From owner-freebsd-hackers Wed Jun 18 14:12:20 1997 Return-Path: Received: (from root@localhost) by hub.freebsd.org (8.8.5/8.8.5) id OAA11441 for hackers-outgoing; Wed, 18 Jun 1997 14:12:20 -0700 (PDT) Received: from verdi.nethelp.no (verdi.nethelp.no [195.1.171.130]) by hub.freebsd.org (8.8.5/8.8.5) with SMTP id OAA11363 for ; Wed, 18 Jun 1997 14:11:59 -0700 (PDT) From: sthaug@nethelp.no Received: (qmail 4915 invoked by uid 1001); 18 Jun 1997 21:11:54 +0000 (GMT) To: tom@sdf.com Cc: ccsanady@scl.ameslab.gov, hackers@FreeBSD.ORG, matt@3am-software.com Subject: Re: Network concurrency problems!? In-Reply-To: Your message of "Wed, 18 Jun 1997 13:48:01 -0700 (PDT)" References: X-Mailer: Mew version 1.05+ on Emacs 19.28.2 Mime-Version: 1.0 Content-Type: Text/Plain; charset=us-ascii Date: Wed, 18 Jun 1997 23:11:54 +0200 Message-ID: <4913.866668314@verdi.nethelp.no> Sender: owner-hackers@FreeBSD.ORG X-Loop: FreeBSD.org Precedence: bulk > > Also, if you expect a PPro-200 to saturate 4 100 Mbps links, I think you > > are a wee bit optimistic. (One link, no problem.) > > Why? As long as the ethernet hw is fast, is should be no problem. It's a problem because of the resource usage of the TCP/IP stack and the driver. The FreeBSD TCP/IP stack is good, but it's not the most efficient. As far as I know, there is still an extra pass over the data to perform the TCP checksum, for instance. > I > can run 2 FW SCSI channels saturated (2 x 20MB/s * 8 = 320mbs), and the > system isn't even working too hard yet. Certainly, full-duplex > connections double the total possible bandwidth (4 x 100 x 2 = 800mbs), > but most servers are mainly outbound. > > However, I really doubt whether most ethernet adapters offload enough > functions from the main CPU. The trend is to make very stupid > controllers, which are slaved to the CPU for everything. There has been a good deal of debate on whether offloading is really the best idea for network protocol implementations. A lot of people have tried it, and a lot of people have failed. If you look at Van Jacobson't work you'll find him arguing in the opposite direction: A "stupid" (in reality: simple and efficient) controller, and a very efficient protocol stack implementation. See http://ee.lbl.gov/nrg-talks.html, in particular "Some Design Issues for High-speed Networks" and "Design Changes to the Kernel Network Architecture for 4.4BSD" Also "The Witless Interface", presented at the 1991 Gigabit Workshop. Steinar Haug, Nethelp consulting, sthaug@nethelp.no