From 43281b779ed0ed9cf69fd030c61fa2012e963b10 Mon Sep 17 00:00:00 2001
From: Moe Jette <jette1@llnl.gov>
Date: Tue, 25 Nov 2008 18:18:33 +0000
Subject: [PATCH] Fix job preemption when sched/gang and select/linear are
configured with non-sharing partitions. preempt_linear.patch from
Chris Holmes
---
NEWS | 2 +
src/plugins/select/cons_res/select_cons_res.c | 2 +-
src/plugins/select/linear/select_linear.c | 71 ++++++++++++++++++-
src/slurmctld/node_scheduler.c | 7 ++
4 files changed, 80 insertions(+), 2 deletions(-)
diff --git a/NEWS b/NEWS
index e4d75a0dc72..a6fd38f1bc6 100644
--- a/NEWS
+++ b/NEWS
@@ -3,6 +3,8 @@ documents those changes that are of interest to users and admins.
* Changes in SLURM 1.4.0-pre6
=============================
+ -- Fix job preemption when sched/gang and select/linear are configured with
+ non-sharing partitions.
* Changes in SLURM 1.4.0-pre5
=============================
diff --git a/src/plugins/select/cons_res/select_cons_res.c b/src/plugins/select/cons_res/select_cons_res.c
index 6e453a80736..4bce32878fa 100644
--- a/src/plugins/select/cons_res/select_cons_res.c
+++ b/src/plugins/select/cons_res/select_cons_res.c
@@ -228,7 +228,7 @@ extern bool cr_priority_selection_enabled()
if (strcmp(sched_type, "sched/gang") == 0)
cr_priority_selection = true;
xfree(sched_type);
- cr_priority_selection = true;
+ cr_priority_test = true;
}
return cr_priority_selection;
diff --git a/src/plugins/select/linear/select_linear.c b/src/plugins/select/linear/select_linear.c
index 99038cce410..f2992bf12a0 100644
--- a/src/plugins/select/linear/select_linear.c
+++ b/src/plugins/select/linear/select_linear.c
@@ -135,6 +135,8 @@ static struct node_record *select_node_ptr = NULL;
static int select_node_cnt = 0;
static uint16_t select_fast_schedule;
static uint16_t cr_type;
+static bool cr_priority_test = false;
+static bool cr_priority_selection = false;
static struct node_cr_record *node_cr_ptr = NULL;
static pthread_mutex_t cr_mutex = PTHREAD_MUTEX_INITIALIZER;
@@ -240,6 +242,19 @@ static int _fini_status_pthread(void)
}
#endif
+static inline bool _cr_priority_selection_enabled(void)
+{
+ if (!cr_priority_test) {
+ char *sched_type = slurm_get_sched_type();
+ if (strcmp(sched_type, "sched/gang") == 0)
+ cr_priority_selection = true;
+ xfree(sched_type);
+ cr_priority_test = true;
+ }
+ return cr_priority_selection;
+
+}
+
static bool _enough_nodes(int avail_nodes, int rem_nodes,
uint32_t min_nodes, uint32_t req_nodes)
{
@@ -556,7 +571,7 @@ extern int select_p_job_test(struct job_record *job_ptr, bitstr_t *bitmap,
}
if (mode != SELECT_MODE_TEST_ONLY) {
- if (job_ptr->details->shared == 1) {
+ if (job_ptr->details->shared) {
max_share = job_ptr->part_ptr->max_share &
~SHARED_FORCE;
} else /* ((shared == 0) || (shared == (uint16_t) NO_VAL)) */
@@ -575,6 +590,8 @@ extern int select_p_job_test(struct job_record *job_ptr, bitstr_t *bitmap,
job_ptr->details->job_min_memory = 0;
}
+ debug3("select/linear: job_test: job %u max_share %d avail nodes %u",
+ job_ptr->job_id, max_share, bit_set_count(bitmap));
orig_map = bit_copy(bitmap);
for (max_run_job=min_share; max_run_job<max_share; max_run_job++) {
bool last_iteration = (max_run_job == (max_share -1));
@@ -586,6 +603,8 @@ extern int select_p_job_test(struct job_record *job_ptr, bitstr_t *bitmap,
orig_map, bitmap,
max_run_job,
max_run_job + sus_jobs);
+ debug3("select/linear: job_test: found %d nodes for %u",
+ j, job_ptr->job_id);
if ((j == prev_cnt) || (j < min_nodes))
continue;
prev_cnt = j;
@@ -697,12 +716,62 @@ static int _job_count_bitmap(struct node_cr_record *node_cr_ptr,
}
if ((run_job_cnt != NO_SHARE_LIMIT) &&
+ (!_cr_priority_selection_enabled()) &&
(node_cr_ptr[i].exclusive_jobid != 0)) {
/* already reserved by some exclusive job */
bit_clear(jobmap, i);
continue;
}
+ if (_cr_priority_selection_enabled()) {
+ /* clear this node if any higher-priority
+ * partitions have existing allocations */
+ total_jobs = 0;
+ part_cr_ptr = node_cr_ptr[i].parts;
+ for( ;part_cr_ptr; part_cr_ptr = part_cr_ptr->next) {
+ if (part_cr_ptr->part_ptr->priority <=
+ job_ptr->part_ptr->priority)
+ continue;
+ total_jobs += part_cr_ptr->tot_job_cnt;
+ }
+ if ((run_job_cnt != NO_SHARE_LIMIT) &&
+ (total_jobs > 0)) {
+ bit_clear(jobmap, i);
+ continue;
+ }
+ /* if not sharing, then check with other partitions
+ * of equal priority. Otherwise, load-balance within
+ * the local partition */
+ total_jobs = 0;
+ total_run_jobs = 0;
+ part_cr_ptr = node_cr_ptr[i].parts;
+ for( ; part_cr_ptr; part_cr_ptr = part_cr_ptr->next) {
+ if (part_cr_ptr->part_ptr->priority !=
+ job_ptr->part_ptr->priority)
+ continue;
+ if (!job_ptr->details->shared) {
+ total_run_jobs +=
+ part_cr_ptr->run_job_cnt;
+ total_jobs += part_cr_ptr->tot_job_cnt;
+ continue;
+ }
+ if (part_cr_ptr->part_ptr == job_ptr->part_ptr){
+ total_run_jobs +=
+ part_cr_ptr->run_job_cnt;
+ total_jobs += part_cr_ptr->tot_job_cnt;
+ break;
+ }
+ }
+ if ((total_run_jobs <= run_job_cnt) &&
+ (total_jobs <= tot_job_cnt)) {
+ bit_set(jobmap, i);
+ count++;
+ } else {
+ bit_clear(jobmap, i);
+ }
+ continue;
+ }
+
total_jobs = 0;
total_run_jobs = 0;
part_cr_ptr = node_cr_ptr[i].parts;
diff --git a/src/slurmctld/node_scheduler.c b/src/slurmctld/node_scheduler.c
index 51260782bcc..1997099601c 100644
--- a/src/slurmctld/node_scheduler.c
+++ b/src/slurmctld/node_scheduler.c
@@ -278,6 +278,13 @@ static int _match_feature(char *seek, struct node_set *node_set_ptr)
* (uint16_t)NO_VAL = default
* 0 = exclusive
* 1 = share=yes
+ *
+ * Return values:
+ * 0 = no sharing
+ * 1 = user requested sharing
+ * 2 = sharing enforced (either by partition or cons_res)
+ * (cons_res plugin needs to distinguish between "enforced" and
+ * "requested" sharing)
*/
static int
_resolve_shared_status(uint16_t user_flag, uint16_t part_max_share,
--
GitLab