Curious about specifying amount of memory for MPI.

Clicked A Few Times

9:24:30 AM PDT - Mon, Jun 19th 2017

Dear all,

I do have two questions in running NWChem. First, I always submit the job on the single node by using following command,

$ mpirun -np N nwchem inputfile.nw >& outputfile.out &

In the specification of memory usage in such input file I use the following line

memory total 8 GB

I just was wondering that 'the real total memory' used for this calculation on this machine would be just 8 GB or 8*N GB. I am not sure that the 'total' keyword in command above is for PER EACH PROCESSOR or PER TOTAL PERCESSOR.

I also just want to know does it sound good to use OpenMP in conjunction with MPI for NWChem calculation, for example

$ mpirun -np N -x OMP_NUM_THREADS=1 nwchem inputfile.nw >& outputfile.out &

Thank you for your suggestion.
RK

Forum Regular

4:34:05 AM PDT - Tue, Jun 20th 2017
All memory specifications are per processor (http://nwchemgit.github.io/index.php/Release66:Top-level#MEMORY). So memory total 8 GB would require N*8 GB of system memory, where N is the number of MPI processes. NWChem itself currently has limited support for OpenMP parallelization (see http://nwchemgit.github.io/index.php/Compiling_NWChem#How-to:_Intel_Xeon_Phi for some details). Whether the threaded versions of the support libraries will offer a performance boost will probably depend on which version of those libraries you are using (e.g. Intel MKL, OpenBLAS, etc.) and the particular calculation you are doing.

Forum >> NWChem's corner >> Running NWChem