[gmx-users] Node assignment using openmpi for multiple simulations in the same submission script in PBS

[Mark Abraham]

himanshu khandelia wrote:
> I get bad scaling with 8 cpus.
> http://www.gromacs.org/index.php?option=com_wrapper&Itemid=84
> 
> 
>> I'm not seeing how openmpi (the implementation) can be
>> allocating cpus "depending on what options gromacs was originally
>> compiled with".
> 
> I do not see immediately how either, but just seemed logical to me. I
> may be wrong

When describing material like this, you'd do well to separate your 
speculation from your knowledge in your description, and limit the 
former. If your speculation was so hot, you'd have probably fixed the 
problem by now :-)

http://www.open-mpi.org/faq/?category=running#mpirun-scheduling suggests 
that the interplay of --byslot and --bynode with your machine will 
depend how the hostfile is set up. You need to find out what the 
defaults for the machine are and how to tweak them to what you want 
under PBS.

hostfile:
node0 slots=4 max_slots=4
node1 slots=4 max_slots=4

and then two "mpirun -np 4 --byslot --hostfile hostfile mdrun" look 
right to me.

Mark