/* gasneti_bootstrapFini
*/
void gasneti_bootstrapFini_pmi(void) {
#if USE_PMI2_API
PMI2_Finalize();
#else
if (PMI_FALSE == gasneti_pmi_initialized) {
PMI_Finalize();
}
#endif
}
void mca_common_pmi_finalize (void) {
if (0 == mca_common_pmi_init_count) {
return;
}
if (0 == --mca_common_pmi_init_count) {
#if WANT_PMI2_SUPPORT
PMI2_Finalize ();
#else
PMI_Finalize ();
#endif
}
}
static int s1_fini(void) {
if (0 == pmix_init_count) {
return OPAL_SUCCESS;
}
if (0 == --pmix_init_count) {
PMI_Finalize ();
}
if (NULL != s1_lranks) {
free(s1_lranks);
}
return OPAL_SUCCESS;
}
static int s1_fini(void) {
if (0 == pmix_init_count) {
return OPAL_SUCCESS;
}
if (0 == --pmix_init_count) {
PMI_Finalize ();
}
// teardown hash table
opal_pmix_base_hash_finalize();
return OPAL_SUCCESS;
}
/*
* No cleanup necessary here.
*/
int pmgr_finalize()
{
/* shut down the PMI library if we're using it */
if (mpirun_pmi_enable) {
#ifdef HAVE_PMI
if (PMI_Finalize() != PMI_SUCCESS) {
pmgr_error("Failed to finalize PMI library @ file %s:%d",
__FILE__, __LINE__
);
}
#endif /* ifdef HAVE_PMI */
}
pmgr_free(mpirun_hostname);
return PMGR_SUCCESS;
}
void mca_common_pmi_finalize (void) {
if (0 == mca_common_pmi_init_count) {
return;
}
if (0 == --mca_common_pmi_init_count) {
#if WANT_PMI2_SUPPORT
if( mca_common_pmi_version == 2){
PMI2_Finalize ();
}
else
#endif
{
PMI_Finalize ();
}
}
}
int main(int argc, char **argv)
{
int spawned;
int rc;
/* init us */
if (PMI_SUCCESS != (rc = PMI_Init(&spawned))) {
fprintf(stderr, "PMI_Init failed: %d\n", rc);
return rc;
}
#if 0
key = "local-key";
PMIX_VAL_SET(&value, int, 12345, rc, kvp_error );
if (PMIX_SUCCESS != (rc = PMIx_Put(PMIX_LOCAL, key, &value))) {
fprintf(stderr, "PMIx_Put failed: %d\n", rc);
}
key = "remote-key";
char *ptr = "Test string";
PMIX_VAL_SET(&value, string, ptr, rc, kvp_error );
if (PMIX_SUCCESS != (rc = PMIx_Put(PMIX_REMOTE, key, &value))) {
fprintf(stderr, "PMIx_Put failed: %d\n", rc);
}
key = "global-key";
PMIX_VAL_SET(&value, float, 10.15, rc, kvp_error );
if (PMIX_SUCCESS != (rc = PMIx_Put(PMIX_GLOBAL, key, &value))) {
fprintf(stderr, "PMIx_Put failed: %d\n", rc);
}
/* Submit the data */
pmix_range_t range;
range.ranks = NULL;
range.nranks = 0;
if (PMIX_SUCCESS != (rc = PMIx_Fence(NULL, 0))) {
fprintf(stderr, "PMIx_Fence failed: %d\n", rc);
return rc;
}
#endif
/* finalize us */
if (PMI_SUCCESS != (rc = PMI_Finalize())) {
fprintf(stderr, "PMI_Finalize failed: %d\n", rc);
}
return rc;
}
int orte_grpcomm_pmi_close(void)
{
#if WANT_CRAY_PMI2_EXT
if (started_by_me && PMI2_Initialized()) {
PMI2_Finalize();
}
#else
PMI_BOOL initialized;
/* if we weren't selected, cleanup */
if (started_by_me &&
PMI_SUCCESS == PMI_Initialized(&initialized) &&
PMI_TRUE == initialized) {
PMI_Finalize();
}
#endif
return ORTE_SUCCESS;
}
static int s1_init(void)
{
PMI_BOOL initialized;
int spawned;
int rc, ret = OPAL_ERROR;
int i, rank, lrank, nrank;
char *pmix_id, tmp[64];
opal_value_t kv;
char *str;
uint32_t ui32;
opal_process_name_t ldr;
char **localranks=NULL;
if (PMI_SUCCESS != (rc = PMI_Initialized(&initialized))) {
OPAL_PMI_ERROR(rc, "PMI_Initialized");
return OPAL_ERROR;
}
if (PMI_TRUE != initialized && PMI_SUCCESS != (rc = PMI_Init(&spawned))) {
OPAL_PMI_ERROR(rc, "PMI_Init");
return OPAL_ERROR;
}
// setup hash table
opal_pmix_base_hash_init();
// Initialize space demands
rc = PMI_KVS_Get_value_length_max(&pmix_vallen_max);
if (PMI_SUCCESS != rc) {
OPAL_PMI_ERROR(rc, "PMI_KVS_Get_value_length_max");
goto err_exit;
}
pmix_vallen_threshold = pmix_vallen_max * 3;
pmix_vallen_threshold >>= 2;
rc = PMI_KVS_Get_name_length_max(&pmix_kvslen_max);
if (PMI_SUCCESS != rc) {
OPAL_PMI_ERROR(rc, "PMI_KVS_Get_name_length_max");
goto err_exit;
}
rc = PMI_KVS_Get_key_length_max(&pmix_keylen_max);
if (PMI_SUCCESS != rc) {
OPAL_PMI_ERROR(rc, "PMI_KVS_Get_key_length_max");
goto err_exit;
}
// Initialize job environment information
pmix_id = (char*)malloc(pmix_vallen_max);
if (pmix_id == NULL) {
ret = OPAL_ERR_OUT_OF_RESOURCE;
goto err_exit;
}
/* Get domain id */
if (PMI_SUCCESS != (rc = PMI_Get_kvs_domain_id(pmix_id, pmix_vallen_max))) {
free(pmix_id);
goto err_exit;
}
/* get our rank */
ret = PMI_Get_rank(&rank);
if( PMI_SUCCESS != ret ) {
OPAL_PMI_ERROR(ret, "PMI_Get_rank");
free(pmix_id);
goto err_exit;
}
/* Slurm PMI provides the job id as an integer followed
* by a '.', followed by essentially a stepid. The first integer
* defines an overall job number. The second integer is the number of
* individual jobs we have run within that allocation. */
s1_pname.jobid = strtoul(pmix_id, &str, 10);
s1_pname.jobid = (s1_pname.jobid << 16) & 0xffff0000;
if (NULL != str) {
ui32 = strtoul(str, NULL, 10);
s1_pname.jobid |= (ui32 & 0x0000ffff);
}
ldr.jobid = s1_pname.jobid;
s1_pname.vpid = rank;
/* store our name in the opal_proc_t so that
* debug messages will make sense - an upper
* layer will eventually overwrite it, but that
* won't do any harm */
opal_proc_set_name(&s1_pname);
opal_output_verbose(2, opal_pmix_base_framework.framework_output,
"%s pmix:s1: assigned tmp name",
OPAL_NAME_PRINT(s1_pname));
OBJ_CONSTRUCT(&kv, opal_value_t);
kv.key = strdup(OPAL_PMIX_JOBID);
kv.type = OPAL_UINT32;
kv.data.uint32 = s1_pname.jobid;
if (OPAL_SUCCESS != (ret = opal_pmix_base_store(&OPAL_PROC_MY_NAME, &kv))) {
OPAL_ERROR_LOG(ret);
OBJ_DESTRUCT(&kv);
goto err_exit;
}
OBJ_DESTRUCT(&kv);
/* save it */
OBJ_CONSTRUCT(&kv, opal_value_t);
kv.key = strdup(OPAL_PMIX_RANK);
kv.type = OPAL_UINT32;
kv.data.uint32 = rank;
if (OPAL_SUCCESS != (ret = opal_pmix_base_store(&OPAL_PROC_MY_NAME, &kv))) {
OPAL_ERROR_LOG(ret);
OBJ_DESTRUCT(&kv);
goto err_exit;
}
OBJ_DESTRUCT(&kv);
pmix_kvs_name = (char*)malloc(pmix_kvslen_max);
if (pmix_kvs_name == NULL) {
ret = OPAL_ERR_OUT_OF_RESOURCE;
goto err_exit;
}
rc = PMI_KVS_Get_my_name(pmix_kvs_name, pmix_kvslen_max);
if (PMI_SUCCESS != rc) {
OPAL_PMI_ERROR(rc, "PMI_KVS_Get_my_name");
goto err_exit;
}
/* get our local proc info to find our local rank */
if (PMI_SUCCESS != (rc = PMI_Get_clique_size(&nlranks))) {
OPAL_PMI_ERROR(rc, "PMI_Get_clique_size");
return rc;
}
/* save the local size */
OBJ_CONSTRUCT(&kv, opal_value_t);
kv.key = strdup(OPAL_PMIX_LOCAL_SIZE);
kv.type = OPAL_UINT32;
kv.data.uint32 = nlranks;
if (OPAL_SUCCESS != (ret = opal_pmix_base_store(&OPAL_PROC_MY_NAME, &kv))) {
OPAL_ERROR_LOG(ret);
OBJ_DESTRUCT(&kv);
goto err_exit;
}
OBJ_DESTRUCT(&kv);
lrank = 0;
nrank = 0;
ldr.vpid = rank;
if (0 < nlranks) {
/* now get the specific ranks */
lranks = (int*)calloc(nlranks, sizeof(int));
if (NULL == lranks) {
rc = OPAL_ERR_OUT_OF_RESOURCE;
OPAL_ERROR_LOG(rc);
return rc;
}
if (PMI_SUCCESS != (rc = PMI_Get_clique_ranks(lranks, nlranks))) {
OPAL_PMI_ERROR(rc, "PMI_Get_clique_ranks");
free(lranks);
return rc;
}
/* note the local ldr */
ldr.vpid = lranks[0];
/* save this */
memset(tmp, 0, 64);
for (i=0; i < nlranks; i++) {
(void)snprintf(tmp, 64, "%d", lranks[i]);
opal_argv_append_nosize(&localranks, tmp);
if (rank == lranks[i]) {
lrank = i;
nrank = i;
}
}
str = opal_argv_join(localranks, ',');
opal_argv_free(localranks);
OBJ_CONSTRUCT(&kv, opal_value_t);
kv.key = strdup(OPAL_PMIX_LOCAL_PEERS);
kv.type = OPAL_STRING;
kv.data.string = str;
if (OPAL_SUCCESS != (ret = opal_pmix_base_store(&OPAL_PROC_MY_NAME, &kv))) {
OPAL_ERROR_LOG(ret);
OBJ_DESTRUCT(&kv);
goto err_exit;
}
OBJ_DESTRUCT(&kv);
}
/* save the local leader */
OBJ_CONSTRUCT(&kv, opal_value_t);
kv.key = strdup(OPAL_PMIX_LOCALLDR);
kv.type = OPAL_UINT64;
kv.data.uint64 = *(uint64_t*)&ldr;
if (OPAL_SUCCESS != (ret = opal_pmix_base_store(&OPAL_PROC_MY_NAME, &kv))) {
OPAL_ERROR_LOG(ret);
OBJ_DESTRUCT(&kv);
goto err_exit;
}
OBJ_DESTRUCT(&kv);
/* save our local rank */
OBJ_CONSTRUCT(&kv, opal_value_t);
kv.key = strdup(OPAL_PMIX_LOCAL_RANK);
kv.type = OPAL_UINT16;
kv.data.uint16 = lrank;
if (OPAL_SUCCESS != (ret = opal_pmix_base_store(&OPAL_PROC_MY_NAME, &kv))) {
OPAL_ERROR_LOG(ret);
OBJ_DESTRUCT(&kv);
goto err_exit;
}
OBJ_DESTRUCT(&kv);
/* and our node rank */
OBJ_CONSTRUCT(&kv, opal_value_t);
kv.key = strdup(OPAL_PMIX_NODE_RANK);
kv.type = OPAL_UINT16;
kv.data.uint16 = nrank;
if (OPAL_SUCCESS != (ret = opal_pmix_base_store(&OPAL_PROC_MY_NAME, &kv))) {
OPAL_ERROR_LOG(ret);
OBJ_DESTRUCT(&kv);
goto err_exit;
}
OBJ_DESTRUCT(&kv);
/* get universe size */
ret = PMI_Get_universe_size(&i);
if (PMI_SUCCESS != ret) {
OPAL_PMI_ERROR(ret, "PMI_Get_universe_size");
goto err_exit;
}
/* push this into the dstore for subsequent fetches */
OBJ_CONSTRUCT(&kv, opal_value_t);
kv.key = strdup(OPAL_PMIX_UNIV_SIZE);
kv.type = OPAL_UINT32;
kv.data.uint32 = i;
if (OPAL_SUCCESS != (ret = opal_pmix_base_store(&OPAL_PROC_MY_NAME, &kv))) {
OPAL_ERROR_LOG(ret);
OBJ_DESTRUCT(&kv);
goto err_exit;
}
OBJ_DESTRUCT(&kv);
/* get job size */
ret = PMI_Get_size(&i);
if (PMI_SUCCESS != ret) {
OPAL_PMI_ERROR(ret, "PMI_Get_size");
goto err_exit;
}
OBJ_CONSTRUCT(&kv, opal_value_t);
kv.key = strdup(OPAL_PMIX_JOB_SIZE);
kv.type = OPAL_UINT32;
kv.data.uint32 = i;
if (OPAL_SUCCESS != (ret = opal_pmix_base_store(&OPAL_PROC_MY_NAME, &kv))) {
OPAL_ERROR_LOG(ret);
OBJ_DESTRUCT(&kv);
goto err_exit;
}
OBJ_DESTRUCT(&kv);
/* get appnum */
ret = PMI_Get_appnum(&i);
if (PMI_SUCCESS != ret) {
OPAL_PMI_ERROR(ret, "PMI_Get_appnum");
goto err_exit;
}
OBJ_CONSTRUCT(&kv, opal_value_t);
kv.key = strdup(OPAL_PMIX_APPNUM);
kv.type = OPAL_UINT32;
kv.data.uint32 = i;
if (OPAL_SUCCESS != (ret = opal_pmix_base_store(&OPAL_PROC_MY_NAME, &kv))) {
OPAL_ERROR_LOG(ret);
OBJ_DESTRUCT(&kv);
goto err_exit;
}
OBJ_DESTRUCT(&kv);
/* increment the init count */
++pmix_init_count;
return OPAL_SUCCESS;
err_exit:
PMI_Finalize();
return ret;
}
static int mca_initialize_pmi_v1(void)
{
PMI_BOOL initialized;
int spawned;
int rc, ret = OPAL_ERROR;
if (PMI_SUCCESS != (rc = PMI_Initialized(&initialized))) {
OPAL_PMI_ERROR(rc, "PMI_Initialized");
return OPAL_ERROR;
}
if( PMI_TRUE != initialized && PMI_SUCCESS != (rc = PMI_Init(&spawned)) ) {
OPAL_PMI_ERROR(rc, "PMI_Init");
return OPAL_ERROR;
}
// Initialize space demands
rc = PMI_KVS_Get_value_length_max(&pmi_vallen_max);
if( PMI_SUCCESS != rc ) {
OPAL_PMI_ERROR(rc, "PMI_KVS_Get_value_length_max");
goto err_exit;
}
rc = PMI_KVS_Get_name_length_max(&pmi_kvslen_max);
if (PMI_SUCCESS != rc ) {
OPAL_PMI_ERROR(rc, "PMI_KVS_Get_name_length_max");
goto err_exit;
}
rc = PMI_KVS_Get_key_length_max(&pmi_keylen_max);
if( PMI_SUCCESS != rc ) {
OPAL_PMI_ERROR(rc, "PMI_KVS_Get_key_length_max");
goto err_exit;
}
// Initialize job environment information
rc = PMI_Get_rank(&pmi_rank);
if( PMI_SUCCESS != rc ) {
OPAL_PMI_ERROR(rc, "PMI_Get_rank");
return OPAL_ERROR;
}
rc = PMI_Get_universe_size(&pmi_usize);
if( PMI_SUCCESS != rc ) {
OPAL_PMI_ERROR(rc, "PMI_Get_universe_size");
goto err_exit;
}
rc = PMI_Get_size(&pmi_size);
if( PMI_SUCCESS != rc ) {
OPAL_PMI_ERROR(rc, "PMI_Get_size");
goto err_exit;
}
rc = PMI_Get_appnum(&pmi_appnum);
if( PMI_SUCCESS != rc ) {
OPAL_PMI_ERROR(rc, "PMI_Get_appnum");
goto err_exit;
}
pmi_kvs_name = (char*)malloc(pmi_kvslen_max);
if( pmi_kvs_name == NULL ){
ret = OPAL_ERR_OUT_OF_RESOURCE;
goto err_exit;
}
rc = PMI_KVS_Get_my_name(pmi_kvs_name,pmi_kvslen_max);
if( PMI_SUCCESS != rc ) {
OPAL_PMI_ERROR(rc, "PMI2_Job_GetId");
goto err_exit;
}
return OPAL_SUCCESS;
err_exit:
PMI_Finalize();
return ret;
}
/*@
MPIDI_PG_Finalize - Finalize the process groups, including freeing all
process group structures
@*/
int MPIDI_PG_Finalize(void)
{
int mpi_errno = MPI_SUCCESS;
MPIDI_PG_t *pg, *pgNext;
MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_MPIDI_PG_FINALIZE);
MPIR_FUNC_VERBOSE_ENTER(MPID_STATE_MPIDI_PG_FINALIZE);
/* Print the state of the process groups */
if (verbose) {
MPIU_PG_Printall( stdout );
}
/* FIXME - straighten out the use of PMI_Finalize - no use after
PG_Finalize */
if (pg_world->connData) {
#ifdef USE_PMI2_API
mpi_errno = PMI2_Finalize();
if (mpi_errno) MPIR_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**ch3|pmi_finalize");
#else
int rc;
rc = PMI_Finalize();
if (rc) {
MPIR_ERR_SET1(mpi_errno,MPI_ERR_OTHER,
"**ch3|pmi_finalize",
"**ch3|pmi_finalize %d", rc);
}
#endif
}
/* Free the storage associated with the process groups */
pg = MPIDI_PG_list;
while (pg) {
pgNext = pg->next;
/* In finalize, we free all process group information, even if
the ref count is not zero. This can happen if the user
fails to use MPI_Comm_disconnect on communicators that
were created with the dynamic process routines.*/
/* XXX DJG FIXME-MT should we be checking this? */
if (MPIR_Object_get_ref(pg) == 0 || 1) {
if (pg == MPIDI_Process.my_pg)
MPIDI_Process.my_pg = NULL;
MPIR_Object_set_ref(pg, 0); /* satisfy assertions in PG_Destroy */
MPIDI_PG_Destroy( pg );
}
pg = pgNext;
}
/* If COMM_WORLD is still around (it normally should be),
try to free it here. The reason that we need to free it at this
point is that comm_world (and comm_self) still exist, and
hence the usual process to free the related VC structures will
not be invoked. */
if (MPIDI_Process.my_pg) {
MPIDI_PG_Destroy(MPIDI_Process.my_pg);
}
MPIDI_Process.my_pg = NULL;
/* ifdefing out this check because the list will not be NULL in
Ch3_finalize because
one additional reference is retained in MPIDI_Process.my_pg.
That reference is released
only after ch3_finalize returns. If I release it before ch3_finalize,
the ssm channel crashes. */
#if 0
if (MPIDI_PG_list != NULL)
{
/* --BEGIN ERROR HANDLING-- */
mpi_errno = MPIR_Err_create_code(MPI_SUCCESS, MPIR_ERR_FATAL, FCNAME, __LINE__, MPI_ERR_INTERN,
"**dev|pg_finalize|list_not_empty", NULL);
/* --END ERROR HANDLING-- */
}
#endif
MPIR_FUNC_VERBOSE_EXIT(MPID_STATE_MPIDI_PG_FINALIZE);
return mpi_errno;
}
ReportHook(POST_ALL)(struct criterion_global_stats *stats)
{
if (pmi_initialized == 1)
PMI_Finalize();
}
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);
}
int main(int argc, char **argv, char **envp)
{
int pmi_rank = -1;
int pmi_process_group_size = -1;
int num_local_procs = 0;
int *local_rank_ids = NULL;
int spawned = PMI_FALSE;
int rc = EXIT_SUCCESS;
char *err = NULL;
PMI_BOOL pmi_initialized = PMI_FALSE;
int pmi_vallen_max, max_length;
char *pmi_kvs_name;
/* sanity */
if (PMI_SUCCESS != PMI_Initialized(&pmi_initialized) ||
PMI_TRUE == pmi_initialized) {
fprintf(stderr, "=== ERROR: PMI sanity failure\n");
return EXIT_FAILURE;
}
if (PMI_SUCCESS != PMI_Init(&spawned)) {
err = "PMI_Init failure!";
goto done;
}
if (PMI_SUCCESS != PMI_Get_size(&pmi_process_group_size)) {
err = "PMI_Get_size failure!";
goto done;
}
if (PMI_SUCCESS != PMI_Get_rank(&pmi_rank)) {
err = "PMI_Get_rank failure!";
goto done;
}
if (PMI_SUCCESS != PMI_Get_clique_size(&num_local_procs)) {
err = "PMI_Get_clique_size failure!";
goto done;
}
if (PMI_SUCCESS != PMI_KVS_Get_value_length_max(&pmi_vallen_max)) {
err = "PMI_KVS_Get_value_length_max failure!";
goto done;
}
if (PMI_SUCCESS != PMI_KVS_Get_name_length_max(&max_length)) {
err = "PMI_KVS_Get_name_length_max failure!";
goto done;
}
pmi_kvs_name = (char*)malloc(max_length);
if (NULL == pmi_kvs_name) {
err = "malloc failure!";
goto done;
}
if (PMI_SUCCESS != PMI_KVS_Get_my_name(pmi_kvs_name,max_length)) {
err = "PMI_KVS_Get_my_name failure!";
goto done;
}
if (NULL == (local_rank_ids = calloc(num_local_procs, sizeof(int)))) {
err = "out of resources";
goto done;
}
if (PMI_SUCCESS != PMI_Get_clique_ranks(local_rank_ids, num_local_procs)) {
err = "PMI_Get_clique_size failure!";
goto done;
}
/* lowest local rank will print env info and tag its output*/
// if (pmi_rank == local_rank_ids[0]) {
// for (; NULL != envp && NULL != *envp; ++envp) {
// printf("===[%d]: %s\n", pmi_rank, *envp);
// }
//}
done:
if (PMI_TRUE == pmi_initialized) {
if (PMI_SUCCESS != PMI_Finalize()) {
err = "PMI_Finalize failure!";
}
}
if (NULL != err) {
fprintf(stderr, "=== ERROR [rank:%d] %s\n", pmi_rank, err);
rc = EXIT_FAILURE;
}
return rc;
}
int main(int argc, char **argv)
{
int ret = 0;
int rc;
char *str = NULL;
int ti = (argc > 1 ? atoi(argv[1]) : 0);
srand(time(NULL));
str = getenv("VERBOSE");
_verbose = (str ? atoi(str) : _verbose);
spawned = random_value(10, 20);
size = random_value(10, 20);
rank = random_value(10, 20);
appnum = random_value(10, 20);
if (PMI_SUCCESS != (rc = PMI_Init(&spawned))) {
log_fatal("PMI_Init failed: %d\n", rc);
return rc;
}
str = getenv("PMIX_NAMESPACE");
_legacy = (str ? 0 : 1);
/* this test should be always run */
if (1) {
rc = test_item1();
ret += (rc ? 1 : 0);
log_info("TI1 : %s\n", (rc ? "FAIL" : "PASS"));
}
if (!ti || 2 == ti) {
rc = test_item2();
ret += (rc ? 1 : 0);
log_info("TI2 : %s\n", (rc ? "FAIL" : "PASS"));
}
if (!ti || 3 == ti) {
rc = test_item3();
ret += (rc ? 1 : 0);
log_info("TI3 : %s\n", (rc ? "FAIL" : "PASS"));
}
if (!ti || 4 == ti) {
rc = test_item4();
ret += (rc ? 1 : 0);
log_info("TI4 : %s\n", (rc ? "FAIL" : "PASS"));
}
if (!ti || 5 == ti) {
rc = test_item5();
ret += (rc ? 1 : 0);
log_info("TI5 : %s\n", (rc ? "FAIL" : "PASS"));
}
if (!ti || 6 == ti) {
rc = test_item6();
ret += (rc ? 1 : 0);
log_info("TI6 : %s\n", (rc ? "FAIL" : "PASS"));
}
if (!ti || 7 == ti) {
rc = test_item7();
ret += (rc ? 1 : 0);
log_info("TI7 : %s\n", (rc ? "FAIL" : "PASS"));
}
if (PMI_SUCCESS != (rc = PMI_Finalize())) {
log_fatal("PMI_Finalize failed: %d\n", rc);
return rc;
}
return ret;
}
static int s1_init(void)
{
PMI_BOOL initialized;
int spawned;
int rc, ret = OPAL_ERROR;
int i;
char *pmix_id, *tmp;
uint32_t jobfam, stepid;
opal_value_t kv;
if (PMI_SUCCESS != (rc = PMI_Initialized(&initialized))) {
OPAL_PMI_ERROR(rc, "PMI_Initialized");
return OPAL_ERROR;
}
if( PMI_TRUE != initialized && PMI_SUCCESS != (rc = PMI_Init(&spawned)) ) {
OPAL_PMI_ERROR(rc, "PMI_Init");
return OPAL_ERROR;
}
// Initialize space demands
rc = PMI_KVS_Get_value_length_max(&pmix_vallen_max);
if( PMI_SUCCESS != rc ) {
OPAL_PMI_ERROR(rc, "PMI_KVS_Get_value_length_max");
goto err_exit;
}
rc = PMI_KVS_Get_name_length_max(&pmix_kvslen_max);
if (PMI_SUCCESS != rc ) {
OPAL_PMI_ERROR(rc, "PMI_KVS_Get_name_length_max");
goto err_exit;
}
rc = PMI_KVS_Get_key_length_max(&pmix_keylen_max);
if( PMI_SUCCESS != rc ) {
OPAL_PMI_ERROR(rc, "PMI_KVS_Get_key_length_max");
goto err_exit;
}
// Initialize job environment information
pmix_id = (char*)malloc(pmix_vallen_max);
if( pmix_id == NULL ){
ret = OPAL_ERR_OUT_OF_RESOURCE;
goto err_exit;
}
/* Get domain id */
if (PMI_SUCCESS != (rc = PMI_Get_kvs_domain_id(pmix_id, pmix_vallen_max))) {
free(pmix_id);
goto err_exit;
}
/* Slurm PMI provides the job id as an integer followed
* by a '.', followed by essentially a stepid. The first integer
* defines an overall job number. The second integer is the number of
* individual jobs we have run within that allocation. So we translate
* this as the overall job number equating to our job family, and
* the individual number equating to our local jobid
*/
jobfam = strtoul(pmix_id, &tmp, 10);
if (NULL == tmp) {
/* hmmm - no '.', so let's just use zero */
stepid = 0;
} else {
tmp++; /* step over the '.' */
stepid = strtoul(tmp, NULL, 10);
}
/* now build the jobid */
s1_jobid = (jobfam << 16) | stepid;
free(pmix_id);
/* get our rank */
ret = PMI_Get_rank(&s1_rank);
if( PMI_SUCCESS != ret ) {
OPAL_PMI_ERROR(ret, "PMI_Get_rank");
goto err_exit;
}
/* store our name in the opal_proc_t so that
* debug messages will make sense - an upper
* layer will eventually overwrite it, but that
* won't do any harm */
s1_pname.jid = s1_jobid;
s1_pname.vid = s1_rank;
opal_proc_set_name((opal_process_name_t*)&s1_pname);
opal_output_verbose(2, opal_pmix_base_framework.framework_output,
"%s pmix:s1: assigned tmp name",
OPAL_NAME_PRINT(*(opal_process_name_t*)&s1_pname));
pmix_kvs_name = (char*)malloc(pmix_kvslen_max);
if( pmix_kvs_name == NULL ){
ret = OPAL_ERR_OUT_OF_RESOURCE;
goto err_exit;
}
rc = PMI_KVS_Get_my_name(pmix_kvs_name, pmix_kvslen_max);
if( PMI_SUCCESS != rc ) {
OPAL_PMI_ERROR(rc, "PMI_KVS_Get_my_name");
goto err_exit;
}
/* get our local proc info to find our local rank */
if (PMI_SUCCESS != (rc = PMI_Get_clique_size(&s1_nlranks))) {
OPAL_PMI_ERROR(rc, "PMI_Get_clique_size");
return rc;
}
/* now get the specific ranks */
s1_lranks = (int*)calloc(s1_nlranks, sizeof(int));
if (NULL == s1_lranks) {
rc = OPAL_ERR_OUT_OF_RESOURCE;
OPAL_ERROR_LOG(rc);
return rc;
}
if (PMI_SUCCESS != (rc = PMI_Get_clique_ranks(s1_lranks, s1_nlranks))) {
OPAL_PMI_ERROR(rc, "PMI_Get_clique_ranks");
free(s1_lranks);
return rc;
}
/* find ourselves */
for (i=0; i < s1_nlranks; i++) {
if (s1_rank == s1_lranks[i]) {
s1_lrank = i;
s1_nrank = i;
break;
}
}
/* get universe size */
ret = PMI_Get_universe_size(&s1_usize);
if (PMI_SUCCESS != ret) {
OPAL_PMI_ERROR(ret, "PMI_Get_universe_size");
goto err_exit;
}
/* push this into the dstore for subsequent fetches */
OBJ_CONSTRUCT(&kv, opal_value_t);
kv.key = strdup(OPAL_DSTORE_UNIV_SIZE);
kv.type = OPAL_UINT32;
kv.data.uint32 = s1_usize;
if (OPAL_SUCCESS != (ret = opal_dstore.store(opal_dstore_internal, &OPAL_PROC_MY_NAME, &kv))) {
OPAL_ERROR_LOG(ret);
OBJ_DESTRUCT(&kv);
goto err_exit;
}
OBJ_DESTRUCT(&kv);
/* get job size */
ret = PMI_Get_size(&s1_jsize);
if (PMI_SUCCESS != ret) {
OPAL_PMI_ERROR(ret, "PMI_Get_size");
goto err_exit;
}
/* get appnum */
ret = PMI_Get_appnum(&s1_appnum);
if (PMI_SUCCESS != ret) {
OPAL_PMI_ERROR(ret, "PMI_Get_appnum");
goto err_exit;
}
return OPAL_SUCCESS;
err_exit:
PMI_Finalize();
return ret;
}
