added zen3 example for multi node batch script

04-zen3_0512-multi-node/01-multi-node-full-node.sh

+#!/bin/bash
+
+##############################################################################
+# User Request:
+#   - run the same program 256 times on multiple nodes
+#   - for each program (aka job step) we will use 1 physical core 
+#   - limit time to 5 minutes
+#
+# Provided Allocation:
+#   - exclusive access
+#   - 2 nodes
+#       - 2x128 physical cores / 2x256 logical cores
+#       - 2x500 GB memory in total
+#       - 128 tasks per node
+#   - 1 physical core bound to each task
+#   - 500 GB memory accessible for each task
+#       - memory access is not restricted by slurm and needs to be managed 
+#         by the application
+#
+# VSC policy:
+#   - shared=0 -> exclusive node access
+#   
+# Accounting:
+#   - 2x 128 = 256 core hours / hour
+##############################################################################
+
+#SBATCH --job-name="multi node; 2 nodes; 256 tasks; full node per task"
+#SBATCH --nodes=2
+#SBATCH --ntasks-per-node=128
+# implicit default: --cpus-per-task=1
+#SBATCH --partition=zen3_0512
+#SBATCH --qos=zen3_0512
+#SBATCH --time=00:05:00
+
+../util/print_job_info.sh
+
+##
+# Either use srun directly
+
+# use --cpus-per-task= if the default was changed
+# see https://slurm.schedmd.com/archive/slurm-22.05.2/sbatch.html#OPT_cpus-per-task
+srun ../util/print_task_info.sh
+
+# OR
+
+# use e.g. mpirun with n = number of tasks per node
+#mpirun --map-by ppr:<n>:node ./my/mpi/application
+##
\ No newline at end of file

04-zen3_0512-multi-node/README.md

+# Run jobs allocating more than one node
+
+In VSC we currently only allow either partial node allocations or full node allocations. So in order to run jobs using more than one node we have the following options:
+- Specify multiple full nodes and configure the tasks accordingly (see [multi-node-full-node.sh](multi-node-full-node.sh)
+- Preferred (unless there is a reason): Use Array Jobs with full node allocations
+
+Note: in theory you can also configure the tasks in the full node allocation to use not all available resources, however for accounting the full node will be considered.