MPI settings

[ams_nanolab]

Dear Sir,

I am submitting a job on parallel on a single machine having 4 processors (6 cores each).

Can I increase the threading further (safely) to maximize memory usage.

Attached is a screenshot of the PIDs running.

[Anders Blom]

No, you can't influence the memory usage by more or less threading.

In general it should only very rarely be necessary to modify the threading settings, except perhaps if you want to limit threading in order not to load the machine too hard. ATK will normally use all available resources on the machine, but the distribution of MPI vs OpenMP is determined within the code, and can't be influenced.

You should be careful not too overload the bus etc with too many MPI processes on a single machine. 4 MPIs seems reasonable on your machine, and then let each MPI thread (it will, automatically). If you see not so strong threading it's because so far only parts of ATK use OpenMP, but we are working actively on increasing this part.

[ams_nanolab]

Thank you  could you elaborate which parts of ATK use MPI.
Also I was wondering if in handling large supercells (as in our case) is it more processor dependent or more memory dependent, we have a local machine with 4 (6-core) processors and 256GB of total memory.

[Anders Blom]

ATK is primarily MPI-parallelized over k-points and (for devices) energy points on the complex contour integration.

This means the MPI advantage is very limited for large supercells since it's a bulk calculation with no k-points. On the other hand threading will be able to work well.

[Nordland]

It is also very parallel in the calculation of the exchange-correlation and calculation of matrix elements of the hamiltonian. And real space operations - however these part is also very fast, so as Blom says the other thing will quickly dominated.