/*
* delete_partition - delete the specified partition (actually leave
* the entry, just flag it as defunct)
* IN job_specs - job specification from RPC
*/
extern int delete_partition(delete_part_msg_t *part_desc_ptr)
{
struct part_record *part_ptr;
part_ptr = find_part_record (part_desc_ptr->name);
if (part_ptr == NULL) /* No such partition */
return ESLURM_INVALID_PARTITION_NAME;
if (partition_in_use(part_desc_ptr->name))
return ESLURM_PARTITION_IN_USE;
if (default_part_loc == part_ptr) {
error("Deleting default partition %s", part_ptr->name);
default_part_loc = NULL;
}
(void) kill_job_by_part_name(part_desc_ptr->name);
list_delete_all(part_list, list_find_part, part_desc_ptr->name);
last_part_update = time(NULL);
slurm_sched_partition_change(); /* notify sched plugin */
select_g_reconfigure(); /* notify select plugin too */
return SLURM_SUCCESS;
}
static int _attempt_backfill(void)
{
DEF_TIMERS;
bool filter_root = false;
List job_queue;
job_queue_rec_t *job_queue_rec;
slurmdb_qos_rec_t *qos_ptr = NULL;
int i, j, node_space_recs;
struct job_record *job_ptr;
struct part_record *part_ptr;
uint32_t end_time, end_reserve;
uint32_t time_limit, comp_time_limit, orig_time_limit;
uint32_t min_nodes, max_nodes, req_nodes;
bitstr_t *avail_bitmap = NULL, *resv_bitmap = NULL;
bitstr_t *exc_core_bitmap = NULL;
time_t now, sched_start, later_start, start_res, resv_end;
node_space_map_t *node_space;
struct timeval bf_time1, bf_time2;
int sched_timeout = 2, yield_sleep = 1;
int rc = 0;
int job_test_count = 0;
uint32_t *uid = NULL, nuser = 0;
uint16_t *njobs = NULL;
bool already_counted;
uint32_t reject_array_job_id = 0;
#ifdef HAVE_CRAY
/*
* Run a Basil Inventory immediately before setting up the schedule
* plan, to avoid race conditions caused by ALPS node state change.
* Needs to be done with the node-state lock taken.
*/
START_TIMER;
if (select_g_reconfigure()) {
debug4("backfill: not scheduling due to ALPS");
return SLURM_SUCCESS;
}
END_TIMER;
if (debug_flags & DEBUG_FLAG_BACKFILL)
info("backfill: ALPS inventory completed, %s", TIME_STR);
/* The Basil inventory can take a long time to complete. Process
* pending RPCs before starting the backfill scheduling logic */
_yield_locks(1);
#endif
START_TIMER;
if (debug_flags & DEBUG_FLAG_BACKFILL)
info("backfill: beginning");
sched_start = now = time(NULL);
if (slurm_get_root_filter())
filter_root = true;
job_queue = build_job_queue(true);
if (list_count(job_queue) == 0) {
debug("backfill: no jobs to backfill");
list_destroy(job_queue);
return 0;
}
gettimeofday(&bf_time1, NULL);
slurmctld_diag_stats.bf_queue_len = list_count(job_queue);
slurmctld_diag_stats.bf_queue_len_sum += slurmctld_diag_stats.
bf_queue_len;
slurmctld_diag_stats.bf_last_depth = 0;
slurmctld_diag_stats.bf_last_depth_try = 0;
slurmctld_diag_stats.bf_when_last_cycle = now;
bf_last_yields = 0;
slurmctld_diag_stats.bf_active = 1;
node_space = xmalloc(sizeof(node_space_map_t) *
(max_backfill_job_cnt + 3));
node_space[0].begin_time = sched_start;
node_space[0].end_time = sched_start + backfill_window;
node_space[0].avail_bitmap = bit_copy(avail_node_bitmap);
node_space[0].next = 0;
node_space_recs = 1;
if (debug_flags & DEBUG_FLAG_BACKFILL)
_dump_node_space_table(node_space);
if (max_backfill_job_per_user) {
uid = xmalloc(BF_MAX_USERS * sizeof(uint32_t));
njobs = xmalloc(BF_MAX_USERS * sizeof(uint16_t));
}
while ((job_queue_rec = (job_queue_rec_t *)
list_pop_bottom(job_queue, sort_job_queue2))) {
job_ptr = job_queue_rec->job_ptr;
orig_time_limit = job_ptr->time_limit;
if ((time(NULL) - sched_start) >= sched_timeout) {
uint32_t save_time_limit = job_ptr->time_limit;
job_ptr->time_limit = orig_time_limit;
if (debug_flags & DEBUG_FLAG_BACKFILL) {
END_TIMER;
info("backfill: completed yielding locks "
"after testing %d jobs, %s",
job_test_count, TIME_STR);
}
if (_yield_locks(yield_sleep) && !backfill_continue) {
if (debug_flags & DEBUG_FLAG_BACKFILL) {
info("backfill: system state changed, "
"breaking out after testing %d "
"jobs", job_test_count);
}
rc = 1;
break;
}
job_ptr->time_limit = save_time_limit;
/* Reset backfill scheduling timers, resume testing */
sched_start = time(NULL);
job_test_count = 0;
START_TIMER;
}
part_ptr = job_queue_rec->part_ptr;
job_test_count++;
xfree(job_queue_rec);
if (!IS_JOB_PENDING(job_ptr))
continue; /* started in other partition */
if (!avail_front_end(job_ptr))
continue; /* No available frontend for this job */
if (job_ptr->array_task_id != (uint16_t) NO_VAL) {
if (reject_array_job_id == job_ptr->array_job_id)
continue; /* already rejected array element */
/* assume reject whole array for now, clear if OK */
reject_array_job_id = job_ptr->array_job_id;
}
job_ptr->part_ptr = part_ptr;
if (debug_flags & DEBUG_FLAG_BACKFILL)
info("backfill test for job %u", job_ptr->job_id);
slurmctld_diag_stats.bf_last_depth++;
already_counted = false;
if (max_backfill_job_per_user) {
for (j = 0; j < nuser; j++) {
if (job_ptr->user_id == uid[j]) {
njobs[j]++;
if (debug_flags & DEBUG_FLAG_BACKFILL)
debug("backfill: user %u: "
"#jobs %u",
uid[j], njobs[j]);
break;
}
}
if (j == nuser) { /* user not found */
if (nuser < BF_MAX_USERS) {
uid[j] = job_ptr->user_id;
njobs[j] = 1;
nuser++;
} else {
error("backfill: too many users in "
"queue. Consider increasing "
"BF_MAX_USERS");
}
if (debug_flags & DEBUG_FLAG_BACKFILL)
debug2("backfill: found new user %u. "
"Total #users now %u",
job_ptr->user_id, nuser);
} else {
if (njobs[j] > max_backfill_job_per_user) {
/* skip job */
if (debug_flags & DEBUG_FLAG_BACKFILL)
debug("backfill: have already "
"checked %u jobs for "
"user %u; skipping "
"job %u",
max_backfill_job_per_user,
job_ptr->user_id,
job_ptr->job_id);
continue;
}
}
}
if (((part_ptr->state_up & PARTITION_SCHED) == 0) ||
(part_ptr->node_bitmap == NULL))
continue;
if ((part_ptr->flags & PART_FLAG_ROOT_ONLY) && filter_root)
continue;
if ((!job_independent(job_ptr, 0)) ||
(license_job_test(job_ptr, time(NULL)) != SLURM_SUCCESS))
continue;
/* 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;
}
/* Determine job's expected completion time */
if (job_ptr->time_limit == NO_VAL) {
if (part_ptr->max_time == INFINITE)
time_limit = 365 * 24 * 60; /* one year */
else
time_limit = part_ptr->max_time;
} else {
if (part_ptr->max_time == INFINITE)
time_limit = job_ptr->time_limit;
else
time_limit = MIN(job_ptr->time_limit,
part_ptr->max_time);
}
comp_time_limit = time_limit;
qos_ptr = job_ptr->qos_ptr;
if (qos_ptr && (qos_ptr->flags & QOS_FLAG_NO_RESERVE) &&
slurm_get_preempt_mode())
time_limit = job_ptr->time_limit = 1;
else if (job_ptr->time_min && (job_ptr->time_min < time_limit))
time_limit = job_ptr->time_limit = job_ptr->time_min;
/* Determine impact of any resource reservations */
later_start = now;
TRY_LATER:
if ((time(NULL) - sched_start) >= sched_timeout) {
uint32_t save_time_limit = job_ptr->time_limit;
job_ptr->time_limit = orig_time_limit;
if (debug_flags & DEBUG_FLAG_BACKFILL) {
END_TIMER;
info("backfill: completed yielding locks 2"
"after testing %d jobs, %s",
job_test_count, TIME_STR);
}
if (_yield_locks(yield_sleep) && !backfill_continue) {
if (debug_flags & DEBUG_FLAG_BACKFILL) {
info("backfill: system state changed, "
"breaking out after testing %d "
"jobs", job_test_count);
}
rc = 1;
break;
}
job_ptr->time_limit = save_time_limit;
/* Reset backfill scheduling timers, resume testing */
sched_start = time(NULL);
job_test_count = 1;
START_TIMER;
}
FREE_NULL_BITMAP(avail_bitmap);
FREE_NULL_BITMAP(exc_core_bitmap);
start_res = later_start;
later_start = 0;
j = job_test_resv(job_ptr, &start_res, true, &avail_bitmap,
&exc_core_bitmap);
if (j != SLURM_SUCCESS) {
job_ptr->time_limit = orig_time_limit;
continue;
}
if (start_res > now)
end_time = (time_limit * 60) + start_res;
else
end_time = (time_limit * 60) + now;
resv_end = find_resv_end(start_res);
/* Identify usable nodes for this job */
bit_and(avail_bitmap, part_ptr->node_bitmap);
bit_and(avail_bitmap, up_node_bitmap);
for (j=0; ; ) {
if ((node_space[j].end_time > start_res) &&
node_space[j].next && (later_start == 0))
later_start = node_space[j].end_time;
if (node_space[j].end_time <= start_res)
;
else if (node_space[j].begin_time <= end_time) {
bit_and(avail_bitmap,
node_space[j].avail_bitmap);
} else
break;
if ((j = node_space[j].next) == 0)
break;
}
if ((resv_end++) &&
((later_start == 0) || (resv_end < later_start))) {
later_start = resv_end;
}
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);
}
/* Test if insufficient nodes remain OR
* required nodes missing OR
* nodes lack features */
if ((bit_set_count(avail_bitmap) < min_nodes) ||
((job_ptr->details->req_node_bitmap) &&
(!bit_super_set(job_ptr->details->req_node_bitmap,
avail_bitmap))) ||
(job_req_node_filter(job_ptr, avail_bitmap))) {
if (later_start) {
job_ptr->start_time = 0;
goto TRY_LATER;
}
/* Job can not start until too far in the future */
job_ptr->time_limit = orig_time_limit;
job_ptr->start_time = sched_start + backfill_window;
continue;
}
/* Identify nodes which are definitely off limits */
FREE_NULL_BITMAP(resv_bitmap);
resv_bitmap = bit_copy(avail_bitmap);
bit_not(resv_bitmap);
/* this is the time consuming operation */
debug2("backfill: entering _try_sched for job %u.",
job_ptr->job_id);
if (!already_counted) {
slurmctld_diag_stats.bf_last_depth_try++;
already_counted = true;
}
j = _try_sched(job_ptr, &avail_bitmap, min_nodes, max_nodes,
req_nodes, exc_core_bitmap);
now = time(NULL);
if (j != SLURM_SUCCESS) {
job_ptr->time_limit = orig_time_limit;
job_ptr->start_time = 0;
continue; /* not runable */
}
if (start_res > job_ptr->start_time) {
job_ptr->start_time = start_res;
last_job_update = now;
}
if (job_ptr->start_time <= now) {
uint32_t save_time_limit = job_ptr->time_limit;
int rc = _start_job(job_ptr, resv_bitmap);
if (qos_ptr && (qos_ptr->flags & QOS_FLAG_NO_RESERVE)) {
if (orig_time_limit == NO_VAL)
job_ptr->time_limit = comp_time_limit;
else
job_ptr->time_limit = orig_time_limit;
job_ptr->end_time = job_ptr->start_time +
(job_ptr->time_limit * 60);
} else if ((rc == SLURM_SUCCESS) && job_ptr->time_min) {
/* Set time limit as high as possible */
job_ptr->time_limit = comp_time_limit;
job_ptr->end_time = job_ptr->start_time +
(comp_time_limit * 60);
_reset_job_time_limit(job_ptr, now,
node_space);
time_limit = job_ptr->time_limit;
} else {
job_ptr->time_limit = orig_time_limit;
}
if (rc == ESLURM_ACCOUNTING_POLICY) {
/* Unknown future start time, just skip job */
job_ptr->start_time = 0;
continue;
} else if (rc != SLURM_SUCCESS) {
/* Planned to start job, but something bad
* happended. */
job_ptr->start_time = 0;
break;
} else {
/* Started this job, move to next one */
reject_array_job_id = 0;
/* Update the database if job time limit
* changed and move to next job */
if (save_time_limit != job_ptr->time_limit)
jobacct_storage_g_job_start(acct_db_conn,
job_ptr);
continue;
}
} else
job_ptr->time_limit = orig_time_limit;
if (later_start && (job_ptr->start_time > later_start)) {
/* Try later when some nodes currently reserved for
* pending jobs are free */
job_ptr->start_time = 0;
goto TRY_LATER;
}
if (job_ptr->start_time > (sched_start + backfill_window)) {
/* Starts too far in the future to worry about */
continue;
}
if (node_space_recs >= max_backfill_job_cnt) {
/* Already have too many jobs to deal with */
break;
}
end_reserve = job_ptr->start_time + (time_limit * 60);
if (_test_resv_overlap(node_space, avail_bitmap,
job_ptr->start_time, end_reserve)) {
/* This job overlaps with an existing reservation for
* job to be backfill scheduled, which the sched
* plugin does not know about. Try again later. */
later_start = job_ptr->start_time;
job_ptr->start_time = 0;
goto TRY_LATER;
}
/*
* Add reservation to scheduling table if appropriate
*/
if (qos_ptr && (qos_ptr->flags & QOS_FLAG_NO_RESERVE))
continue;
reject_array_job_id = 0;
bit_not(avail_bitmap);
_add_reservation(job_ptr->start_time, end_reserve,
avail_bitmap, node_space, &node_space_recs);
if (debug_flags & DEBUG_FLAG_BACKFILL)
_dump_node_space_table(node_space);
}
xfree(uid);
xfree(njobs);
FREE_NULL_BITMAP(avail_bitmap);
FREE_NULL_BITMAP(exc_core_bitmap);
FREE_NULL_BITMAP(resv_bitmap);
for (i=0; ; ) {
FREE_NULL_BITMAP(node_space[i].avail_bitmap);
if ((i = node_space[i].next) == 0)
break;
}
xfree(node_space);
list_destroy(job_queue);
gettimeofday(&bf_time2, NULL);
_do_diag_stats(&bf_time1, &bf_time2, yield_sleep);
if (debug_flags & DEBUG_FLAG_BACKFILL) {
END_TIMER;
info("backfill: completed testing %d jobs, %s",
job_test_count, TIME_STR);
}
return rc;
}
/*
* update_part - create or update a partition's configuration data
* IN part_desc - description of partition changes
* IN create_flag - create a new partition
* RET 0 or an error code
* global: part_list - list of partition entries
* last_part_update - update time of partition records
*/
extern int update_part (update_part_msg_t * part_desc, bool create_flag)
{
int error_code;
struct part_record *part_ptr;
if (part_desc->name == NULL) {
info("update_part: invalid partition name, NULL");
return ESLURM_INVALID_PARTITION_NAME;
}
error_code = SLURM_SUCCESS;
part_ptr = list_find_first(part_list, &list_find_part,
part_desc->name);
if (create_flag) {
if (part_ptr) {
verbose("Duplicate partition name for create (%s)",
part_desc->name);
return ESLURM_INVALID_PARTITION_NAME;
}
info("update_part: partition %s being created",
part_desc->name);
part_ptr = create_part_record();
xfree(part_ptr->name);
part_ptr->name = xstrdup(part_desc->name);
} else {
if (!part_ptr) {
verbose("Update for partition not found (%s)",
part_desc->name);
return ESLURM_INVALID_PARTITION_NAME;
}
}
last_part_update = time(NULL);
if (part_desc->max_cpus_per_node != NO_VAL) {
info("update_part: setting MaxCPUsPerNode to %u for partition %s",
part_desc->max_cpus_per_node, part_desc->name);
part_ptr->max_cpus_per_node = part_desc->max_cpus_per_node;
}
if (part_desc->max_time != NO_VAL) {
info("update_part: setting max_time to %u for partition %s",
part_desc->max_time, part_desc->name);
part_ptr->max_time = part_desc->max_time;
}
if ((part_desc->default_time != NO_VAL) &&
(part_desc->default_time > part_ptr->max_time)) {
info("update_part: DefaultTime would exceed MaxTime for "
"partition %s", part_desc->name);
} else if (part_desc->default_time != NO_VAL) {
info("update_part: setting default_time to %u "
"for partition %s",
part_desc->default_time, part_desc->name);
part_ptr->default_time = part_desc->default_time;
}
if (part_desc->max_nodes != NO_VAL) {
info("update_part: setting max_nodes to %u for partition %s",
part_desc->max_nodes, part_desc->name);
part_ptr->max_nodes = part_desc->max_nodes;
part_ptr->max_nodes_orig = part_desc->max_nodes;
select_g_alter_node_cnt(SELECT_SET_MP_CNT,
&part_ptr->max_nodes);
}
if (part_desc->min_nodes != NO_VAL) {
info("update_part: setting min_nodes to %u for partition %s",
part_desc->min_nodes, part_desc->name);
part_ptr->min_nodes = part_desc->min_nodes;
part_ptr->min_nodes_orig = part_desc->min_nodes;
select_g_alter_node_cnt(SELECT_SET_MP_CNT,
&part_ptr->min_nodes);
}
if (part_desc->grace_time != NO_VAL) {
info("update_part: setting grace_time to %u for partition %s",
part_desc->grace_time, part_desc->name);
part_ptr->grace_time = part_desc->grace_time;
}
if (part_desc->flags & PART_FLAG_HIDDEN) {
info("update_part: setting hidden for partition %s",
part_desc->name);
part_ptr->flags |= PART_FLAG_HIDDEN;
} else if (part_desc->flags & PART_FLAG_HIDDEN_CLR) {
info("update_part: clearing hidden for partition %s",
part_desc->name);
part_ptr->flags &= (~PART_FLAG_HIDDEN);
}
if (part_desc->flags & PART_FLAG_REQ_RESV) {
info("update_part: setting req_resv for partition %s",
part_desc->name);
part_ptr->flags |= PART_FLAG_REQ_RESV;
} else if (part_desc->flags & PART_FLAG_REQ_RESV_CLR) {
info("update_part: clearing req_resv for partition %s",
part_desc->name);
part_ptr->flags &= (~PART_FLAG_REQ_RESV);
}
if (part_desc->flags & PART_FLAG_ROOT_ONLY) {
info("update_part: setting root_only for partition %s",
part_desc->name);
part_ptr->flags |= PART_FLAG_ROOT_ONLY;
} else if (part_desc->flags & PART_FLAG_ROOT_ONLY_CLR) {
info("update_part: clearing root_only for partition %s",
part_desc->name);
part_ptr->flags &= (~PART_FLAG_ROOT_ONLY);
}
if (part_desc->flags & PART_FLAG_NO_ROOT) {
info("update_part: setting no_root for partition %s",
part_desc->name);
part_ptr->flags |= PART_FLAG_NO_ROOT;
} else if (part_desc->flags & PART_FLAG_NO_ROOT_CLR) {
info("update_part: clearing no_root for partition %s",
part_desc->name);
part_ptr->flags &= (~PART_FLAG_NO_ROOT);
}
if (part_desc->flags & PART_FLAG_DEFAULT) {
if (default_part_name == NULL) {
info("update_part: setting default partition to %s",
part_desc->name);
} else if (strcmp(default_part_name, part_desc->name) != 0) {
info("update_part: changing default "
"partition from %s to %s",
default_part_name, part_desc->name);
}
xfree(default_part_name);
default_part_name = xstrdup(part_desc->name);
default_part_loc = part_ptr;
part_ptr->flags |= PART_FLAG_DEFAULT;
} else if ((part_desc->flags & PART_FLAG_DEFAULT_CLR) &&
(default_part_loc == part_ptr)) {
info("update_part: clearing default partition from %s",
part_desc->name);
xfree(default_part_name);
default_part_loc = NULL;
part_ptr->flags &= (~PART_FLAG_DEFAULT);
}
if (part_desc->state_up != (uint16_t) NO_VAL) {
info("update_part: setting state_up to %u for partition %s",
part_desc->state_up, part_desc->name);
part_ptr->state_up = part_desc->state_up;
}
if (part_desc->max_share != (uint16_t) NO_VAL) {
uint16_t force = part_desc->max_share & SHARED_FORCE;
uint16_t val = part_desc->max_share & (~SHARED_FORCE);
char tmp_str[24];
if (val == 0)
snprintf(tmp_str, sizeof(tmp_str), "EXCLUSIVE");
else if (force)
snprintf(tmp_str, sizeof(tmp_str), "FORCE:%u", val);
else if (val == 1)
snprintf(tmp_str, sizeof(tmp_str), "NO");
else
snprintf(tmp_str, sizeof(tmp_str), "YES:%u", val);
info("update_part: setting share to %s for partition %s",
tmp_str, part_desc->name);
part_ptr->max_share = part_desc->max_share;
}
if (part_desc->preempt_mode != (uint16_t) NO_VAL) {
uint16_t new_mode;
new_mode = part_desc->preempt_mode & (~PREEMPT_MODE_GANG);
if (new_mode <= PREEMPT_MODE_CANCEL) {
info("update_part: setting preempt_mode to %s for "
"partition %s",
preempt_mode_string(new_mode), part_desc->name);
part_ptr->preempt_mode = new_mode;
} else {
info("update_part: invalid preempt_mode %u", new_mode);
}
}
if (part_desc->priority != (uint16_t) NO_VAL) {
info("update_part: setting priority to %u for partition %s",
part_desc->priority, part_desc->name);
part_ptr->priority = part_desc->priority;
/* If the max_priority changes we need to change all
* the normalized priorities of all the other
* partitions. If not then just set this partition.
*/
if (part_ptr->priority > part_max_priority) {
ListIterator itr = list_iterator_create(part_list);
struct part_record *part2 = NULL;
part_max_priority = part_ptr->priority;
while((part2 = list_next(itr))) {
part2->norm_priority = (double)part2->priority
/ (double)part_max_priority;
}
list_iterator_destroy(itr);
} else {
part_ptr->norm_priority = (double)part_ptr->priority
/ (double)part_max_priority;
}
}
if (part_desc->allow_groups != NULL) {
xfree(part_ptr->allow_groups);
xfree(part_ptr->allow_uids);
if ((strcasecmp(part_desc->allow_groups, "ALL") == 0) ||
(part_desc->allow_groups[0] == '\0')) {
info("update_part: setting allow_groups to ALL for "
"partition %s",
part_desc->name);
} else {
part_ptr->allow_groups = part_desc->allow_groups;
part_desc->allow_groups = NULL;
info("update_part: setting allow_groups to %s for "
"partition %s",
part_ptr->allow_groups, part_desc->name);
part_ptr->allow_uids =
_get_groups_members(part_ptr->allow_groups);
clear_group_cache();
}
}
if (part_desc->allow_alloc_nodes != NULL) {
xfree(part_ptr->allow_alloc_nodes);
if ((part_desc->allow_alloc_nodes[0] == '\0') ||
(strcasecmp(part_desc->allow_alloc_nodes, "ALL") == 0)) {
part_ptr->allow_alloc_nodes = NULL;
info("update_part: setting allow_alloc_nodes to ALL"
" for partition %s",part_desc->name);
}
else {
part_ptr->allow_alloc_nodes = part_desc->
allow_alloc_nodes;
part_desc->allow_alloc_nodes = NULL;
info("update_part: setting allow_alloc_nodes to %s for "
"partition %s",
part_ptr->allow_alloc_nodes, part_desc->name);
}
}
if (part_desc->alternate != NULL) {
xfree(part_ptr->alternate);
if ((strcasecmp(part_desc->alternate, "NONE") == 0) ||
(part_desc->alternate[0] == '\0'))
part_ptr->alternate = NULL;
else
part_ptr->alternate = xstrdup(part_desc->alternate);
part_desc->alternate = NULL;
info("update_part: setting alternate to %s for "
"partition %s",
part_ptr->alternate, part_desc->name);
}
if (part_desc->def_mem_per_cpu != NO_VAL) {
char *key;
uint32_t value;
if (part_desc->def_mem_per_cpu & MEM_PER_CPU) {
key = "DefMemPerCpu";
value = part_desc->def_mem_per_cpu & (~MEM_PER_CPU);
} else {
key = "DefMemPerNode";
value = part_desc->def_mem_per_cpu;
}
info("update_part: setting %s to %u for partition %s",
key, value, part_desc->name);
part_ptr->def_mem_per_cpu = part_desc->def_mem_per_cpu;
}
if (part_desc->max_mem_per_cpu != NO_VAL) {
char *key;
uint32_t value;
if (part_desc->max_mem_per_cpu & MEM_PER_CPU) {
key = "MaxMemPerCpu";
value = part_desc->max_mem_per_cpu & (~MEM_PER_CPU);
} else {
key = "MaxMemPerNode";
value = part_desc->max_mem_per_cpu;
}
info("update_part: setting %s to %u for partition %s",
key, value, part_desc->name);
part_ptr->max_mem_per_cpu = part_desc->max_mem_per_cpu;
}
if (part_desc->nodes != NULL) {
char *backup_node_list = part_ptr->nodes;
if (part_desc->nodes[0] == '\0')
part_ptr->nodes = NULL; /* avoid empty string */
else {
int i;
part_ptr->nodes = xstrdup(part_desc->nodes);
for (i=0; part_ptr->nodes[i]; i++) {
if (isspace(part_ptr->nodes[i]))
part_ptr->nodes[i] = ',';
}
}
error_code = _build_part_bitmap(part_ptr);
if (error_code) {
xfree(part_ptr->nodes);
part_ptr->nodes = backup_node_list;
} else {
info("update_part: setting nodes to %s "
"for partition %s",
part_ptr->nodes, part_desc->name);
xfree(backup_node_list);
}
update_part_nodes_in_resv(part_ptr);
} else if (part_ptr->node_bitmap == NULL) {
/* Newly created partition needs a bitmap, even if empty */
part_ptr->node_bitmap = bit_alloc(node_record_count);
}
if (error_code == SLURM_SUCCESS) {
slurm_sched_partition_change(); /* notify sched plugin */
select_g_reconfigure(); /* notify select plugin too */
}
return error_code;
}
/*
* get_nodes - get information on specific node(s) changed since some time
* cmd_ptr IN - CMD=GETNODES ARG=[<UPDATETIME>:<NODEID>[:<NODEID>]...]
* [<UPDATETIME>:ALL]
* err_code OUT - 0 or an error code
* err_msg OUT - response message
* NOTE: xfree() err_msg if err_code is zero
* RET 0 on success, -1 on failure
*
* Response format
* ARG=<cnt>#<NODEID>:
* STATE=<state>; Moab equivalent node state
* [CAT=<reason>]; Reason for a node being down or drained
* colon separator
* CCLASS=<[part:cpus]>; SLURM partition with CPU count of node,
* make have more than one partition
* [CPULOAD=<load_ave>;] One minute BSD load average
* [ARCH=<architecture>;] Computer architecture
* [OS=<operating_system>;] Operating system
* CMEMORY=<MB>; MB of memory on node
* CDISK=<MB>; MB of disk space on node
* CPROC=<cpus>; CPU count on node
* [FEATURE=<feature>;] Features associated with node, if any
* [GRES=<name>[:<count>],...;] generic resources on the node
* [#<NODEID>:...];
*/
extern int get_nodes(char *cmd_ptr, int *err_code, char **err_msg)
{
char *arg_ptr = NULL, *tmp_char = NULL, *tmp_buf = NULL, *buf = NULL;
time_t update_time;
/* Locks: read node, read partition */
slurmctld_lock_t node_read_lock = {
NO_LOCK, NO_LOCK, READ_LOCK, READ_LOCK };
int node_rec_cnt = 0, buf_size = 0;
#ifdef HAVE_ALPS_CRAY
/* Locks: write node */
slurmctld_lock_t node_write_lock = {
NO_LOCK, NO_LOCK, WRITE_LOCK, NO_LOCK };
/*
* Run a Basil Inventory immediately before scheduling, to avoid
* race conditions caused by ALPS node state change (caused e.g.
* by the node health checker).
* This relies on the above write lock for the node state.
*/
lock_slurmctld(node_write_lock);
if (select_g_reconfigure()) {
unlock_slurmctld(node_write_lock);
*err_code = -720;
*err_msg = "Unable to run ALPS inventory";
error("wiki: Unable to run ALPS inventory");
return -1;
}
unlock_slurmctld(node_write_lock);
#endif
arg_ptr = strstr(cmd_ptr, "ARG=");
if (arg_ptr == NULL) {
*err_code = -300;
*err_msg = "GETNODES lacks ARG";
error("wiki: GETNODES lacks ARG");
return -1;
}
update_time = (time_t) strtoul(arg_ptr+4, &tmp_char, 10);
if (tmp_char[0] != ':') {
*err_code = -300;
*err_msg = "Invalid ARG value";
error("wiki: GETNODES has invalid ARG value");
return -1;
}
tmp_char++;
lock_slurmctld(node_read_lock);
if (strncmp(tmp_char, "ALL", 3) == 0) {
/* report all nodes */
buf = _dump_all_nodes(&node_rec_cnt, update_time);
} else {
struct node_record *node_ptr = NULL;
char *node_name, *slurm_hosts;
int node_cnt;
hostset_t slurm_hostset;
slurm_hosts = moab2slurm_task_list(tmp_char, &node_cnt);
if ((slurm_hostset = hostset_create(slurm_hosts))) {
while ((node_name = hostset_shift(slurm_hostset))) {
node_ptr = find_node_record(node_name);
if (node_ptr == NULL) {
error("sched/wiki2: bad hostname %s",
node_name);
continue;
}
if (_hidden_node(node_ptr))
continue;
tmp_buf = _dump_node(node_ptr, NULL,
update_time);
if (node_rec_cnt > 0)
xstrcat(buf, "#");
xstrcat(buf, tmp_buf);
xfree(tmp_buf);
node_rec_cnt++;
}
hostset_destroy(slurm_hostset);
} else {
error("hostset_create(%s): %m", slurm_hosts);
}
xfree(slurm_hosts);
}
unlock_slurmctld(node_read_lock);
/* Prepend ("ARG=%d", node_rec_cnt) to reply message */
if (buf)
buf_size = strlen(buf);
tmp_buf = xmalloc(buf_size + 32);
if (node_rec_cnt)
sprintf(tmp_buf, "SC=0 ARG=%d#%s", node_rec_cnt, buf);
else
sprintf(tmp_buf, "SC=0 ARG=0#");
xfree(buf);
*err_code = 0;
*err_msg = tmp_buf;
return 0;
}
static int _attempt_backfill(void)
{
DEF_TIMERS;
bool filter_root = false;
List job_queue;
job_queue_rec_t *job_queue_rec;
slurmdb_qos_rec_t *qos_ptr = NULL;
int i, j, node_space_recs;
struct job_record *job_ptr;
struct part_record *part_ptr, **bf_part_ptr = NULL;
uint32_t end_time, end_reserve;
uint32_t time_limit, comp_time_limit, orig_time_limit, part_time_limit;
uint32_t min_nodes, max_nodes, req_nodes;
bitstr_t *avail_bitmap = NULL, *resv_bitmap = NULL;
bitstr_t *exc_core_bitmap = NULL, *non_cg_bitmap = NULL;
time_t now, sched_start, later_start, start_res, resv_end, window_end;
node_space_map_t *node_space;
struct timeval bf_time1, bf_time2;
int rc = 0;
int job_test_count = 0;
uint32_t *uid = NULL, nuser = 0, bf_parts = 0, *bf_part_jobs = NULL;
uint16_t *njobs = NULL;
bool already_counted;
uint32_t reject_array_job_id = 0;
struct part_record *reject_array_part = NULL;
uint32_t job_start_cnt = 0, start_time;
time_t config_update = slurmctld_conf.last_update;
time_t part_update = last_part_update;
struct timeval start_tv;
bf_last_yields = 0;
#ifdef HAVE_ALPS_CRAY
/*
* Run a Basil Inventory immediately before setting up the schedule
* plan, to avoid race conditions caused by ALPS node state change.
* Needs to be done with the node-state lock taken.
*/
START_TIMER;
if (select_g_reconfigure()) {
debug4("backfill: not scheduling due to ALPS");
return SLURM_SUCCESS;
}
END_TIMER;
if (debug_flags & DEBUG_FLAG_BACKFILL)
info("backfill: ALPS inventory completed, %s", TIME_STR);
/* The Basil inventory can take a long time to complete. Process
* pending RPCs before starting the backfill scheduling logic */
_yield_locks(1000000);
#endif
START_TIMER;
if (debug_flags & DEBUG_FLAG_BACKFILL)
info("backfill: beginning");
else
debug("backfill: beginning");
sched_start = now = time(NULL);
gettimeofday(&start_tv, NULL);
if (slurm_get_root_filter())
filter_root = true;
job_queue = build_job_queue(true, true);
if (list_count(job_queue) == 0) {
if (debug_flags & DEBUG_FLAG_BACKFILL)
info("backfill: no jobs to backfill");
else
debug("backfill: no jobs to backfill");
list_destroy(job_queue);
return 0;
}
gettimeofday(&bf_time1, NULL);
non_cg_bitmap = bit_copy(cg_node_bitmap);
bit_not(non_cg_bitmap);
slurmctld_diag_stats.bf_queue_len = list_count(job_queue);
slurmctld_diag_stats.bf_queue_len_sum += slurmctld_diag_stats.
bf_queue_len;
slurmctld_diag_stats.bf_last_depth = 0;
slurmctld_diag_stats.bf_last_depth_try = 0;
slurmctld_diag_stats.bf_when_last_cycle = now;
slurmctld_diag_stats.bf_active = 1;
node_space = xmalloc(sizeof(node_space_map_t) *
(max_backfill_job_cnt * 2 + 1));
node_space[0].begin_time = sched_start;
window_end = sched_start + backfill_window;
node_space[0].end_time = window_end;
node_space[0].avail_bitmap = bit_copy(avail_node_bitmap);
node_space[0].next = 0;
node_space_recs = 1;
if (debug_flags & DEBUG_FLAG_BACKFILL_MAP)
_dump_node_space_table(node_space);
if (max_backfill_job_per_part) {
ListIterator part_iterator;
struct part_record *part_ptr;
bf_parts = list_count(part_list);
bf_part_ptr = xmalloc(sizeof(struct part_record *) * bf_parts);
bf_part_jobs = xmalloc(sizeof(int) * bf_parts);
part_iterator = list_iterator_create(part_list);
i = 0;
while ((part_ptr = (struct part_record *)
list_next(part_iterator))) {
bf_part_ptr[i++] = part_ptr;
}
list_iterator_destroy(part_iterator);
}
if (max_backfill_job_per_user) {
uid = xmalloc(BF_MAX_USERS * sizeof(uint32_t));
njobs = xmalloc(BF_MAX_USERS * sizeof(uint16_t));
}
sort_job_queue(job_queue);
while (1) {
job_queue_rec = (job_queue_rec_t *) list_pop(job_queue);
if (!job_queue_rec) {
if (debug_flags & DEBUG_FLAG_BACKFILL)
info("backfill: reached end of job queue");
break;
}
if (slurmctld_config.shutdown_time)
break;
if (((defer_rpc_cnt > 0) &&
(slurmctld_config.server_thread_count >= defer_rpc_cnt)) ||
(_delta_tv(&start_tv) >= sched_timeout)) {
if (debug_flags & DEBUG_FLAG_BACKFILL) {
END_TIMER;
info("backfill: completed yielding locks "
"after testing %u(%d) jobs, %s",
slurmctld_diag_stats.bf_last_depth,
job_test_count, TIME_STR);
}
if ((_yield_locks(yield_sleep) && !backfill_continue) ||
(slurmctld_conf.last_update != config_update) ||
(last_part_update != part_update)) {
if (debug_flags & DEBUG_FLAG_BACKFILL) {
info("backfill: system state changed, "
"breaking out after testing "
"%u(%d) jobs",
slurmctld_diag_stats.bf_last_depth,
job_test_count);
}
rc = 1;
xfree(job_queue_rec);
break;
}
/* cg_node_bitmap may be changed */
bit_copybits(non_cg_bitmap, cg_node_bitmap);
bit_not(non_cg_bitmap);
/* Reset backfill scheduling timers, resume testing */
sched_start = time(NULL);
gettimeofday(&start_tv, NULL);
job_test_count = 0;
START_TIMER;
}
job_ptr = job_queue_rec->job_ptr;
/* With bf_continue configured, the original job could have
* been cancelled and purged. Validate pointer here. */
if ((job_ptr->magic != JOB_MAGIC) ||
(job_ptr->job_id != job_queue_rec->job_id)) {
xfree(job_queue_rec);
continue;
}
orig_time_limit = job_ptr->time_limit;
part_ptr = job_queue_rec->part_ptr;
job_test_count++;
slurmctld_diag_stats.bf_last_depth++;
already_counted = false;
xfree(job_queue_rec);
if (!IS_JOB_PENDING(job_ptr))
continue; /* started in other partition */
if (!avail_front_end(job_ptr))
continue; /* No available frontend for this job */
if (job_ptr->array_task_id != NO_VAL) {
if ((reject_array_job_id == job_ptr->array_job_id) &&
(reject_array_part == part_ptr))
continue; /* already rejected array element */
/* assume reject whole array for now, clear if OK */
reject_array_job_id = job_ptr->array_job_id;
reject_array_part = part_ptr;
}
job_ptr->part_ptr = part_ptr;
if (debug_flags & DEBUG_FLAG_BACKFILL) {
info("backfill test for JobID=%u Prio=%u Partition=%s",
job_ptr->job_id, job_ptr->priority,
job_ptr->part_ptr->name);
}
if (max_backfill_job_per_part) {
bool skip_job = false;
for (j = 0; j < bf_parts; j++) {
if (bf_part_ptr[j] != job_ptr->part_ptr)
continue;
if (bf_part_jobs[j]++ >=
max_backfill_job_per_part)
skip_job = true;
break;
}
if (skip_job) {
if (debug_flags & DEBUG_FLAG_BACKFILL)
info("backfill: have already "
"checked %u jobs for "
"partition %s; skipping "
"job %u",
max_backfill_job_per_part,
job_ptr->part_ptr->name,
job_ptr->job_id);
continue;
}
}
if (max_backfill_job_per_user) {
for (j = 0; j < nuser; j++) {
if (job_ptr->user_id == uid[j]) {
njobs[j]++;
if (debug_flags & DEBUG_FLAG_BACKFILL)
debug("backfill: user %u: "
"#jobs %u",
uid[j], njobs[j]);
break;
}
}
if (j == nuser) { /* user not found */
static bool bf_max_user_msg = true;
if (nuser < BF_MAX_USERS) {
uid[j] = job_ptr->user_id;
njobs[j] = 1;
nuser++;
} else if (bf_max_user_msg) {
bf_max_user_msg = false;
error("backfill: too many users in "
"queue. Consider increasing "
"BF_MAX_USERS");
}
if (debug_flags & DEBUG_FLAG_BACKFILL)
debug2("backfill: found new user %u. "
"Total #users now %u",
job_ptr->user_id, nuser);
} else {
if (njobs[j] >= max_backfill_job_per_user) {
/* skip job */
if (debug_flags & DEBUG_FLAG_BACKFILL)
info("backfill: have already "
"checked %u jobs for "
"user %u; skipping "
"job %u",
max_backfill_job_per_user,
job_ptr->user_id,
job_ptr->job_id);
continue;
}
}
}
if (((part_ptr->state_up & PARTITION_SCHED) == 0) ||
(part_ptr->node_bitmap == NULL) ||
((part_ptr->flags & PART_FLAG_ROOT_ONLY) && filter_root)) {
if (debug_flags & DEBUG_FLAG_BACKFILL)
info("backfill: partition %s not usable",
job_ptr->part_ptr->name);
continue;
}
if ((!job_independent(job_ptr, 0)) ||
(license_job_test(job_ptr, time(NULL)) != SLURM_SUCCESS)) {
if (debug_flags & DEBUG_FLAG_BACKFILL)
info("backfill: job %u not runable now",
job_ptr->job_id);
continue;
}
/* 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) {
if (debug_flags & DEBUG_FLAG_BACKFILL)
info("backfill: job %u node count too high",
job_ptr->job_id);
continue;
}
/* Determine job's expected completion time */
if (part_ptr->max_time == INFINITE)
part_time_limit = 365 * 24 * 60; /* one year */
else
part_time_limit = part_ptr->max_time;
if (job_ptr->time_limit == NO_VAL) {
time_limit = part_time_limit;
} else {
if (part_ptr->max_time == INFINITE)
time_limit = job_ptr->time_limit;
else
time_limit = MIN(job_ptr->time_limit,
part_time_limit);
}
comp_time_limit = time_limit;
qos_ptr = job_ptr->qos_ptr;
if (qos_ptr && (qos_ptr->flags & QOS_FLAG_NO_RESERVE) &&
slurm_get_preempt_mode())
time_limit = job_ptr->time_limit = 1;
else if (job_ptr->time_min && (job_ptr->time_min < time_limit))
time_limit = job_ptr->time_limit = job_ptr->time_min;
/* Determine impact of any resource reservations */
later_start = now;
TRY_LATER:
if (slurmctld_config.shutdown_time)
break;
if (((defer_rpc_cnt > 0) &&
(slurmctld_config.server_thread_count >= defer_rpc_cnt)) ||
(_delta_tv(&start_tv) >= sched_timeout)) {
uint32_t save_job_id = job_ptr->job_id;
uint32_t save_time_limit = job_ptr->time_limit;
job_ptr->time_limit = orig_time_limit;
if (debug_flags & DEBUG_FLAG_BACKFILL) {
END_TIMER;
info("backfill: completed yielding locks "
"after testing %u(%d) jobs, %s",
slurmctld_diag_stats.bf_last_depth,
job_test_count, TIME_STR);
}
if ((_yield_locks(yield_sleep) && !backfill_continue) ||
(slurmctld_conf.last_update != config_update) ||
(last_part_update != part_update)) {
if (debug_flags & DEBUG_FLAG_BACKFILL) {
info("backfill: system state changed, "
"breaking out after testing "
"%u(%d) jobs",
slurmctld_diag_stats.bf_last_depth,
job_test_count);
}
rc = 1;
break;
}
/* cg_node_bitmap may be changed */
bit_copybits(non_cg_bitmap, cg_node_bitmap);
bit_not(non_cg_bitmap);
/* With bf_continue configured, the original job could
* have been scheduled or cancelled and purged.
* Revalidate job the record here. */
if ((job_ptr->magic != JOB_MAGIC) ||
(job_ptr->job_id != save_job_id))
continue;
if (!IS_JOB_PENDING(job_ptr))
continue;
if (!avail_front_end(job_ptr))
continue; /* No available frontend */
job_ptr->time_limit = save_time_limit;
/* Reset backfill scheduling timers, resume testing */
sched_start = time(NULL);
gettimeofday(&start_tv, NULL);
job_test_count = 1;
START_TIMER;
}
FREE_NULL_BITMAP(avail_bitmap);
FREE_NULL_BITMAP(exc_core_bitmap);
start_res = later_start;
later_start = 0;
j = job_test_resv(job_ptr, &start_res, true, &avail_bitmap,
&exc_core_bitmap);
if (j != SLURM_SUCCESS) {
if (debug_flags & DEBUG_FLAG_BACKFILL)
info("backfill: job %u reservation defer",
job_ptr->job_id);
job_ptr->time_limit = orig_time_limit;
continue;
}
if (start_res > now)
end_time = (time_limit * 60) + start_res;
else
end_time = (time_limit * 60) + now;
resv_end = find_resv_end(start_res);
/* Identify usable nodes for this job */
bit_and(avail_bitmap, part_ptr->node_bitmap);
bit_and(avail_bitmap, up_node_bitmap);
bit_and(avail_bitmap, non_cg_bitmap);
for (j=0; ; ) {
if ((node_space[j].end_time > start_res) &&
node_space[j].next && (later_start == 0))
later_start = node_space[j].end_time;
if (node_space[j].end_time <= start_res)
;
else if (node_space[j].begin_time <= end_time) {
bit_and(avail_bitmap,
node_space[j].avail_bitmap);
} else
break;
if ((j = node_space[j].next) == 0)
break;
}
if (resv_end && (++resv_end < window_end) &&
((later_start == 0) || (resv_end < later_start))) {
later_start = resv_end;
}
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);
}
/* Test if insufficient nodes remain OR
* required nodes missing OR
* nodes lack features OR
* no change since previously tested nodes (only changes
* in other partition nodes) */
if ((bit_set_count(avail_bitmap) < min_nodes) ||
((job_ptr->details->req_node_bitmap) &&
(!bit_super_set(job_ptr->details->req_node_bitmap,
avail_bitmap))) ||
(job_req_node_filter(job_ptr, avail_bitmap))) {
if (later_start) {
job_ptr->start_time = 0;
goto TRY_LATER;
}
/* Job can not start until too far in the future */
job_ptr->time_limit = orig_time_limit;
job_ptr->start_time = sched_start + backfill_window;
continue;
}
/* Identify nodes which are definitely off limits */
FREE_NULL_BITMAP(resv_bitmap);
resv_bitmap = bit_copy(avail_bitmap);
bit_not(resv_bitmap);
/* this is the time consuming operation */
debug2("backfill: entering _try_sched for job %u.",
job_ptr->job_id);
if (!already_counted) {
slurmctld_diag_stats.bf_last_depth_try++;
already_counted = true;
}
if (debug_flags & DEBUG_FLAG_BACKFILL_MAP)
_dump_job_test(job_ptr, avail_bitmap, start_res);
j = _try_sched(job_ptr, &avail_bitmap, min_nodes, max_nodes,
req_nodes, exc_core_bitmap);
now = time(NULL);
if (j != SLURM_SUCCESS) {
job_ptr->time_limit = orig_time_limit;
job_ptr->start_time = 0;
continue; /* not runable */
}
if (start_res > job_ptr->start_time) {
job_ptr->start_time = start_res;
last_job_update = now;
}
if (job_ptr->start_time <= now) { /* Can start now */
uint32_t save_time_limit = job_ptr->time_limit;
uint32_t hard_limit;
bool reset_time = false;
int rc = _start_job(job_ptr, resv_bitmap);
if (qos_ptr && (qos_ptr->flags & QOS_FLAG_NO_RESERVE)) {
if (orig_time_limit == NO_VAL) {
acct_policy_alter_job(
job_ptr, comp_time_limit);
job_ptr->time_limit = comp_time_limit;
} else {
acct_policy_alter_job(
job_ptr, orig_time_limit);
job_ptr->time_limit = orig_time_limit;
}
} else if ((rc == SLURM_SUCCESS) && job_ptr->time_min) {
/* Set time limit as high as possible */
acct_policy_alter_job(job_ptr, comp_time_limit);
job_ptr->time_limit = comp_time_limit;
reset_time = true;
} else if (orig_time_limit == NO_VAL) {
acct_policy_alter_job(job_ptr, comp_time_limit);
job_ptr->time_limit = comp_time_limit;
} else {
acct_policy_alter_job(job_ptr, orig_time_limit);
job_ptr->time_limit = orig_time_limit;
}
if (job_ptr->time_limit == INFINITE)
hard_limit = 365 * 24 * 60; /* one year */
else
hard_limit = job_ptr->time_limit;
job_ptr->end_time = job_ptr->start_time +
(hard_limit * 60);
if (reset_time) {
_reset_job_time_limit(job_ptr, now,
node_space);
time_limit = job_ptr->time_limit;
}
if (rc == ESLURM_ACCOUNTING_POLICY) {
/* Unknown future start time, just skip job */
job_ptr->start_time = 0;
continue;
} else if (rc != SLURM_SUCCESS) {
if (debug_flags & DEBUG_FLAG_BACKFILL) {
info("backfill: planned start of job %u"
" failed: %s", job_ptr->job_id,
slurm_strerror(rc));
}
/* Drop through and reserve these resources.
* Likely due to state changes during sleep.
* Make best-effort based upon original state */
job_ptr->time_limit = orig_time_limit;
later_start = 0;
} else {
/* Started this job, move to next one */
reject_array_job_id = 0;
reject_array_part = NULL;
/* Update the database if job time limit
* changed and move to next job */
if (save_time_limit != job_ptr->time_limit)
jobacct_storage_g_job_start(acct_db_conn,
job_ptr);
job_start_cnt++;
if (max_backfill_jobs_start &&
(job_start_cnt >= max_backfill_jobs_start)){
if (debug_flags & DEBUG_FLAG_BACKFILL) {
info("backfill: bf_max_job_start"
" limit of %d reached",
max_backfill_jobs_start);
}
break;
}
continue;
}
} else {
job_ptr->time_limit = orig_time_limit;
}
start_time = job_ptr->start_time;
end_reserve = job_ptr->start_time + (time_limit * 60);
start_time = (start_time / backfill_resolution) *
backfill_resolution;
end_reserve = (end_reserve / backfill_resolution) *
backfill_resolution;
if (later_start && (start_time > later_start)) {
/* Try later when some nodes currently reserved for
* pending jobs are free */
job_ptr->start_time = 0;
goto TRY_LATER;
}
if (job_ptr->start_time > (sched_start + backfill_window)) {
/* Starts too far in the future to worry about */
if (debug_flags & DEBUG_FLAG_BACKFILL)
_dump_job_sched(job_ptr, end_reserve,
avail_bitmap);
continue;
}
if (node_space_recs >= max_backfill_job_cnt) {
if (debug_flags & DEBUG_FLAG_BACKFILL) {
info("backfill: table size limit of %u reached",
max_backfill_job_cnt);
}
break;
}
if ((job_ptr->start_time > now) &&
_test_resv_overlap(node_space, avail_bitmap,
start_time, end_reserve)) {
/* This job overlaps with an existing reservation for
* job to be backfill scheduled, which the sched
* plugin does not know about. Try again later. */
later_start = job_ptr->start_time;
job_ptr->start_time = 0;
goto TRY_LATER;
}
/*
* Add reservation to scheduling table if appropriate
*/
if (debug_flags & DEBUG_FLAG_BACKFILL)
_dump_job_sched(job_ptr, end_reserve, avail_bitmap);
if (qos_ptr && (qos_ptr->flags & QOS_FLAG_NO_RESERVE))
continue;
reject_array_job_id = 0;
reject_array_part = NULL;
if (debug_flags & DEBUG_FLAG_BACKFILL)
_dump_job_sched(job_ptr, end_reserve, avail_bitmap);
xfree(job_ptr->sched_nodes);
job_ptr->sched_nodes = bitmap2node_name(avail_bitmap);
bit_not(avail_bitmap);
_add_reservation(start_time, end_reserve,
avail_bitmap, node_space, &node_space_recs);
if (debug_flags & DEBUG_FLAG_BACKFILL_MAP)
_dump_node_space_table(node_space);
}
xfree(bf_part_jobs);
xfree(bf_part_ptr);
xfree(uid);
xfree(njobs);
FREE_NULL_BITMAP(avail_bitmap);
FREE_NULL_BITMAP(exc_core_bitmap);
FREE_NULL_BITMAP(resv_bitmap);
FREE_NULL_BITMAP(non_cg_bitmap);
for (i=0; ; ) {
FREE_NULL_BITMAP(node_space[i].avail_bitmap);
if ((i = node_space[i].next) == 0)
break;
}
xfree(node_space);
list_destroy(job_queue);
gettimeofday(&bf_time2, NULL);
_do_diag_stats(&bf_time1, &bf_time2, yield_sleep);
if (debug_flags & DEBUG_FLAG_BACKFILL) {
END_TIMER;
info("backfill: completed testing %u(%d) jobs, %s",
slurmctld_diag_stats.bf_last_depth,
job_test_count, TIME_STR);
}
return rc;
}
static int _attempt_backfill(void)
{
bool filter_root = false;
List job_queue;
job_queue_rec_t *job_queue_rec;
slurmdb_qos_rec_t *qos_ptr = NULL;
int i, j, node_space_recs;
struct job_record *job_ptr;
struct part_record *part_ptr;
uint32_t end_time, end_reserve;
uint32_t time_limit, comp_time_limit, orig_time_limit;
uint32_t min_nodes, max_nodes, req_nodes;
bitstr_t *avail_bitmap = NULL, *resv_bitmap = NULL;
time_t now = time(NULL), sched_start, later_start, start_res;
node_space_map_t *node_space;
static int sched_timeout = 0;
int this_sched_timeout = 0, rc = 0;
sched_start = now;
if (sched_timeout == 0) {
sched_timeout = slurm_get_msg_timeout() / 2;
sched_timeout = MAX(sched_timeout, 1);
sched_timeout = MIN(sched_timeout, 10);
}
this_sched_timeout = sched_timeout;
#ifdef HAVE_CRAY
/*
* Run a Basil Inventory immediately before setting up the schedule
* plan, to avoid race conditions caused by ALPS node state change.
* Needs to be done with the node-state lock taken.
*/
if (select_g_reconfigure()) {
debug4("backfill: not scheduling due to ALPS");
return SLURM_SUCCESS;
}
#endif
if (slurm_get_root_filter())
filter_root = true;
job_queue = build_job_queue(true);
if (list_count(job_queue) <= 1) {
debug("backfill: no jobs to backfill");
list_destroy(job_queue);
return 0;
}
node_space = xmalloc(sizeof(node_space_map_t) *
(max_backfill_job_cnt + 3));
node_space[0].begin_time = sched_start;
node_space[0].end_time = sched_start + backfill_window;
node_space[0].avail_bitmap = bit_copy(avail_node_bitmap);
node_space[0].next = 0;
node_space_recs = 1;
if (debug_flags & DEBUG_FLAG_BACKFILL)
_dump_node_space_table(node_space);
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 (!IS_JOB_PENDING(job_ptr))
continue; /* started in other partition */
job_ptr->part_ptr = part_ptr;
if (debug_flags & DEBUG_FLAG_BACKFILL)
info("backfill test for job %u", job_ptr->job_id);
if ((job_ptr->state_reason == WAIT_ASSOC_JOB_LIMIT) ||
(job_ptr->state_reason == WAIT_ASSOC_RESOURCE_LIMIT) ||
(job_ptr->state_reason == WAIT_ASSOC_TIME_LIMIT) ||
(job_ptr->state_reason == WAIT_QOS_JOB_LIMIT) ||
(job_ptr->state_reason == WAIT_QOS_RESOURCE_LIMIT) ||
(job_ptr->state_reason == WAIT_QOS_TIME_LIMIT) ||
!acct_policy_job_runnable(job_ptr)) {
debug2("backfill: job %u is not allowed to run now. "
"Skipping it. State=%s. Reason=%s. Priority=%u",
job_ptr->job_id,
job_state_string(job_ptr->job_state),
job_reason_string(job_ptr->state_reason),
job_ptr->priority);
continue;
}
if (((part_ptr->state_up & PARTITION_SCHED) == 0) ||
(part_ptr->node_bitmap == NULL))
continue;
if ((part_ptr->flags & PART_FLAG_ROOT_ONLY) && filter_root)
continue;
if ((!job_independent(job_ptr, 0)) ||
(license_job_test(job_ptr, time(NULL)) != SLURM_SUCCESS))
continue;
/* 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;
}
/* Determine job's expected completion time */
if (job_ptr->time_limit == NO_VAL) {
if (part_ptr->max_time == INFINITE)
time_limit = 365 * 24 * 60; /* one year */
else
time_limit = part_ptr->max_time;
} else {
if (part_ptr->max_time == INFINITE)
time_limit = job_ptr->time_limit;
else
time_limit = MIN(job_ptr->time_limit,
part_ptr->max_time);
}
comp_time_limit = time_limit;
orig_time_limit = job_ptr->time_limit;
if (qos_ptr && (qos_ptr->flags & QOS_FLAG_NO_RESERVE))
time_limit = job_ptr->time_limit = 1;
else if (job_ptr->time_min && (job_ptr->time_min < time_limit))
time_limit = job_ptr->time_limit = job_ptr->time_min;
/* Determine impact of any resource reservations */
later_start = now;
TRY_LATER:
FREE_NULL_BITMAP(avail_bitmap);
start_res = later_start;
later_start = 0;
j = job_test_resv(job_ptr, &start_res, true, &avail_bitmap);
if (j != SLURM_SUCCESS) {
job_ptr->time_limit = orig_time_limit;
continue;
}
if (start_res > now)
end_time = (time_limit * 60) + start_res;
else
end_time = (time_limit * 60) + now;
/* Identify usable nodes for this job */
bit_and(avail_bitmap, part_ptr->node_bitmap);
bit_and(avail_bitmap, up_node_bitmap);
for (j=0; ; ) {
if ((node_space[j].end_time > start_res) &&
node_space[j].next && (later_start == 0))
later_start = node_space[j].end_time;
if (node_space[j].end_time <= start_res)
;
else if (node_space[j].begin_time <= end_time) {
bit_and(avail_bitmap,
node_space[j].avail_bitmap);
} else
break;
if ((j = node_space[j].next) == 0)
break;
}
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);
}
/* Test if insufficient nodes remain OR
* required nodes missing OR
* nodes lack features */
if ((bit_set_count(avail_bitmap) < min_nodes) ||
((job_ptr->details->req_node_bitmap) &&
(!bit_super_set(job_ptr->details->req_node_bitmap,
avail_bitmap))) ||
(job_req_node_filter(job_ptr, avail_bitmap))) {
if (later_start) {
job_ptr->start_time = 0;
goto TRY_LATER;
}
job_ptr->time_limit = orig_time_limit;
continue;
}
/* Identify nodes which are definitely off limits */
FREE_NULL_BITMAP(resv_bitmap);
resv_bitmap = bit_copy(avail_bitmap);
bit_not(resv_bitmap);
if ((time(NULL) - sched_start) >= this_sched_timeout) {
debug("backfill: loop taking too long, yielding locks");
if (_yield_locks()) {
debug("backfill: system state changed, "
"breaking out");
rc = 1;
break;
} else {
this_sched_timeout += sched_timeout;
}
}
/* this is the time consuming operation */
debug2("backfill: entering _try_sched for job %u.",
job_ptr->job_id);
j = _try_sched(job_ptr, &avail_bitmap,
min_nodes, max_nodes, req_nodes);
debug2("backfill: finished _try_sched for job %u.",
job_ptr->job_id);
now = time(NULL);
if (j != SLURM_SUCCESS) {
job_ptr->time_limit = orig_time_limit;
job_ptr->start_time = 0;
continue; /* not runable */
}
if (start_res > job_ptr->start_time) {
job_ptr->start_time = start_res;
last_job_update = now;
}
if (job_ptr->start_time <= now) {
int rc = _start_job(job_ptr, resv_bitmap);
if (qos_ptr && (qos_ptr->flags & QOS_FLAG_NO_RESERVE))
job_ptr->time_limit = orig_time_limit;
else if ((rc == SLURM_SUCCESS) && job_ptr->time_min) {
/* Set time limit as high as possible */
job_ptr->time_limit = comp_time_limit;
job_ptr->end_time = job_ptr->start_time +
(comp_time_limit * 60);
_reset_job_time_limit(job_ptr, now,
node_space);
time_limit = job_ptr->time_limit;
} else {
job_ptr->time_limit = orig_time_limit;
}
if (rc == ESLURM_ACCOUNTING_POLICY) {
/* Unknown future start time, just skip job */
job_ptr->start_time = 0;
continue;
} else if (rc != SLURM_SUCCESS) {
/* Planned to start job, but something bad
* happended. */
job_ptr->start_time = 0;
break;
} else {
/* Started this job, move to next one */
continue;
}
} else
job_ptr->time_limit = orig_time_limit;
if (later_start && (job_ptr->start_time > later_start)) {
/* Try later when some nodes currently reserved for
* pending jobs are free */
job_ptr->start_time = 0;
goto TRY_LATER;
}
if (job_ptr->start_time > (sched_start + backfill_window)) {
/* Starts too far in the future to worry about */
continue;
}
if (node_space_recs >= max_backfill_job_cnt) {
/* Already have too many jobs to deal with */
break;
}
end_reserve = job_ptr->start_time + (time_limit * 60);
if (_test_resv_overlap(node_space, avail_bitmap,
job_ptr->start_time, end_reserve)) {
/* This job overlaps with an existing reservation for
* job to be backfill scheduled, which the sched
* plugin does not know about. Try again later. */
later_start = job_ptr->start_time;
job_ptr->start_time = 0;
goto TRY_LATER;
}
/*
* Add reservation to scheduling table if appropriate
*/
qos_ptr = job_ptr->qos_ptr;
if (qos_ptr && (qos_ptr->flags & QOS_FLAG_NO_RESERVE))
continue;
bit_not(avail_bitmap);
_add_reservation(job_ptr->start_time, end_reserve,
avail_bitmap, node_space, &node_space_recs);
if (debug_flags & DEBUG_FLAG_BACKFILL)
_dump_node_space_table(node_space);
}
FREE_NULL_BITMAP(avail_bitmap);
FREE_NULL_BITMAP(resv_bitmap);
for (i=0; ; ) {
FREE_NULL_BITMAP(node_space[i].avail_bitmap);
if ((i = node_space[i].next) == 0)
break;
}
xfree(node_space);
list_destroy(job_queue);
return rc;
}
