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 PrgEnv-gnu cudatoolkit 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@cori07:~$ pip install libensemble

  1. Install via conda:

(my_env) user@cori07:~$ conda config --add channels conda-forge
(my_env) user@cori07:~$ conda install -c conda-forge libensemble

See here for more information on advanced options for installing libEnsemble, including using Spack.

Installing mpi4py

If using mpi4py for communications (optional), it is recommended that you install using the following line (having installed the cudatoolkit module):

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.

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

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 or mpi4py.

If running more than one worker per node, the following is recommended to prevent resource conflicts:

export SLURM_EXACT=1
export SLURM_MEM_PER_NODE=0

Alternatively, the --exact option to srun, along with other relevant options can be given on any srun lines (including executor submission lines via the extra_args option).

Example

A simple example batch script for a libEnsemble use case that runs 5 workers (one generator and four simulators) on one node:

 1#!/bin/bash
 2#SBATCH -J libE_small_test
 3#SBATCH -A <myproject_g>
 4#SBATCH -C gpu
 5#SBATCH --time 15
 6#SBATCH --nodes 1
 7
 8export MPICH_GPU_SUPPORT_ENABLED=1
 9export SLURM_EXACT=1
10export SLURM_MEM_PER_NODE=0
11
12python libe_calling_script.py --comms local --nworkers 5

Note

Any loaded modules and environment variables (including conda environments) are inherited by the job on Perlmutter.

This example calling script has the following line so the node is divided into four resource sets (the example generator does not need dedicated resources):

libE_specs['zero_resource_workers'] = [1]

The MPIExecutor is also initiated in the calling script, ensuring that srun is picked up:

from libensemble.executors.mpi_executor import MPIExecutor
exctr = MPIExecutor(custom_info={'mpi_runner':'srun'})

Each worker runs a simulator function that uses the MPIExecutor submit function, including the argument --gpus-per-task=1:

from libensemble.executors.executor import Executor
exctr = Executor.executor
task = exctr.submit(app_name='sim1',
                    num_procs=n_rsets,
                    app_args='input.txt',
                    extra_args='--gpus-per-task=1'
                    )

If running using variable resource workers, between one and four-way MPI runs may be issued by any of the workers (with each MPI task using a GPU). libEnsemble’s resource manager automatically disables workers whose resources are being used by another worker.

Example submission scripts are also given in the examples.

Perlmutter FAQ

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

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 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"

Additional Information

See the NERSC Perlmutter docs for more information about Perlmutter.