samba - May 1999 - Bad NICs

Steve Arnold wrote:
[snip]
[Lots of bad performance stats for Realtek PCI ethernet
cards]
> I guess I just don't have a head for business; why anyone would 
> want to buy or sell a deficient piece of hardware just because it 
> costs a few percent less, as opposed to a product that ACTUALLY 
> WORKS for just a few bucks more, is beyond my meager comprehension. 
[snip]
Steve's tests are interesting and I'm not saying he's wrong, but you
can
get some very strange performance variations moving big chunks of data
between a slow and fast computer. This is caused by the underlying TCP link
protocol trying to flow control the data stream - basically, the TCP window
size drops right down and the slow computer is only receiving a tiny data
chunk before sending an ack. This causes huge protocol overhead and the
data rate falls through the floor. I suspect a number of the reported
performance problems in this list are related to this - people tend to
stick Linux and Samba on an old 486 then use this to serve an office full
of brand new PIIs running WNT.

Why the performance figures for the 3c509 were so much better I haven't a
clue, unless the fact that it was an ISA card slowed the data rate
sufficiently to minimise the windowing problem?

Best regards, Paul

Paul Sherwin Consulting    22 Monmouth Road, Oxford OX1 4TD, UK
Phone +44 (0)1865 721438   http://www.telinco.co.uk/psherwin/index.htm  
Fax   +44 (0)1865 434331   mailto:psherwin@telinco.co.uk
Pager +44 (0)7666 797228

When the world was young, Paul Sherwin carved some runes like this:
> I suspect a number of the reported performance problems in this list
> are related to this - people tend to stick Linux and Samba on an old 486
> then use this to serve an office full of brand new PIIs running WNT.
> 
> Why the performance figures for the 3c509 were so much better I haven't
a
> clue, unless the fact that it was an ISA card slowed the data rate
> sufficiently to minimise the windowing problem?
I hate to disagree (wait a minute - no I don't) but I like the fact 
that a real OS (like linux) can clearly show when you have a bad 
piece of hardware (whether failing or marginally designed).  You 
may be right about some of the performance problems (in a sense) 
but the fact is that it's due to bad hardware, NOT just because 
you're running linux on a 486.  If the 486 hardware is good, then 
there should not be any problems such as you describe.  Certainly a 
486 running linux will not be able to serve an unlimited number of 
samba connections, but that's not what we're talking about here.

An example:

At work we have a mixture of newer Dell PII-266 machines running 
(of course) win95 (with DEC 21040 based PCI NICs) and a similar 
number of older machines of various types (486-pentium) with all 
different ISA NICs (eg, WD 8*13, ne2000 clones, Intel, SMC, etc - 
even a real Novell ne2000 card).  My linux boxes are a 486-100 and 
a 586-100.  We even have a half coax and half UTP ethernet setup.  
Performance is pretty consistent across all machines (around 850-
950 k/sec for both smb and ftp).  The best throughput I measured 
was an ftp transfer between the NT box (PII-450) and my 486 linux 
server; about 980 k/sec.  Besides the samba stuff, that box also 
runs apache, MySQL, etc.  I have seen nothing in this environment 
that supports your analysis shown above.  Nice try though ;-)

I stand by my evaluation of the RealTek 8029/WinBond chips as crap, 
while the DEC and AMD chips are known to work fine (and the price 
difference is negligible).  I'm not asking for bleeding edge, 
overclock-your-300a-to-450MHz performance.  I'm only asking for 
nominal 10 Mbs ethernet performance; if they can do it over the ISA 
bus, they can certainly do it over the PCI bus...

Steve

******************************************************************
Stephen L Arnold                      http://www.rain.org/~sarnold
#include <std_disclaimer.h>
******************************************************************