[4013] in linux-net channel archive
Re: Question on performance
daemon@ATHENA.MIT.EDU (David C Niemi)
Thu Aug 15 04:44:19 1996
Date: Tue, 13 Aug 1996 17:22:34 -0400 (EDT)
From: David C Niemi <niemi@wauug.erols.com>
To: Kristian Elof Soerensen <kris10an@internet.dk>
cc: netmail <linux-net@vger.rutgers.edu>
In-Reply-To: <Pine.LNX.3.91.960812144953.312B-100000@hugin.bifrost.dk>
On Mon, 12 Aug 1996, Kristian Elof Soerensen wrote:
> On Thu, 8 Aug 1996, Emanuele Buttice wrote:
> > I have never been able to get more than 500KB transfer rate between any
> > of my PC's, Is this normal ?
> >
> > I have about 5 pc's running Linux 1.2.13 mostly P5's with >= 16meg Ram
> > and ISA 3ccom 3c509TP cards. The network traffic is very light and my
> > tests are done when no one is working. I have done transfers using ftp
> > and tcpspray but the results are simular.
>
> That's very normal, actually it's pretty high.
> About 300KB/s is normal for a network in good condition, with some
> traffic.
> In no cases schould it be slover than that, and on a small network with
> little or no traffic, schould you be able to see higher rates.
> Getting near 10Mbit/s will only happen with absolutely no disturbing
> traffic, and short exellent connections. Like two machines connected
> without a hub.
>
> In real world cases ethernet isn't a 10Mbit net, more like a 3-4 Mbit net.
Perhaps on a very large subnet with lots of small bursts of traffic. On
real networks like the one here I use, with many dozens of chattering
Novell clients and NT servers, I still routinely get 800-950 KB/sec FTPs
using ISA AMD PCNET Ethernet cards on a 486 DX2-66. To get this
throughput, you *must* be using binary mode with hash turned off (or
changed to tic mode); the file must be quickly gotten, i.e. in the cache
or at least on a fast disk; and the other end has to be fast too. I find
FTP is usually a more meaningful test than tcpspray, but if you use large
blocks with tcpspray the results start to converge.
At home I routinely get 1.1 MB/sec under no load between 2 PCs with PCI
Tulip cards, and between either of them and a SPARCstation 5 running SunOS
4. For some reason the Suns at work running Solaris2 are slower (700-800
KB/sec maximum even during quiet times).
A final note: 3c509s are not the fastest cards around. I consistently got
5-20% better throughput from the cheap PCNET cards, which use
bus-mastering (the 3c5x9s do not).
Now if I can only get both my 100 Mb Tulip cards to go into 100Mb mode
over a Cat5 Null Ethernet cable...if anyone has accomplished this please
let me know.
David
Niemi@linux.wauug.org 703-810-5538 Reston, Virginia, USA
------ Money talks, but it is wrong half of the time. -----
