int main( int argc, char *argv[] )
{
int errs = 0;
int wrank, wsize, mrank, msize, inter_rank;
int np = 2;
int errcodes[2];
int rrank = -1;
MPI_Comm parentcomm, intercomm, intercomm2, even_odd_comm, merged_world;
int can_spawn;
MTest_Init( &argc, &argv );
errs += MTestSpawnPossible(&can_spawn);
if (can_spawn) {
MPI_Comm_rank( MPI_COMM_WORLD, &wrank );
MPI_Comm_size( MPI_COMM_WORLD, &wsize );
if (wsize != 2) {
printf( "world size != 2, this test will not work correctly\n" );
errs++;
}
MPI_Comm_get_parent( &parentcomm );
if (parentcomm == MPI_COMM_NULL) {
MPI_Comm_spawn( (char*)"./spaiccreate2", MPI_ARGV_NULL, np,
MPI_INFO_NULL, 0, MPI_COMM_WORLD,
&intercomm, errcodes );
}
else {
intercomm = parentcomm;
}
MPI_Intercomm_merge( intercomm, (parentcomm == MPI_COMM_NULL ? 0 : 1), &merged_world );
MPI_Comm_rank( merged_world, &mrank );
MPI_Comm_size( merged_world, &msize );
MPI_Comm_split( merged_world, mrank % 2, wrank, &even_odd_comm );
MPI_Intercomm_create( even_odd_comm, 0, merged_world, (mrank + 1) % 2, 123, &intercomm2 );
MPI_Comm_rank( intercomm2, &inter_rank );
/* odds receive from evens */
MPI_Sendrecv( &inter_rank, 1, MPI_INT, inter_rank, 456,
&rrank, 1, MPI_INT, inter_rank, 456, intercomm2, MPI_STATUS_IGNORE );
if (rrank != inter_rank) {
printf( "Received %d from %d; expected %d\n",
rrank, inter_rank, inter_rank );
errs++;
}
MPI_Barrier( intercomm2 );
MPI_Comm_free( &intercomm );
MPI_Comm_free( &intercomm2 );
MPI_Comm_free( &merged_world );
MPI_Comm_free( &even_odd_comm );
/* Note that the MTest_Finalize get errs only over COMM_WORLD */
/* Note also that both the parent and child will generate "No Errors"
if both call MTest_Finalize */
if (parentcomm == MPI_COMM_NULL) {
MTest_Finalize( errs );
}
} else {
MTest_Finalize( errs );
}
MPI_Finalize();
return 0;
}
int main(int argc, char *argv[])
{
char str[10];
MPI_Comm intercomm1, intracomm, intercomm2;
int err, errcodes[256], rank;
MPI_Init(&argc, &argv);
/* printf("Child out of MPI_Init\n");
fflush(stdout);
*/
MPI_Comm_rank(MPI_COMM_WORLD, &rank);
MPI_Comm_get_parent(&intercomm1);
MPI_Intercomm_merge(intercomm1, 1, &intracomm);
err = MPI_Comm_spawn("spawn_merge_child2", MPI_ARGV_NULL, 2,
MPI_INFO_NULL, 2, intracomm, &intercomm2, errcodes);
if (err)
printf("Error in MPI_Comm_spawn\n");
MPI_Comm_rank(intercomm2, &rank);
if (rank == 3) {
err = MPI_Recv(str, 3, MPI_CHAR, 1, 0, intercomm2, MPI_STATUS_IGNORE);
printf("Parent (first child) received from child 2: %s\n", str);
fflush(stdout);
err = MPI_Send("bye", 4, MPI_CHAR, 1, 0, intercomm2);
}
MPI_Finalize();
return 0;
}
int main(int argc, char **argv)
{
int iter, err, rank, size;
MPI_Comm comm, merged;
/* MPI environnement */
printf("parent*******************************\n");
printf("parent: Launching MPI*\n");
MPI_Init( &argc, &argv);
for (iter = 0; iter < 1000; ++iter) {
MPI_Comm_spawn(EXE_TEST, NULL, 1, MPI_INFO_NULL,
0, MPI_COMM_WORLD, &comm, &err);
printf("parent: MPI_Comm_spawn #%d return : %d\n", iter, err);
MPI_Intercomm_merge(comm, 0, &merged);
MPI_Comm_rank(merged, &rank);
MPI_Comm_size(merged, &size);
printf("parent: MPI_Comm_spawn #%d rank %d, size %d\n",
iter, rank, size);
// sleep(2);
MPI_Comm_free(&merged);
}
MPI_Finalize();
printf("parent: End .\n" );
return 0;
}
static int
spawn_and_merge( char* argv[], char* arg, int count,
MPI_Comm* inter, MPI_Comm* intra )
{
int *errcode, err, i;
char *spawn_argv[2];
errcode = malloc(sizeof(int) * count);
if (errcode == NULL)
ompitest_error(__FILE__, __LINE__, "Doh! Rank %d was not able to allocate enough memory. MPI test aborted!\n", 0);
memset(errcode, -1, count);
/*MPI_Comm_set_errhandler(MPI_COMM_WORLD, MPI_ERRORS_RETURN);*/
spawn_argv[0] = arg;
spawn_argv[1] = NULL;
err = MPI_Comm_spawn(argv[0], spawn_argv, count, MPI_INFO_NULL, 0,
MPI_COMM_WORLD, inter, errcode);
for (i = 0; i < count; i++)
if (errcode[i] != MPI_SUCCESS)
ompitest_error(__FILE__, __LINE__,
"ERROR: MPI_Comm_spawn returned errcode[%d] = %d\n",
i, errcode[i]);
if (err != MPI_SUCCESS)
ompitest_error(__FILE__, __LINE__,
"ERROR: MPI_Comm_spawn returned errcode = %d\n", err);
err = MPI_Intercomm_merge( *inter, 0, intra );
free(errcode);
return err;
}
int main(int argc, char *argv[])
{
int rank;
MPI_Init(&argc, &argv);
MPI_Comm_rank(MPI_COMM_WORLD,&rank);
MPI_Comm parentcomm, intercomm;
/* If child, finalize */
MPI_Comm_get_parent(&parentcomm);
if (parentcomm != MPI_COMM_NULL)
goto out;
/* Set add-host info */
MPI_Info info;
MPI_Info_create(&info);
MPI_Info_set(info,"add-host","grsacc18");
/* Spawn the children */
printf("all info set, ready to spawn\n");
MPI_Comm_spawn("/home/grsprabh/resmgmt/branches/Malleability/workdir/addhosttest",MPI_ARGV_NULL,1,info,0,MPI_COMM_WORLD,&intercomm,MPI_ERRCODES_IGNORE);
printf("spawn returned successfully\n");
out:
printf("spawn completed successfully\n");
MPI_Finalize();
}
/*@C
PetscHMPISpawn - Initialize additional processes to be used as "worker" processes. This is not generally
called by users. One should use -hmpi_spawn_size <n> to indicate that you wish to have n-1 new MPI
processes spawned for each current process.
Not Collective (could make collective on MPI_COMM_WORLD, generate one huge comm and then split it up)
Input Parameter:
. nodesize - size of each compute node that will share processors
Options Database:
. -hmpi_spawn_size nodesize
Notes: This is only supported on systems with an MPI 2 implementation that includes the MPI_Comm_Spawn() routine.
$ Comparison of two approaches for HMPI usage (MPI started with N processes)
$
$ -hmpi_spawn_size <n> requires MPI 2, results in n*N total processes with N directly used by application code
$ and n-1 worker processes (used by PETSc) for each application node.
$ You MUST launch MPI so that only ONE MPI process is created for each hardware node.
$
$ -hmpi_merge_size <n> results in N total processes, N/n used by the application code and the rest worker processes
$ (used by PETSc)
$ You MUST launch MPI so that n MPI processes are created for each hardware node.
$
$ petscmpiexec -n 2 ./ex1 -hmpi_spawn_size 3 gives 2 application nodes (and 4 PETSc worker nodes)
$ petscmpiexec -n 6 ./ex1 -hmpi_merge_size 3 gives the SAME 2 application nodes and 4 PETSc worker nodes
$ This is what would use if each of the computers hardware nodes had 3 CPUs.
$
$ These are intended to be used in conjunction with USER HMPI code. The user will have 1 process per
$ computer (hardware) node (where the computer node has p cpus), the user's code will use threads to fully
$ utilize all the CPUs on the node. The PETSc code will have p processes to fully use the compute node for
$ PETSc calculations. The user THREADS and PETSc PROCESSES will NEVER run at the same time so the p CPUs
$ are always working on p task, never more than p.
$
$ See PCHMPI for a PETSc preconditioner that can use this functionality
$
For both PetscHMPISpawn() and PetscHMPIMerge() PETSC_COMM_WORLD consists of one process per "node", PETSC_COMM_LOCAL_WORLD
consists of all the processes in a "node."
In both cases the user's code is running ONLY on PETSC_COMM_WORLD (that was newly generated by running this command).
Level: developer
Concepts: HMPI
.seealso: PetscFinalize(), PetscInitializeFortran(), PetscGetArgs(), PetscHMPIFinalize(), PetscInitialize(), PetscHMPIMerge(), PetscHMPIRun()
@*/
PetscErrorCode PetscHMPISpawn(PetscMPIInt nodesize)
{
PetscErrorCode ierr;
PetscMPIInt size;
MPI_Comm parent,children;
PetscFunctionBegin;
ierr = MPI_Comm_get_parent(&parent);CHKERRQ(ierr);
if (parent == MPI_COMM_NULL) { /* the original processes started by user */
char programname[PETSC_MAX_PATH_LEN];
char **argv;
ierr = PetscGetProgramName(programname,PETSC_MAX_PATH_LEN);CHKERRQ(ierr);
ierr = PetscGetArguments(&argv);CHKERRQ(ierr);
ierr = MPI_Comm_spawn(programname,argv,nodesize-1,MPI_INFO_NULL,0,PETSC_COMM_SELF,&children,MPI_ERRCODES_IGNORE);CHKERRQ(ierr);
ierr = PetscFreeArguments(argv);CHKERRQ(ierr);
ierr = MPI_Intercomm_merge(children,0,&PETSC_COMM_LOCAL_WORLD);CHKERRQ(ierr);
ierr = MPI_Comm_size(PETSC_COMM_WORLD,&size);CHKERRQ(ierr);
ierr = PetscInfo2(0,"PETSc HMPI successfully spawned: number of nodes = %d node size = %d\n",size,nodesize);CHKERRQ(ierr);
saved_PETSC_COMM_WORLD = PETSC_COMM_WORLD;
} else { /* worker nodes that get spawned */
ierr = MPI_Intercomm_merge(parent,1,&PETSC_COMM_LOCAL_WORLD);CHKERRQ(ierr);
ierr = PetscHMPIHandle(PETSC_COMM_LOCAL_WORLD);CHKERRQ(ierr);
PetscHMPIWorker = PETSC_TRUE; /* so that PetscHMPIFinalize() will not attempt a broadcast from this process */
PetscEnd(); /* cannot continue into user code */
}
PetscFunctionReturn(0);
}
int main (int argc, const char *argv[])
{
double pi,
avepi = 0.;
int taskid,
size,
rc;
MPI_Comm everyone;
MPI_Status status;
MPI_Init(&argc, (char***)&argv);
MPI_Comm_size(MPI_COMM_WORLD, &size);
MPI_Comm_rank(MPI_COMM_WORLD, &taskid);
MPI_Comm_spawn("COMP428-A1-Parallel-Spawn-Slave", MPI_ARGV_NULL, MAXPROCS, MPI_INFO_NULL, 0, MPI_COMM_SELF, &everyone, MPI_ERRCODES_IGNORE);
int i;
for (i = 0; i < MAXPROCS; i++) {
MPI_Recv(&pi, 1, MPI_DOUBLE, MPI_ANY_SOURCE, MPI_ANY_TAG, everyone, &status);
printf("Process Master: received %.10f by %d\n", pi, status.MPI_SOURCE);
}
MPI_Finalize();
return (EXIT_SUCCESS);
}
int
main (int argc, char **argv)
{
int rank;
int size;
int *error_codes;
int spawn_counter = 0;
char *slave_argv[] = { "arg1", "arg2", 0 };
MPI_Comm spawn;
MPI_Init(&argc, &argv);
MPI_Comm_rank(MPI_COMM_WORLD, &rank);
MPI_Comm_size(MPI_COMM_WORLD, &size);
if (rank == 0)
{
printf("[master] running on %i processors\n", size);
while (1)
{
printf("[master] (%i) forking processes\n", spawn_counter++);
error_codes = (int*) malloc(sizeof(int)*size);
MPI_Comm_spawn("./slave", slave_argv, size, MPI_INFO_NULL, 0, MPI_COMM_SELF, &spawn, error_codes);
printf("[master] waiting at barrier\n");
MPI_Barrier(spawn);
free(error_codes);
}
}
MPI_Finalize();
}
int main( int argc, char *argv[] )
{
char str[10];
int err=0, errcodes[256], rank, nprocs;
MPI_Comm intercomm;
MPI_Init(&argc, &argv);
MPI_Comm_size(MPI_COMM_WORLD,&nprocs);
if (nprocs != 4) {
printf("Run this program with 4 processes\n");
MPI_Abort(MPI_COMM_WORLD,1);
}
err = MPI_Comm_spawn("child", MPI_ARGV_NULL, 4,
MPI_INFO_NULL, 1, MPI_COMM_WORLD,
&intercomm, errcodes);
if (err) printf("Error in MPI_Comm_spawn\n");
MPI_Comm_rank(MPI_COMM_WORLD, &rank);
if (rank == 3) {
err = MPI_Recv(str, 3, MPI_CHAR, 3, 0, intercomm, MPI_STATUS_IGNORE);
printf("Parent received from child: %s\n", str);
fflush(stdout);
err = MPI_Send("bye", 4, MPI_CHAR, 3, 0, intercomm);
}
MPI_Finalize();
return 0;
}
void mpi_comm_spawn_f(char *command, char *argv, MPI_Fint *maxprocs,
MPI_Fint *info, MPI_Fint *root, MPI_Fint *comm,
MPI_Fint *intercomm, MPI_Fint *array_of_errcodes,
MPI_Fint *ierr, int cmd_len, int string_len)
{
MPI_Comm c_comm, c_new_comm;
MPI_Info c_info;
int size;
int *c_errs;
char **c_argv;
char *c_command;
OMPI_ARRAY_NAME_DECL(array_of_errcodes);
c_comm = MPI_Comm_f2c(*comm);
c_info = MPI_Info_f2c(*info);
MPI_Comm_size(c_comm, &size);
ompi_fortran_string_f2c(command, cmd_len, &c_command);
/* It's allowed to ignore the errcodes */
if (OMPI_IS_FORTRAN_ERRCODES_IGNORE(array_of_errcodes)) {
c_errs = MPI_ERRCODES_IGNORE;
} else {
OMPI_ARRAY_FINT_2_INT_ALLOC(array_of_errcodes, size);
c_errs = OMPI_ARRAY_NAME_CONVERT(array_of_errcodes);
}
/* It's allowed to have no argv */
if (OMPI_IS_FORTRAN_ARGV_NULL(argv)) {
c_argv = MPI_ARGV_NULL;
} else {
ompi_fortran_argv_f2c(argv, string_len, string_len, &c_argv);
}
*ierr = OMPI_INT_2_FINT(MPI_Comm_spawn(c_command, c_argv,
OMPI_FINT_2_INT(*maxprocs),
c_info,
OMPI_FINT_2_INT(*root),
c_comm, &c_new_comm, c_errs));
if (MPI_SUCCESS == OMPI_FINT_2_INT(*ierr)) {
*intercomm = MPI_Comm_c2f(c_new_comm);
}
free(c_command);
if (MPI_ARGV_NULL != c_argv && NULL != c_argv) {
opal_argv_free(c_argv);
}
if (!OMPI_IS_FORTRAN_ERRCODES_IGNORE(array_of_errcodes)) {
OMPI_ARRAY_INT_2_FINT(array_of_errcodes, size);
} else {
OMPI_ARRAY_FINT_2_INT_CLEANUP(array_of_errcodes);
}
}
/**
* Crea una nueva instancia de proceo de RNA de RP con los argumentos proporcionados usando mpi
*
* Parametros:
* const char *comandoRNA - Comando para la ejecucion de la RNA
* int cant_neu1 - Cantidad de neuronas en la capa oculta 1
* int cant_neu2 - Cantidad de neuronas en la capa oculta 2
* float const_apr - Valor de la constante de aprendizaje
* float raz_mom - Valor de la razon de momentum
* int max_iter - Maxima cantidad de iteraciones para el entrenamiento
* int tipo_fun_o - Tipo de funcion de activacion a usar por las neuronas en las capas ocultas
* int tipo_fun_s - Tipo de funcion de activacion a usar por las neuronas en la capa de salida
* int cant_rep - Cantidad de repeticiones de entrenamiento e interrogatorio
* const char *ruta - Ruta al directorio donde se encuentra el archivo con los patrones
* const char *nombre - Nombre del archivo con los patrones (que se asume tiene sufijo .dat)
* int pos_hijo - Posicion asignada a este proceso en el vector de comunicadores (se asume es unico)
*
* Salida
* int - Entero con el valor 1 si pudo crearse la instancia, 0 en caso contrario
*/
int instancia_rna_mpi(char *comandoRNA, int cant_neu1, int cant_neu2, float const_apr, float raz_mom, int max_iter,
int tipo_fun_o, int tipo_fun_s, int cant_rep, const char *ruta, const char *nombre, int pos_hijo){
//Variables
int result = 0; //Indica si la instancia fue creada exitosamente
//Se verifican que los datos se hayan preparado previamente
if (preparado_ && esquema_ == MPI){
#ifdef HAVE_MPI
int ii; //Contador
char **params; //Cadena de caracteres con los parametros
int tam_int = 0; //Cantidad maxima de bits en un int
int num = INT_MAX; //Valor auxiliar utilizado para determinar la cantidad maxima de bits
//Cantidad maxima de bits de un entero
do{
tam_int++;
num /= 10;
}while (num > 0);
//Conversion del valor de los parametros a cadenas de caracteres para la ejecucion del proceso
params = calloc(15, sizeof(char *));
params[0] = (char *) calloc(20, sizeof(char)); sprintf(params[0], "--debug=n");
params[1] = (char *) calloc(20, sizeof(char)); sprintf(params[1], "--mensajes=n");
params[2] = (char *) calloc(tam_int, sizeof(char)); sprintf(params[2], "1");
params[3] = (char *) calloc(tam_int, sizeof(char)); sprintf(params[3], "%d", cant_neu1);
params[4] = (char *) calloc(tam_int, sizeof(char)); sprintf(params[4], "%d", cant_neu2);
params[5] = (char *) calloc(tam_int + 8, sizeof(char)); sprintf(params[5], "%.5f", const_apr);
params[6] = (char *) calloc(tam_int + 8, sizeof(char)); sprintf(params[6], "%.3f", raz_mom);
params[7] = (char *) calloc(tam_int, sizeof(char)); sprintf(params[7], "%d", max_iter);
params[8] = (char *) calloc(tam_int, sizeof(char)); sprintf(params[8], "%d", tipo_fun_o);
params[9] = (char *) calloc(tam_int, sizeof(char)); sprintf(params[9], "%d", tipo_fun_s);
params[10] = (char *) calloc(tam_int, sizeof(char)); sprintf(params[10], "%d", cant_rep);
params[11] = (char *) calloc(strlen(ruta) + 5, sizeof(char)); sprintf(params[11], "%s", ruta);
params[12] = (char *) calloc(strlen(nombre) + 5, sizeof(char)); sprintf(params[12], "%s", nombre);
params[13] = (char *) calloc(tam_int, sizeof(char)); sprintf(params[13], "3");
params[14] = NULL;
//Se crea de forma dinamica un nuevo proceso que ejecutara la RNA con los parametros dados
if (MPI_Comm_spawn(comandoRNA, params, 1, MPI_INFO_NULL, 0, MPI_COMM_SELF, &(datos_mpi.vec_hijos[pos_hijo]),
MPI_ERRCODES_IGNORE) == MPI_SUCCESS){
result = 1;
}
//Liberacion de memoria del vector de parametros
for (ii = 0; ii < 13; ii++)
free(params[ii]);
free(params);
#endif
}
//Resultado
return result;
}
int main(int argc, char* argv[])
{
int msg, rc;
MPI_Comm parent, child;
int rank, size;
char hostname[512];
pid_t pid;
pid = getpid();
printf("[pid %ld] starting up!\n", (long)pid);
MPI_Init(NULL, NULL);
MPI_Comm_rank(MPI_COMM_WORLD, &rank);
printf("%d completed MPI_Init\n", rank);
MPI_Comm_size(MPI_COMM_WORLD, &size);
MPI_Comm_get_parent(&parent);
/* If we get COMM_NULL back, then we're the parent */
if (MPI_COMM_NULL == parent) {
pid = getpid();
printf("Parent [pid %ld] about to spawn!\n", (long)pid);
if (MPI_SUCCESS != (rc = MPI_Comm_spawn(argv[0], MPI_ARGV_NULL, 3, MPI_INFO_NULL,
0, MPI_COMM_WORLD, &child, MPI_ERRCODES_IGNORE))) {
printf("Child failed to spawn\n");
return rc;
}
printf("Parent done with spawn\n");
if (0 == rank) {
msg = 38;
printf("Parent sending message to child\n");
MPI_Send(&msg, 1, MPI_INT, 0, 1, child);
}
MPI_Comm_disconnect(&child);
printf("Parent disconnected\n");
}
/* Otherwise, we're the child */
else {
MPI_Comm_rank(MPI_COMM_WORLD, &rank);
MPI_Comm_size(MPI_COMM_WORLD, &size);
gethostname(hostname, 512);
pid = getpid();
printf("Hello from the child %d of %d on host %s pid %ld\n", rank, 3, hostname, (long)pid);
if (0 == rank) {
MPI_Recv(&msg, 1, MPI_INT, 0, 1, parent, MPI_STATUS_IGNORE);
printf("Child %d received msg: %d\n", rank, msg);
}
MPI_Comm_disconnect(&parent);
printf("Child %d disconnected\n", rank);
}
MPI_Finalize();
fprintf(stderr, "%d: exiting\n", pid);
return 0;
}
int main(int argc, char **argv) {
int rank, size, version, subversion, namelen, universe_size;
char processor_name[MPI_MAX_PROCESSOR_NAME], worker_program[100];
MPI_Comm esclavos_comm;
MPI_Init(&argc, &argv); /* starts MPI */
MPI_Comm_rank(MPI_COMM_WORLD, &rank); /* get current process id */
MPI_Comm_size(MPI_COMM_WORLD, &size); /* get number of processes */
MPI_Get_processor_name(processor_name, &namelen);
MPI_Get_version(&version, &subversion);
printf("[maestro] Iniciado proceso maestro %d de %d en %s ejecutando MPI %d.%d\n", rank, size, processor_name,
version,
subversion);
strcpy(worker_program, "./Debug/esclavo");
MPI_Comm_spawn(worker_program, MPI_ARGV_NULL,ESCLAVOS, MPI_INFO_NULL, 0, MPI_COMM_SELF, &esclavos_comm,
MPI_ERRCODES_IGNORE);
/* Cálculo del número de puntos de intervalo por esclavo */
int n_esclavo = (N +1) / ESCLAVOS;
printf("PUNTOS POR ESCLAVO: %d\n", n_esclavo);
/* Cálculo de vector de valores de la función */
double dx = (double)(B - A) / (double)N ;
double h = ((double)B - (double)A) / (2*(double)N);
printf("DIFERENCIAL DE X: %f\n", dx);
int i = 0;
double y[N+1], y_esclavo[n_esclavo]; // número de puntos = número de intervalos + 1
double x = (double) A;
for(i=0;i<N+1;i++){
y[i]= x * x;
x+=dx;
printf("VALOR DE F(X) EN PUNTO i %d: %f\n", i, y[i]);
}
MPI_Bcast(&n_esclavo, 1, MPI_INT, MPI_ROOT, esclavos_comm);
MPI_Scatter(y, n_esclavo, MPI_DOUBLE, y_esclavo, n_esclavo, MPI_DOUBLE, MPI_ROOT, esclavos_comm);
double suma;
MPI_Reduce(NULL, &suma, 1, MPI_DOUBLE, MPI_SUM, MPI_ROOT, esclavos_comm);
printf("SUMA REDUCIDA ES: %f\n", suma);
double integral = (double) dx/3 * ((double)A*(double)A + suma + (double) B * (double) B);
printf("RESULTADO DE LA INTEGRAL: %f\n", integral);
MPI_Comm_disconnect(&esclavos_comm);
MPI_Finalize();
return 0;
}
int main(int argc, char *argv[])
{
int rank, size;
int universe_size, *universe_sizep, flag;
int lsize, rsize;
int grank, gsize;
MPI_Comm everyone, global; /* intercommunicator */
char worker_program[100];
MPI_Init(&argc, &argv);
MPI_Comm_size(MPI_COMM_WORLD, &size);
MPI_Comm_rank(MPI_COMM_WORLD, &rank);
MPI_Comm_get_attr(MPI_COMM_WORLD, MPI_UNIVERSE_SIZE,
&universe_sizep, &flag);
if (!flag) {
universe_size = 8;
} else
universe_size = *universe_sizep;
if( rank == 0 ) {
printf("univ size = %d\n", universe_size);
}
sprintf(worker_program, "./slave");
MPI_Comm_spawn(worker_program, MPI_ARGV_NULL, 6,
MPI_INFO_NULL, 0, MPI_COMM_WORLD, &everyone,
MPI_ERRCODES_IGNORE);
MPI_Comm_size(everyone, &lsize);
MPI_Comm_remote_size(everyone, &rsize);
MPI_Intercomm_merge(everyone, 1, &global);
MPI_Comm_rank(global, &grank);
MPI_Comm_size(global, &gsize);
printf("parent %d: lsize=%d, rsize=%d, grank=%d, gsize=%d\n",
rank, lsize, rsize, grank, gsize);
MPI_Barrier(global);
printf("%d: after Barrier\n", grank);
MPI_Comm_free(&global);
MPI_Finalize();
return 0;
}
void LeaderNode::spawn_swing_nodes(MPI_Comm parent, MPI_Comm *child,
uint16_t count) {
MPI_ASSERT(parent != MPI_COMM_NULL);
MPI_ASSERT(child != NULL);
// Duplicate the leader comm so that the original doesn't get cluttered.
MPI_Comm_dup(parent, child);
// Spawn the swing node process group.
MPI_Comm_spawn("./swing_layer", MPI_ARGV_NULL, count, MPI_INFO_NULL, 0,
*child, child, MPI_ERRCODES_IGNORE);
// Add the new swing node set to the listing of swing nodes for easy cleanup
// leater.
swing_nodes.push_back(std::make_pair(*child, count));
// Update the number of swing nodes.
num_swing_nodes += count;
}
int main( int argc, char *argv[] ) {
int np = NUM_SPAWNS;
int my_rank, size;
int errcodes[NUM_SPAWNS];
MPI_Comm allcomm;
MPI_Comm intercomm;
MPI_Init( &argc, &argv );
MPI_Comm_rank(MPI_COMM_WORLD, &my_rank);
MPI_Comm_size(MPI_COMM_WORLD, &size);
MPI_Comm_spawn( (char*)"./spawntest_child", MPI_ARGV_NULL, np,
MPI_INFO_NULL, 0, MPI_COMM_WORLD, &intercomm, errcodes );
if ( intercomm == MPI_COMM_NULL ) {
fprintf(stdout, "intercomm is null\n");
}
MPI_Intercomm_merge(intercomm, 0, &allcomm);
MPI_Comm_rank(allcomm, &my_rank);
MPI_Comm_size(allcomm, &size);
/* Without the Free of allcomm, the children *must not exit* until the
master calls MPI_Finalize. */
MPI_Barrier( allcomm );
/* According to 10.5.4, case 1b in MPI2.2, the children and master are
still connected unless MPI_Comm_disconnect is used with allcomm.
MPI_Comm_free is not sufficient */
MPI_Comm_free( &allcomm );
MPI_Comm_disconnect( &intercomm );
fprintf(stdout, "%s:%d: Sleep starting; children should exit\n",
__FILE__, __LINE__ );fflush(stdout);
sleep(30);
fprintf(stdout,
"%s:%d: Sleep done; all children should have already exited\n",
__FILE__, __LINE__ );fflush(stdout);
MPI_Finalize();
return 0;
}
int
main(int argc, char* argv[])
{
MPI_Init(0, 0);
MPI_Comm_rank(MPI_COMM_WORLD, &rank);
MPI_Comm_size(MPI_COMM_WORLD, &size);
printf("MAIN: %i/%i\n", rank, size);
int maxprocs = 1;
int codes[maxprocs];
MPI_Comm comm;
MPI_Comm_spawn("./child.x", NULL, maxprocs, MPI_INFO_NULL, 0,
MPI_COMM_WORLD, &comm, codes);
MPI_Finalize();
return 0;
}
void mpif_comm_spawn_(char *command, char **argv, int *maxprocs,
MPI_Fint *info, int *root, MPI_Fint *comm,
MPI_Fint *intercomm, int *array_of_errcodes, int*error)
{
MPI_Comm c_comm = MPI_Comm_f2c(*comm);
MPI_Info c_info = MPI_Info_f2c(*info);
ALLOCATE_ITEMS(MPI_Comm, *maxprocs, c_intercomm, p_intercomm);
int i;
for(i=0; i<*maxprocs;i++)
p_intercomm[i] = MPI_Comm_f2c(intercomm[i]);
*error = MPI_Comm_spawn(command, argv, *maxprocs,
c_info, *root, c_comm,
p_intercomm, array_of_errcodes);
for(i=0; i<*maxprocs;i++)
intercomm[i] = MPI_Comm_c2f(p_intercomm[i]);
FREE_ITEMS(*maxprocs, p_intercomm);
}
Albany::MesoScaleLinkProblem::
MesoScaleLinkProblem(const Teuchos::RCP<Teuchos::ParameterList>& params_,
const Teuchos::RCP<ParamLib>& paramLib_,
const int numDim_,
Teuchos::RCP<const Teuchos::Comm<int> >& commT_):
Albany::AbstractProblem(params_, paramLib_, numDim_),
haveSource(false),
numDim(numDim_),
commT(commT_),
mpi_comm(Albany::getMpiCommFromTeuchosComm(commT_)) {
TEUCHOS_TEST_FOR_EXCEPTION(commT->getSize() != 1, std::logic_error,
"MesoScale bridge only supports 1 master processor currently:\n\tRun with \"mpirun -np 1 Albany\"");
std::string& method = params->get("Name", "MesoScaleLink ");
*out << "Problem Name = " << method << std::endl;
haveSource = params->isSublist("Source Functions");
matModel = params->sublist("Material Model").get("Model Name", "LinearMesoScaleLink");
TEUCHOS_TEST_FOR_EXCEPTION(matModel != "Bridge", std::logic_error,
"Must specify \"Bridge\" for the material model in the input file.");
exeName = params->sublist("Material Model").get("Executable", "zzz");
*out << "Establishing MPI bridging link to: " << exeName << std::endl;
numMesoPEs = params->sublist("Material Model").get("Num Meso PEs", 1);
interCommunicator = Teuchos::rcp(new MPI_Comm());
// Fire off the remote processes
MPI_Comm_spawn(&exeName[0], MPI_ARGV_NULL, numMesoPEs,
MPI_INFO_NULL, 0, mpi_comm, interCommunicator.get(), MPI_ERRCODES_IGNORE);
}
/**
* Main loop of the function.
*/
int main(int argc, char **argv) {
int rank, size, num_workers;
char worker_name[] = "./demo/worker";
MPI_Comm everyone;
MPI_Init(&argc, &argv);
MPI_Comm_rank(MPI_COMM_WORLD, &rank);
MPI_Comm_size(MPI_COMM_WORLD, &size);
num_workers = atoi(*++argv);
MPI_Comm_spawn(worker_name, MPI_ARGV_NULL, num_workers,
MPI_INFO_NULL, 0, MPI_COMM_SELF, &everyone,
MPI_ERRCODES_IGNORE);
sleep(3);
printf("Master woke up.\n");
MPI_Finalize();
return 0;
}
/* **** PA_Exec ****
* The ParallelAgent's equivelant of "main". This function spawns the
* children processes, sends them the data, and gets back the results.
*/
SEXP PA_Exec(SEXP scriptLocn, SEXP sxInputVector) {
int iFunction;
int iSpawnFlag = 1;
int iNumProcs;
int ipDims[10]= { 0,0,0,0,0,0,0,0,0,0 };
double *dpA = NULL;
double *dpB = NULL;
int iStatus;
int returnValue;
SEXP sRet;
#ifndef DONT_SPAWN_R
char *cpProgram = "R"; /* Program to call; Usually "R" */
char *child_args[] = {
"BATCH",
"--no-save",
CHAR(STRING_ELT((scriptLocn),0)),
"abc.out",
NULL };
#else
/* There are four ways to call the child processes (all go through
* MPI_COMM_SPAWN):
* 1. Spawn a copy of R and let it call the child process function through
* an R script. This approach simplified the MPI-BLACS interation,
* at the cost of major overhead.
* 2. Spawn a shell script (dfCRDriver) which runs the child process
* driver. Spawning the program directly is better, but there is a
* problem with needing the shared library in the library path.
* 3. Spawn a shell (sh) which first adds the path to the scalapack.so
* library to LD_LIBRARY_PATH and then calls the driver program.
* This eliminates the need for the extra script (dfCRDriver), while
* also being several thousandths of a second faster.
* 4. Spawn the driver program directly. This requires building the
* program differently so that it doesn't need to scalapack.so
* shared library. This is the preferred/current method.
*/
char *cpProgram; /* =R_PACKAGE_DIR"/exec/dfCRDriver.sh"; */
char *child_args[] = { NULL, NULL };
int iLength;
/* The scriptLocn (script location) variable contains the path to the
* executable directory (followed by the script name). Extract the path,
* and use it for the executable's path.
*/
cpProgram = (char *) (CHAR(STRING_ELT((scriptLocn), 0)));
iLength = strrchr(cpProgram, '/') - cpProgram;
if (iLength < 0) {
Rprintf("Path to script is not complete. Unable to continue.\n");
return R_NilValue;
}
cpProgram = (char *) malloc(sizeof(char) * (iLength + 12));
if (cpProgram == NULL) {
Rprintf("Memory allocation (%d bytes) failed!\n", sizeof(char) *
(iLength + 12));
return R_NilValue;
}
*(cpProgram) = '\0';
strncat(cpProgram, CHAR(STRING_ELT((scriptLocn),0)), iLength);
strncat(cpProgram, "/CRDriver", 10);
D_Rprintf(("Child process: \"%s\" \"%s\" \"%s\"\n", cpProgram, child_args[0], child_args[1]));
#endif /* Endof If DONT_SPAWN_R is defined */
/* Begin by unpacking the input vector into all of the seperate variables
* that get their values from it */
if (PA_UnpackInput(sxInputVector, ipDims, &dpA, &dpB, &iNumProcs,
&iFunction, &iSpawnFlag) != 0) {
free(cpProgram);
return R_NilValue;
}
#if 0
// Guru
int ig;
for ( ig = 0; ig < 10; ig++ )
{
fprintf(stdout, "ipDims[%d] = %d\n", ig, ipDims[ig] );fflush(stdout);
}
#endif
/* Initialize MPI (if it is already initialized, it won't be
* initialized again). */
if (PA_Init() != 0){
Rprintf(" ERROR[1]: Failed while intializing MPI \n");
free(cpProgram);
return R_NilValue;
}
if (iSpawnFlag != 0 && iGlobalNumChildren != 0) {
Rprintf(" Error: Attempt to spawn a new grid without releasing the previous grid.\n");
return R_NilValue;
}
if(iSpawnFlag == 0 && iGlobalNumChildren == 0){
Rprintf(" Error: Process Grid not present and Spawn option is set FALSE \n");
return R_NilValue;
}
int *ipErrcodes = (int *)Calloc(iNumProcs, int);
/* Begin: Spawn the child processes */
if ( iSpawnFlag != 0 ) {
/* Begin: Spawn the child processes */
D_Rprintf(("PA: Preparing to spawn %d child processes.\n", iNumProcs)); fflush(stdout);
iStatus = MPI_Comm_spawn(cpProgram, child_args, iNumProcs, MPI_INFO_NULL,
0, MPI_COMM_WORLD, &childComm, ipErrcodes);
free(cpProgram);
if (iStatus != MPI_SUCCESS) {
Rprintf(" ERROR: Failed to spawn (%d) child processes.\n", iNumProcs);
return R_NilValue;
}
D_Rprintf(("SPAWNING SUCCESSFUL\n")); fflush(stdout);
/* End: Spawn the child processes */
iGlobalNumChildren = iNumProcs;
iNprows = ipDims[6];
iNpcols = ipDims[7];
}
/* SPECIAL for SVD */
/* If the function is SVD, the child process needs to know the nu,nv
* parameters.
*/
if (iFunction == 2) {
ipDims[2] = (int) dpB[0];
ipDims[3] = (int) dpB[1];
}
/* DATA DISTRIBUTION */
/* The data is distributed by the PA to all of the child processes. */
if ((returnValue = PA_SendData(ipDims, dpA, dpB)) == 0) {
D_Rprintf(("PA: DATA SENT TO CHILD PROCESSES.\n")); fflush(stdout);
} else { /* The send data failed, */
Rprintf("ERROR [1] : DATA COULD NOT BE SENT TO CHILD PROCESSES.\n");
iGlobalNumChildren = 0;
iNprows = 0;
iNpcols = 0;
return R_NilValue;
}
D_Rprintf(("PA: back to exec.\n")); fflush(stdout);
/* If the release flag == 1, then the grid will be (or was) released. */
if (ipDims[9] == 1)
{
iGlobalNumChildren = 0;
iNprows = 0;
iNpcols = 0;
}
/* If the function is sla.gridInit or sla.gridExit, just return. */
if (iFunction == 0)
{
// MPI_Comm_free(&intercomm);
return R_NilValue;
}
/* GET BACK THE RESULT */
sRet = PA_RecvResult(ipDims);
return sRet;
}
int main(int argc, char *argv[])
{
int error;
int rank, size;
char *argv1[2] = { (char*)"connector", NULL };
char *argv2[2] = { (char*)"acceptor", NULL };
MPI_Comm comm_connector, comm_acceptor, comm_parent, comm;
char port[MPI_MAX_PORT_NAME];
MPI_Status status;
MPI_Info spawn_path = MPI_INFO_NULL;
int verbose = 0;
if (getenv("MPITEST_VERBOSE"))
{
verbose = 1;
}
IF_VERBOSE(("init.\n"));
error = MPI_Init(&argc, &argv);
check_error(error, "MPI_Init");
/* To improve reporting of problems about operations, we
change the error handler to errors return */
MPI_Comm_set_errhandler( MPI_COMM_WORLD, MPI_ERRORS_RETURN );
MPI_Comm_set_errhandler( MPI_COMM_SELF, MPI_ERRORS_RETURN );
IF_VERBOSE(("size.\n"));
error = MPI_Comm_size(MPI_COMM_WORLD, &size);
check_error(error, "MPI_Comm_size");
IF_VERBOSE(("rank.\n"));
error = MPI_Comm_rank(MPI_COMM_WORLD, &rank);
check_error(error, "MPI_Comm_rank");
if (argc == 1)
{
/* Make sure that the current directory is in the path.
Not all implementations may honor or understand this, but
it is highly recommended as it gives users a clean way
to specify the location of the executable without
specifying a particular directory format (e.g., this
should work with both Windows and Unix implementations) */
error = MPI_Info_create( &spawn_path );
check_error( error, "MPI_Info_create" );
error = MPI_Info_set( spawn_path, (char*)"path", (char*)"." );
check_error( error, "MPI_Info_set" );
IF_VERBOSE(("spawn connector.\n"));
error = MPI_Comm_spawn((char*)"spaconacc", argv1, 1, spawn_path, 0,
MPI_COMM_SELF, &comm_connector,
MPI_ERRCODES_IGNORE);
check_error(error, "MPI_Comm_spawn");
IF_VERBOSE(("spawn acceptor.\n"));
error = MPI_Comm_spawn((char*)"spaconacc", argv2, 1, spawn_path, 0,
MPI_COMM_SELF, &comm_acceptor,
MPI_ERRCODES_IGNORE);
check_error(error, "MPI_Comm_spawn");
error = MPI_Info_free( &spawn_path );
check_error( error, "MPI_Info_free" );
MPI_Comm_set_errhandler( comm_connector, MPI_ERRORS_RETURN );
MPI_Comm_set_errhandler( comm_acceptor, MPI_ERRORS_RETURN );
IF_VERBOSE(("recv port.\n"));
error = MPI_Recv(port, MPI_MAX_PORT_NAME, MPI_CHAR, 0, 0,
comm_acceptor, &status);
check_error(error, "MPI_Recv");
IF_VERBOSE(("send port.\n"));
error = MPI_Send(port, MPI_MAX_PORT_NAME, MPI_CHAR, 0, 0,
comm_connector);
check_error(error, "MPI_Send");
IF_VERBOSE(("barrier acceptor.\n"));
error = MPI_Barrier(comm_acceptor);
check_error(error, "MPI_Barrier");
IF_VERBOSE(("barrier connector.\n"));
error = MPI_Barrier(comm_connector);
check_error(error, "MPI_Barrier");
error = MPI_Comm_free(&comm_acceptor);
check_error(error, "MPI_Comm_free");
error = MPI_Comm_free(&comm_connector);
check_error(error, "MPI_Comm_free");
printf(" No Errors\n");
}
else if ((argc == 2) && (strcmp(argv[1], "acceptor") == 0))
{
IF_VERBOSE(("get_parent.\n"));
error = MPI_Comm_get_parent(&comm_parent);
check_error(error, "MPI_Comm_get_parent");
if (comm_parent == MPI_COMM_NULL)
{
printf("acceptor's parent is NULL.\n");fflush(stdout);
MPI_Abort(MPI_COMM_WORLD, -1);
}
IF_VERBOSE(("open_port.\n"));
error = MPI_Open_port(MPI_INFO_NULL, port);
check_error(error, "MPI_Open_port");
MPI_Comm_set_errhandler( comm_parent, MPI_ERRORS_RETURN );
IF_VERBOSE(("0: opened port: <%s>\n", port));
IF_VERBOSE(("send.\n"));
error = MPI_Send(port, MPI_MAX_PORT_NAME, MPI_CHAR, 0, 0, comm_parent);
check_error(error, "MPI_Send");
IF_VERBOSE(("accept.\n"));
error = MPI_Comm_accept(port, MPI_INFO_NULL, 0, MPI_COMM_SELF, &comm);
check_error(error, "MPI_Comm_accept");
IF_VERBOSE(("close_port.\n"));
error = MPI_Close_port(port);
check_error(error, "MPI_Close_port");
IF_VERBOSE(("disconnect.\n"));
error = MPI_Comm_disconnect(&comm);
check_error(error, "MPI_Comm_disconnect");
IF_VERBOSE(("barrier.\n"));
error = MPI_Barrier(comm_parent);
check_error(error, "MPI_Barrier");
MPI_Comm_free( &comm_parent );
}
else if ((argc == 2) && (strcmp(argv[1], "connector") == 0))
{
IF_VERBOSE(("get_parent.\n"));
error = MPI_Comm_get_parent(&comm_parent);
check_error(error, "MPI_Comm_get_parent");
if (comm_parent == MPI_COMM_NULL)
{
printf("acceptor's parent is NULL.\n");fflush(stdout);
MPI_Abort(MPI_COMM_WORLD, -1);
}
MPI_Comm_set_errhandler( comm_parent, MPI_ERRORS_RETURN );
IF_VERBOSE(("recv.\n"));
error = MPI_Recv(port, MPI_MAX_PORT_NAME, MPI_CHAR, 0, 0,
comm_parent, &status);
check_error(error, "MPI_Recv");
IF_VERBOSE(("1: received port: <%s>\n", port));
IF_VERBOSE(("connect.\n"));
error = MPI_Comm_connect(port, MPI_INFO_NULL, 0, MPI_COMM_SELF, &comm);
check_error(error, "MPI_Comm_connect");
MPI_Comm_set_errhandler( comm, MPI_ERRORS_RETURN );
IF_VERBOSE(("disconnect.\n"));
error = MPI_Comm_disconnect(&comm);
check_error(error, "MPI_Comm_disconnect");
IF_VERBOSE(("barrier.\n"));
error = MPI_Barrier(comm_parent);
check_error(error, "MPI_Barrier");
MPI_Comm_free( &comm_parent );
}
else
{
printf("invalid command line.\n");fflush(stdout);
{
int i;
for (i=0; i<argc; i++)
{
printf("argv[%d] = <%s>\n", i, argv[i]);
}
}
fflush(stdout);
MPI_Abort(MPI_COMM_WORLD, -2);
}
MPI_Finalize();
return 0;
}
int MpiCommunicator::init( int minId, long thecomm_ )
{
VT_FUNC_I( "MpiCommunicator::init" );
assert( sizeof(thecomm_) >= sizeof(MPI_Comm) );
MPI_Comm thecomm = (MPI_Comm)thecomm_;
// turn wait mode on for intel mpi if possible
// this should greatly improve performance for intel mpi
PAL_SetEnvVar( "I_MPI_WAIT_MODE", "enable", 0);
int flag;
MPI_Initialized( &flag );
if ( ! flag ) {
int p;
//!! FIXME passing NULL ptr breaks mvapich1 mpi implementation
MPI_Init_thread( 0, NULL, MPI_THREAD_MULTIPLE, &p );
if( p != MPI_THREAD_MULTIPLE ) {
// can't use Speaker yet, need Channels to be inited
std::cerr << "[CnC] Warning: not MPI_THREAD_MULTIPLE (" << MPI_THREAD_MULTIPLE << "), but " << p << std::endl;
}
} else if( thecomm == 0 ) {
CNC_ABORT( "Process has already been initialized" );
}
MPI_Comm myComm = MPI_COMM_WORLD;
int rank;
MPI_Comm parentComm;
if( thecomm == 0 ) {
MPI_Comm_get_parent( &parentComm );
} else {
m_customComm = true;
m_exit0CallOk = false;
myComm = thecomm;
}
MPI_Comm_rank( myComm, &rank );
// father of all checks if he's requested to spawn processes:
if ( rank == 0 && parentComm == MPI_COMM_NULL ) {
// Ok, let's spawn the clients.
// I need some information for the startup.
// 1. Name of the executable (default is the current exe)
const char * _tmp = getenv( "CNC_MPI_SPAWN" );
if ( _tmp ) {
int nClientsToSpawn = atol( _tmp );
_tmp = getenv( "CNC_MPI_EXECUTABLE" );
std::string clientExe( _tmp ? _tmp : "" );
if( clientExe.empty() ) clientExe = PAL_GetProgname();
CNC_ASSERT( ! clientExe.empty() );
// 3. Special setting for MPI_Info: hosts
const char * clientHost = getenv( "CNC_MPI_HOSTS" );
// Prepare MPI_Info object:
MPI_Info clientInfo = MPI_INFO_NULL;
if ( clientHost ) {
MPI_Info_create( &clientInfo );
if ( clientHost ) {
MPI_Info_set( clientInfo, const_cast< char * >( "host" ), const_cast< char * >( clientHost ) );
// can't use Speaker yet, need Channels to be inited
std::cerr << "[CnC " << rank << "] Set MPI_Info_set( \"host\", \"" << clientHost << "\" )\n";
}
}
// Now spawn the client processes:
// can't use Speaker yet, need Channels to be inited
std::cerr << "[CnC " << rank << "] Spawning " << nClientsToSpawn << " MPI processes" << std::endl;
int* errCodes = new int[nClientsToSpawn];
MPI_Comm interComm;
int err = MPI_Comm_spawn( const_cast< char * >( clientExe.c_str() ),
MPI_ARGV_NULL, nClientsToSpawn,
clientInfo, 0, MPI_COMM_WORLD,
&interComm, errCodes );
delete [] errCodes;
if ( err ) {
// can't use Speaker yet, need Channels to be inited
std::cerr << "[CnC " << rank << "] Error in MPI_Comm_spawn. Skipping process spawning";
} else {
MPI_Intercomm_merge( interComm, 0, &myComm );
}
} // else {
// No process spawning
// MPI-1 situation: all clients to be started by mpiexec
// myComm = MPI_COMM_WORLD;
//}
}
if ( thecomm == 0 && parentComm != MPI_COMM_NULL ) {
// I am a child. Build intra-comm to the parent.
MPI_Intercomm_merge( parentComm, 1, &myComm );
}
MPI_Comm_rank( myComm, &rank );
CNC_ASSERT( m_channel == NULL );
MpiChannelInterface* myChannel = new MpiChannelInterface( use_crc(), myComm );
m_channel = myChannel;
int size;
MPI_Comm_size( myComm, &size );
// Are we on the host or on the remote side?
if ( rank == 0 ) {
if( size <= 1 ) {
Speaker oss( std::cerr ); oss << "Warning: no clients avabilable. Forgot to set CNC_MPI_SPAWN?";
}
// ==> HOST startup:
// This initializes the mpi environment in myChannel.
MpiHostInitializer hostInitializer( *myChannel );
hostInitializer.init_mpi_comm( myComm );
} else {
// ==> CLIENT startup:
// This initializes the mpi environment in myChannel.
MpiClientInitializer clientInitializer( *myChannel );
clientInitializer.init_mpi_comm( myComm );
}
{ Speaker oss( std::cerr ); oss << "MPI initialization complete (rank " << rank << ")."; }
// MPI_Barrier( myComm );
// Now the mpi specific setup is finished.
// Do the generic initialization stuff.
GenericCommunicator::init( minId );
return 0;
}
int main(int argc, char *argv[])
{
complex_t coord_point, julia_constant;
double x_max, x_min, y_max, y_min, x_resolution, y_resolution;
double divergent_limit;
char file_message[160];
char filename[100];
int icount, imax_iterations;
int ipixels_across, ipixels_down;
int i, j, k, julia, alternate_equation;
int imin, imax, jmin, jmax;
int work[5];
header_t *result = NULL;
/* make an integer array of size [N x M] to hold answers. */
int *in_grid_array, *out_grid_array = NULL;
int numprocs;
int namelen;
char processor_name[MPI_MAX_PROCESSOR_NAME];
int num_colors;
color_t *colors = NULL;
MPI_Status status;
int listener;
int save_image = 0;
int optval;
int num_children = DEFAULT_NUM_SLAVES;
int master = 1;
MPI_Comm parent, *child_comm = NULL;
MPI_Request *child_request = NULL;
int error_code, error;
char error_str[MPI_MAX_ERROR_STRING];
int length;
int index;
int pid;
MPI_Init(&argc, &argv);
MPI_Comm_size(MPI_COMM_WORLD, &numprocs);
MPI_Comm_rank(MPI_COMM_WORLD, &myid);
MPI_Comm_get_parent(&parent);
MPI_Get_processor_name(processor_name, &namelen);
pid = getpid();
if (parent == MPI_COMM_NULL)
{
if (numprocs > 1)
{
printf("Error: only one process allowed for the master.\n");
PrintUsage();
error_code = MPI_Abort(MPI_COMM_WORLD, -1);
exit(error_code);
}
printf("Welcome to the Mandelbrot/Julia set explorer.\n");
master = 1;
/* Get inputs-- region to view (must be within x/ymin to x/ymax, make sure
xmax>xmin and ymax>ymin) and resolution (number of pixels along an edge,
N x M, i.e. 256x256)
*/
read_mand_args(argc, argv, &imax_iterations, &ipixels_across, &ipixels_down,
&x_min, &x_max, &y_min, &y_max, &julia, &julia_constant.real,
&julia_constant.imaginary, &divergent_limit,
&alternate_equation, filename, &num_colors, &use_stdin, &save_image,
&num_children);
check_mand_params(&imax_iterations, &ipixels_across, &ipixels_down,
&x_min, &x_max, &y_min, &y_max, &divergent_limit, &num_children);
if (julia == 1) /* we're doing a julia figure */
check_julia_params(&julia_constant.real, &julia_constant.imaginary);
/* spawn slaves */
child_comm = (MPI_Comm*)malloc(num_children * sizeof(MPI_Comm));
child_request = (MPI_Request*)malloc(num_children * sizeof(MPI_Request));
result = (header_t*)malloc(num_children * sizeof(header_t));
if (child_comm == NULL || child_request == NULL || result == NULL)
{
printf("Error: unable to allocate an array of %d communicators, requests and work objects for the slaves.\n", num_children);
error_code = MPI_Abort(MPI_COMM_WORLD, -1);
exit(error_code);
}
printf("Spawning %d slaves.\n", num_children);
for (i=0; i<num_children; i++)
{
error = MPI_Comm_spawn(argv[0], MPI_ARGV_NULL, 1,
MPI_INFO_NULL, 0, MPI_COMM_WORLD, &child_comm[i], &error_code);
if (error != MPI_SUCCESS)
{
error_str[0] = '\0';
length = MPI_MAX_ERROR_STRING;
MPI_Error_string(error, error_str, &length);
printf("Error: MPI_Comm_spawn failed: %s\n", error_str);
error_code = MPI_Abort(MPI_COMM_WORLD, -1);
exit(error_code);
}
}
/* send out parameters */
for (i=0; i<num_children; i++)
{
MPI_Send(&num_colors, 1, MPI_INT, 0, 0, child_comm[i]);
MPI_Send(&imax_iterations, 1, MPI_INT, 0, 0, child_comm[i]);
MPI_Send(&ipixels_across, 1, MPI_INT, 0, 0, child_comm[i]);
MPI_Send(&ipixels_down, 1, MPI_INT, 0, 0, child_comm[i]);
MPI_Send(&divergent_limit, 1, MPI_DOUBLE, 0, 0, child_comm[i]);
MPI_Send(&julia, 1, MPI_INT, 0, 0, child_comm[i]);
MPI_Send(&julia_constant.real, 1, MPI_DOUBLE, 0, 0, child_comm[i]);
MPI_Send(&julia_constant.imaginary, 1, MPI_DOUBLE, 0, 0, child_comm[i]);
MPI_Send(&alternate_equation, 1, MPI_INT, 0, 0, child_comm[i]);
}
}
else
{
master = 0;
MPI_Recv(&num_colors, 1, MPI_INT, 0, 0, parent, MPI_STATUS_IGNORE);
MPI_Recv(&imax_iterations, 1, MPI_INT, 0, 0, parent, MPI_STATUS_IGNORE);
MPI_Recv(&ipixels_across, 1, MPI_INT, 0, 0, parent, MPI_STATUS_IGNORE);
MPI_Recv(&ipixels_down, 1, MPI_INT, 0, 0, parent, MPI_STATUS_IGNORE);
MPI_Recv(&divergent_limit, 1, MPI_DOUBLE, 0, 0, parent, MPI_STATUS_IGNORE);
MPI_Recv(&julia, 1, MPI_INT, 0, 0, parent, MPI_STATUS_IGNORE);
MPI_Recv(&julia_constant.real, 1, MPI_DOUBLE, 0, 0, parent, MPI_STATUS_IGNORE);
MPI_Recv(&julia_constant.imaginary, 1, MPI_DOUBLE, 0, 0, parent, MPI_STATUS_IGNORE);
MPI_Recv(&alternate_equation, 1, MPI_INT, 0, 0, parent, MPI_STATUS_IGNORE);
}
if (master)
{
colors = malloc((num_colors+1)* sizeof(color_t));
if (colors == NULL)
{
MPI_Abort(MPI_COMM_WORLD, -1);
exit(-1);
}
Make_color_array(num_colors, colors);
colors[num_colors] = 0; /* add one on the top to avoid edge errors */
}
/* allocate memory */
if ( (in_grid_array = (int *)calloc(ipixels_across * ipixels_down, sizeof(int))) == NULL)
{
printf("Memory allocation failed for data array, aborting.\n");
MPI_Abort(MPI_COMM_WORLD, -1);
exit(-1);
}
if (master)
{
int istep, jstep;
int i1[400], i2[400], j1[400], j2[400];
int ii, jj;
struct sockaddr_in addr;
int len;
char line[1024], *token;
if ( (out_grid_array = (int *)calloc(ipixels_across * ipixels_down, sizeof(int))) == NULL)
{
printf("Memory allocation failed for data array, aborting.\n");
MPI_Abort(MPI_COMM_WORLD, -1);
exit(-1);
}
srand(getpid());
if (!use_stdin)
{
addr.sin_family = AF_INET;
addr.sin_addr.s_addr = INADDR_ANY;
addr.sin_port = htons(DEFAULT_PORT);
listener = socket(AF_INET, SOCK_STREAM, 0);
if (listener == -1)
{
printf("unable to create a listener socket.\n");
MPI_Abort(MPI_COMM_WORLD, -1);
exit(-1);
}
if (bind(listener, &addr, sizeof(addr)) == -1)
{
addr.sin_port = 0;
if (bind(listener, &addr, sizeof(addr)) == -1)
{
printf("unable to create a listener socket.\n");
MPI_Abort(MPI_COMM_WORLD, -1);
exit(-1);
}
}
if (listen(listener, 1) == -1)
{
printf("unable to listen.\n");
MPI_Abort(MPI_COMM_WORLD, -1);
exit(-1);
}
len = sizeof(addr);
getsockname(listener, &addr, &len);
printf("%s listening on port %d\n", processor_name, ntohs(addr.sin_port));
fflush(stdout);
sock = accept(listener, NULL, NULL);
if (sock == -1)
{
printf("unable to accept a socket connection.\n");
MPI_Abort(MPI_COMM_WORLD, -1);
exit(-1);
}
printf("accepted connection from visualization program.\n");
fflush(stdout);
#ifdef HAVE_WINDOWS_H
optval = 1;
setsockopt(sock, IPPROTO_TCP, TCP_NODELAY, (char *)&optval, sizeof(optval));
#endif
printf("sending image size to visualizer.\n");
sock_write(sock, &ipixels_across, sizeof(int));
sock_write(sock, &ipixels_down, sizeof(int));
sock_write(sock, &num_colors, sizeof(int));
sock_write(sock, &imax_iterations, sizeof(int));
}
for (;;)
{
/* get x_min, x_max, y_min, and y_max */
if (use_stdin)
{
printf("input xmin ymin xmax ymax max_iter, (0 0 0 0 0 to quit):\n");fflush(stdout);
fgets(line, 1024, stdin);
printf("read <%s> from stdin\n", line);fflush(stdout);
token = strtok(line, " \n");
x_min = atof(token);
token = strtok(NULL, " \n");
y_min = atof(token);
token = strtok(NULL, " \n");
x_max = atof(token);
token = strtok(NULL, " \n");
y_max = atof(token);
token = strtok(NULL, " \n");
imax_iterations = atoi(token);
/*sscanf(line, "%g %g %g %g", &x_min, &y_min, &x_max, &y_max);*/
/*scanf("%g %g %g %g", &x_min, &y_min, &x_max, &y_max);*/
}
else
{
printf("reading xmin,ymin,xmax,ymax.\n");fflush(stdout);
sock_read(sock, &x_min, sizeof(double));
sock_read(sock, &y_min, sizeof(double));
sock_read(sock, &x_max, sizeof(double));
sock_read(sock, &y_max, sizeof(double));
sock_read(sock, &imax_iterations, sizeof(int));
}
printf("x0,y0 = (%f, %f) x1,y1 = (%f,%f) max_iter = %d\n", x_min, y_min, x_max, y_max, imax_iterations);fflush(stdout);
/*printf("sending the limits: (%f,%f)(%f,%f)\n", x_min, y_min, x_max, y_max);fflush(stdout);*/
/* let everyone know the limits */
for (i=0; i<num_children; i++)
{
MPI_Send(&x_min, 1, MPI_DOUBLE, 0, 0, child_comm[i]);
MPI_Send(&x_max, 1, MPI_DOUBLE, 0, 0, child_comm[i]);
MPI_Send(&y_min, 1, MPI_DOUBLE, 0, 0, child_comm[i]);
MPI_Send(&y_max, 1, MPI_DOUBLE, 0, 0, child_comm[i]);
MPI_Send(&imax_iterations, 1, MPI_INT, 0, 0, child_comm[i]);
}
/* check for the end condition */
if (x_min == x_max && y_min == y_max)
{
/*printf("root bailing.\n");fflush(stdout);*/
break;
}
/* break the work up into 400 pieces */
istep = ipixels_across / 20;
jstep = ipixels_down / 20;
if (istep < 1)
istep = 1;
if (jstep < 1)
jstep = 1;
k = 0;
for (i=0; i<20; i++)
{
for (j=0; j<20; j++)
{
i1[k] = MIN(istep * i, ipixels_across - 1);
i2[k] = MIN((istep * (i+1)) - 1, ipixels_across - 1);
j1[k] = MIN(jstep * j, ipixels_down - 1);
j2[k] = MIN((jstep * (j+1)) - 1, ipixels_down - 1);
k++;
}
}
/* shuffle the work */
for (i=0; i<500; i++)
{
ii = rand() % 400;
jj = rand() % 400;
swap(&i1[ii], &i1[jj]);
swap(&i2[ii], &i2[jj]);
swap(&j1[ii], &j1[jj]);
swap(&j2[ii], &j2[jj]);
}
/* send a piece of work to each worker (there must be more work than workers) */
k = 0;
for (i=0; i<num_children; i++)
{
work[0] = k+1;
work[1] = i1[k]; /* imin */
work[2] = i2[k]; /* imax */
work[3] = j1[k]; /* jmin */
work[4] = j2[k]; /* jmax */
/*printf("sending work(%d) to %d\n", k+1, i);fflush(stdout);*/
MPI_Send(work, 5, MPI_INT, 0, 100, child_comm[i]);
MPI_Irecv(&result[i], 5, MPI_INT, 0, 200, child_comm[i], &child_request[i]);
k++;
}
/* receive the results and hand out more work until the image is complete */
while (k<400)
{
MPI_Waitany(num_children, child_request, &index, &status);
memcpy(work, &result[index], 5 * sizeof(int));
/*printf("master receiving data in k<400 loop.\n");fflush(stdout);*/
MPI_Recv(in_grid_array, (work[2] + 1 - work[1]) * (work[4] + 1 - work[3]), MPI_INT, 0, 201, child_comm[index], &status);
/* draw data */
output_data(in_grid_array, &work[1], out_grid_array, ipixels_across, ipixels_down);
work[0] = k+1;
work[1] = i1[k];
work[2] = i2[k];
work[3] = j1[k];
work[4] = j2[k];
/*printf("sending work(%d) to %d\n", k+1, index);fflush(stdout);*/
MPI_Send(work, 5, MPI_INT, 0, 100, child_comm[index]);
MPI_Irecv(&result[index], 5, MPI_INT, 0, 200, child_comm[index], &child_request[index]);
k++;
}
/* receive the last pieces of work */
/* and tell everyone to stop */
for (i=0; i<num_children; i++)
{
MPI_Wait(&child_request[i], &status);
memcpy(work, &result[i], 5 * sizeof(int));
/*printf("master receiving data in tail loop.\n");fflush(stdout);*/
MPI_Recv(in_grid_array, (work[2] + 1 - work[1]) * (work[4] + 1 - work[3]), MPI_INT, 0, 201, child_comm[i], &status);
/* draw data */
output_data(in_grid_array, &work[1], out_grid_array, ipixels_across, ipixels_down);
work[0] = 0;
work[1] = 0;
work[2] = 0;
work[3] = 0;
work[4] = 0;
/*printf("sending %d to tell %d to stop\n", work[0], i);fflush(stdout);*/
MPI_Send(work, 5, MPI_INT, 0, 100, child_comm[i]);
}
/* tell the visualizer the image is done */
if (!use_stdin)
{
work[0] = 0;
work[1] = 0;
work[2] = 0;
work[3] = 0;
sock_write(sock, work, 4 * sizeof(int));
}
}
}
else
{
for (;;)
{
/*printf("slave[%d] receiveing bounds.\n", pid);fflush(stdout);*/
MPI_Recv(&x_min, 1, MPI_DOUBLE, 0, 0, parent, MPI_STATUS_IGNORE);
MPI_Recv(&x_max, 1, MPI_DOUBLE, 0, 0, parent, MPI_STATUS_IGNORE);
MPI_Recv(&y_min, 1, MPI_DOUBLE, 0, 0, parent, MPI_STATUS_IGNORE);
MPI_Recv(&y_max, 1, MPI_DOUBLE, 0, 0, parent, MPI_STATUS_IGNORE);
MPI_Recv(&imax_iterations, 1, MPI_INT, 0, 0, parent, MPI_STATUS_IGNORE);
/*printf("slave[%d] received bounding box: (%f,%f)(%f,%f)\n", pid, x_min, y_min, x_max, y_max);fflush(stdout);*/
/* check for the end condition */
if (x_min == x_max && y_min == y_max)
{
/*printf("slave[%d] done.\n", pid);fflush(stdout);*/
break;
}
x_resolution = (x_max-x_min)/ ((double)ipixels_across);
y_resolution = (y_max-y_min)/ ((double)ipixels_down);
MPI_Recv(work, 5, MPI_INT, 0, 100, parent, &status);
/*printf("slave[%d] received work: %d, (%d,%d)(%d,%d)\n", pid, work[0], work[1], work[2], work[3], work[4]);fflush(stdout);*/
while (work[0] != 0)
{
imin = work[1];
imax = work[2];
jmin = work[3];
jmax = work[4];
k = 0;
for (j=jmin; j<=jmax; ++j)
{
coord_point.imaginary = y_max - j*y_resolution; /* go top to bottom */
for (i=imin; i<=imax; ++i)
{
/* Call Mandelbrot routine for each code, fill array with number of iterations. */
coord_point.real = x_min + i*x_resolution; /* go left to right */
if (julia == 1)
{
/* doing Julia set */
/* julia eq: z = z^2 + c, z_0 = grid coordinate, c = constant */
icount = single_mandelbrot_point(coord_point, julia_constant, imax_iterations, divergent_limit);
}
else if (alternate_equation == 1)
{
/* doing experimental form 1 */
icount = subtractive_mandelbrot_point(coord_point, julia_constant, imax_iterations, divergent_limit);
}
else if (alternate_equation == 2)
{
/* doing experimental form 2 */
icount = additive_mandelbrot_point(coord_point, julia_constant, imax_iterations, divergent_limit);
}
else
{
/* default to doing Mandelbrot set */
/* mandelbrot eq: z = z^2 + c, z_0 = c, c = grid coordinate */
icount = single_mandelbrot_point(coord_point, coord_point, imax_iterations, divergent_limit);
}
in_grid_array[k] = icount;
++k;
}
}
/* send the result to the root */
/*printf("slave[%d] sending work %d back.\n", pid, work[0]);fflush(stdout);*/
MPI_Send(work, 5, MPI_INT, 0, 200, parent);
/*printf("slave[%d] sending work %d data.\n", pid, work[0]);fflush(stdout);*/
MPI_Send(in_grid_array, (work[2] + 1 - work[1]) * (work[4] + 1 - work[3]), MPI_INT, 0, 201, parent);
/* get the next piece of work */
/*printf("slave[%d] receiving new work.\n", pid);fflush(stdout);*/
MPI_Recv(work, 5, MPI_INT, 0, 100, parent, &status);
/*printf("slave[%d] received work: %d, (%d,%d)(%d,%d)\n", pid, work[0], work[1], work[2], work[3], work[4]);fflush(stdout);*/
}
}
}
if (master && save_image)
{
imax_iterations = 0;
for (i=0; i<ipixels_across * ipixels_down; ++i)
{
/* look for "brightest" pixel value, for image use */
if (out_grid_array[i] > imax_iterations)
imax_iterations = out_grid_array[i];
}
if (julia == 0)
printf("Done calculating mandelbrot, now creating file\n");
else
printf("Done calculating julia, now creating file\n");
fflush(stdout);
/* Print out the array in some appropriate form. */
if (julia == 0)
{
/* it's a mandelbrot */
sprintf(file_message, "Mandelbrot over (%lf-%lf,%lf-%lf), size %d x %d",
x_min, x_max, y_min, y_max, ipixels_across, ipixels_down);
}
else
{
/* it's a julia */
sprintf(file_message, "Julia over (%lf-%lf,%lf-%lf), size %d x %d, center (%lf, %lf)",
x_min, x_max, y_min, y_max, ipixels_across, ipixels_down,
julia_constant.real, julia_constant.imaginary);
}
dumpimage(filename, out_grid_array, ipixels_across, ipixels_down, imax_iterations, file_message, num_colors, colors);
}
if (master)
{
for (i=0; i<num_children; i++)
{
MPI_Comm_disconnect(&child_comm[i]);
}
free(child_comm);
free(child_request);
free(colors);
}
MPI_Finalize();
return 0;
}
int main(int argc, char *argv[])
{
int errs = 0;
int rank, size, rsize;
int np = 3;
MPI_Comm parentcomm, intercomm;
int verbose = 0;
char *env;
int can_spawn;
env = getenv("MPITEST_VERBOSE");
if (env) {
if (*env != '0')
verbose = 1;
}
MTest_Init(&argc, &argv);
errs += MTestSpawnPossible(&can_spawn);
if (can_spawn) {
MPI_Comm_get_parent(&parentcomm);
if (parentcomm == MPI_COMM_NULL) {
IF_VERBOSE(("spawning %d processes\n", np));
/* Create 3 more processes */
MPI_Comm_spawn((char *) "./disconnect", MPI_ARGV_NULL, np,
MPI_INFO_NULL, 0, MPI_COMM_WORLD, &intercomm, MPI_ERRCODES_IGNORE);
} else {
intercomm = parentcomm;
}
/* We now have a valid intercomm */
MPI_Comm_remote_size(intercomm, &rsize);
MPI_Comm_size(intercomm, &size);
MPI_Comm_rank(intercomm, &rank);
if (parentcomm == MPI_COMM_NULL) {
IF_VERBOSE(("parent rank %d alive.\n", rank));
/* Parent */
if (rsize != np) {
errs++;
printf("Did not create %d processes (got %d)\n", np, rsize);
fflush(stdout);
}
IF_VERBOSE(("disconnecting child communicator\n"));
MPI_Comm_disconnect(&intercomm);
/* Errors cannot be sent back to the parent because there is no
* communicator connected to the children
* for (i=0; i<rsize; i++)
* {
* MPI_Recv(&err, 1, MPI_INT, i, 1, intercomm, MPI_STATUS_IGNORE);
* errs += err;
* }
*/
} else {
IF_VERBOSE(("child rank %d alive.\n", rank));
/* Child */
if (size != np) {
errs++;
printf("(Child) Did not create %d processes (got %d)\n", np, size);
fflush(stdout);
}
IF_VERBOSE(("disconnecting communicator\n"));
MPI_Comm_disconnect(&intercomm);
/* Send the errs back to the master process */
/* Errors cannot be sent back to the parent because there is no
* communicator connected to the parent */
/*MPI_Ssend(&errs, 1, MPI_INT, 0, 1, intercomm); */
}
/* Note that the MTest_Finalize get errs only over COMM_WORLD */
/* Note also that both the parent and child will generate "No Errors"
* if both call MTest_Finalize */
if (parentcomm == MPI_COMM_NULL) {
MTest_Finalize(errs);
} else {
MPI_Finalize();
}
} else {
MTest_Finalize(errs);
}
IF_VERBOSE(("calling finalize\n"));
return MTestReturnValue(errs);
}
// Método que constroi a árvore de eliminação de threads
ThreadTree* buildBNThreadTree(int nxq, int* xq, int nxe, int* xe, Potential** findings,BayesNet* bayesnet, int* elmorder, Graph* moral, Graph* elmtree,MPI_Comm *everyone) {
int msg = -1;
MPI_Comm t;
ThreadTree* threadtree = NULL; // Árvore de Threads
ThreadVertex* thread = NULL; // Variável auxiliar que instancia um vértive da árvore de threads
VertexNode* child; // Nó de uma lista ligada simples
VertexNode* current = NULL; // Variável auxiliar para percorrer o Moral Graph
Potential* result = NULL; // Potencial resultante da sequencia de operações
Potential* potential = NULL; // Variável potencial auxiliar
int varelmorder[bnsize(bayesnet)]; // Mapa o id da variável para a ordem de eliminação
int used[bnsize(bayesnet)]; // Vetor de flag para indicar se um potencial foi utilizado
Elimination** eliminations = NULL;
int nvars, ct_used;;
int i,j,id;;
// Número de variáveis envolvidas, pode não ser toda a rede
nvars = graphsize(elmtree);
//MPI_Comm_spawn(char *command, char *argv[], int maxprocs, MPI_Info info, int root, MPI_Comm comm, MPI_Comm *intercomm, int array_of_errcodes[])
MPI_Comm_spawn("./mpitask",MPI_ARGV_NULL,nvars,MPI_INFO_NULL,0,MPI_COMM_SELF,everyone,MPI_ERRCODES_IGNORE);
//MPI_Bcast(&msg,1,MPI_INT,MPI_ROOT,t);
// Monta a estrutura que será utilizada para passar dados para as threads
eliminations = malloc(nvars*sizeof(Elimination*));
// Constrói a árvore de thread
threadtree = buildThreadTree();
// Cria-se as threads vértices
for (i=0;i<nvars;i++) {
// Inicializa o mapa
varelmorder[elmorder[i]] = i;
// Inicializa a flag
used[elmorder[i]] = 0;
// Monta as eliminações e popula varelmorder
eliminations[i] = (Elimination*) buildElimination(i,bayesnet->variables[elmorder[i]],1);
// Constrói uma thread da árvore de thread
thread = buildThreadVertex(taskVariableElimination,(void*)(eliminations));
// Atribui o vértice à árvore de threads
addThreadVertex(&threadtree,&thread);
}
// Adiciona os arcos de dependência
for (i=0;i<nvars;i++) {
child = (getVertexChildren(getVertex(&elmtree,i)))->first;
while (child!=NULL) {
addTreeArcByIds(&threadtree,i,child->vertex->id);
child = child->next;
}
}
//---
// Tomando Xq, se não NULL, marca as operações que não devem ser marginalizadas
if (xq!=NULL) { // Apenas por segurança
for (i=0;i<nxq;i++) eliminations[varelmorder[xq[i]]]->marginalize = 0;
}
for (i=0, ct_used=0;i<nvars && ct_used<nvars;i++) {
// Atribui os potenciais da operação de eliminação
// 1. Verifica o nó da variável "da vez"
if (!used[elmorder[i]]) {
id = elmorder[i];
if (findings && findings[id]) {
// Marginaliza a distribuição de probabilidade se necessário
potential =(bayesnet->potentials[id]->nvars>1)?marginalizeNotRequired(moral,bayesnet->potentials[id]):NULL;
// Produto do finding pela distribuição de probabilidade
sllappend(&eliminations[i]->potentials,multPotentials((potential?potential:bayesnet->potentials[id]),findings[id]));
// Libera o espaço alocado (se necessário)
if (potential) destroyPotential(&potential);
} else {
sllappend(&eliminations[i]->potentials,bayesnet->potentials[id]);
}
used[id] = 1;
ct_used++;
}
// 2. Verifica os filhos da variável "da vez"
current = getVertexChildren(getVertexById(&bayesnet->graph,elmorder[i]))->first;
while (current!=NULL) {
if (getVertexById(&moral,current->vertex->id) && !used[current->vertex->id]) {
id = current->vertex->id;
if (findings && findings[id]) {
// Marginaliza a distribuição de probabilidade se necessário
potential =(bayesnet->potentials[id]->nvars>1)?marginalizeNotRequired(moral,bayesnet->potentials[id]):NULL;
// Produto do finding pela distribuição de probabilidade
sllappend(&eliminations[i]->potentials,multPotentials((potential?potential:bayesnet->potentials[id]),findings[id]));
// Libera o espaço alocado (se necessário)
if (potential) destroyPotential(&potential);
} else {
sllappend(&eliminations[i]->potentials,bayesnet->potentials[id]);
}
used[id] = 1;
ct_used++;
}
current = current->next;
}
// Fixa o número de potenciais fixos na elimiantion
eliminations[i]->nconst = slllength(eliminations[i]->potentials);
}
//---
// Tem que atribuir NULL para liberar a memória pois eu fiz um malloc explicito para ela
eliminations = NULL;
return threadtree;
}
int mca_sharedfp_addproc_file_open (struct ompi_communicator_t *comm,
const char* filename,
int amode,
struct ompi_info_t *info,
mca_io_ompio_file_t *fh)
{
int ret = OMPI_SUCCESS, err;
int rank;
struct mca_sharedfp_base_data_t* sh;
mca_io_ompio_file_t * shfileHandle, *ompio_fh;
MPI_Comm newInterComm;
struct mca_sharedfp_addproc_data * addproc_data = NULL;
mca_io_ompio_data_t *data;
/*-------------------------------------------------*/
/*Open the same file again without shared file pointer*/
/*-------------------------------------------------*/
shfileHandle = (mca_io_ompio_file_t *)malloc(sizeof(mca_io_ompio_file_t));
ret = mca_common_ompio_file_open(comm,filename,amode,info,shfileHandle,false);
if ( OMPI_SUCCESS != ret) {
opal_output(0, "mca_sharedfp_addproc_file_open: Error during file open\n");
return ret;
}
shfileHandle->f_fh = fh->f_fh;
data = (mca_io_ompio_data_t *) fh->f_fh->f_io_selected_data;
ompio_fh = &data->ompio_fh;
err = mca_common_ompio_set_view (shfileHandle,
ompio_fh->f_disp,
ompio_fh->f_etype,
ompio_fh->f_orig_filetype,
ompio_fh->f_datarep,
MPI_INFO_NULL);
/*Memory is allocated here for the sh structure*/
if ( mca_sharedfp_addproc_verbose ) {
opal_output(ompi_sharedfp_base_framework.framework_output,
"mca_sharedfp_addproc_file_open: malloc f_sharedfp_ptr struct\n");
}
sh = (struct mca_sharedfp_base_data_t*)malloc(sizeof(struct mca_sharedfp_base_data_t));
if ( NULL == sh ){
opal_output(ompi_sharedfp_base_framework.framework_output,
"mca_sharedfp_addproc_file_open: Error, unable to malloc f_sharedfp_ptr struct\n");
return OMPI_ERR_OUT_OF_RESOURCE;
}
/*Populate the sh file structure based on the implementation*/
sh->sharedfh = shfileHandle; /* Shared file pointer*/
sh->global_offset = 0; /* Global Offset*/
sh->comm = comm; /* Communicator*/
sh->selected_module_data = NULL;
rank = ompi_comm_rank ( sh->comm );
if ( mca_sharedfp_addproc_verbose ) {
opal_output(ompi_sharedfp_base_framework.framework_output,
"mca_sharedfp_addproc_file_open: START spawn by rank=%d\n",rank);
}
/*Spawn a new process which will maintain the offsets for this file open*/
ret = MPI_Comm_spawn("mca_sharedfp_addproc_control", MPI_ARGV_NULL, 1, MPI_INFO_NULL,
0, sh->comm, &newInterComm, &err);
if ( OMPI_SUCCESS != ret ) {
opal_output(0, "mca_sharedfp_addproc_file_open: error spawning control process ret=%d\n",
ret);
}
/*If spawning successful*/
if (newInterComm) {
addproc_data = (struct mca_sharedfp_addproc_data*)malloc(sizeof(struct mca_sharedfp_addproc_data));
if ( NULL == addproc_data ){
opal_output (0,"mca_sharedfp_addproc_file_open: Error, unable to malloc addproc_data struct\n");
return OMPI_ERR_OUT_OF_RESOURCE;
}
/*Store the new Intercommunicator*/
addproc_data->intercom = newInterComm;
/*save the addproc data*/
sh->selected_module_data = addproc_data;
/*remember the shared file handle*/
fh->f_sharedfp_data = sh;
}
else{
opal_output(ompi_sharedfp_base_framework.framework_output,
"mca_sharedfp_addproc_file_open: DONE spawn by rank=%d, errcode[success=%d, err=%d]=%d\n",
rank, MPI_SUCCESS, MPI_ERR_SPAWN, ret);
ret = OMPI_ERROR;
}
return ret;
}
int
main (int argc, char **argv)
{
int rank, size;
int *error_codes;
int i;
char *buffer;
int *flag;
int all_done;
MPI_Comm intercomm;
MPI_Request *request;
MPI_Status *status;
struct timespec nanoseconds;
MPI_Init(&argc, &argv);
MPI_Comm_rank(MPI_COMM_WORLD, &rank);
MPI_Comm_size(MPI_COMM_WORLD, &size);
if (rank == 0)
{
error_codes = (int*) malloc(sizeof(int)*size);
buffer = (char*) malloc(sizeof(char)*size);
request = (MPI_Request*) malloc(sizeof(MPI_Request)*size);
status = (MPI_Status*) malloc(sizeof(MPI_Status)*size);
flag = (int*) malloc(sizeof(int)*size);
while (1)
{
printf("[master] spawning %i processes\n", size);
MPI_Comm_spawn("./other", argv, size, MPI_INFO_NULL, 0, MPI_COMM_SELF, &intercomm, error_codes);
/* Wait for children to finish. */
for (i = 0; i < size; ++i)
{
MPI_Irecv(buffer, 1, MPI_CHAR, MPI_ANY_SOURCE, 1, intercomm, &request[i]);
flag[i] = 0;
}
all_done = 0;
while (! all_done)
{
all_done = 1;
for (i = 0; i < size; ++i)
{
if (! flag[i])
{
MPI_Test(&request[i], &flag[i], &status[i]);
if (! flag[i])
{
all_done = 0;
}
}
}
/* Sleep a little. */
nanoseconds.tv_sec = 0;
nanoseconds.tv_nsec = 0.5e9;
nanosleep(&nanoseconds, NULL);
}
}
}
printf("[master (%i)] waiting at barrier\n", rank);
MPI_Barrier(MPI_COMM_WORLD);
printf("[master (%i)] done\n", rank);
MPI_Finalize();
}
int main( int argc, char *argv[] )
{
int errs = 0, err;
int rank, size, rsize, i;
int np = 2;
int errcodes[2];
MPI_Comm parentcomm, intercomm, intracomm, intracomm2, intracomm3;
int isChild = 0;
MPI_Status status;
MTest_Init( &argc, &argv );
MPI_Comm_get_parent( &parentcomm );
if (parentcomm == MPI_COMM_NULL) {
/* Create 2 more processes */
MPI_Comm_spawn( (char*)"./spawnintra", MPI_ARGV_NULL, np,
MPI_INFO_NULL, 0, MPI_COMM_WORLD,
&intercomm, errcodes );
}
else
intercomm = parentcomm;
/* We now have a valid intercomm */
MPI_Comm_remote_size( intercomm, &rsize );
MPI_Comm_size( intercomm, &size );
MPI_Comm_rank( intercomm, &rank );
if (parentcomm == MPI_COMM_NULL) {
/* Master */
if (rsize != np) {
errs++;
printf( "Did not create %d processes (got %d)\n", np, rsize );
}
if (rank == 0) {
for (i=0; i<rsize; i++) {
MPI_Send( &i, 1, MPI_INT, i, 0, intercomm );
}
}
}
else {
/* Child */
isChild = 1;
if (size != np) {
errs++;
printf( "(Child) Did not create %d processes (got %d)\n",
np, size );
}
MPI_Recv( &i, 1, MPI_INT, 0, 0, intercomm, &status );
if (i != rank) {
errs++;
printf( "Unexpected rank on child %d (%d)\n", rank, i );
}
}
/* At this point, try to form the intracommunicator */
MPI_Intercomm_merge( intercomm, isChild, &intracomm );
/* Check on the intra comm */
{
int icsize, icrank, wrank;
MPI_Comm_size( intracomm, &icsize );
MPI_Comm_rank( intracomm, &icrank );
MPI_Comm_rank( MPI_COMM_WORLD, &wrank );
if (icsize != rsize + size) {
errs++;
printf( "Intracomm rank %d thinks size is %d, not %d\n",
icrank, icsize, rsize + size );
}
/* Make sure that the processes are ordered correctly */
if (isChild) {
int psize;
MPI_Comm_remote_size( parentcomm, &psize );
if (icrank != psize + wrank ) {
errs++;
printf( "Intracomm rank %d (from child) should have rank %d\n",
icrank, psize + wrank );
}
}
else {
if (icrank != wrank) {
errs++;
printf( "Intracomm rank %d (from parent) should have rank %d\n",
icrank, wrank );
}
}
}
/* At this point, try to form the intracommunicator, with the other
processes first */
MPI_Intercomm_merge( intercomm, !isChild, &intracomm2 );
/* Check on the intra comm */
{
int icsize, icrank, wrank;
MPI_Comm_size( intracomm2, &icsize );
MPI_Comm_rank( intracomm2, &icrank );
MPI_Comm_rank( MPI_COMM_WORLD, &wrank );
if (icsize != rsize + size) {
errs++;
printf( "(2)Intracomm rank %d thinks size is %d, not %d\n",
icrank, icsize, rsize + size );
}
/* Make sure that the processes are ordered correctly */
if (isChild) {
if (icrank != wrank ) {
errs++;
printf( "(2)Intracomm rank %d (from child) should have rank %d\n",
icrank, wrank );
}
}
else {
int csize;
MPI_Comm_remote_size( intercomm, &csize );
if (icrank != wrank + csize) {
errs++;
printf( "(2)Intracomm rank %d (from parent) should have rank %d\n",
icrank, wrank + csize );
}
}
}
/* At this point, try to form the intracommunicator, with an
arbitrary choice for the first group of processes */
MPI_Intercomm_merge( intercomm, 0, &intracomm3 );
/* Check on the intra comm */
{
int icsize, icrank, wrank;
MPI_Comm_size( intracomm3, &icsize );
MPI_Comm_rank( intracomm3, &icrank );
MPI_Comm_rank( MPI_COMM_WORLD, &wrank );
if (icsize != rsize + size) {
errs++;
printf( "(3)Intracomm rank %d thinks size is %d, not %d\n",
icrank, icsize, rsize + size );
}
/* Eventually, we should test that the processes are ordered
correctly, by groups (must be one of the two cases above) */
}
/* Update error count */
if (isChild) {
/* Send the errs back to the master process */
MPI_Ssend( &errs, 1, MPI_INT, 0, 1, intercomm );
}
else {
if (rank == 0) {
/* We could use intercomm reduce to get the errors from the
children, but we'll use a simpler loop to make sure that
we get valid data */
for (i=0; i<rsize; i++) {
MPI_Recv( &err, 1, MPI_INT, i, 1, intercomm, MPI_STATUS_IGNORE );
errs += err;
}
}
}
/* It isn't necessary to free the intracomms, but it should not hurt */
MPI_Comm_free( &intracomm );
MPI_Comm_free( &intracomm2 );
MPI_Comm_free( &intracomm3 );
/* It isn't necessary to free the intercomm, but it should not hurt */
MPI_Comm_free( &intercomm );
/* Note that the MTest_Finalize get errs only over COMM_WORLD */
/* Note also that both the parent and child will generate "No Errors"
if both call MTest_Finalize */
if (parentcomm == MPI_COMM_NULL) {
MTest_Finalize( errs );
}
MPI_Finalize();
return 0;
}
int
main(int ac, char *av[])
{
int rank, size;
char name[MPI_MAX_PROCESSOR_NAME];
int nameLen;
int n = 1, i;
int slave = 0;
int *errs;
char *args[] = { "-W", NULL};
MPI_Comm intercomm, icomm;
int err;
char *line;
char *buff;
int buffSize;
int one_int;
char who[1024];
memset(name, sizeof(name), 0);
for(i=1; i<ac; i++){
if (av[i] == NULL)
continue;
if (strcmp(av[i],"-W") == 0){
slave = 1;
}
else if (strcmp(av[i],"-n") == 0){
n = atoi(av[i+1]);
av[i+1] = NULL;
}
}
if (n <= 0){
fprintf(stderr, "n=%d has an illegal value.\n", n);
return -1;
}
sprintf(who, "%s[%d]", slave? " slave": "master", getpid());
if (!slave)
printf("Generating %d slave processes\n", n);
errs = (int *)alloca(sizeof(int)*n);
fprintf(stderr, "%s before MPI_Init()\n", who);
MPI_Init(&ac, &av);
fprintf(stderr, "%s after MPI_Init()\n", who);
MPI_Comm_rank(MPI_COMM_WORLD, &rank);
MPI_Comm_size(MPI_COMM_WORLD, &size);
MPI_Get_processor_name(name, &nameLen);
sprintf(&who[strlen(who)], " (%s) %d/%d", name, rank, size);
fprintf(stderr, "%s\n", who);
if (!slave){
err = MPI_Comm_spawn(av[0], args, n, MPI_INFO_NULL, 0, MPI_COMM_SELF, &intercomm, errs);
if (err){
fprintf(stderr, "MPI_Comm_spawn generated error %d.\n", err);
}
}
else {
fprintf(stderr, "%s before MPI_Comm_get_parent()\n", who);
MPI_Comm_get_parent(&intercomm);
}
fprintf(stderr, "%s before MPI_Bcast()\n", who);
MPI_Bcast(&one_int, 1, MPI_INT, 0, intercomm);
fprintf(stderr, "%s after MPI_Bcast()\n", who);
fprintf(stderr, "%s before MPI_Barrier()\n", who);
MPI_Barrier(intercomm);
fprintf(stderr, "%s after MPI_Barrier()\n", who);
fprintf(stderr, "%s before MPI_Finalize()\n", who);
MPI_Finalize();
return 0;
}
