becker at scyld.com
Thu Oct 24 13:07:11 PDT 2002
On Thu, 24 Oct 2002, Michael Stein wrote:
> > I recently had two IBM Linux clusters installed at my work site. Due to
> > certain customer issues, our cisco switches have all ports set to 100Mbits
> > Full Duplex for all server connections.
> My understanding is that setting ports to 100 Mbits/full duplex will
> PREVENT the NICs from negotiating thus requiring specifying the matching
> values on every server NIC.
Correct. To repeat the point: forcing full duplex will turn off
autonegotiation, and often link speed sensing.
> If not specified on the server the server should, without negotiating,
> default to half-duplex, thus resulting in a duplex mismatch.
> There might have been a time years ago when FE negotiating was a problem,
> but shouldn't we be way past that by now?
I answer questions relating to this issue every day.
The problem sites are those with Cisco switches and administrators with
Cisco switches had a broken autonegotiation implementation. Rather than
fix the deployed hardware, Cisco taught that autonegotiation was
unreliable and should be turned off. OK. But they went further and
recommended forcing full duplex.
> This maximizes the effort/support needed on each and every server.
That's the big problem. The result is that every connected machine must
now be specifically configured. Rather than having plug-and-play
computing resources that work out of the box, you need installation
people that understand the configuration specifics of each device.
Donald Becker becker at scyld.com
Scyld Computing Corporation http://www.scyld.com
410 Severn Ave. Suite 210 Scyld Beowulf cluster system
Annapolis MD 21403 410-990-9993
More information about the Beowulf