From 89500f1b18d0e9d083350691e62d2b90da9428b5 Mon Sep 17 00:00:00 2001
From: Danny Auble <da@schedmd.com>
Date: Wed, 15 Jan 2014 16:04:11 -0800
Subject: [PATCH] CRAY - Add node flag noting it is in a blade using Network
performance counters
---
slurm/slurm.h.in | 9 +++++-
src/common/slurm_protocol_defs.c | 8 +++++-
src/plugins/select/cray/select_cray.c | 41 +++++++++++++++++++++++++++
3 files changed, 56 insertions(+), 2 deletions(-)
diff --git a/slurm/slurm.h.in b/slurm/slurm.h.in
index 362e71b2e31..d763e945c21 100644
--- a/slurm/slurm.h.in
+++ b/slurm/slurm.h.in
@@ -584,7 +584,14 @@ enum node_states {
};
#define NODE_STATE_BASE 0x000f
#define NODE_STATE_FLAGS 0xfff0
-#define NODE_STATE_RES 0x0020
+#define NODE_STATE_NET 0x0010 /* If a node is using Cray's
+ * Network Performance
+ * Counters but isn't in a
+ * allocation. */
+#define NODE_STATE_RES 0x0020 /* If a node is in a
+ * reservation (used primarily
+ * to note a node isn't idle
+ * for non-reservation jobs) */
#define NODE_STATE_UNDRAIN 0x0040 /* Clear DRAIN flag for a node */
#define NODE_STATE_CLOUD 0x0080 /* node comes from cloud */
#define NODE_RESUME 0x0100 /* Restore a DRAINED, DRAINING, DOWN
diff --git a/src/common/slurm_protocol_defs.c b/src/common/slurm_protocol_defs.c
index e241abbe985..a2f3b3c7293 100644
--- a/src/common/slurm_protocol_defs.c
+++ b/src/common/slurm_protocol_defs.c
@@ -1506,6 +1506,7 @@ extern char *node_state_string(uint16_t inx)
bool drain_flag = (inx & NODE_STATE_DRAIN);
bool fail_flag = (inx & NODE_STATE_FAIL);
bool maint_flag = (inx & NODE_STATE_MAINT);
+ bool net_flag = (inx & NODE_STATE_NET);
bool res_flag = (inx & NODE_STATE_RES);
bool resume_flag = (inx & NODE_RESUME);
bool no_resp_flag = (inx & NODE_STATE_NO_RESPOND);
@@ -1585,6 +1586,8 @@ extern char *node_state_string(uint16_t inx)
return "IDLE~";
if (no_resp_flag)
return "IDLE*";
+ if (net_flag)
+ return "NETPERF";
if (res_flag)
return "RESERVED";
return "IDLE";
@@ -1628,8 +1631,9 @@ extern char *node_state_string_compact(uint16_t inx)
bool drain_flag = (inx & NODE_STATE_DRAIN);
bool fail_flag = (inx & NODE_STATE_FAIL);
bool maint_flag = (inx & NODE_STATE_MAINT);
- bool resume_flag = (inx & NODE_RESUME);
+ bool net_flag = (inx & NODE_STATE_NET);
bool res_flag = (inx & NODE_STATE_RES);
+ bool resume_flag = (inx & NODE_RESUME);
bool no_resp_flag = (inx & NODE_STATE_NO_RESPOND);
bool power_down_flag = (inx & NODE_STATE_POWER_SAVE);
bool power_up_flag = (inx & NODE_STATE_POWER_UP);
@@ -1708,6 +1712,8 @@ extern char *node_state_string_compact(uint16_t inx)
return "IDLE~";
if (no_resp_flag)
return "IDLE*";
+ if (net_flag)
+ return "NPC";
if (res_flag)
return "RESV";
return "IDLE";
diff --git a/src/plugins/select/cray/select_cray.c b/src/plugins/select/cray/select_cray.c
index a727d0ed44d..9032eab8816 100644
--- a/src/plugins/select/cray/select_cray.c
+++ b/src/plugins/select/cray/select_cray.c
@@ -142,12 +142,14 @@ int bg_recover __attribute__((weak_import)) = NOT_FROM_CONTROLLER;
slurmdb_cluster_rec_t *working_cluster_rec __attribute__((weak_import)) = NULL;
struct node_record *node_record_table_ptr __attribute__((weak_import));
int node_record_count __attribute__((weak_import));
+time_t last_node_update __attribute__((weak_import));
#else
slurm_ctl_conf_t slurmctld_conf;
int bg_recover = NOT_FROM_CONTROLLER;
slurmdb_cluster_rec_t *working_cluster_rec = NULL;
struct node_record *node_record_table_ptr;
int node_record_count;
+time_t last_node_update;
#endif
static blade_info_t *blade_array = NULL;
@@ -155,6 +157,7 @@ static bitstr_t *blades_running_jobs = NULL;
static bitstr_t *blades_running_npc = NULL;
static uint32_t blade_cnt = 0;
static pthread_mutex_t blade_mutex = PTHREAD_MUTEX_INITIALIZER;
+static time_t last_npc_update;
#ifdef HAVE_NATIVE_CRAY
@@ -743,6 +746,10 @@ static void _remove_job_from_blades(select_jobinfo_t *jobinfo)
bit_not(blades_running_jobs);
}
}
+
+ if (jobinfo->npc)
+ last_npc_update = time(NULL);
+
slurm_mutex_unlock(&blade_mutex);
}
@@ -802,6 +809,9 @@ static void _set_job_running(struct job_record *job_ptr)
node_bitmap);
}
}
+
+ if (jobinfo->npc)
+ last_npc_update = time(NULL);
}
/* job_write and blade_mutex must be locked before calling */
@@ -824,6 +834,9 @@ static void _set_job_running_restore(select_jobinfo_t *jobinfo)
if (jobinfo->npc)
bit_or(blades_running_npc, blade_array[i].node_bitmap);
}
+
+ if (jobinfo->npc)
+ last_npc_update = time(NULL);
}
static void *_job_fini(void *args)
@@ -1804,6 +1817,34 @@ unpack_error:
extern int select_p_select_nodeinfo_set_all(void)
{
+ int i;
+ static time_t last_set_all = 0;
+
+ /* only set this once when the last_bg_update is newer than
+ the last time we set things up. */
+ if (last_set_all && (last_npc_update-1 < last_set_all)) {
+ debug3("Node select info for set all hasn't "
+ "changed since %ld",
+ last_set_all);
+ return SLURM_NO_CHANGE_IN_DATA;
+ }
+ last_set_all = last_npc_update;
+
+ /* set this here so we know things have changed */
+ last_node_update = time(NULL);
+
+ slurm_mutex_lock(&blade_mutex);
+ /* clear all marks */
+ for (i=0; i<node_record_count; i++) {
+ struct node_record *node_ptr = &(node_record_table_ptr[i]);
+ if (bit_test(blades_running_npc, i))
+ node_ptr->node_state |= NODE_STATE_NET;
+ else
+ node_ptr->node_state &= (~NODE_STATE_NET);
+ }
+
+ slurm_mutex_unlock(&blade_mutex);
+
return other_select_nodeinfo_set_all();
}
--
GitLab