tulip vs. Lite On

Robert G. Brown rgb@phy.duke.edu
Tue Dec 8 09:03:37 1998


On Tue, 8 Dec 1998, Douglas Eadline wrote:

> 
> There was some discussion on the linux smp list about
> the genuine tulip vs. the lite on (now used in netgear) chips.
> 
> I mentioned that the performance as determined by netperf, was about 75%
> for the lite on chips set. I thought the tulip list would 
> be interested in these  numbers.
> 
> Tulip driver is .90
 (deleted stuff)
> ===================
> 
> UDP UNIDIRECTIONAL SEND TEST to remote2
> Socket  Message  Elapsed      Messages                
> Size    Size     Time         Okay Errors   Throughput
> bytes   bytes    secs            #      #   10^6bits/sec
> 
>  65535    1472   60.00      363918      0      71.43
>  65535           60.00      363918             71.43
> 
> TCP STREAM TEST to remote2
> Recv   Send    Send                          
> Socket Socket  Message  Elapsed              
> Size   Size    Size     Time     Throughput  
> bytes  bytes   bytes    secs.    10^6bits/sec  
> 
>  65535  65535  65535    60.00      71.04   

These are very interesting numbers, Doug.  Have you tried tulip->lite-on
or lite-on->tulip?  They are interesting because the card is handling
only ~6000 pps on the transmission side!  This is absurdly low (0.17 ms
between packets) since the transmissions ought not to be blocked by wire
bandwidth.

Another question:  have you tried a couple or three points along the
small-message part of the curve?  You might try (UDP_STREAMS) m = 1, m =
512, m = 1024 to sort of sketch out how the NIC handles small packets --
on a real tulip the interface should handle 20K-50K pps for small
packets (depending on host CPU speed) until the interface starts to
saturate the available bandwidth and has to cut back on the number of
packets sent.  One is curious -- if the lite-on can only send 6K pps
with optimal packet sizes (which is far from saturating the bandwidth),
can it only send 6K pps for small packets?  

If so, this would make the NIC absurdly poor, far worse than only "75%
of a tulip".  In fact, I would call it openly "broken" unless there is a
lite-on-specific card hack unknown (so far) to Don that might make it
perform normally.  Alas, all these tulip-chip clones and even many real
tulip cards layer on the nonsense that makes it difficult to impossible
to write a "generic" tulip driver; Don has expressed his frustration on
this very subject but what can you do?

   rgb

Robert G. Brown	                       http://www.phy.duke.edu/~rgb/
Duke University Dept. of Physics, Box 90305
Durham, N.C. 27708-0305
Phone: 1-919-660-2567  Fax: 919-660-2525     email:rgb@phy.duke.edu