Ethernet math (Was Re: MWC's Coherent - A Lemon...)

Davin Yap davin at me.utoronto.ca
Wed May 30 03:37:36 AEST 1990


In article <2871 at crash.cts.com> jca at pnet01.cts.com (John C. Archambeau) writes:

>My data is my findings via experience in hooking machines to an ethernet.  I
>have yet to see an ethernet perform a consistant (key word here) throughput of
>more than 3 Mbit per second.  Last October I added a Sun 386i/250 to an
>existing ethernet of less than 10 machines.  The best throughput I've seen on
>that wire was 4 Mbit per second.  You tell me why it was dropping packets like
>crazy.  

Are you sure it was dropping packets?  What were the other machines on
the net and how were you testing the ethernet _interface_.  A lot depends
on the combination of computer/ethernet controller, the smarter the
controller, the less the cpu has to do, the greater your throughput.  If
the controller doesn't handle the IP and the transport layers, the cpu
must; if the card isn't a bus master, you lose anyway.  3 - 4 Mbit/s
sounds right if you've got a bunch of Sun 3s and pc's on your net, the
individual ethernet interfaces of these machines max out around here; the
Suns because of the nonoptimal drivers that came with SunOS, the pc's
more likely due to hardware limits.  I know about the Sun 3 numbers, I'm
speculating on the pc numbers.  
   On the other hand, I've come across machines/ethernet controllers
that'll handle 10 Mbit/s easy.  A MIPS RC3260 with a block mode ethernet
controller comes to mind: over NFS (i.e. UDP) I was transfering 1 MByte/s.
On a Sun 4/490 I was getting 800+ kBytes/s.  Over TCP on a cluster of
Sparcstation 1(s) I was getting over 700 kBytes/s.  These were all casual
tests, I wasn't out to find the ultimate performance, just close to it :-).
The tests were constructed so that the ethernet interface was always the
rate limiting step, the data to be transfered was always from memory not
from disk.
   BTW, I'm told the 'le' (lance ethernet) controller has been enhanced
in SunOS 4.1 (apparently they've incorporated Van Jacobsons improvements)
so that a Sparcstation 1 can now saturate the ethernet.  Putting multiple
ethernet cards in an SS1 leads to a truly inspiring router :-)
   I wonder, though, about putting slow machines on the same ethernet as
these faster machines, it'll be difficult for them to get a word in
edgewise :-).  A collision of wills (like wit) where the quickest
prevails :-).

>When *I* start seeing consistant throughputs of more than 3 or 4 Mbits per
>second on an ethernet then I'll agree with you, but until then I write all of
>this off as the overhead of ethernet.

It isn't ethernet overhead that's bogging you down, depending on you
packet size there's precious little of that.  Your problem is that your
ethernet interfaces suck, the combination of computer, ethernet
controller and driver.  Unfortunately, you're in the same boat as most
people out there, and for you 3-4 Mbit/s is what you're gonna see.

Davin
--
Davin Yap, Center for Computer Integrated Engineering, University of Toronto
davin at me.utoronto.ca  davin at me.utoronto.bitnet | 5 King's College Rd., Toronto
  ...{pyramid,uunet}!utai!utme!davin           |    Ontario, CANADA, M5S 1A4



More information about the Comp.unix.xenix mailing list