Changeset 3817

Timestamp:

Jun 27, 2025, 6:59:42 PM (45 hours ago)

Author:

emillour

Message:

Mars PCM:
Move adastra_cpu_binding.sh to "adastra" subforlder of deftank.
And adapt scripts and README there.
EM

Location:

trunk/LMDZ.MARS

Files:

• changelog.txt (modified) (1 diff)
• deftank/adastra/README (modified) (2 diffs)
• deftank/adastra/adastra_cpu_binding.sh (moved) (moved from trunk/LMDZ.MARS/deftank/adastra_cpu_binding.sh)
• deftank/adastra/run0 (modified) (1 diff)
• deftank/adastra/run_month1 (modified) (1 diff)

trunk/LMDZ.MARS/changelog.txt

@@ -4879 +4879 @@
   - More interactive (keep asking the user to plot variables until he/she quits for ex);
   - Simplification and generalization of the script to get user answers.
+
+== 27/06/2025 == EM
+Move adastra_cpu_binding.sh to "adastra" subforlder of deftank.
+And adapt scripts and README there.
+

trunk/LMDZ.MARS/deftank/adastra/README

@@ -1 @@
->> EXAMPLES FOR OCCIGEN CLUSTER
+>> EXAMPLES FOR ADASTRA CLUSTER
 Job "run_month1" runs a simulation for month #1 and then possibly launches
 a job "run_month2" which runs a simulation for month #2 which then possibly
@@ -8 +8 @@
 >> to match your case
 
->> You might also need to adapt "run_month1" to match number of OpenMP threads
->> which MUST be specified both as
-#SBATCH --cpus-per-task=...
->> in the job header
->> AND via the environment variable
-export OMP_NUM_THREADS=...
->> in the script
-
+>> You might also need to adapt "run_month1" if you run with less than 192
+cores (ie. 24 MPI times 8 OpenMP).
+>> If using 192 cores then you should aslo have the companion script
+adastra_cpu_binding.sh
+>> alongside "run_month1" 

trunk/LMDZ.MARS/deftank/adastra/run0

@@ -55 +55 @@
 
 
-# Run GCM
-srun --cpu-bind=threads --label -c${OMP_NUM_THREADS:=1} $gcm > lrun${numnew} 2>&1
+# Run GCM using srun (no cpu binding... OK for MPI only)
+#srun --cpu-bind=threads --label -c${OMP_NUM_THREADS:=1} $gcm > lrun${numnew} 2>&1
+# Ideally using adequate cpu binding, especially if using mixed MPI/OpenMP on a full node
+srun --ntasks-per-node=${SLURM_NTASKS_PER_NODE} --cpu-bind=none --mem-bind=none --label -- ./adastra_cpu_binding.sh ./$gcm > lrun${numnew} 2>&1
 
 

trunk/LMDZ.MARS/deftank/adastra/run_month1

@@ -5 +5 @@
 #SBATCH --nodes=1
 #SBATCH --ntasks-per-node=24
-#SBATCH --cpus-per-task=4
+#SBATCH --cpus-per-task=8
 #SBATCH --threads-per-core=1 # --hint=nomultithread
 #SBATCH --exclusive