From owner-freebsd-hackers Wed Jan 11 09:33:02 1995 Return-Path: hackers-owner Received: (from root@localhost) by freefall.cdrom.com (8.6.9/8.6.6) id JAA05824 for hackers-outgoing; Wed, 11 Jan 1995 09:33:02 -0800 Received: from cs.weber.edu (cs.weber.edu [137.190.16.16]) by freefall.cdrom.com (8.6.9/8.6.6) with SMTP id JAA05818 for ; Wed, 11 Jan 1995 09:33:01 -0800 Received: by cs.weber.edu (4.1/SMI-4.1.1) id AA22967; Wed, 11 Jan 95 10:26:45 MST From: terry@cs.weber.edu (Terry Lambert) Message-Id: <9501111726.AA22967@cs.weber.edu> Subject: Re: FreeBSD as an IP Router To: bde@zeta.org.au (Bruce Evans) Date: Wed, 11 Jan 95 10:26:45 MST Cc: hsu@cs.hut.fi, wollman@halloran-eldar.lcs.mit.edu, freebsd-hackers@freefall.cdrom.com In-Reply-To: <199501110553.QAA02180@godzilla.zeta.org.au> from "Bruce Evans" at Jan 11, 95 04:53:11 pm X-Mailer: ELM [version 2.4dev PL52] Sender: hackers-owner@FreeBSD.org Precedence: bulk > >The speed limit for ISA network interfaces is the ISA bus speed. You > >are actually CPU limited, but because the CPU has to reach out over > >the ISA bus to pull data off the interface, it takes much longer than > >would a memory access to data in system RAM. > > Actually, on my DX2/66 with a WD8013EBT (16 bit shared memory), the ISA > overhead for a single interface running `ttcp -t' at full speed is about > 35% while the general system overhead is also about 35%. On slower > systems, the ISA overhead is almost the same while the general system > overhead increases. The system doesn't have to get much slower before > it cannot saturate the ethernet. My DX33 with a WD8013EBT can barely > saturate the ethernet, but when the card is in NE2000 (PIO) mode, it > cannot. I have to agree; the limiting factor is the transport bandwidth, since it is less than the bus bandwidth. I can prove this by running a Novell NetWare sever on a split vs. a single net (2 cards vs. 1) with 125 clients on either net -- and have done so in Novell's SuperLab. Whether the drivers are too compute intensive to make this work well is another question (the FreeBSD drivers are not, the UnixWare drivers are borderline above the 250 client range due to [unidentified] propagation latencies in the Streams implementation). By the way, judging Streams by USL's implementation is equivalent to judging rock music by Primus. 8^). Terry Lambert terry@cs.weber.edu --- Any opinions in this posting are my own and not those of my present or previous employers.