As "Open Internet Access by Molecular Dynamics Simulations", a couple of various codes are shown. This page is discussed on the parallel and vector-parallel SIESTA codes.
"Ab initio SIESTA simulation code" is implemented for electronic structure calculations and ab-initio molecular dynamics simulations of molecules and solids by the spanish authors, in https://departments.icmab.es/ (Ref. 1, 2). It is compiled by the gfortran compiler and for MPI parallel environments. It is also compiled by Intel's vector-and-parallel compiler, where the points of arch.make in CC, FC and LIBS are shown here in our PDF file. More things have to be modified and added due to vectorized complexity.
We can compile the siesta-4.1-b4 directory by gfortran for the parallel version with mpich, ScaLapack, and OpenBLAS. The file named "siesta-4.1-b4gcc.tar.gz" is unzipped, do "sh ../Src/obj_setup.sh", and copy "arch.make" of "arch.make-2" (1) or (2) script to your machine. The three packages including mpich above must be installed to your system bofore the "make" steps of SIESTA are executed.
The zipped files of mpi-4.0.2, scalapack-2.2.0, and OpenBLAS-0.3.21 (now in winter, 2022) are downloaded at the internet sites. If they are not yet installed in your system, unzip and "make", and "make install" separately for MPI, ScaLapack and OpenBLAS directories. To compile the MPICH, one may use the configure script: "./configure --prefix=/opt/mpich-4.0.2" and go to the "make" step. In the ScaLapack, "SLmake.inc" in that directory may be changed to your PC environments. In the OpenBLAS, "Makefile.rule" may be configured manually before the make step.
After the installation step, one should test which choice of MPI or OMP is most efficient in the gfortran run. It is very important that generic gfortran compiler must be used throughout the configure and make steps. The PGI fortran does not compile the SIESTA code properly.
For the Intel's vector and parallel compiler supplied by NEC machines, it uses the MPI and Scalapack packages. But, it needs to rewrite the code on more steps due to exact NEC7s fortran coding. The arch.make script is shown as (3) of "arch.make-2", and the modified changes are summarized in "Annual Report of Tanaka and Zempo (2022)" (PDF) of this page. All the changes and figures are shown in English, so you may not miss the points. The files are already corrected in "arch.make" and siesta-4.1-b4-LX.tar.gz. It is unzipped, does the shell script, and the "make" step is executed in ~/siesta-4.1-b4-LX/Obj (Ref. 3). The run of CH4 molecules is tested with the vector-parallel computer of 48 cores in the directory ~/siesta-4.1-b4-LX/Examples/C96H384-MD35 (figures in Ref. 3).
First, we download the SIESTA-4.1b code by internet. On our Linux, we do 'tar -xfzv siesta-4.1-b4.tar.gz'. Under the NEC's compiler, we invoke the 'module load intel-lx', and do 'sh ../Src/obj_setup.sh' for SIESTA Obj directory. The siesta-4.1b's MPICH+OMP script should be,
CC= mpiicc -O2 -qopenmp
FPP= $(FC) -E -P -x c
FC= mpiifort
MPI_INTERFACE = libmpi_f90.a
MPI_INCLUDE = .
FC_SERIAL= ifort
FFLAGS = -O2 -fPIC -qopenmp
LIBS = -L${MKLROOT}/lib/intel64 -lmkl_scalapack_lp64 -lmkl_intel_lp64 -lmkl_sequential -lmkl_core -lmkl_blacs_intelmpi_lp64 -mkl -qopenmp -lpthread -lm -ldl
Then, we proceed to the 'make' step. For NEC's own compiler problems, we must add additional terms. For the six files including "iokp.f", "m_mixing.F90", "m_ts_contour_neq.f90", "m_ts_electype.F90", "m_ts_weight.F90" and " ofc.f90", we change the correct statement 'e13.6' from 'e12.6'. Next point is that we must comment out the $OMP lines of "inal_H_f_stress.F" as,
!!$OMP parallel default(shared)
!!$OMP workshare
H_tmp = 0.0_dp
!!$OMP end workshare nowait
!!$OMP single
! Initialize forces and stress ...................
nullify(fal)
call re_alloc( fal, 1, 3, 1, na_u, 'fal', 'final_H_f_stress' )
!!$OMP end single
!!$OMP workshare
fa(1:3,1:na_u) = 0.0_dp
fal(1:3,1:na_u) = 0.0_dp
stress(1:3,1:3) = 0.0_dp
stressl(1:3,1:3) = 0.0_dp
!!$OMP end workshare nowait
!!$OMP end parallel
The vector lines must be changed as "novector" in the "old_atmfuncs.f" file:
!NEC$ novector
do 5 izeta=1,nzetasave(l,nsm,is)
norb=norb+(2*l+1)
indx=indx+1
if(norb.ge.io) goto 30
5 continue
They are the lines at 426, 436, 492, 502, 523, 570, 580, 605, 666, 712, 724 and 756 of "old_atmfuncs.f" file. The file "normalize_dm.F90" undergoes an error, thus we just skips as '! call die(msg)' at the line 95. We compile the rest of the code.
Finally for execution, we may write:
#PBS -T intmpi
#PBS -v NQSV_MPI_VER= 2020update0
module load intel-lx/$NQSV_MPI_VER
MPI and ScaLapack by gfortran; configure, make, and make install (packages are in winter 2022). Not compatible with the PGI fortran.
mpich-4.0.2: ./configure --prefix=/opt/mpich-4.0.2 2>&1 | tee conf.txt
OpenBLAS-0.3.21: in Makefile.rule, VERSION= 0.3.21 LIBNAMESUFFIX= omp ...
ScaLapack-2.2.0: in SLmake.inc, CDEFS= -DAdd_ FC= mpifort CC= mpicc ...
For an example:
#!/bin/bash; mpiexec -n 6 ~/siesta-4.1-b4gcc/Obj/siesta c12h48.fdf > c12h48.out &; exit 0
- J. M. Soler et al., J. Phys. Cond. Matt. 14, 2745 (2002).
- A. García et al., Chem. Phys. 152, 204108 (2020).
- M. Tanaka and Y. Zempo, Annual Report of National Institute for Fusion Science, R03/275 (2021) *).