Skip to end of metadata
Go to start of metadata

FHI-aims is an all-electron, full-potential electronic structure code package for computational materials science, see https://aimsclub.fhi-berlin.mpg.de/ for more details. FHI-aims can be compiled on Sisu using either gfortran or ifort. For gfortran one should load the correct environment modules and use the following setting in the Makefile:

 

FC = ftn
MPIFC = ftn
FFLAGS = -O3 -ffast-math -funroll-loops -ffree-line-length-none
F90FLAGS = $(FFLAGS)
FFLAGS_NO = -O0 -ffree-line-length-none
USE_MPI = yes

And for ifort with the proper modules

FC = ftn
MPIFC = ftn
FFLAGS = -O3 -ip -xHost -mkl=sequential -dynamic -opt-prefetch -unroll-aggressive
F90FLAGS = $(FFLAGS)
FFLAGS_NO = -O0 -mkl=sequential -dynamic
USE_MPI = yes
MKLROOT=/opt/intel/composer_xe_2013.1.117/mkl
LAPACKBLAS = -L$(MKLROOT)/lib/intel64 -lmkl_scalapack_lp64 \
-lmkl_intel_lp64 -lmkl_sequential -lmkl_core \
-lmkl_blacs_intelmpi_lp64 -lpthread -lm

that produces a working binary linking to Intel's MKL.

The Cray compiler can't be used since it optimizes a segfault into the code. This is probably a bug in the current compiler version since on Louhi exactly the same code works with the Cray compiler.

The overall performance and scalability was tested with an organic molecule, so called A beta oligomer consisting of 3456 atoms.

 

The figure above shows the time spent in one iteration of the s.c.f. cycle for 64 - 4096 CPUs. The timings are subdivided into the basic tasks of a DFT cycle: update of the charge density, solution of the corresponding Hartree potential, integration of the Hamilton matrix and solution of the eigenvalue problem. The solver for the most time consuming step, i.e., the Kohn-Sham eigenvalue problem is the ELPA library, see http://elpa.rzg.mpg.de/. The results are obtained with gfortran since for some reason it produces a slightly faster binary than ifort.

  • No labels