/* Attempt to schedule a specific job on specific available nodes
* IN job_ptr - job to schedule
* IN/OUT avail_bitmap - nodes available/selected to use
* IN exc_core_bitmap - cores which can not be used
* RET SLURM_SUCCESS on success, otherwise an error code
*/
static int _try_sched(struct job_record *job_ptr, bitstr_t **avail_bitmap,
uint32_t min_nodes, uint32_t max_nodes,
uint32_t req_nodes, bitstr_t *exc_core_bitmap)
{
bitstr_t *tmp_bitmap;
int rc = SLURM_SUCCESS;
int feat_cnt = _num_feature_count(job_ptr);
List preemptee_candidates = NULL;
if (feat_cnt) {
/* Ideally schedule the job feature by feature,
* but I don't want to add that complexity here
* right now, so clear the feature counts and try
* to schedule. This will work if there is only
* one feature count. It should work fairly well
* in cases where there are multiple feature
* counts. */
struct job_details *detail_ptr = job_ptr->details;
ListIterator feat_iter;
struct feature_record *feat_ptr;
int i = 0, list_size;
uint16_t *feat_cnt_orig = NULL, high_cnt = 0;
/* Clear the feature counts */
list_size = list_count(detail_ptr->feature_list);
feat_cnt_orig = xmalloc(sizeof(uint16_t) * list_size);
feat_iter = list_iterator_create(detail_ptr->feature_list);
while ((feat_ptr =
(struct feature_record *) list_next(feat_iter))) {
high_cnt = MAX(high_cnt, feat_ptr->count);
feat_cnt_orig[i++] = feat_ptr->count;
feat_ptr->count = 0;
}
list_iterator_destroy(feat_iter);
if ((job_req_node_filter(job_ptr, *avail_bitmap) !=
SLURM_SUCCESS) ||
(bit_set_count(*avail_bitmap) < high_cnt)) {
rc = ESLURM_NODES_BUSY;
} else {
preemptee_candidates =
slurm_find_preemptable_jobs(job_ptr);
rc = select_g_job_test(job_ptr, *avail_bitmap,
high_cnt, max_nodes, req_nodes,
SELECT_MODE_WILL_RUN,
preemptee_candidates, NULL,
exc_core_bitmap);
}
/* Restore the feature counts */
i = 0;
feat_iter = list_iterator_create(detail_ptr->feature_list);
while ((feat_ptr =
(struct feature_record *) list_next(feat_iter))) {
feat_ptr->count = feat_cnt_orig[i++];
}
list_iterator_destroy(feat_iter);
xfree(feat_cnt_orig);
} else {
/* Try to schedule the job. First on dedicated nodes
* then on shared nodes (if so configured). */
uint16_t orig_shared;
time_t now = time(NULL);
char str[100];
preemptee_candidates = slurm_find_preemptable_jobs(job_ptr);
orig_shared = job_ptr->details->shared;
job_ptr->details->shared = 0;
tmp_bitmap = bit_copy(*avail_bitmap);
if (exc_core_bitmap) {
bit_fmt(str, (sizeof(str) - 1), exc_core_bitmap);
debug2(" _try_sched with exclude core bitmap: %s",str);
}
rc = select_g_job_test(job_ptr, *avail_bitmap, min_nodes,
max_nodes, req_nodes,
SELECT_MODE_WILL_RUN,
preemptee_candidates, NULL,
exc_core_bitmap);
job_ptr->details->shared = orig_shared;
if (((rc != SLURM_SUCCESS) || (job_ptr->start_time > now)) &&
(orig_shared != 0)) {
FREE_NULL_BITMAP(*avail_bitmap);
*avail_bitmap= tmp_bitmap;
rc = select_g_job_test(job_ptr, *avail_bitmap,
min_nodes, max_nodes, req_nodes,
SELECT_MODE_WILL_RUN,
preemptee_candidates, NULL,
exc_core_bitmap);
} else
FREE_NULL_BITMAP(tmp_bitmap);
}
if (preemptee_candidates)
list_destroy(preemptee_candidates);
return rc;
}
static void _compute_start_times(void)
{
int j, rc = SLURM_SUCCESS, job_cnt = 0;
List job_queue;
job_queue_rec_t *job_queue_rec;
List preemptee_candidates = NULL;
struct job_record *job_ptr;
struct part_record *part_ptr;
bitstr_t *alloc_bitmap = NULL, *avail_bitmap = NULL;
bitstr_t *exc_core_bitmap = NULL;
uint32_t max_nodes, min_nodes, req_nodes, time_limit;
time_t now = time(NULL), sched_start, last_job_alloc;
bool resv_overlap = false;
sched_start = now;
last_job_alloc = now - 1;
alloc_bitmap = bit_alloc(node_record_count);
job_queue = build_job_queue(true, false);
sort_job_queue(job_queue);
while ((job_queue_rec = (job_queue_rec_t *) list_pop(job_queue))) {
job_ptr = job_queue_rec->job_ptr;
part_ptr = job_queue_rec->part_ptr;
xfree(job_queue_rec);
if (part_ptr != job_ptr->part_ptr)
continue; /* Only test one partition */
if (job_cnt++ > max_sched_job_cnt) {
debug2("scheduling loop exiting after %d jobs",
max_sched_job_cnt);
break;
}
/* Determine minimum and maximum node counts */
/* On BlueGene systems don't adjust the min/max node limits
here. We are working on midplane values. */
min_nodes = MAX(job_ptr->details->min_nodes,
part_ptr->min_nodes);
if (job_ptr->details->max_nodes == 0)
max_nodes = part_ptr->max_nodes;
else
max_nodes = MIN(job_ptr->details->max_nodes,
part_ptr->max_nodes);
max_nodes = MIN(max_nodes, 500000); /* prevent overflows */
if (job_ptr->details->max_nodes)
req_nodes = max_nodes;
else
req_nodes = min_nodes;
if (min_nodes > max_nodes) {
/* job's min_nodes exceeds partition's max_nodes */
continue;
}
j = job_test_resv(job_ptr, &now, true, &avail_bitmap,
&exc_core_bitmap, &resv_overlap, false);
if (j != SLURM_SUCCESS) {
FREE_NULL_BITMAP(avail_bitmap);
FREE_NULL_BITMAP(exc_core_bitmap);
continue;
}
rc = select_g_job_test(job_ptr, avail_bitmap,
min_nodes, max_nodes, req_nodes,
SELECT_MODE_WILL_RUN,
preemptee_candidates, NULL,
exc_core_bitmap);
if (rc == SLURM_SUCCESS) {
last_job_update = now;
if (job_ptr->time_limit == INFINITE)
time_limit = 365 * 24 * 60 * 60;
else if (job_ptr->time_limit != NO_VAL)
time_limit = job_ptr->time_limit * 60;
else if (job_ptr->part_ptr &&
(job_ptr->part_ptr->max_time != INFINITE))
time_limit = job_ptr->part_ptr->max_time * 60;
else
time_limit = 365 * 24 * 60 * 60;
if (bit_overlap(alloc_bitmap, avail_bitmap) &&
(job_ptr->start_time <= last_job_alloc)) {
job_ptr->start_time = last_job_alloc;
}
bit_or(alloc_bitmap, avail_bitmap);
last_job_alloc = job_ptr->start_time + time_limit;
}
FREE_NULL_BITMAP(avail_bitmap);
FREE_NULL_BITMAP(exc_core_bitmap);
if ((time(NULL) - sched_start) >= sched_timeout) {
debug2("scheduling loop exiting after %d jobs",
max_sched_job_cnt);
break;
}
}
FREE_NULL_LIST(job_queue);
FREE_NULL_BITMAP(alloc_bitmap);
}
static char * _will_run_test(uint32_t jobid, time_t start_time,
char *node_list, int *err_code, char **err_msg)
{
struct job_record *job_ptr = NULL;
struct part_record *part_ptr;
bitstr_t *avail_bitmap = NULL, *resv_bitmap = NULL;
bitstr_t *exc_core_bitmap = NULL;
char *hostlist, *reply_msg = NULL;
uint32_t min_nodes, max_nodes, req_nodes;
int rc;
time_t start_res, orig_start_time;
List preemptee_candidates;
debug2("wiki2: will_run job_id=%u start_time=%u node_list=%s",
jobid, (uint32_t)start_time, node_list);
job_ptr = find_job_record(jobid);
if (job_ptr == NULL) {
*err_code = -700;
*err_msg = "No such job";
error("wiki: Failed to find job %u", jobid);
return NULL;
}
if ((job_ptr->details == NULL) || (!IS_JOB_PENDING(job_ptr))) {
*err_code = -700;
*err_msg = "WillRun not applicable to non-pending job";
error("wiki: WillRun on non-pending job %u", jobid);
return NULL;
}
part_ptr = job_ptr->part_ptr;
if (part_ptr == NULL) {
*err_code = -700;
*err_msg = "Job lacks a partition";
error("wiki: Job %u lacks a partition", jobid);
return NULL;
}
if ((node_list == NULL) || (node_list[0] == '\0')) {
/* assume all nodes available to job for testing */
avail_bitmap = bit_copy(avail_node_bitmap);
} else if (node_name2bitmap(node_list, false, &avail_bitmap) != 0) {
*err_code = -700;
*err_msg = "Invalid available nodes value";
error("wiki: Attempt to set invalid available node "
"list for job %u, %s", jobid, node_list);
return NULL;
}
/* Enforce reservation: access control, time and nodes */
start_res = start_time;
rc = job_test_resv(job_ptr, &start_res, true, &resv_bitmap,
&exc_core_bitmap);
if (rc != SLURM_SUCCESS) {
*err_code = -730;
*err_msg = "Job denied access to reservation";
error("wiki: reservation access denied for job %u", jobid);
FREE_NULL_BITMAP(avail_bitmap);
return NULL;
}
start_time = MAX(start_time, start_res);
bit_and(avail_bitmap, resv_bitmap);
FREE_NULL_BITMAP(resv_bitmap);
/* Only consider nodes that are not DOWN or DRAINED */
bit_and(avail_bitmap, avail_node_bitmap);
/* Consider only nodes in this job's partition */
if (part_ptr->node_bitmap)
bit_and(avail_bitmap, part_ptr->node_bitmap);
else {
*err_code = -730;
*err_msg = "Job's partition has no nodes";
error("wiki: no nodes in partition %s for job %u",
part_ptr->name, jobid);
FREE_NULL_BITMAP(avail_bitmap);
return NULL;
}
if (job_req_node_filter(job_ptr, avail_bitmap) != SLURM_SUCCESS) {
/* Job probably has invalid feature list */
*err_code = -730;
*err_msg = "Job's required features not available "
"on selected nodes";
error("wiki: job %u not runnable on hosts=%s",
jobid, node_list);
FREE_NULL_BITMAP(avail_bitmap);
return NULL;
}
if (job_ptr->details->exc_node_bitmap) {
bit_not(job_ptr->details->exc_node_bitmap);
bit_and(avail_bitmap, job_ptr->details->exc_node_bitmap);
bit_not(job_ptr->details->exc_node_bitmap);
}
if ((job_ptr->details->req_node_bitmap) &&
(!bit_super_set(job_ptr->details->req_node_bitmap,
avail_bitmap))) {
*err_code = -730;
*err_msg = "Job's required nodes not available";
error("wiki: job %u not runnable on hosts=%s",
jobid, node_list);
FREE_NULL_BITMAP(avail_bitmap);
return NULL;
}
min_nodes = MAX(job_ptr->details->min_nodes, part_ptr->min_nodes);
if (job_ptr->details->max_nodes == 0)
max_nodes = part_ptr->max_nodes;
else
max_nodes = MIN(job_ptr->details->max_nodes,
part_ptr->max_nodes);
max_nodes = MIN(max_nodes, 500000); /* prevent overflows */
if (job_ptr->details->max_nodes)
req_nodes = max_nodes;
else
req_nodes = min_nodes;
if (min_nodes > max_nodes) {
/* job's min_nodes exceeds partitions max_nodes */
*err_code = -730;
*err_msg = "Job's min_nodes > max_nodes";
error("wiki: job %u not runnable on hosts=%s",
jobid, node_list);
FREE_NULL_BITMAP(avail_bitmap);
return NULL;
}
preemptee_candidates = slurm_find_preemptable_jobs(job_ptr);
orig_start_time = job_ptr->start_time;
rc = select_g_job_test(job_ptr, avail_bitmap,
min_nodes, max_nodes, req_nodes,
SELECT_MODE_WILL_RUN,
preemptee_candidates, NULL, exc_core_bitmap);
if (preemptee_candidates)
list_destroy(preemptee_candidates);
if (rc == SLURM_SUCCESS) {
char tmp_str[128];
*err_code = 0;
uint32_t proc_cnt = 0;
xstrcat(reply_msg, "STARTINFO=");
#ifdef HAVE_BG
select_g_select_jobinfo_get(job_ptr->select_jobinfo,
SELECT_JOBDATA_NODE_CNT,
&proc_cnt);
#else
proc_cnt = job_ptr->total_cpus;
#endif
snprintf(tmp_str, sizeof(tmp_str), "%u:%u@%u,",
jobid, proc_cnt, (uint32_t) job_ptr->start_time);
xstrcat(reply_msg, tmp_str);
hostlist = bitmap2node_name(avail_bitmap);
xstrcat(reply_msg, hostlist);
xfree(hostlist);
} else {
xstrcat(reply_msg, "Jobs not runable on selected nodes");
error("wiki: jobs not runnable on nodes");
}
/* Restore pending job's expected start time */
job_ptr->start_time = orig_start_time;
FREE_NULL_BITMAP(avail_bitmap);
return reply_msg;
}
static char * _will_run_test2(uint32_t jobid, time_t start_time,
char *node_list,
uint32_t *preemptee, int preemptee_cnt,
int *err_code, char **err_msg)
{
struct job_record *job_ptr = NULL, *pre_ptr;
struct part_record *part_ptr;
bitstr_t *avail_bitmap = NULL, *resv_bitmap = NULL;
bitstr_t *exc_core_bitmap = NULL;
time_t start_res;
uint32_t min_nodes, max_nodes, req_nodes;
List preemptee_candidates = NULL, preempted_jobs = NULL;
time_t orig_start_time;
char *reply_msg = NULL;
int i, rc;
bool resv_overlap = false;
xassert(node_list);
debug2("wiki2: will_run2 job_id=%u start_time=%u node_list=%s",
jobid, (uint32_t)start_time, node_list);
job_ptr = find_job_record(jobid);
if (job_ptr == NULL) {
*err_code = -700;
*err_msg = "No such job";
error("wiki: Failed to find job %u", jobid);
return NULL;
}
if ((job_ptr->details == NULL) || (!IS_JOB_PENDING(job_ptr))) {
*err_code = -700;
*err_msg = "WillRun not applicable to non-pending job";
error("wiki: WillRun on non-pending job %u", jobid);
return NULL;
}
part_ptr = job_ptr->part_ptr;
if (part_ptr == NULL) {
*err_code = -700;
*err_msg = "Job lacks a partition";
error("wiki: Job %u lacks a partition", jobid);
return NULL;
}
if (node_name2bitmap(node_list, false, &avail_bitmap) != 0) {
*err_code = -700;
*err_msg = "Invalid available nodes value";
error("wiki: Attempt to set invalid available node "
"list for job %u, %s", jobid, node_list);
return NULL;
}
/* Enforce reservation: access control, time and nodes */
start_res = start_time;
rc = job_test_resv(job_ptr, &start_res, true, &resv_bitmap,
&exc_core_bitmap, &resv_overlap);
if (rc != SLURM_SUCCESS) {
*err_code = -730;
*err_msg = "Job denied access to reservation";
error("wiki: reservation access denied for job %u", jobid);
FREE_NULL_BITMAP(avail_bitmap);
FREE_NULL_BITMAP(exc_core_bitmap);
return NULL;
}
bit_and(avail_bitmap, resv_bitmap);
FREE_NULL_BITMAP(resv_bitmap);
/* Only consider nodes that are not DOWN or DRAINED */
bit_and(avail_bitmap, avail_node_bitmap);
/* Consider only nodes in this job's partition */
if (part_ptr->node_bitmap)
bit_and(avail_bitmap, part_ptr->node_bitmap);
else {
*err_code = -730;
*err_msg = "Job's partition has no nodes";
error("wiki: no nodes in partition %s for job %u",
part_ptr->name, jobid);
FREE_NULL_BITMAP(avail_bitmap);
FREE_NULL_BITMAP(exc_core_bitmap);
return NULL;
}
if (job_req_node_filter(job_ptr, avail_bitmap) != SLURM_SUCCESS) {
/* Job probably has invalid feature list */
*err_code = -730;
*err_msg = "Job's required features not available "
"on selected nodes";
error("wiki: job %u not runnable on hosts=%s",
jobid, node_list);
FREE_NULL_BITMAP(avail_bitmap);
FREE_NULL_BITMAP(exc_core_bitmap);
return NULL;
}
if (job_ptr->details->exc_node_bitmap) {
bit_not(job_ptr->details->exc_node_bitmap);
bit_and(avail_bitmap, job_ptr->details->exc_node_bitmap);
bit_not(job_ptr->details->exc_node_bitmap);
}
if ((job_ptr->details->req_node_bitmap) &&
(!bit_super_set(job_ptr->details->req_node_bitmap,
avail_bitmap))) {
*err_code = -730;
*err_msg = "Job's required nodes not available";
error("wiki: job %u not runnable on hosts=%s",
jobid, node_list);
FREE_NULL_BITMAP(avail_bitmap);
FREE_NULL_BITMAP(exc_core_bitmap);
return NULL;
}
min_nodes = MAX(job_ptr->details->min_nodes, part_ptr->min_nodes);
if (job_ptr->details->max_nodes == 0)
max_nodes = part_ptr->max_nodes;
else
max_nodes = MIN(job_ptr->details->max_nodes,
part_ptr->max_nodes);
max_nodes = MIN(max_nodes, 500000); /* prevent overflows */
if (job_ptr->details->max_nodes)
req_nodes = max_nodes;
else
req_nodes = min_nodes;
if (min_nodes > max_nodes) {
/* job's min_nodes exceeds partitions max_nodes */
*err_code = -730;
*err_msg = "Job's min_nodes > max_nodes";
error("wiki: job %u not runnable on hosts=%s",
jobid, node_list);
FREE_NULL_BITMAP(avail_bitmap);
FREE_NULL_BITMAP(exc_core_bitmap);
return NULL;
}
if (preemptee_cnt) {
preemptee_candidates = list_create(NULL);
for (i=0; i<preemptee_cnt; i++) {
if ((pre_ptr = find_job_record(preemptee[i])))
list_append(preemptee_candidates, pre_ptr);
}
}
orig_start_time = job_ptr->start_time;
rc = select_g_job_test(job_ptr, avail_bitmap, min_nodes, max_nodes,
req_nodes, SELECT_MODE_WILL_RUN,
preemptee_candidates, &preempted_jobs,
exc_core_bitmap);
FREE_NULL_LIST(preemptee_candidates);
if (rc == SLURM_SUCCESS) {
char *hostlist, *sep, tmp_str[128];
uint32_t pre_cnt = 0, proc_cnt = 0;
#ifdef HAVE_BG
select_g_select_jobinfo_get(job_ptr->select_jobinfo,
SELECT_JOBDATA_NODE_CNT, &proc_cnt);
#else
proc_cnt = job_ptr->total_cpus;
#endif
snprintf(tmp_str, sizeof(tmp_str),
"STARTINFO=%u TASKS=%u STARTTIME=%u NODES=",
job_ptr->job_id, proc_cnt,
(uint32_t) job_ptr->start_time);
xstrcat(reply_msg, tmp_str);
hostlist = bitmap2node_name(avail_bitmap);
xstrcat(reply_msg, hostlist);
xfree(hostlist);
if (preempted_jobs) {
while ((pre_ptr = list_pop(preempted_jobs))) {
if (pre_cnt++)
sep = ",";
else
sep = " PREEMPT=";
snprintf(tmp_str, sizeof(tmp_str), "%s%u",
sep, pre_ptr->job_id);
xstrcat(reply_msg, tmp_str);
}
FREE_NULL_LIST(preempted_jobs);
}
} else {
xstrcat(reply_msg, "Jobs not runable on selected nodes");
error("wiki: jobs not runnable on nodes");
}
/* Restore pending job's expected start time */
job_ptr->start_time = orig_start_time;
FREE_NULL_BITMAP(avail_bitmap);
FREE_NULL_BITMAP(exc_core_bitmap);
return reply_msg;
}
static void _compute_start_times(void)
{
int j, rc = SLURM_SUCCESS, job_cnt = 0;
List job_queue;
job_queue_rec_t *job_queue_rec;
List preemptee_candidates = NULL;
struct job_record *job_ptr;
struct part_record *part_ptr;
bitstr_t *alloc_bitmap = NULL, *avail_bitmap = NULL;
uint32_t max_nodes, min_nodes, req_nodes, time_limit;
time_t now = time(NULL), sched_start, last_job_alloc;
sched_start = now;
last_job_alloc = now - 1;
alloc_bitmap = bit_alloc(node_record_count);
if (alloc_bitmap == NULL)
fatal("bit_alloc: malloc failure");
job_queue = build_job_queue(true);
while ((job_queue_rec = (job_queue_rec_t *)
list_pop_bottom(job_queue, sort_job_queue2))) {
job_ptr = job_queue_rec->job_ptr;
part_ptr = job_queue_rec->part_ptr;
xfree(job_queue_rec);
if (part_ptr != job_ptr->part_ptr)
continue; /* Only test one partition */
if (job_cnt++ > max_backfill_job_cnt) {
debug("backfill: loop taking to long, breaking out");
break;
}
/* Determine minimum and maximum node counts */
min_nodes = MAX(job_ptr->details->min_nodes,
part_ptr->min_nodes);
if (job_ptr->details->max_nodes == 0)
max_nodes = part_ptr->max_nodes;
else
max_nodes = MIN(job_ptr->details->max_nodes,
part_ptr->max_nodes);
max_nodes = MIN(max_nodes, 500000); /* prevent overflows */
if (job_ptr->details->max_nodes)
req_nodes = max_nodes;
else
req_nodes = min_nodes;
if (min_nodes > max_nodes) {
/* job's min_nodes exceeds partition's max_nodes */
continue;
}
j = job_test_resv(job_ptr, &now, true, &avail_bitmap);
if (j != SLURM_SUCCESS)
continue;
rc = select_g_job_test(job_ptr, avail_bitmap,
min_nodes, max_nodes, req_nodes,
SELECT_MODE_WILL_RUN,
preemptee_candidates, NULL);
last_job_update = now;
if (job_ptr->time_limit == INFINITE)
time_limit = 365 * 24 * 60 * 60;
else if (job_ptr->time_limit != NO_VAL)
time_limit = job_ptr->time_limit * 60;
else if (job_ptr->part_ptr &&
(job_ptr->part_ptr->max_time != INFINITE))
time_limit = job_ptr->part_ptr->max_time * 60;
else
time_limit = 365 * 24 * 60 * 60;
if (bit_overlap(alloc_bitmap, avail_bitmap) &&
(job_ptr->start_time <= last_job_alloc)) {
job_ptr->start_time = last_job_alloc;
}
bit_or(alloc_bitmap, avail_bitmap);
last_job_alloc = job_ptr->start_time + time_limit;
FREE_NULL_BITMAP(avail_bitmap);
if ((time(NULL) - sched_start) >= sched_timeout) {
debug("backfill: loop taking to long, breaking out");
break;
}
}
list_destroy(job_queue);
FREE_NULL_BITMAP(alloc_bitmap);
}
