extern int launch_p_step_launch(
srun_job_t *job, slurm_step_io_fds_t *cio_fds, uint32_t *global_rc,
slurm_step_launch_callbacks_t *step_callbacks)
{
slurm_step_launch_params_t launch_params;
slurm_step_launch_callbacks_t callbacks;
int rc = 0;
bool first_launch = 0;
slurm_step_launch_params_t_init(&launch_params);
memcpy(&callbacks, step_callbacks, sizeof(callbacks));
if (!task_state) {
task_state = task_state_create(job->ntasks);
local_srun_job = job;
local_global_rc = global_rc;
first_launch = 1;
} else
task_state_alter(task_state, job->ntasks);
launch_params.gid = opt.gid;
launch_params.alias_list = job->alias_list;
launch_params.argc = opt.argc;
launch_params.argv = opt.argv;
launch_params.multi_prog = opt.multi_prog ? true : false;
launch_params.cwd = opt.cwd;
launch_params.slurmd_debug = opt.slurmd_debug;
launch_params.buffered_stdio = !opt.unbuffered;
launch_params.labelio = opt.labelio ? true : false;
launch_params.remote_output_filename =fname_remote_string(job->ofname);
launch_params.remote_input_filename = fname_remote_string(job->ifname);
launch_params.remote_error_filename = fname_remote_string(job->efname);
launch_params.task_prolog = opt.task_prolog;
launch_params.task_epilog = opt.task_epilog;
launch_params.cpu_bind = opt.cpu_bind;
launch_params.cpu_bind_type = opt.cpu_bind_type;
launch_params.mem_bind = opt.mem_bind;
launch_params.mem_bind_type = opt.mem_bind_type;
launch_params.open_mode = opt.open_mode;
if (opt.acctg_freq >= 0)
launch_params.acctg_freq = opt.acctg_freq;
launch_params.pty = opt.pty;
if (opt.cpus_set)
launch_params.cpus_per_task = opt.cpus_per_task;
else
launch_params.cpus_per_task = 1;
launch_params.cpu_freq = opt.cpu_freq;
launch_params.task_dist = opt.distribution;
launch_params.ckpt_dir = opt.ckpt_dir;
launch_params.restart_dir = opt.restart_dir;
launch_params.preserve_env = opt.preserve_env;
launch_params.spank_job_env = opt.spank_job_env;
launch_params.spank_job_env_size = opt.spank_job_env_size;
launch_params.user_managed_io = opt.user_managed_io;
memcpy(&launch_params.local_fds, cio_fds, sizeof(slurm_step_io_fds_t));
if (MPIR_being_debugged) {
launch_params.parallel_debug = true;
pmi_server_max_threads(1);
} else {
launch_params.parallel_debug = false;
}
/* Normally this isn't used, but if an outside process (other
than srun (poe) is using this logic to launch tasks then we
can use this to signal the step.
*/
callbacks.task_start = _task_start;
/* If poe is using this code with multi-prog it always returns
1 for each task which could be confusing since no real
error happened.
*/
if (!launch_params.multi_prog
|| (!callbacks.step_signal
|| (callbacks.step_signal == launch_g_fwd_signal))) {
callbacks.task_finish = _task_finish;
}
mpir_init(job->ctx_params.task_count);
update_job_state(job, SRUN_JOB_LAUNCHING);
launch_start_time = time(NULL);
if (first_launch) {
if (slurm_step_launch(job->step_ctx, &launch_params,
&callbacks) != SLURM_SUCCESS) {
rc = errno;
*local_global_rc = errno;
error("Application launch failed: %m");
slurm_step_launch_abort(job->step_ctx);
slurm_step_launch_wait_finish(job->step_ctx);
goto cleanup;
}
} else {
if (slurm_step_launch_add(job->step_ctx, &launch_params,
job->nodelist, job->fir_nodeid)
!= SLURM_SUCCESS) {
rc = errno;
*local_global_rc = errno;
error("Application launch add failed: %m");
slurm_step_launch_abort(job->step_ctx);
slurm_step_launch_wait_finish(job->step_ctx);
goto cleanup;
}
}
update_job_state(job, SRUN_JOB_STARTING);
if (slurm_step_launch_wait_start(job->step_ctx) == SLURM_SUCCESS) {
update_job_state(job, SRUN_JOB_RUNNING);
/* Only set up MPIR structures if the step launched
* correctly. */
if (opt.multi_prog)
mpir_set_multi_name(job->ctx_params.task_count,
launch_params.argv[0]);
else
mpir_set_executable_names(launch_params.argv[0]);
MPIR_debug_state = MPIR_DEBUG_SPAWNED;
if (opt.debugger_test)
mpir_dump_proctable();
else
MPIR_Breakpoint(job);
} else {
info("Job step %u.%u aborted before step completely launched.",
job->jobid, job->stepid);
}
cleanup:
return rc;
}
int main (int argc, char *argv[])
{
int i, min_nodes = 1, max_nodes = 1, nodes, tasks = 0, rc = 0;
job_desc_msg_t job_req;
resource_allocation_response_msg_t *job_resp;
slurm_step_ctx_params_t step_params[1];
slurm_step_ctx_t *ctx = NULL;
slurm_step_launch_params_t launch[1];
char *task_argv[3];
int *fd_array = NULL;
int num_fd;
if (argc > 1) {
i = atoi(argv[1]);
if (i > 0)
min_nodes = i;
}
if (argc > 2) {
i = atoi(argv[2]);
if (i > 0)
max_nodes = i;
}
if (max_nodes < min_nodes)
max_nodes = min_nodes;
/* Create a job allocation */
slurm_init_job_desc_msg( &job_req );
job_req.min_nodes = min_nodes;
job_req.max_nodes = max_nodes;
job_req.user_id = getuid();
job_req.group_id = getgid();
job_req.time_limit = 1;
if (slurm_allocate_resources(&job_req, &job_resp)) {
slurm_perror ("slurm_allocate_resources");
printf("INFO: min_nodes=%u max_nodes=%u user_id=%u group_id=%u",
job_req.min_nodes, job_req.max_nodes,
job_req.user_id, job_req.group_id);
exit(0);
}
printf("job_id %u\n", job_resp->job_id);
fflush(stdout);
/* Wait for allocation request to be satisfied */
if ((job_resp->node_list == NULL) ||
(strlen(job_resp->node_list) == 0)) {
printf("Waiting for resource allocation\n");
fflush(stdout);
while ((job_resp->node_list == NULL) ||
(strlen(job_resp->node_list) == 0)) {
sleep(5);
if (slurm_allocation_lookup_lite(job_resp->job_id,
&job_resp) &&
(slurm_get_errno() != ESLURM_JOB_PENDING)) {
slurm_perror("slurm_confirm_allocation");
exit(0);
}
}
}
nodes = job_resp->node_cnt;
if (argc > 3)
tasks = atoi(argv[3]);
if (tasks < 1)
tasks = nodes * TASKS_PER_NODE;
if (tasks < nodes) {
fprintf(stderr, "Invalid task count argument\n");
exit(1);
}
printf("Starting %d tasks on %d nodes\n", tasks, nodes);
fflush(stdout);
/*
* Create a job step context.
*/
slurm_step_ctx_params_t_init(step_params);
step_params->job_id = job_resp->job_id;
step_params->min_nodes = nodes;
step_params->task_count = tasks;
ctx = slurm_step_ctx_create(step_params);
if ((ctx == NULL) &&
(slurm_get_errno() == ESLURM_PROLOG_RUNNING)) {
printf("SlurmctldProlog is still running, "
"sleep and try again\n");
sleep(10);
ctx = slurm_step_ctx_create(step_params);
}
if (ctx == NULL) {
slurm_perror("slurm_step_ctx_create");
rc = 1;
goto done;
}
/*
* Hack to run one task per node, regardless of what we set up
* when we created the job step context.
*/
if (slurm_step_ctx_daemon_per_node_hack(ctx) != SLURM_SUCCESS) {
slurm_perror("slurm_step_ctx_daemon_per_node_hack");
rc = 1;
goto done;
}
/*
* Launch the tasks using "user managed" IO.
* "user managed" IO means a TCP stream for each task, directly
* connected to the stdin, stdout, and stderr the task.
*/
slurm_step_launch_params_t_init(launch);
task_argv[0] = "./test7.3.io";
launch->argv = task_argv;
launch->argc = 1;
launch->user_managed_io = true; /* This is the key to using
"user managed" IO */
if (slurm_step_launch(ctx, launch, NULL) != SLURM_SUCCESS) {
slurm_perror("slurm_step_launch");
rc = 1;
goto done;
}
if (slurm_step_launch_wait_start(ctx) != SLURM_SUCCESS) {
slurm_perror("slurm_step_launch_wait_start");
rc =1;
goto done;
}
slurm_step_ctx_get(ctx, SLURM_STEP_CTX_USER_MANAGED_SOCKETS,
&num_fd, &fd_array);
/* Interact with launched tasks as desired */
_do_task_work(fd_array, tasks);
for (i = 0; i < tasks; i++) {
close(fd_array[i]);
}
slurm_step_launch_wait_finish(ctx);
/* Terminate the job killing all tasks */
done: slurm_kill_job(job_resp->job_id, SIGKILL, 0);
/* clean up storage */
slurm_free_resource_allocation_response_msg(job_resp);
if (ctx)
slurm_step_ctx_destroy(ctx);
exit(0);
}
