[Wien] Parallel running

[Peter Blaha]

Hi,
I'm sorry, but I don't quite understand your last email.


> is that this cluster is also used by too many people from all the
> University campus with all kind of programs, sequential and parallel at
> the same time. When I start running the Wien2k, for example, for 4
> computers, the PBS allow me the free computers and by a script a file
> “.machines” is created and submit runsp_lapw –p –cc 0.0001 –NI. Depending
> on the computer available the program lapw1 runs with a relevant time
> difference of other that is running a program of other user, with an
> additional drawback, the amount of RAM memory available in the computer
> also influences in this time. After a long time (not finishing the lapw1)
> the PBS shows an error and aborts the program. I would be most grateful
> if you had a suggestion to help me to solve this problem. So I could talk
> to the system administrator to implement the possible solution and finish
> my PhD. Please get back to me if you have further questions.

When you start your pbs job, you shoud get assigned some hosts for your job
AND ONLY you should be allowed to use these nodes.

Are you telling me that during your job, other users can also start jobs
on some of these nodes,....

If so, than you can forget any finegrain mpi parallelization, because
synchronization will always be the problem.

On the other hand you mentioned, the job does not finish (or gets killed
by PBS).  What is your test case. You should have at least some
reasonable matrix size (maybe >3000). Don't try TiC !

And in this case the sequential run gives you an upper time limit, the
parallel run on eg. 4 cpus should NOT be slower.

Regards


                                      P.Blaha
--------------------------------------------------------------------------
Peter BLAHA, Inst.f. Materials Chemistry, TU Vienna, A-1060 Vienna
Phone: +43-1-58801-15671             FAX: +43-1-58801-15698
Email: blaha at theochem.tuwien.ac.at    WWW: http://info.tuwien.ac.at/theochem/
--------------------------------------------------------------------------