Next: Automatic force tester Up: User's manual of OpenMX Previous: Interface for developers Contents Index
OpenMX can be utilized as library or computational engine from your program using MPI_Comm_spawn. You may mpirun your program and may want to call OpenMX with different input files from the program in different MPI groups at the same time. In such cases the functionality may be useful. Let us illustrate the functionality by introducing 'example_mpi_spawn.c' which is stored in the directory 'source'. The code 'example_mpi_spawn.c' can be compiled as
% make example_mpi_spawnThen, you may obtain the executable file of 'example_mpi_spawn' in the directory 'work'. Please move to the directory 'work' and perform as
% mpirun -np 24 -machinefile machine0 --oversubscribe ./example_mpi_spawnSetting the number of MPI processes, the machinefile, and the parameters may be changed depending on your computational environment. In some enviroment, you may need to attach '-oversubscribe' as option for mpirun in order to access the full physical cores on your machine. In the test calculation, the MPI processes of 24 are divided to three groups, each of which consists of 8 MPI processes, and in each group having 8 MPI processes OpenMX runs with an input file: 'Methane.dat', 'C60.dat', or 'Fe2.dat'. Starting from the example code as shown below, you may be able to develop a host program which calls OpenMX as child process. In the code 'Make_Comm_Worlds' creates three MPI communicaition groups, and 'MPI_Comm_spawn' calls OpenMX with one of the input files as argument. Note that OpenMX called by 'MPI_Comm_spawn' will not display most of standard output but write them in 'System.Name.std' to avoid appearence of too much information as standard output in such a case that OpenMX is multiply called by 'MPI_Comm_spawn' at the same time.
/**********************************************************************
example_mpi_spawn.c:
An example of calling OpenMX as library by MPI_Comm_spawn, where
a given MPI processes are grouped to three new MPI communication
groups and OpenMX runs with an input file: 'Methane.dat', 'C60.dat',
or 'Fe2.dat' in each MPI group.
Log of example_mpi_spawn.c:
25/Sep./2019 Released by Taisuke Ozaki
***********************************************************************/
#include "mpi.h"
#include <stdio.h>
#include <stdlib.h>
void Make_Comm_Worlds(
MPI_Comm MPI_Curret_Comm_WD,
int myid0,
int numprocs0,
int Num_Comm_World,
int *myworld1,
MPI_Comm *MPI_CommWD, /* size: Num_Comm_World */
int *NPROCS1_ID, /* size: numprocs0 */
int *Comm_World1, /* size: numprocs0 */
int *NPROCS1_WD, /* size: Num_Comm_World */
int *Comm_World_StartID /* size: Num_Comm_World */
);
int main(int argc, char *argv[])
{
int i,j;
int numprocs0,myid0,ID0;
int numprocs1,myid1;
int num;
int Num_Comm_World1;
int myworld1;
int *NPROCS1_ID,*NPROCS1_WD;
int *Comm_World1;
int *Comm_World_StartID;
MPI_Comm *MPI_CommWD;
MPI_Comm comm;
MPI_Comm *intercomm;
MPI_Init(&argc,&argv);
MPI_Comm_size(MPI_COMM_WORLD,&numprocs0);
MPI_Comm_rank(MPI_COMM_WORLD,&myid0);
/* set Num_Comm_World1 */
Num_Comm_World1 = 3;
/* allocation of arrays */
NPROCS1_ID = (int*)malloc(sizeof(int)*numprocs0);
Comm_World1 = (int*)malloc(sizeof(int)*numprocs0);
NPROCS1_WD = (int*)malloc(sizeof(int)*Num_Comm_World1);
Comm_World_StartID = (int*)malloc(sizeof(int)*Num_Comm_World1);
MPI_CommWD = (MPI_Comm*)malloc(sizeof(MPI_Comm)*Num_Comm_World1);
intercomm = (MPI_Comm*)malloc(sizeof(MPI_Comm)*Num_Comm_World1);
/* Make_Comm_Worlds */
Make_Comm_Worlds(MPI_COMM_WORLD, myid0, numprocs0, Num_Comm_World1,
&myworld1, MPI_CommWD,
NPROCS1_ID, Comm_World1, NPROCS1_WD, Comm_World_StartID);
/* get numprocs1 and myid1 */
MPI_Comm_size(MPI_CommWD[myworld1],&numprocs1);
MPI_Comm_rank(MPI_CommWD[myworld1],&myid1);
/*
printf("numprocs0=%2d myid0=%2d myworld1=%2d numprocs1=%2d myid1=%2d\n",
numprocs0,myid0,myworld1,numprocs1,myid1);
*/
/* MPI_Comm_spawn */
char command[] = "./openmx";
char **argvin;
char *inputfiles[] = { "Methane.dat", "C60.dat", "Fe2.dat" };
argvin=(char **)malloc(2 * sizeof(char *));
argvin[0] = inputfiles[myworld1];
argvin[1] = NULL;
MPI_Comm_spawn( command, argvin, numprocs1, MPI_INFO_NULL, 0,
MPI_CommWD[myworld1], &intercomm[myworld1], MPI_ERRCODES_IGNORE );
/* MPI_Barrier */
MPI_Barrier(MPI_COMM_WORLD);
/* freeing of arrays */
free(NPROCS1_ID);
free(Comm_World1);
free(NPROCS1_WD);
free(Comm_World_StartID);
free(MPI_CommWD);
free(intercomm);
/* MPI_Finalize() */
fflush(stdout);
MPI_Finalize();
return 0;
}
void Make_Comm_Worlds(
MPI_Comm MPI_Curret_Comm_WD,
int myid0,
int numprocs0,
int Num_Comm_World,
int *myworld1,
MPI_Comm *MPI_CommWD, /* size: Num_Comm_World */
int *NPROCS1_ID, /* size: numprocs0 */
int *Comm_World1, /* size: numprocs0 */
int *NPROCS1_WD, /* size: Num_Comm_World */
int *Comm_World_StartID /* size: Num_Comm_World */
)
{
int i,j,is,ie;
int ID0;
int numprocs1,myid1;
int num;
double avnum;
int *new_ranks;
MPI_Group new_group,old_group;
.....
}