Cluster switch changes

Stefan Savage (savage@cs.washington.edu)
Mon, 26 Apr 1999 15:12:15 -0700

Messages sorted by: [ date ][ thread ][ subject ][ author ]
Next message: tom@emigrant: "meeting this week"
Previous message: Stefan Savage: "A new bit of evidence in the DSL vs cable debate..."

It turns out that some NICs (including all the cluster machines from
212-235) and some switches (when using "short" cables) and some hubs, do not
"cooperate" correctly with Intel's autonegotiation procedure. Consequently
the two sides will arrive at conflicting decisions about whether the channel
is full-duplex or half-duplex. This results in a performance degradation
between 2x and 100x. Consequently, I have reconfigured the switch to
default to half-duplex and I've saved this configuration in its flash.
There are three exceptions. I left in the incoming port (1) in autoneg
(since one hopes that the dept hooked us up point-to-point), I've put the
DEC Alpha port (5) in 10Mbps mode and I've put the mm-studio port (20) in
full-duplex since its supposed to be a server. Please let me know if any of
you experience problems with this new configuration.

- Stefan

P.S. It is not clear to me whether any of the NICs in the old Intel boxes
can successfully support full duplex.

Next message: tom@emigrant: "meeting this week"
Previous message: Stefan Savage: "A new bit of evidence in the DSL vs cable debate..."