int mca_common_pmi_put(const char *kvs_name,
const char *key, const char *value)
{
int rc;
#if WANT_PMI2_SUPPORT
if( mca_common_pmi_version == 2 ){
if( PMI2_SUCCESS != PMI2_KVS_Put(key, value) ){
// FIXME: OPAL_PMI2_ERROR(rc, "PMI2_KVS_Put");
return OPAL_ERROR;
}
}
else
#endif
{
rc = PMI_KVS_Put(kvs_name, key, value);
if( PMI_SUCCESS != rc ){
OPAL_PMI_ERROR(rc, "PMI_KVS_Put");
return OPAL_ERROR;
}
}
return OPAL_SUCCESS;
}
static int test_item6(void)
{
int rc = 0;
char val[100];
const char *tkey = __func__;
const char *tval = __FILE__;
if (PMI_SUCCESS != (rc = PMI_KVS_Put(jobid, tkey, tval))) {
log_fatal("PMI_KVS_Put %d\n", rc);
return rc;
}
if (PMI_SUCCESS != (rc = PMI_KVS_Get(jobid, tkey, val, sizeof(val)))) {
log_fatal("PMI_KVS_Get %d\n", rc);
return rc;
}
log_info("tkey=%s tval=%s val=%s\n", tkey, tval, val);
log_assert(!strcmp(tval, val), "value does not meet expectation");
return rc;
}
static
int InitPscomConnections(pscom_socket_t *socket) {
char key[50];
unsigned long guard_pmi_key = MAGIC_PMI_KEY;
int i;
int mpi_errno = MPI_SUCCESS;
int pg_rank = MPIDI_Process.my_pg_rank;
int pg_size = MPIDI_Process.my_pg_size;
char *pg_id = MPIDI_Process.pg_id_name;
char *listen_socket;
char **psp_port = NULL;
/* Distribute my contact information */
snprintf(key, sizeof(key), "pscom%d", pg_rank);
listen_socket = MPL_strdup(pscom_listen_socket_ondemand_str(socket));
PMICALL(PMI_KVS_Put(pg_id, key, listen_socket));
#define IPSCOM_VERSION "pscom_v5.0"
i_version_set(pg_id, pg_rank, IPSCOM_VERSION);
PMICALL(PMI_KVS_Commit(pg_id));
PMICALL(PMI_Barrier());
i_version_check(pg_id, pg_rank, IPSCOM_VERSION);
init_grank_port_mapping();
/* Get portlist */
psp_port = MPL_malloc(pg_size * sizeof(*psp_port), MPL_MEM_OBJECT);
assert(psp_port);
for (i = 0; i < pg_size; i++) {
char val[100];
unsigned long guard_pmi_value = MAGIC_PMI_VALUE;
if (i != pg_rank) {
snprintf(key, sizeof(key), "pscom%d", i);
checked_PMI_KVS_Get(pg_id, key, val, sizeof(val));
/* simple_pmi.c has a bug.(fixed in mpich2-1.0.5)
Test for the bugfix: */
assert(guard_pmi_value == MAGIC_PMI_VALUE);
assert(guard_pmi_key == MAGIC_PMI_KEY);
} else {
/* myself: Dont use PMI_KVS_Get, because this fail
in the case of no pm (SINGLETON_INIT_BUT_NO_PM) */
strcpy(val, listen_socket);
}
psp_port[i] = MPL_strdup(val);
}
/* Create all connections */
for (i = 0; i < pg_size; i++) {
pscom_connection_t *con;
pscom_err_t rc;
const char *dest;
dest = psp_port[i];
con = pscom_open_connection(socket);
rc = pscom_connect_socket_str(con, dest);
if (rc != PSCOM_SUCCESS) {
PRINTERROR("Connecting %s to %s (rank %d to %d) failed : %s",
listen_socket, dest, pg_rank, i, pscom_err_str(rc));
goto fn_fail;
}
grank2con_set(i, con);
}
pscom_stop_listen(socket);
fn_exit:
if (psp_port) {
for (i = 0; i < pg_size; i++) {
MPL_free(psp_port[i]);
psp_port[i] = NULL;
}
MPL_free(psp_port);
}
MPL_free(listen_socket);
return mpi_errno;
/* --- */
fn_fail:
mpi_errno = MPIR_Err_create_code(MPI_SUCCESS, MPIR_ERR_FATAL,
"InitPscomConnections", __LINE__, MPI_ERR_OTHER, "**connfailed", 0);
goto fn_exit;
}
static
int InitPortConnections(pscom_socket_t *socket) {
char key[50];
unsigned long guard_pmi_key = MAGIC_PMI_KEY;
int i;
int mpi_errno = MPI_SUCCESS;
int pg_rank = MPIDI_Process.my_pg_rank;
int pg_size = MPIDI_Process.my_pg_size;
char *pg_id = MPIDI_Process.pg_id_name;
char *listen_socket;
char **psp_port = NULL;
/* Distribute my contact information */
snprintf(key, sizeof(key), "psp%d", pg_rank);
listen_socket = MPL_strdup(pscom_listen_socket_str(socket));
PMICALL(PMI_KVS_Put(pg_id, key, listen_socket));
#define INIT_VERSION "ps_v5.0"
i_version_set(pg_id, pg_rank, INIT_VERSION);
PMICALL(PMI_KVS_Commit(pg_id));
PMICALL(PMI_Barrier());
i_version_check(pg_id, pg_rank, INIT_VERSION);
init_grank_port_mapping();
/* Get portlist */
psp_port = MPL_malloc(pg_size * sizeof(*psp_port), MPL_MEM_OBJECT);
assert(psp_port);
for (i = 0; i < pg_size; i++) {
char val[100];
unsigned long guard_pmi_value = MAGIC_PMI_VALUE;
if (i != pg_rank) {
snprintf(key, sizeof(key), "psp%d", i);
checked_PMI_KVS_Get(pg_id, key, val, sizeof(val));
/* simple_pmi.c has a bug.(fixed in mpich2-1.0.5)
Test for the bugfix: */
assert(guard_pmi_value == MAGIC_PMI_VALUE);
assert(guard_pmi_key == MAGIC_PMI_KEY);
} else {
/* myself: Dont use PMI_KVS_Get, because this fail
in the case of no pm (SINGLETON_INIT_BUT_NO_PM) */
strcpy(val, listen_socket);
}
psp_port[i] = MPL_strdup(val);
}
/* connect ranks pg_rank..(pg_rank + pg_size/2) */
for (i = 0; i <= pg_size / 2; i++) {
int dest = (pg_rank + i) % pg_size;
int src = (pg_rank + pg_size - i) % pg_size;
if (!i || (pg_rank / i) % 2) {
/* connect, accept */
if (do_connect(socket, pg_rank, dest, psp_port[dest])) goto fn_fail;
if (!i || src != dest) {
do_wait(pg_rank, src);
}
} else {
/* accept, connect */
do_wait(pg_rank, src);
if (src != dest) {
if (do_connect(socket, pg_rank, dest, psp_port[dest])) goto fn_fail;
}
}
}
/* Wait for all connections: (already done?) */
for (i = 0; i < pg_size; i++) {
while (!grank2con_get(i)) {
pscom_wait_any();
}
}
/* ToDo: */
pscom_stop_listen(socket);
fn_exit:
if (psp_port) {
for (i = 0; i < pg_size; i++) {
MPL_free(psp_port[i]);
psp_port[i] = NULL;
}
MPL_free(psp_port);
}
MPL_free(listen_socket);
return mpi_errno;
/* --- */
fn_fail:
mpi_errno = MPIR_Err_create_code(MPI_SUCCESS, MPIR_ERR_FATAL,
"InitPortConnections", __LINE__, MPI_ERR_OTHER, "**connfailed", 0);
goto fn_exit;
}
int MPIDI_PG_SetConnInfo( int rank, const char *connString )
{
#ifdef USE_PMI2_API
int mpi_errno = MPI_SUCCESS;
int len;
char key[PMI2_MAX_KEYLEN];
MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_MPIDI_PG_SetConnInfo);
MPIR_FUNC_VERBOSE_ENTER(MPID_STATE_MPIDI_PG_SetConnInfo);
len = MPL_snprintf(key, sizeof(key), "P%d-businesscard", rank);
MPIR_ERR_CHKANDJUMP1(len < 0 || len > sizeof(key), mpi_errno, MPI_ERR_OTHER, "**snprintf", "**snprintf %d", len);
mpi_errno = PMI2_KVS_Put(key, connString);
if (mpi_errno) MPIR_ERR_POP(mpi_errno);
mpi_errno = PMI2_KVS_Fence();
if (mpi_errno) MPIR_ERR_POP(mpi_errno);
fn_exit:
MPIR_FUNC_VERBOSE_EXIT(MPID_STATE_MPIDI_PG_SetConnInfo);
return mpi_errno;
fn_fail:
goto fn_exit;
#else
int mpi_errno = MPI_SUCCESS;
int pmi_errno;
int len;
char key[128];
MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_MPIDI_PG_SetConnInfo);
MPIR_FUNC_VERBOSE_ENTER(MPID_STATE_MPIDI_PG_SetConnInfo);
MPIR_Assert(pg_world->connData);
len = MPL_snprintf(key, sizeof(key), "P%d-businesscard", rank);
if (len < 0 || len > sizeof(key)) {
MPIR_ERR_SETANDJUMP1(mpi_errno,MPI_ERR_OTHER, "**snprintf",
"**snprintf %d", len);
}
pmi_errno = PMI_KVS_Put(pg_world->connData, key, connString );
if (pmi_errno != PMI_SUCCESS) {
MPIR_ERR_SETANDJUMP1(mpi_errno,MPI_ERR_OTHER, "**pmi_kvs_put",
"**pmi_kvs_put %d", pmi_errno);
}
pmi_errno = PMI_KVS_Commit(pg_world->connData);
if (pmi_errno != PMI_SUCCESS) {
MPIR_ERR_SETANDJUMP1(mpi_errno,MPI_ERR_OTHER, "**pmi_kvs_commit",
"**pmi_kvs_commit %d", pmi_errno);
}
pmi_errno = PMI_Barrier();
if (pmi_errno != PMI_SUCCESS) {
MPIR_ERR_SETANDJUMP1(mpi_errno,MPI_ERR_OTHER, "**pmi_barrier",
"**pmi_barrier %d", pmi_errno);
}
fn_exit:
MPIR_FUNC_VERBOSE_EXIT(MPID_STATE_MPIDI_PG_SetConnInfo);
return mpi_errno;
fn_fail:
goto fn_exit;
#endif
}
int
main(int argc,
char *argv[])
{
int initialized, rank, size;
int i, max_name_len, max_key_len, max_val_len;
char *name, *key, *val;
if (PMI_SUCCESS != PMI_Initialized(&initialized)) {
return 1;
}
if (0 == initialized) {
if (PMI_SUCCESS != PMI_Init(&initialized)) {
return 1;
}
}
if (PMI_SUCCESS != PMI_Get_rank(&rank)) {
return 1;
}
if (PMI_SUCCESS != PMI_Get_size(&size)) {
return 1;
}
printf("Hello, World. I am %d of %d\n", rank, size);
if (PMI_SUCCESS != PMI_KVS_Get_name_length_max(&max_name_len)) {
return 1;
}
name = (char*) malloc(max_name_len);
if (NULL == name) return 1;
if (PMI_SUCCESS != PMI_KVS_Get_key_length_max(&max_key_len)) {
return 1;
}
key = (char*) malloc(max_key_len);
if (NULL == key) return 1;
if (PMI_SUCCESS != PMI_KVS_Get_value_length_max(&max_val_len)) {
return 1;
}
val = (char*) malloc(max_val_len);
if (NULL == val) return 1;
if (PMI_SUCCESS != PMI_KVS_Get_my_name(name, max_name_len)) {
return 1;
}
/* put my information */
snprintf(key, max_key_len, "pmi_hello-%lu-test", (long unsigned) rank);
snprintf(val, max_val_len, "%lu", (long unsigned) rank);
if (PMI_SUCCESS != PMI_KVS_Put(name, key, val)) {
return 1;
}
if (PMI_SUCCESS != PMI_KVS_Commit(name)) {
return 1;
}
if (PMI_SUCCESS != PMI_Barrier()) {
return 1;
}
/* verify everyone's information */
for (i = 0 ; i < size ; ++i) {
snprintf(key, max_key_len, "pmi_hello-%lu-test", (long unsigned) i);
if (PMI_SUCCESS != PMI_KVS_Get(name, key, val, max_val_len)) {
return 1;
}
if (i != strtol(val, NULL, 0)) {
fprintf(stderr, "%d: Error: Expected %d, got %d\n", rank, i, (int) strtol(val, NULL, 0));
return 1;
}
}
PMI_Finalize();
return 0;
}
main (int argc, char **argv)
{
int i, j, rc;
int nprocs, procid;
int clique_size, *clique_ranks = NULL;
char *jobid_ptr, *nprocs_ptr, *procid_ptr;
int pmi_rank, pmi_size, kvs_name_len, key_len, val_len;
PMI_BOOL initialized;
char *key, *val, *kvs_name;
struct timeval tv1, tv2;
long delta_t;
char tv_str[20];
gettimeofday(&tv1, NULL);
/* Get process count and our id from environment variables */
jobid_ptr = getenv("SLURM_JOB_ID");
nprocs_ptr = getenv("SLURM_NPROCS");
procid_ptr = getenv("SLURM_PROCID");
if (jobid_ptr == NULL) {
printf("WARNING: PMI test not run under SLURM\n");
nprocs = 1;
procid = 0;
} else if ((nprocs_ptr == NULL) || (procid_ptr == NULL)) {
printf("FAILURE: SLURM environment variables not set\n");
exit(1);
} else {
nprocs = atoi(nprocs_ptr);
procid = atoi(procid_ptr);
}
/* Validate process count and our id */
if ((nprocs < 1) || (nprocs > 9999)) {
printf("FAILURE: Invalid nprocs %s\n", nprocs_ptr);
exit(1);
}
if ((procid < 0) || (procid > 9999)) {
printf("FAILURE: Invalid procid %s\n", procid_ptr);
exit(1);
}
/* Get process count and size from PMI and validate */
if ((rc = PMI_Init(&i)) != PMI_SUCCESS) {
printf("FAILURE: PMI_Init: %d\n", rc);
exit(1);
}
initialized = PMI_FALSE;
if ((rc = PMI_Initialized(&initialized)) != PMI_SUCCESS) {
printf("FAILURE: PMI_Initialized: %d\n", rc);
exit(1);
}
if (initialized != PMI_TRUE) {
printf("FAILURE: PMI_Initialized returned false\n");
exit(1);
}
if ((rc = PMI_Get_rank(&pmi_rank)) != PMI_SUCCESS) {
printf("FAILURE: PMI_Get_rank: %d\n", rc);
exit(1);
}
#if _DEBUG
printf("PMI_Get_rank = %d\n", pmi_rank);
#endif
if ((rc = PMI_Get_size(&pmi_size)) != PMI_SUCCESS) {
printf("FAILURE: PMI_Get_size: %d, task %d\n", rc, pmi_rank);
exit(1);
}
#if _DEBUG
printf("PMI_Get_size = %d\n", pmi_size);
#endif
if (pmi_rank != procid) {
printf("FAILURE: Rank(%d) != PROCID(%d)\n",
pmi_rank, procid);
exit(1);
}
if (pmi_size != nprocs) {
printf("FAILURE: Size(%d) != NPROCS(%d), task %d\n",
pmi_size, nprocs, pmi_rank);
exit(1);
}
if ((rc = PMI_Get_clique_size(&clique_size)) != PMI_SUCCESS) {
printf("FAILURE: PMI_Get_clique_size: %d, task %d\n",
rc, pmi_rank);
exit(1);
}
clique_ranks = malloc(sizeof(int) * clique_size);
if ((rc = PMI_Get_clique_ranks(clique_ranks, clique_size)) !=
PMI_SUCCESS) {
printf("FAILURE: PMI_Get_clique_ranks: %d, task %d\n",
rc, pmi_rank);
exit(1);
}
#if _DEBUG
for (i=0; i<clique_size; i++)
printf("PMI_Get_clique_ranks[%d]=%d\n", i, clique_ranks[i]);
#endif
free(clique_ranks);
if ((rc = PMI_KVS_Get_name_length_max(&kvs_name_len)) != PMI_SUCCESS) {
printf("FAILURE: PMI_KVS_Get_name_length_max: %d, task %d\n",
rc, pmi_rank);
exit(1);
}
#if _DEBUG
printf("PMI_KVS_Get_name_length_max = %d\n", kvs_name_len);
#endif
kvs_name = malloc(kvs_name_len);
if ((rc = PMI_KVS_Get_my_name(kvs_name, kvs_name_len)) != PMI_SUCCESS) {
printf("FAILURE: PMI_KVS_Get_my_name: %d, task %d\n", rc,
pmi_rank);
exit(1);
}
#if _DEBUG
printf("PMI_KVS_Get_my_name = %s\n", kvs_name);
#endif
if ((rc = PMI_KVS_Get_key_length_max(&key_len)) != PMI_SUCCESS) {
printf("FAILURE: PMI_KVS_Get_key_length_max: %d, task %d\n",
rc, pmi_rank);
exit(1);
}
key = malloc(key_len);
if ((rc = PMI_KVS_Get_value_length_max(&val_len)) != PMI_SUCCESS) {
printf("FAILURE: PMI_KVS_Get_value_length_max: %d, task %d\n",
rc, pmi_rank);
exit(1);
}
#if _DEBUG
printf("PMI_KVS_Get_value_length_max = %d\n", val_len);
#endif
val = malloc(val_len);
/* Build and set some key=val pairs */
snprintf(key, key_len, "ATTR_1_%d", procid);
snprintf(val, val_len, "A%d", procid+OFFSET_1);
if ((rc = PMI_KVS_Put(kvs_name, key, val)) != PMI_SUCCESS) {
printf("FAILURE: PMI_KVS_Put(%s,%s,%s): %d, task %d\n",
kvs_name, key, val, rc, pmi_rank);
exit(1);
}
#if _DEBUG
printf("PMI_KVS_Put(%s,%s,%s)\n", kvs_name, key, val);
#endif
snprintf(key, key_len, "attr_2_%d", procid);
snprintf(val, val_len, "B%d", procid+OFFSET_2);
if ((rc = PMI_KVS_Put(kvs_name, key, val)) != PMI_SUCCESS) {
printf("FAILURE: PMI_KVS_Put(%s,%s,%s): %d, task %d\n",
kvs_name, key, val, rc, pmi_rank);
exit(1);
}
#if _DEBUG
printf("PMI_KVS_Put(%s,%s,%s)\n", kvs_name, key, val);
#endif
/* Sync KVS across all tasks */
if ((rc = PMI_KVS_Commit(kvs_name)) != PMI_SUCCESS) {
printf("FAILURE: PMI_KVS_Commit: %d, task %d\n", rc, pmi_rank);
exit(1);
}
#if _DEBUG
printf("PMI_KVS_Commit completed\n");
#endif
if ((rc = PMI_Barrier()) != PMI_SUCCESS) {
printf("FAILURE: PMI_Barrier: %d, task %d\n", rc, pmi_rank);
exit(1);
}
#if _DEBUG
printf("PMI_Barrier completed\n");
#endif
/* Now lets get all keypairs and validate */
for (i=0; i<pmi_size; i++) {
snprintf(key, key_len, "ATTR_1_%d", i);
if ((rc = PMI_KVS_Get(kvs_name, key, val, val_len))
!= PMI_SUCCESS) {
printf("FAILURE: PMI_KVS_Get(%s): %d, task %d\n",
key, rc, pmi_rank);
exit(1);
}
if ((val[0] != 'A')
|| ((atoi(&val[1])-OFFSET_1) != i)) {
printf("FAILURE: Bad keypair %s=%s, task %d\n",
key, val, pmi_rank);
exit(1);
}
#if _DEBUG
if ((pmi_size <= 8) && (pmi_rank == 0)) /* limit output */
printf("PMI_KVS_Get(%s,%s) %s\n", kvs_name, key, val);
#endif
snprintf(key, key_len, "attr_2_%d", i);
if ((rc = PMI_KVS_Get(kvs_name, key, val, val_len))
!= PMI_SUCCESS) {
printf("FAILURE: PMI_KVS_Get(%s): %d, task %d\n",
key, rc, pmi_rank);
exit(1);
}
if ((val[0] != 'B')
|| ((atoi(&val[1])-OFFSET_2) != i)) {
printf("FAILURE: Bad keypair %s=%s, task %d\n",
key,val, pmi_rank);
exit(1);
}
#if _DEBUG
if ((pmi_size <= 8) && (pmi_rank == 1)) /* limit output */
printf("PMI_KVS_Get(%s,%s) %s\n", kvs_name, key, val);
#endif
}
/* use iterator */
if ((rc = PMI_KVS_Iter_first(kvs_name, key, key_len, val,
val_len)) != PMI_SUCCESS) {
printf("FAILURE: PMI_KVS_iter_first: %d, task %d\n", rc,
pmi_rank);
exit(1);
}
for (i=0; ; i++) {
if (key[0] == '\0') {
if (i != (pmi_size * 2)) {
printf("FAILURE: PMI_KVS_iter_next "
"cycle count(%d, %d), task %d\n",
i, pmi_size, pmi_rank);
}
break;
}
#if _DEBUG
if ((pmi_size <= 8) && (pmi_rank == 1)) { /* limit output */
printf("PMI_KVS_Iter_next(%s,%d): %s=%s\n", kvs_name,
i, key, val);
}
#endif
if ((rc = PMI_KVS_Iter_next(kvs_name, key, key_len,
val, val_len)) != PMI_SUCCESS) {
printf("FAILURE: PMI_KVS_iter_next: %d, task %d\n",
rc, pmi_rank);
exit(1);
}
}
/* Build some more key=val pairs */
snprintf(key, key_len, "ATTR_3_%d", procid);
snprintf(val, val_len, "C%d", procid+OFFSET_1);
if ((rc = PMI_KVS_Put(kvs_name, key, val)) != PMI_SUCCESS) {
printf("FAILURE: PMI_KVS_Put(%s,%s,%s): %d, task %d\n",
kvs_name, key, val, rc, pmi_rank);
exit(1);
}
#if _DEBUG
printf("PMI_KVS_Put(%s,%s,%s)\n", kvs_name, key, val);
#endif
snprintf(key, key_len, "attr_4_%d", procid);
snprintf(val, val_len, "D%d", procid+OFFSET_2);
if ((rc = PMI_KVS_Put(kvs_name, key, val)) != PMI_SUCCESS) {
printf("FAILURE: PMI_KVS_Put(%s,%s,%s): %d, task %d\n",
kvs_name, key, val, rc, pmi_rank);
exit(1);
}
#if _DEBUG
printf("PMI_KVS_Put(%s,%s,%s)\n", kvs_name, key, val);
#endif
/* Sync KVS across all tasks */
if ((rc = PMI_KVS_Commit(kvs_name)) != PMI_SUCCESS) {
printf("FAILURE: PMI_KVS_Commit: %d, task %d\n", rc, pmi_rank);
exit(1);
}
#if _DEBUG
printf("PMI_KVS_Commit completed\n");
#endif
if ((rc = PMI_Barrier()) != PMI_SUCCESS) {
printf("FAILURE: PMI_Barrier: %d, task %d\n", rc, pmi_rank);
exit(1);
}
#if _DEBUG
printf("PMI_Barrier completed\n");
#endif
/* Now lets get some keypairs and validate */
for (i=0; i<pmi_size; i++) {
snprintf(key, key_len, "ATTR_1_%d", i);
if ((rc = PMI_KVS_Get(kvs_name, key, val, val_len))
!= PMI_SUCCESS) {
printf("FAILURE: PMI_KVS_Get(%s): %d, task %d\n",
key, rc, pmi_rank);
exit(1);
}
if ((val[0] != 'A')
|| ((atoi(&val[1])-OFFSET_1) != i)) {
printf("FAILURE: Bad keypair %s=%s, task %d\n",
key, val, pmi_rank);
exit(1);
}
#if _DEBUG
if ((pmi_size <= 8) && (pmi_rank == 1)) /* limit output */
printf("PMI_KVS_Get(%s,%s) %s\n", kvs_name, key, val);
#endif
snprintf(key, key_len, "attr_4_%d", i);
if ((rc = PMI_KVS_Get(kvs_name, key, val, val_len))
!= PMI_SUCCESS) {
printf("FAILURE: PMI_KVS_Get(%s): %d, task %d\n",
key, rc, pmi_rank);
exit(1);
}
if ((val[0] != 'D')
|| ((atoi(&val[1])-OFFSET_2) != i)) {
printf("FAILURE: Bad keypair %s=%s, task %d\n",
key,val, pmi_rank);
exit(1);
}
#if _DEBUG
if ((pmi_size <= 8) && (pmi_rank == 1)) /* limit output */
printf("PMI_KVS_Get(%s,%s) %s\n", kvs_name, key, val);
#endif
}
/* Replicate the very heavy load that MVAPICH2 puts on PMI
* This load exceeds that of MPICH2 by a very wide margin */
#if _DEBUG
printf("Starting %d iterations each with %d PMI_KVS_Put and \n"
" one each PMI_KVS_Commit and KVS_Barrier\n",
BARRIER_CNT, PUTS_PER_BARRIER);
fflush(stdout);
#endif
for (i=0; i<BARRIER_CNT; i++) {
for (j=0; j<PUTS_PER_BARRIER; j++) {
snprintf(key, key_len, "ATTR_%d_%d_%d", i, j, procid);
snprintf(val, val_len, "C%d", procid+OFFSET_1);
if ((rc = PMI_KVS_Put(kvs_name, key, val)) !=
PMI_SUCCESS) {
printf("FAILURE: PMI_KVS_Put(%s,%s,%s): "
"%d, task %d\n",
kvs_name, key, val, rc, pmi_rank);
exit(1);
}
}
if ((rc= PMI_KVS_Commit(kvs_name)) != PMI_SUCCESS) {
printf("FAILURE: PMI_KVS_Commit: %d, task %d\n",
rc, pmi_rank);
exit(1);
}
if ((rc = PMI_Barrier()) != PMI_SUCCESS) {
printf("FAILURE: PMI_Barrier: %d, task %d\n",
rc, pmi_rank);
exit(1);
}
/* Don't bother with PMI_KVS_Get as those are all local
* and do not put a real load on srun or the network */
}
#if _DEBUG
printf("Interative PMI calls successful\n");
#endif
/* create new keyspace and test it */
if ((rc = PMI_KVS_Create(kvs_name, kvs_name_len)) != PMI_SUCCESS) {
printf("FAILURE: PMI_KVS_Create: %d, task %d\n", rc, pmi_rank);
exit(1);
}
#if _DEBUG
printf("PMI_KVS_Create %s\n", kvs_name);
#endif
if ((rc = PMI_KVS_Put(kvs_name, "KVS_KEY", "KVS_VAL")) != PMI_SUCCESS) {
printf("FAILURE: PMI_KVS_Put: %d, task %d\n", rc, pmi_rank);
exit(1);
}
#if _DEBUG
printf("PMI_KVS_Put(%s,KVS_KEY,KVS_VAL)\n", kvs_name);
#endif
if ((rc = PMI_KVS_Get(kvs_name, "KVS_KEY", val, val_len)) !=
PMI_SUCCESS) {
printf("FAILURE: PMI_KVS_Get(%s, KVS_KEY): %d, task %d\n",
kvs_name, rc, pmi_rank);
exit(1);
}
#if _DEBUG
printf("PMI_KVS_Get(%s,%s) %s\n", kvs_name, "KVS_KEY", val);
#endif
if ((rc = PMI_KVS_Destroy(kvs_name)) != PMI_SUCCESS) {
printf("FAILURE: PMI_KVS_Destroy(%s): %d, task %d\n",
kvs_name, rc, pmi_rank);
exit(1);
}
if ((rc = PMI_KVS_Get(kvs_name, "KVS_KEY", val, val_len)) !=
PMI_ERR_INVALID_KVS) {
printf("FAILURE: PMI_KVS_Get(%s, KVS_KEY): %d, task %d\n",
kvs_name, rc, pmi_rank);
exit(1);
}
if ((rc = PMI_Finalize()) != PMI_SUCCESS) {
printf("FAILURE: PMI_Finalize: %d, task %d\n", rc, pmi_rank);
exit(1);
}
if (_DEBUG || (pmi_rank < 4)) {
gettimeofday(&tv2, NULL);
delta_t = (tv2.tv_sec - tv1.tv_sec) * 1000000;
delta_t += tv2.tv_usec - tv1.tv_usec;
snprintf(tv_str, sizeof(tv_str), "usec=%ld", delta_t);
printf("PMI test ran successfully, for task %d, %s\n",
pmi_rank, tv_str);
}
if (pmi_rank == 0) {
printf("NOTE: All failures reported, ");
printf("but only first four successes reported\n");
}
exit(0);
}
/*** MODEX SECTION ***/
static int modex(opal_list_t *procs)
{
int rc;
char *rml_uri, *attr;
orte_vpid_t v;
orte_process_name_t name;
OPAL_OUTPUT_VERBOSE((1, orte_grpcomm_base.output,
"%s grpcomm:pmi: modex entered",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME)));
/* add our oob endpoint info so that oob communications
* can be supported
*/
rml_uri = orte_rml.get_contact_info();
if (strlen(rml_uri) > (size_t)pmi_vallen_max) {
opal_output(0, "grpcomm:pmi: RML uri length is too long\n");
return ORTE_ERROR;
}
if (0 > asprintf(&attr, "%s-RMLURI", ORTE_NAME_PRINT(ORTE_PROC_MY_NAME))) {
free(rml_uri);
return ORTE_ERR_OUT_OF_RESOURCE;
}
rc = PMI_KVS_Put(pmi_kvs_name, attr, rml_uri);
if (PMI_SUCCESS != rc) {
ORTE_PMI_ERROR(rc, "PMI_KVS_Put");
free(rml_uri);
free(attr);
return ORTE_ERROR;
}
free(rml_uri);
free(attr);
/* commit our modex info */
if (PMI_SUCCESS != (rc = PMI_KVS_Commit(pmi_kvs_name))) {
ORTE_PMI_ERROR(rc, "PMI_KVS_Commit failed");
return ORTE_ERROR;
}
/* Barrier here to ensure all other procs have committed */
if (ORTE_SUCCESS != (rc = pmi_barrier())) {
return rc;
}
/* harvest the oob endpoint info for all other procs
* in our job so oob wireup can be completed
*/
rml_uri = malloc(pmi_vallen_max);
if (NULL == rml_uri) {
return ORTE_ERR_OUT_OF_RESOURCE;
}
name.jobid = ORTE_PROC_MY_NAME->jobid;
for (v=0; v < orte_process_info.num_procs; v++) {
if (v == ORTE_PROC_MY_NAME->vpid) {
continue;
}
name.vpid = v;
if (0 > asprintf(&attr, "%s-RMLURI", ORTE_NAME_PRINT(&name))) {
free(rml_uri);
return ORTE_ERR_OUT_OF_RESOURCE;
}
rc = PMI_KVS_Get(pmi_kvs_name, attr, rml_uri, pmi_vallen_max);
if (PMI_SUCCESS != rc) {
ORTE_PMI_ERROR(rc, "PMI_KVS_Get");
free(rml_uri);
free(attr);
return ORTE_ERROR;
}
free(attr);
OPAL_OUTPUT_VERBOSE((5, orte_grpcomm_base.output,
"%s grpcomm:pmi: proc %s oob endpoint %s",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME),
ORTE_NAME_PRINT(&name), rml_uri));
/* set the contact info into the hash table */
if (ORTE_SUCCESS != (rc = orte_rml.set_contact_info(rml_uri))) {
free(rml_uri);
return rc;
}
}
free(rml_uri);
OPAL_OUTPUT_VERBOSE((1, orte_grpcomm_base.output,
"%s grpcomm:pmi: modex completed",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME)));
return rc;
}
int MPIDU_bc_table_create(int rank, int size, int *nodemap, void *bc, int bc_len, int same_len,
int roots_only, void **bc_table, size_t ** bc_indices)
{
int rc, mpi_errno = MPI_SUCCESS;
int start, end, i;
int key_max, val_max, name_max, out_len, rem;
char *kvsname = NULL, *key = NULL, *val = NULL, *val_p;
int local_rank = -1, local_leader = -1;
size_t my_bc_len = bc_len;
MPIR_NODEMAP_get_local_info(rank, size, nodemap, &local_size, &local_rank, &local_leader);
rc = PMI_KVS_Get_name_length_max(&name_max);
MPIR_ERR_CHKANDJUMP(rc, mpi_errno, MPI_ERR_OTHER, "**pmi_kvs_get_name_length_max");
rc = PMI_KVS_Get_key_length_max(&key_max);
MPIR_ERR_CHKANDJUMP(rc, mpi_errno, MPI_ERR_OTHER, "**pmi_kvs_get_key_length_max");
rc = PMI_KVS_Get_value_length_max(&val_max);
MPIR_ERR_CHKANDJUMP(rc, mpi_errno, MPI_ERR_OTHER, "**pmi_kvs_get_value_length_max");
/* if business cards can be different length, use the max value length */
if (!same_len)
bc_len = val_max;
mpi_errno = MPIDU_shm_seg_alloc(bc_len * size, (void **) &segment, MPL_MEM_ADDRESS);
if (mpi_errno)
MPIR_ERR_POP(mpi_errno);
mpi_errno =
MPIDU_shm_seg_commit(&memory, &barrier, local_size, local_rank, local_leader, rank,
MPL_MEM_ADDRESS);
if (mpi_errno)
MPIR_ERR_POP(mpi_errno);
if (size == 1) {
memcpy(segment, bc, my_bc_len);
goto single;
}
kvsname = MPL_malloc(name_max, MPL_MEM_ADDRESS);
MPIR_Assert(kvsname);
rc = PMI_KVS_Get_my_name(kvsname, name_max);
MPIR_ERR_CHKANDJUMP(rc, mpi_errno, MPI_ERR_OTHER, "**pmi_kvs_get_my_name");
val = MPL_malloc(val_max, MPL_MEM_ADDRESS);
memset(val, 0, val_max);
val_p = val;
rem = val_max;
rc = MPL_str_add_binary_arg(&val_p, &rem, "mpi", (char *) bc, my_bc_len);
MPIR_ERR_CHKANDJUMP(rc, mpi_errno, MPI_ERR_OTHER, "**buscard");
MPIR_Assert(rem >= 0);
key = MPL_malloc(key_max, MPL_MEM_ADDRESS);
MPIR_Assert(key);
if (!roots_only || rank == local_leader) {
sprintf(key, "bc-%d", rank);
rc = PMI_KVS_Put(kvsname, key, val);
MPIR_ERR_CHKANDJUMP(rc, mpi_errno, MPI_ERR_OTHER, "**pmi_kvs_put");
rc = PMI_KVS_Commit(kvsname);
MPIR_ERR_CHKANDJUMP(rc, mpi_errno, MPI_ERR_OTHER, "**pmi_kvs_commit");
}
rc = PMI_Barrier();
MPIR_ERR_CHKANDJUMP(rc, mpi_errno, MPI_ERR_OTHER, "**pmi_barrier");
if (!roots_only) {
start = local_rank * (size / local_size);
end = start + (size / local_size);
if (local_rank == local_size - 1)
end += size % local_size;
for (i = start; i < end; i++) {
sprintf(key, "bc-%d", i);
rc = PMI_KVS_Get(kvsname, key, val, val_max);
MPIR_ERR_CHKANDJUMP(rc, mpi_errno, MPI_ERR_OTHER, "**pmi_kvs_get");
rc = MPL_str_get_binary_arg(val, "mpi", &segment[i * bc_len], bc_len, &out_len);
MPIR_ERR_CHKANDJUMP(rc, mpi_errno, MPI_ERR_OTHER, "**argstr_missinghost");
}
} else {
int num_nodes, *node_roots;
MPIR_NODEMAP_get_node_roots(nodemap, size, &node_roots, &num_nodes);
start = local_rank * (num_nodes / local_size);
end = start + (num_nodes / local_size);
if (local_rank == local_size - 1)
end += num_nodes % local_size;
for (i = start; i < end; i++) {
sprintf(key, "bc-%d", node_roots[i]);
rc = PMI_KVS_Get(kvsname, key, val, val_max);
MPIR_ERR_CHKANDJUMP(rc, mpi_errno, MPI_ERR_OTHER, "**pmi_kvs_get");
rc = MPL_str_get_binary_arg(val, "mpi", &segment[i * bc_len], bc_len, &out_len);
MPIR_ERR_CHKANDJUMP(rc, mpi_errno, MPI_ERR_OTHER, "**argstr_missinghost");
}
MPL_free(node_roots);
}
mpi_errno = MPIDU_shm_barrier(barrier, local_size);
if (mpi_errno)
MPIR_ERR_POP(mpi_errno);
single:
if (!same_len) {
indices = MPL_malloc(size * sizeof(size_t), MPL_MEM_ADDRESS);
MPIR_Assert(indices);
for (i = 0; i < size; i++)
indices[i] = bc_len * i;
*bc_indices = indices;
}
fn_exit:
MPL_free(kvsname);
MPL_free(key);
MPL_free(val);
*bc_table = segment;
return mpi_errno;
fn_fail:
goto fn_exit;
}