[CSC 435] How to play nice...

[Andrew J. Pounds]

Several of you will start your cluster testing today.  I will boot the
last two system in GSC 218 into the cluster, but I recommend you start
small.   To play "nice" I recommend that you break your jobs up as much
as possible to give everyone equal access to the queues.  Normally I
would say do somethings like this in a PBS script...


mpirun -np 4 ...  mmm_hybrid 2000 1
mpirun -np 4 ...  mmm_hybrid 2000 1
mpirun -np 4 ...  mmm_hybrid 2000 1
mpirun -np 4 ...  mmm_hybrid 2000 2
mpirun -np 4 ...  mmm_hybrid 2000 2
mpirun -np 4 ...  mmm_hybrid 2000 2
.
.
.
mpirun -np 4 ...  mmm_hybrid 2000 20
mpirun -np 4 ...  mmm_hybrid 2000 20
mpirun -np 4 ...  mmm_hybrid 2000 20


But I would say in this case do something like the following....


mpirun -np 4 ...  mmm_hybrid 2000 1
mpirun -np 4 ...  mmm_hybrid 2000 3
mpirun -np 4 ...  mmm_hybrid 2000 5
mpirun -np 4 ...  mmm_hybrid 2000 7
.
.
.
mpirun -np 4 ...  mmm_hybrid 2000 15

in one job and then in another job pick up the even number of threads.  
If you see your performance improving near the high end of your range,
add more threads in another job -- but don't repeat all of the low order
threads.  Once you know where the maximum is then you can do another
test where you run all of the tests twice.  In this job you will include
all of off the threads up to your maximum and the runs use 1,2,3, and 4
threads beyond your maximum.   This should give you enough data to get 
"reasonable" benchmarks so you can see the maxima on your 3D plots.

   

-- 
Andrew J. Pounds, Ph.D.  (pounds_aj at mercer.edu)
Professor of Chemistry and Computer Science
Director of the Computational Science Program
Mercer University,  Macon, GA 31207   (478) 301-5627

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://theochem.mercer.edu/pipermail/csc435/attachments/20200429/02a6eec5/attachment-0001.html>