03-hpc-tuning: Hands-on B: Add example jobscript

03-hpc-tuning/B.md

@@ -73,6 +73,37 @@ Use the OpenMP environment variables and/or `numactl` to configure the pinning o
 
 * What is the inter-socket memory-bandwidth in GB/s?
 
+Here is an example jobscript for the fourth case with pinning 8 threads to the first socket and data placement on the second socket.
+Note that `numactl` expects "NUMA nodes" of which there are 4 per socket on the compute nodes.
+
+```bash
+#!/usr/bin/env bash
+#SBATCH --job-name=stream
+#SBATCH --nodes=1
+#SBATCH --ntasks-per-node=1
+#SBATCH --time=00:30:00
+#SBATCH --account=hpc-lco-usrtr
+#SBATCH --exclusive
+
+module reset
+module load toolchain/intel/2022a
+
+export OMP_DISPLAY_AFFINITY=true
+
+# set OPENMP to
+# - use 8 worker threads
+# - spread threads over allowed cores
+# - pin threads to single cores
+export OMP_NUM_THREADS=8
+export OMP_PROC_BIND=spread
+export OMP_PLACES=cores
+
+# mask threads to use NUMA domains on first socket (0,1,2,3) and
+# mask memory allocation to use NUMA domains on second socket (4,5,6,7)
+numactl --cpunodebind=0,1,2,3 --membind=4,5,6,7 ./stream_c.exe
+```
+
+
 ---
 
 Find the solution to these exercises [in the solution branch](/../solution/03-hpc-tuning).