov <opolyakov@yahoo.com>, <freebsd-net@FreeBSD.ORG> Subject: Re: Initial congestion window increase Message-ID: <20020814233935.F97690-100000@patrocles.silby.com> In-Reply-To: <20020814121701.GA27934@tp.databus.com>
next in thread | previous in thread | raw e-mail | index | archive | help
On Wed, 14 Aug 2002, Barney Wolff wrote: > You're assuming that the jumbo will be the successful MTU. But at > the start of a connection PMTUD has yet to run, and you could be > sending jumbos into a choppy link somewhere on the path. > > The tcp-impl IETF WG had (and the email list still has) a very smart > bunch of people with decades of experience with TCP. Those RFCs > didn't just come out of somebody's idle thought. > > Slowstart flightsize doesn't matter a whole lot on a lan (as long > as it's at least 2 to compensate for delayed ack) other than for > locker-room comparisons with Linux. But it does matter a lot on > long pipes, whether fat or thin, and that's where the risk of > an overaggressive strategy is that you can congest the Internet. Hrm, I'm not sure that PMTUD is a strong enough argument against X*MSS slowstart for gigabit networks. Think about the following cases: 1. Server with MTU 1500, client with MTU 1480 (they're going over PPPoE or something similar.) - All four 1500 byte packets sent back to back, all 4 bounced with ICMP too big messages. Bandwidth wasted: All 4 packets traversing the net, all 4 icmps coming back across the net. 2. Server with MTU 9000, client with MTU 1500. - All four 9000 byte packets sent back to back, bounced back at local border router with MTU of 1500. Bandwidth wasted: Internal network bandwidth only. Perhaps less than 4 packets, if all data fit into a single 9000 byte packet. Considering this, I don't believe that the gigabit host using jumbo frames would be any more harmful than a 100mbps host using normal ethernet frames. Mike "Silby" Silbersack To Unsubscribe: send mail to majordomo@FreeBSD.org with "unsubscribe freebsd-net" in the body of the message
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?20020814233935.F97690-100000>