Curious phenomenon

[Robert G. Brown]

On Wed, 26 Aug 1998, Greg Lindahl wrote:

> I have sort-of verified the cat5000 autosense weirdness reported
> earlier.
> 
> I have 2 alphas with DE500 cards (modern 21143-based cards) and
> driver v0.83, wired to 100 megabits at compile time.
> #define TULIP_DEFAULT_MEDIA 3
> 
> If my cat5000 is set to 100/full, I get 5% to 10% packet loss with
> 8000 byte packets between the 2 machines. They get lots of nfs errors
> in general and their clocks jump 10 to 20 milliseconds every time they
> get one.
> 
> If my cat5000 is set to autonegotiate, I get 0% loss.
> 
> Now, if you read what I said above, you'll note that I made a
> mistake. I wired the driver to half-duplex and the cat5000 to
> full-duplex. Is this the behavior I should expect? Or am I reading
> something wrong? Is this what rgb was seeing, too?

The only difference is that I wasn't hotwiring the drivers -- I was
letting them choose their own best speed.  I was also seeing more than
5-10% loss -- loss rates depended on message size but for reliable
protocols TCP rates went down from 93 or so Mbps to 5 Mbps in netperf
runs.  NFS (client or service) REALLY sucked, but even tcp-based
protocols like scp visibly suffered.  Losses were bidirectional; I
didn't do any better with incoming netperfs than with outgoing
netperfs.

I have different (tulip KNE100 and eepro100) cards and (single/dual)
(p6 or PII's).  Not alpha/intel specific.  Not card/driver specific.
Not known whether it is Cat5000 specific.  Anybody out there with a
non-Cisco 100BT switch that can be hard-set to 100/full who cares to
make a measurement or two?

   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