/* Transmit PMI Keyval space data */
int slurm_send_kvs_comm_set(struct kvs_comm_set *kvs_set_ptr,
int pmi_rank, int pmi_size)
{
slurm_msg_t msg_send;
int rc, retries = 0, timeout = 0;
if (kvs_set_ptr == NULL)
return EINVAL;
if ((rc = _get_addr()) != SLURM_SUCCESS)
return rc;
_set_pmi_time();
slurm_msg_t_init(&msg_send);
msg_send.address = srun_addr;
msg_send.msg_type = PMI_KVS_PUT_REQ;
msg_send.data = (void *) kvs_set_ptr;
/* Send the RPC to the local srun communcation manager.
* Since the srun can be sent thousands of messages at
* the same time and refuse some connections, retry as
* needed. Spread out messages by task's rank. Also
* increase the timeout if many tasks since the srun
* command is very overloaded.
* We also increase the timeout (default timeout is
* 10 secs). */
_delay_rpc(pmi_rank, pmi_size);
if (pmi_size > 4000) /* 240 secs */
timeout = slurm_get_msg_timeout() * 24000;
else if (pmi_size > 1000) /* 120 secs */
timeout = slurm_get_msg_timeout() * 12000;
else if (pmi_size > 100) /* 50 secs */
timeout = slurm_get_msg_timeout() * 5000;
else if (pmi_size > 10) /* 20 secs */
timeout = slurm_get_msg_timeout() * 2000;
while (slurm_send_recv_rc_msg_only_one(&msg_send, &rc, timeout) < 0) {
if (retries++ > MAX_RETRIES) {
error("slurm_send_kvs_comm_set: %m");
return SLURM_ERROR;
} else
debug("send_kvs retry %d", retries);
_delay_rpc(pmi_rank, pmi_size);
}
return rc;
}
/* Wait for barrier and get full PMI Keyval space data */
int slurm_get_kvs_comm_set(struct kvs_comm_set **kvs_set_ptr,
int pmi_rank, int pmi_size)
{
int rc, srun_fd, retries = 0, timeout = 0;
slurm_msg_t msg_send, msg_rcv;
slurm_addr_t slurm_addr, srun_reply_addr;
char hostname[64];
uint16_t port;
kvs_get_msg_t data;
char *env_pmi_ifhn;
if (kvs_set_ptr == NULL)
return EINVAL;
*kvs_set_ptr = NULL; /* initialization */
if ((rc = _get_addr()) != SLURM_SUCCESS) {
error("_get_addr: %m");
return rc;
}
_set_pmi_time();
if (pmi_fd < 0) {
if ((pmi_fd = slurm_init_msg_engine_port(0)) < 0) {
error("slurm_init_msg_engine_port: %m");
return SLURM_ERROR;
}
fd_set_blocking(pmi_fd);
}
if (slurm_get_stream_addr(pmi_fd, &slurm_addr) < 0) {
error("slurm_get_stream_addr: %m");
return SLURM_ERROR;
}
/* hostname is not set here, so slurm_get_addr fails
slurm_get_addr(&slurm_addr, &port, hostname, sizeof(hostname)); */
port = ntohs(slurm_addr.sin_port);
if ((env_pmi_ifhn = getenv("SLURM_PMI_RESP_IFHN"))) {
strncpy(hostname, env_pmi_ifhn, sizeof(hostname));
hostname[sizeof(hostname)-1] = 0;
} else
gethostname_short(hostname, sizeof(hostname));
data.task_id = pmi_rank;
data.size = pmi_size;
data.port = port;
data.hostname = hostname;
slurm_msg_t_init(&msg_send);
slurm_msg_t_init(&msg_rcv);
msg_send.address = srun_addr;
msg_send.msg_type = PMI_KVS_GET_REQ;
msg_send.data = &data;
/* Send the RPC to the local srun communcation manager.
* Since the srun can be sent thousands of messages at
* the same time and refuse some connections, retry as
* needed. Wait until all key-pairs have been sent by
* all tasks then spread out messages by task's rank.
* Also increase the message timeout if many tasks
* since the srun command can get very overloaded (the
* default timeout is 10 secs).
*/
_delay_rpc(pmi_rank, pmi_size);
if (pmi_size > 4000) /* 240 secs */
timeout = slurm_get_msg_timeout() * 24000;
else if (pmi_size > 1000) /* 120 secs */
timeout = slurm_get_msg_timeout() * 12000;
else if (pmi_size > 100) /* 60 secs */
timeout = slurm_get_msg_timeout() * 6000;
else if (pmi_size > 10) /* 20 secs */
timeout = slurm_get_msg_timeout() * 2000;
while (slurm_send_recv_rc_msg_only_one(&msg_send, &rc, timeout) < 0) {
if (retries++ > MAX_RETRIES) {
error("slurm_get_kvs_comm_set: %m");
return SLURM_ERROR;
} else
debug("get kvs retry %d", retries);
_delay_rpc(pmi_rank, pmi_size);
}
if (rc != SLURM_SUCCESS) {
error("slurm_get_kvs_comm_set error_code=%d", rc);
return rc;
}
/* get the message after all tasks reach the barrier */
srun_fd = slurm_accept_msg_conn(pmi_fd, &srun_reply_addr);
if (srun_fd < 0) {
error("slurm_accept_msg_conn: %m");
return errno;
}
while ((rc = slurm_receive_msg(srun_fd, &msg_rcv, timeout)) != 0) {
if (errno == EINTR)
continue;
error("slurm_receive_msg: %m");
slurm_close(srun_fd);
return errno;
}
if (msg_rcv.auth_cred)
(void)g_slurm_auth_destroy(msg_rcv.auth_cred);
if (msg_rcv.msg_type != PMI_KVS_GET_RESP) {
error("slurm_get_kvs_comm_set msg_type=%d", msg_rcv.msg_type);
slurm_close(srun_fd);
return SLURM_UNEXPECTED_MSG_ERROR;
}
if (slurm_send_rc_msg(&msg_rcv, SLURM_SUCCESS) < 0)
error("slurm_send_rc_msg: %m");
slurm_close(srun_fd);
*kvs_set_ptr = msg_rcv.data;
rc = _forward_comm_set(*kvs_set_ptr);
return rc;
}
