static int
_setup_srun_kvs(const mpi_plugin_client_info_t *job)
{
int rc;
rc = temp_kvs_init();
return rc;
}
static int
_setup_stepd_kvs(const stepd_step_rec_t *job, char ***env)
{
int rc = SLURM_SUCCESS, i = 0, pp_cnt = 0;
char *p, env_key[32], *ppkey, *ppval;
kvs_seq = 1;
rc = temp_kvs_init();
if (rc != SLURM_SUCCESS)
return rc;
rc = kvs_init();
if (rc != SLURM_SUCCESS)
return rc;
/* preput */
p = getenvp(*env, PMI2_PREPUT_CNT_ENV);
if (p) {
pp_cnt = atoi(p);
}
for (i = 0; i < pp_cnt; i ++) {
snprintf(env_key, 32, PMI2_PPKEY_ENV"%d", i);
p = getenvp(*env, env_key);
ppkey = p; /* getenvp will not modify p */
snprintf(env_key, 32, PMI2_PPVAL_ENV"%d", i);
p = getenvp(*env, env_key);
ppval = p;
kvs_put(ppkey, ppval);
}
/*
* For PMI11.
* A better logic would be to put PMI_process_mapping in KVS only if
* the task distribution method is not "arbitrary", because in
* "arbitrary" distribution the process mapping varible is not correct.
* MPICH2 may deduce the clique info from the hostnames. But that
* is rather costly.
*/
kvs_put("PMI_process_mapping", job_info.proc_mapping);
return SLURM_SUCCESS;
}
extern int
temp_kvs_send(void)
{
int rc = SLURM_ERROR, retry = 0;
unsigned int delay = 1;
char *nodelist = NULL;
if (!in_stepd()) /* srun */
nodelist = xstrdup(job_info.step_nodelist);
else if (tree_info.parent_node)
nodelist = xstrdup(tree_info.parent_node);
/* cmd included in temp_kvs_buf */
kvs_seq++; /* expecting new kvs after now */
while (1) {
if (retry == 1)
verbose("failed to send temp kvs, rc=%d, retrying", rc);
if (nodelist)
/* srun or non-first-level stepds */
rc = slurm_forward_data(&nodelist,
tree_sock_addr,
temp_kvs_cnt,
temp_kvs_buf);
else /* first level stepds */
rc = tree_msg_to_srun(temp_kvs_cnt, temp_kvs_buf);
if (rc == SLURM_SUCCESS)
break;
if (++retry >= MAX_RETRIES)
break;
/* wait, in case parent stepd / srun not ready */
sleep(delay);
delay *= 2;
}
temp_kvs_init(); /* clear old temp kvs */
xfree(nodelist);
return rc;
}
extern int
temp_kvs_send(void)
{
int rc = SLURM_ERROR, retry = 0;
unsigned int delay = 1;
/* cmd included in temp_kvs_buf */
kvs_seq ++; /* expecting new kvs after now */
while (1) {
if (retry == 1) {
verbose("failed to send temp kvs, rc=%d, retrying", rc);
}
if (! in_stepd()) { /* srun */
rc = tree_msg_to_stepds(job_info.step_nodelist,
temp_kvs_cnt,
temp_kvs_buf);
} else if (tree_info.parent_node != NULL) {
/* non-first-level stepds */
rc = tree_msg_to_stepds(tree_info.parent_node,
temp_kvs_cnt,
temp_kvs_buf);
} else { /* first level stepds */
rc = tree_msg_to_srun(temp_kvs_cnt, temp_kvs_buf);
}
if (rc == SLURM_SUCCESS)
break;
retry ++;
if (retry >= MAX_RETRIES)
break;
/* wait, in case parent stepd / srun not ready */
sleep(delay);
delay *= 2;
}
temp_kvs_init(); /* clear old temp kvs */
return rc;
}
