Guy Teverovsky wrote:
[Guy] One of my biggest concerns is the ease of use. Reading the Oscar's
documentation I noticed that the job submission/distribution process is
far from being intuitive. What I have is a bunch of researchers who are
very smart at what they do, but most of them will scold me if the job
submission involves some interaction with additional tools. Using
pvm/mpi in Oscar (at least according to the docs) would require some
manual fiddling on the user side. With *Mosix the process is much more
user friendly.
If ease of use is such a concern, pretty anything is going to be a
hassle for your researchers. A reasonable compromise would be a batch
system (i.e. NQS, DQS, OpenPBS) - which is what I suggest in this case.
In my case, what we produced for a client who was not too fond of weird
tools (and had some 48 2-CPU compute nodes) was a batch dispatcher.
I have no intension putting the nodes on a separate network segment/VLAN
In your case, I do not see a need for a separate network for the nodes -
if you are indeed compute and not I/O and/or inter-node communication
intensive.
I think I'll try installing OpenMosix kernel, disabling OSCAR's HPC
related modules and using OSCAR only for nodes
deployment/management/image updates, but I'm open to suggestions.
I suggest trying one of the simple batch dispatchers or - perish the
thought - writing one yourself.
In most cases, it is extremely not difficult, if you're willing to
assume a number of things.
M
=================================================================
To unsubscribe, send mail to [EMAIL PROTECTED] with
the word "unsubscribe" in the message body, e.g., run the command
echo unsubscribe | mail [EMAIL PROTECTED]
