On Nov 26, 2007, at 5:59 AM, scott fletcher (BITS) wrote:
Problem 1
=========
...
At this point even if we revert to submitting jobs directly to
Condor we
get the same message, the only thing that seems to fix it is a reboot.
I don't have an idea about this one, and I suspect you'll have better
luck with it in a Condor forum. I am surprised, however, because I
know that in their architecture there's a separate daemon called the
GAHP they use to offload their interactions with things like Globus
into a separate daemon. The only thing I can think to suggest is to
look if a GAHP is up and running at the time you experience this
problem and try killing it.
Problem 2
=========
When we submit a job to the master node it gets there and runs as you
would expect and then exits, however on the submission node the job
appears idle until about a minute after the job has actually finished
(on short jobs lasting 10 secs, we have not really tried any long ones
yet), it then shows status as running (which takes several times
the job
actually took to run) and then exits.
This has to do with the architecture of GRAM2. It polls for job
completion, and does so at a one minute interval. Condor-G is
meddles with it to try to improve things, which I believe is the
poll_fast output you're seeing. It sounds like the poll_fast isn't
speeding things up, and you're instead getting the default one-minute
interval polling. If you set up the more recent implementation of
GRAM, which also works with Condor, you will get near-instantaneous
notification of job completion.
Charles
