/**
* @brief
* executes a job
*
* @param[in] job - The fully qualified job id.
* @param[in] server - The name of the server that manages the job.
* @param[in] location - location indicating where to run job
*
* @return - Void
*
* @File Variables:
* exitstatus Set to two if an error occurs.
*
*/
static void
execute(char *job, char *server, char *location)
{
int ct; /* Connection to the server */
int err; /* Error return from pbs_run */
int out; /* Stores the size of err_msg_buf*/
int located = FALSE;
char *errmsg;
char err_msg_buf[COMMENT_BUF_SIZE] = {'\0'}; /* generic buffer - comments & logging*/
char rmt_server[MAXSERVERNAME];
cnt:
if ((ct = cnt2server(server)) > 0) {
if (async)
err = pbs_asyrunjob(ct, job, location, NULL);
else
err = pbs_runjob(ct, job, location, NULL);
if (err && (pbs_errno != PBSE_UNKJOBID)) {
errmsg = pbs_geterrmsg(ct);
if (errmsg != NULL) {
if (pbs_errno == PBSE_UNKNODE) {
out = snprintf(err_msg_buf, sizeof(err_msg_buf),"qrun: %s %s",errmsg, location);
if (out >= sizeof(err_msg_buf)) {
fprintf(stderr,"%s...\n", err_msg_buf);
} else {
fprintf(stderr, "%s\n", err_msg_buf);
}
} else {
prt_job_err("qrun", ct, job);
}
} else {
fprintf(stderr, "qrun : Server returned error %d for job ", pbs_errno);
}
exitstatus = 2;
} else if (err && (pbs_errno == PBSE_UNKJOBID) && !located) {
located = TRUE;
if (locate_job(job, server, rmt_server)) {
pbs_disconnect(ct);
strcpy(server, rmt_server);
goto cnt;
}
prt_job_err("qrun", ct, job);
exitstatus = 2;
}
pbs_disconnect(ct);
} else {
fprintf(stderr,
"qrun: could not connect to server %s (%d)\n", server, pbs_errno);
exitstatus = 2;
}
}
/*
*
* run_update_job - run the job and update the job information
*
* pbs_sd - connection to pbs_server
* sinfo - server job is on
* qinfo - queue job resides in
* jinfo - the job to run
*
* returns success/failure - see pbs_errno for more info
*
*/
int run_update_job(int pbs_sd, server_info *sinfo, queue_info *qinfo,
job_info *jinfo)
{
int ret; /* return code from pbs_runjob() */
node_info *best_node = NULL; /* best node to run job on */
char *best_node_name = NULL; /* name of best node */
char buf[256] = {'\0'}; /* generic buffer - comments & logging*/
char timebuf[128]; /* buffer to hold the time and date */
resource_req *res; /* ptr to the resource of ncpus */
int ncpus; /* numeric amount of resource ncpus */
char *errmsg; /* used for pbs_geterrmsg() */
strftime(timebuf, 128, "started on %a %b %d at %H:%M", localtime(&cstat.current_time));
if (cstat.load_balancing || cstat.load_balancing_rr)
{
best_node = find_best_node(jinfo, sinfo -> timesharing_nodes);
if (best_node != NULL)
{
best_node_name = best_node -> name;
sprintf(buf, "Job run on node %s - %s", best_node_name, timebuf);
}
}
if (best_node == NULL)
sprintf(buf, "Job %s", timebuf);
update_job_comment(pbs_sd, jinfo, buf);
buf[0] = '\0';
ret = pbs_runjob(pbs_sd, jinfo -> name, best_node_name, NULL);
if (ret == 0)
{
/* If a job is 100% efficent, it will raise the load average by 1 per
* cpu is uses. Temporarly inflate load average by that value
*/
if (cstat.load_balancing && best_node != NULL)
{
if ((res = find_resource_req(jinfo -> resreq, "ncpus")) == NULL)
ncpus = 1;
else
ncpus = res -> amount;
best_node -> loadave += ncpus;
}
if (cstat.help_starving_jobs && jinfo == cstat.starving_job)
jinfo -> sch_priority = 0;
sched_log(PBSEVENT_SCHED, PBS_EVENTCLASS_JOB, jinfo -> name, "Job Run");
update_server_on_run(sinfo, qinfo, jinfo);
update_queue_on_run(qinfo, jinfo);
update_job_on_run(pbs_sd, jinfo);
if (cstat.fair_share)
update_usage_on_run(jinfo);
free(sinfo -> running_jobs);
sinfo -> running_jobs = job_filter(sinfo -> jobs, sinfo -> sc.total,
check_run_job, NULL);
free(qinfo -> running_jobs);
qinfo -> running_jobs = job_filter(qinfo -> jobs, qinfo -> sc.total,
check_run_job, NULL);
}
else
{
errmsg = pbs_geterrmsg(pbs_sd);
sprintf(buf, "Not Running - PBS Error: %s", errmsg);
update_job_comment(pbs_sd, jinfo, buf);
}
return ret;
}
int
schd_run_job_on(Job *job, Queue *destq, char *exechost, int set_comment)
{
char *id = "schd_run_job_on";
char reason[128], tmp_word[20];
char *date;
Queue *srcq = NULL;
int ret = 0;
/* Get the datestamp from 'ctime()'. Remove the trailing '\n'. */
date = ctime(&schd_TimeNow);
date[strlen(date) - 1] = '\0';
if (set_comment)
{
sprintf(reason, "Started on %s", date);
if (job->flags & JFLAGS_PRIORITY)
{
strcat(reason, " (EXPRESS/high priority job)");
}
if (job->flags & JFLAGS_WAITING)
{
strcat(reason, " (long-waiting job)");
}
schd_comment_job(job, reason, JOB_COMMENT_REQUIRED);
}
/* If this is NOT a suspended job... */
if (!(job->flags & JFLAGS_SUSPENDED))
{
/*
* If a destination Queue is provided, and it is different from the
* source queue, then ask PBS to move the job to that queue before
* running it.
*/
srcq = job->queue;
/*
* Move the job from its queue to the specified run queue.
*/
if ((destq != NULL) && (strcmp(destq->qname, srcq->qname) != 0))
{
if (pbs_movejob(connector, job->jobid, destq->qname, NULL))
{
(void)sprintf(log_buffer, "move job %s to queue %s failed, %d",
job->jobid, destq->qname, pbs_errno);
log_record(PBSEVENT_SYSTEM, PBS_EVENTCLASS_SERVER,
id, log_buffer);
DBPRT(("%s: %s\n", id, log_buffer));
return (-1);
}
schd_move_job_to(job, destq);
}
/*
* Give the job handle (JOBID) to PBS to run.
*/
if (pbs_runjob(connector, job->jobid, exechost, NULL))
{
(void)sprintf(log_buffer, "failed start job %s on queue %s@%s, %d",
job->jobid, destq->qname, exechost, pbs_errno);
log_record(PBSEVENT_SYSTEM, PBS_EVENTCLASS_SERVER, id, log_buffer);
DBPRT(("%s: %s\n", id, log_buffer));
/*
* Running failed! Move the job back to the source queue (if
* applicable) before returning. This prevents jobs being left
* in execution queues.
*/
if (srcq)
{
DBPRT(("Attempting to move job %s back to queue %s\n",
job->jobid, srcq->qname));
if (pbs_movejob(connector, job->jobid, srcq->qname, NULL))
{
(void)sprintf(log_buffer,
"failed to move job %s back to queue %s, %d",
job->jobid, srcq->qname, pbs_errno);
log_record(PBSEVENT_SYSTEM, PBS_EVENTCLASS_SERVER, id,
log_buffer);
DBPRT(("%s: %s\n", id, log_buffer));
}
schd_move_job_to(job, srcq);
}
return (-1);
}
strcpy(tmp_word, "started");
}
else /* it IS a suspended job */
{
schd_move_job_to(job, destq);
ret = pbs_sigjob(connector, job->jobid, "resume", NULL);
if (ret)
{
sprintf(log_buffer, "resume of job %s FAILED (%d)",
job->jobid, ret);
return (-1);
}
job->flags &= ~JFLAGS_SUSPENDED;
strcpy(tmp_word, "resumed");
}
/* PBS accepted the job (and presumably will run it). Log the fact. */
(void)sprintf(log_buffer, "job %s %s on %s@%s", job->jobid, tmp_word,
destq->qname, exechost);
log_record(PBSEVENT_SYSTEM, PBS_EVENTCLASS_SERVER, id, log_buffer);
DBPRT(("%s: %s\n", id, log_buffer));
/*
* Change the state of the local representation of the job to "Running".
*/
job->state = 'R';
/*
* Account for the job on this queue's statistics. 'queued' will be
* bumped up if the queued job was moved to a new destination queue.
*/
job->queue->queued --;
job->queue->running ++;
/* The queue is no longer idle. Unset the idle timer. */
job->queue->idle_since = 0;
return (0); /* Job successfully started. */
}
