From:	SMTP%"CARROLL@pppl.gov"  5-DEC-2001 09:16:25.01
To:	RONEY@pppl.gov, TGIBNEY@pppl.gov
CC:	TCARROLL@pppl.gov
Subj:	Notes for Europa's ethernet interface FASTFD 100MBIT Full Duplex

Phyllis and Tom,

	This is for your notes file:

The reboot of Europa yesterday after The cpu's were swapped and cards
reseated created a new and mysterious problem.

The console ethernet device eia0 was set FASTFD but the PPLCC hub changed
its speed to 100 half duplex. This created a very slow cluster and a Europa
that ran at one tenth it normal speed. Many tasks on all cluster nodes in
rwscs state(lock management wait) for very long periods of time. Many Disks
were going into mount verify timeout also across the cluster. The command

On Europa:

$ mcr ncp show known line counters


Displayed send failures at a rate of about 1/sec.

$ mcr ncp zero known line counters     ! to zap them to get the current

				       !situation

Bob Persing and Ken Tindall worked on this and said the problem was that
Europa was now allowing auto-negociate on its nic card. Compaq did update
the console firmware to version 6 but they did not think it would change the
nic card settings. Display of those setting looked normal from the Europa
console point of view.

After various tests the PPPLCC hub was changed to make the card always
auto-neg the nic card. Europa was still at fastfd on its console and was
unchanged in all these tests.

We powered off Europa and rebooted several times and it always came up ok if
the hub was auto-neg. In tests where the pplcc hub was 100 full duplex it
failed every time.

We have decided to leave the PPLCC hub with the auto-neg setting for the
Europa port.

This is a new problem so it may happen with other cluster nodes. I suspect
the new cabletron hubs installed on Nov 19 created this new behavior
because we never had a issue like this before with Europa. We did have a
issue with IO where it changed its speed while IO was up and running. Tom
Gibney was involved with this issue when I was out some time ago. After a
week it was traced to the network speed issue.

If this occurs when the experiment is running it will degrade the cluster
into a non-function state because the lock management can't be done in a
timely way.

Or in other words we need reliable networking to run.

---Tom

     **************************************************************
     *       Tom Carroll        |   Princeton Plasma Physics Lab  *
     *     Analyst/Programmer   !       Computer Division         *
     * Email: tcarroll@pppl.gov !   Office phone (609) 243-3419   *
     **************************************************************