Perlmutter

Perlmutter is an HPE Cray “Shasta” system located at NERSC. Its compute nodes are equipped with four A100 NVIDIA GPUs.

It uses the SLURM scheduler to submit jobs from login nodes to run on the compute nodes.

Configuring Python and Installation

Begin by loading the python module. The following modules are recommended:

module load python

Create a conda environment

You can create a conda environment in which to install libEnsemble and all dependencies. For example:

conda create -n libe-pm python=3.9 -y

As Perlmutter has a shared HOME filesystem with other clusters, using the -pm suffix (for Perlmutter) is good practice.

Activate your virtual environment with:

export PYTHONNOUSERSITE=1
conda activate libe-pm

Installing libEnsemble and dependencies

Having loaded the Anaconda Python module, libEnsemble can be installed by one of the following ways.

  1. Install via pip into the environment.

(my_env) user@perlmutter07:~$ pip install libensemble

  1. Install via conda:

(my_env) user@perlmutter07:~$ conda config --add channels conda-forge
(my_env) user@perlmutter07:~$ conda install -c conda-forge libensemble

See advanced installation for other installation options.

Job Submission

Perlmutter uses Slurm for job submission and management. The two most common commands for initiating jobs are salloc and sbatch for running in interactive and batch modes, respectively. libEnsemble runs on the compute nodes on Perlmutter using either multi-processing (recommended) or mpi4py.

Example

To run the forces_gpu tutorial on Perlmutter.

To obtain the example you can git clone libEnsemble - although only the forces sub-directory is needed:

git clone https://github.com/Libensemble/libensemble
cd libensemble/libensemble/tests/scaling_tests/forces/forces_app

To compile forces:

module load PrgEnv-nvidia cudatoolkit craype-accel-nvidia80
cc -DGPU -O3 -fopenmp -mp=gpu -target-accel=nvidia80 -o forces.x forces.c

Now go to forces_gpu directory:

cd ../forces_gpu

Now grab an interactive session on one node:

salloc -N 1 -t 20 -C gpu -q interactive -A <project_id>

Then in the session run:

python run_libe_forces.py --comms local --nworkers 4

To see GPU usage, ssh into the node you are on in another window and run:

watch -n 0.1 nvidia-smi

To watch video

There is a video demonstration of the forces example on Perlmutter.

Note

The video uses libEnsemble version 0.9.3, where some adjustments of the scripts are needed to run on Perlmutter. These are no longer necessary. libEnsemble now correctly detects MPI runner and GPU setting on Perlmutter and the GPU code runs with many more particles than the CPU version (forces_simple).

Example submission scripts are also given in the examples.

Running libEnsemble with mpi4py

Running libEnsemble with local comms is usually sufficient on Perlmutter. However, if you need to use mpi4py, you should install and run as follows:

module load PrgEnv-gnu cudatoolkit
MPICC="cc -target-accel=nvidia80 -shared" pip install --force --no-cache-dir --no-binary=mpi4py mpi4py

This line will build mpi4py on top of a CUDA-aware Cray MPICH.

To run using 4 workers (one manager):

export SLURM_EXACT=1
srun -n 5 python my_script.py

More information on using Python and mpi4py on Perlmutter can be found in the Python on Perlmutter documentation.

Perlmutter FAQ

Some FAQs specific to Perlmutter. See more on the FAQ page.

srun: Job ****** step creation temporarily disabled, retrying (Requested nodes are busy)

Having created a dir /ccs/proj/<project_id>/libensemble:

You may also see: srun: Job ****** step creation still disabled, retrying (Requested nodes are busy)

This error has been encountered on Perlmutter. It is recommended to add these lines to submission scripts:

export SLURM_EXACT=1
export SLURM_MEM_PER_NODE=0

and to avoid using #SBATCH commands that may limit resources to srun job steps such as:

#SBATCH --ntasks-per-node=4
#SBATCH --gpus-per-task=1

Instead provide these to sub-tasks via the extra_args option to the MPIExecutor submit function.

GTL_DEBUG: [0] cudaHostRegister: no CUDA-capable device is detected

If using the environment variable MPICH_GPU_SUPPORT_ENABLED, then srun commands, at time of writing, expect an option for allocating GPUs (e.g.~ --gpus-per-task=1 would allocate one GPU to each MPI task of the MPI run). It is recommended that tasks submitted via the MPIExecutor specify this in the extra_args option to the submit function (rather than using an #SBATCH command). This is needed even when using setting CUDA_VISIBLE_DEVICES or other options.

If running the libEnsemble user calling script with srun, then it is recommended that MPICH_GPU_SUPPORT_ENABLED is set in the user sim_f or gen_f function where GPU runs will be submitted, instead of in the batch script. E.g:

os.environ["MPICH_GPU_SUPPORT_ENABLED"] = "1"
warning: /tmp/pgcudafatYDO6wtSva6K2.o: missing .note.GNU-stack section implies executable stack

Recently this warning has been encountered when compiling the forces example on Perlmutter. This does not affect the run, but can be supressed by adding -Wl,-znoexecstack to the build line.

Additional Information

See the NERSC Perlmutter docs for more information about Perlmutter.