Content
Code execution
When using OpenMPI, binding is controlled using the –-bind-to parameter. To bind processes to cores, use --bind-to core. Possible other values can be found in the man page.
| Codeblock |
|---|
mpirun --bind-to core ./yourprogram |
Our hardware supports hyperthreading, allowing you start 192 processes on Cascade Lake machines (*96 partitions) and 80 on Skylake machines.
If no specific requests regarding the number of tasks has been done, mpirun defaults to hyperthreading and starts cores*2 processes. If a number of tasks has been specified (with -N and/or --tasks-per-node), mpirun honors this via the flag -map-by. For example:
...
For examples for code execution, please visit Slurm partition CPU CLX.
Code compilation
For code compilation please use gnu compiler.
| Codeblock |
|---|
| title | MPI, gnu |
|---|
| collapse | true |
|---|
|
module load gcc/13.3.0
module load openmpi/gcc/5.0.3
mpicc -Wl,-rpath,$LD_RUN_PATH -o hello.bin hello.c |
| Codeblock |
|---|
| title | MPI, OpenMP, gnu |
|---|
| collapse | true |
|---|
|
module load gcc/13.3.0
module load openmpi/gcc/5.0.3
mpicc -fopenmp -Wl,-rpath,$LD_RUN_PATH -o hello.bin hello.c |
Slurm job script
A slurm script is submitted to the job scheduler slurm. It contains
- the request for compute nodes of a Slurm partition CPU CLX and
- commands to start your binary. You have two options to start an MPI binary.
Using mpirun
Using mpirun (from the MPI library) to start the binary you need to switch off slurm binding by adding export SLURM_CPU_BIND=none.
| Codeblock |
|---|
| title | MPI, full node |
|---|
| collapse | true |
|---|
|
#!/bin/bash
#SBATCH --nodes=2
#SBATCH --partition=cpu-clx:test
module load openmpi/gcc/5.0.3
export SLURM_CPU_BIND=none
mpirun -np 192 --map-by ppr:96:node ./hello.bin |
| Codeblock |
|---|
| title | MPI, half node |
|---|
| collapse | true |
|---|
|
#!/bin/bash
#SBATCH - |
...
-nodes=2
#SBATCH --partition=cpu-clx:test
module load |
...
...
...
...
...
...
...
...
BIND=none
mpirun -np 96 --map-by ppr: |
...
...
You can run one code compiled with MPI and OpenMP. The example covers the setup
- 2 nodes,
- 4 processes per node, 24 threads per process.
| Codeblock |
|---|
| title | MPI, OpenMPI, full node |
|---|
| collapse | true |
|---|
|
#!/bin/bash
#SBATCH -N 4-nodes=2
#SBATCH --tasks-per-node 96
partition=cpu-clx:test
module load gcc/9.3.0 openmpi/gcc.9/45.10.43
export tasksSLURM_perCPU_node=${SLURM_TASKS_PER_NODE%\(*}
mpirun -BIND=none
export OMP_NUM_THREADS=24
mpirun -np 8 --map-by ppr:$tasks_per_node:node ./yourexe:4:node:pe=24 ./hello.bin |
Using srun
| Codeblock |
|---|
| title | MPI, full node |
|---|
| collapse | true |
|---|
|
#!/bin/bash
#SBATCH --nodes=2
#SBATCH --partition=cpu-clx:test
srun --ntasks-per-node=96 ./hello.bin |
You can run one code compiled with MPI and OpenMP. The example covers the setup
- 2 nodes,
- 4 processes per node, 24 threads per process.
| Codeblock |
|---|
| title | MPI, OpenMP, full node |
|---|
| collapse | true |
|---|
|
#!/bin/bash
#SBATCH --nodes=2
#SBATCH --partition=cpu-clx:test
export OMP_PROC_BIND=spread
export OMP_NUM_THREADS=24
srun --ntasks-per-node=4 --cpus-per-task=48 ./hello.bin |