( ESNUG 310 Item 11 ) ---------------------------------------------- [2/7/99]
Subject: (ESNUG 308 #7) All In All, I Was Rather Disappointed With LSF
> Overall, I don't think I can say enough good things about LSF - but don't
> tell them that, they'll just want to raise their prices.
>
> - [ The Cisco Kid ]
From: "David C. Hoffmeister" <dch@eng.umd.edu>
John,
I've been reading your ESNUG column for three years now and this is the
first time I have felt the need to contribute. I just had to respond to
the comments praising LSF. While finishing my research for my Ph.D.
dissertation recently I was able to gain access to a cluster of Sparcs
at the Naval Research Lab. I was not using it for ASIC design but for
running discrete-event simulations of a computer network. I can only say
that either the LSF system at NRL was not configured and maintained well
or else the software is not all that reliable. I do know that NRL claimed
to be having problems integrating LSF with AFS and Kerberos, both of which
they were using. They also only had one queue set up for submitting jobs
to. So, the overall configuration may have been causing some problems.
My main gripe is the control of job submission. I think they could improve
the default job submission to be more intelligent. I had jobs that required
very little memory and were completely cpu bound. So, it would have been
nice if the default job control chose the most lightly loaded cpu and
dispatched my job there. This was not the case. Often my jobs were
submitted to hosts that had more than one job per cpu already while other
hosts had multiple idle cpus. It took a significant amount of work on my
part writing resource requirements for my jobs before I got approximately
what I wanted. On top of that, if I submitted too many jobs at once, even
with resource requirements, it allowed all of them to run and swamped the
system.
I was forced to use a Perl script (written by a fellow graduate student) to
check the number of jobs I was running and submit jobs as old jobs
completed. It was not convenient at all.
I also found that some of my jobs would continue running for more than a
full day. When I killed these jobs they would often report results as if
they had finished normally. I don't know how long they would have remained
in limbo if I had not killed them manually.
All in all, I was rather disappointed with LSF. Maybe I expect too much
functionality without a great deal of work. And, as I said, I cannot
guarantee that the system was installed and maintained properly.
- David C. Hoffmeister
University of Maryland
|
|
