Skip to content

Argonne Leadership Computing Facility


Julia is a high-level, high-performance dynamic programming language for technical computing. It has a syntax familiar to users of many other technical computing environments. Designed at MIT to tackle large-scale partial-differential equation simulation and distributed linear algebra, Julia features a robust ecosystem of tools for optimization, statistics, parallel programming, and data visualization. Julia is actively developed by the Julia Labs team at MIT and in industry, along with hundreds of domain-expert scientists and programmers worldwide.


This guide is a first draft of the Julia documentation for Polaris. If you have any suggestions or contributions, please open a pull request or contact us by opening a ticket at the ALCF Helpdesk.

Julia Installation

Using the official Julia 1.9 binaries from the Julia webpage is recommended. Juliaup provides a convenient way to install Julia and manage the various Julia versions.

curl -fsSL | sh

You may then list the available Julia versions with juliaup list and install a specific version with juliaup install <version>. You can then activate a specific version with juliaup use <version> and set the default version with juliaup default <version>. juliaup update will update the installed Julia versions. In general, the latest stable release of Julia should be used.

juliaup add release

Julia Project Environment

The Julia built-in package manager allows you to create a project and enable project-specific dependencies. Julia manages packages in the Julia depot located by default in ~/.julia. However, that NFS filesystem is not meant for high-speed access. Therefore, this Julia depot folder should be located on a fast filesystem of your choice (grand, eagle). The Julia depot directory is set via the environment variable JULIA_DEPOT_PATH. For example, you can set the Julia depot to a directory on Polaris grand filesystem by adding the following line to your ~/.bashrc file:

export /lus/grand/projects/$PROJECT/$USER/julia_depot

Programming Julia on Polaris

There are three key components to using Julia for large-scale computations:

  1. MPI support through MPI.jl
  2. GPU support through CUDA.jl
  3. HDF5 support through HDF5.jl

In addition, we recommend VSCode with the Julia extension for a modern IDE experience, together with the ssh-remote extension for remote interactive development.

MPI Support

MPI support is provided through the MPI.jl.

julia> ] add MPI
This will install the MPI.jl package and default MPI prebuilt binaries provided by an artifact. For on-node debugging purposes the default artifact is sufficient. However, for large-scale computations, it is to use the system MPI library that is loaded via module. As of MPI.jl v0.20 this is handled through MPIPrefences.jl.
julia --project -e 'using MPIPreferences; MPIPreferences.use_system_binary()'

Check that the correct MPI library is targeted with Julia.

julia --project -e 'using MPI; MPI.versioninfo()'
  binary:  system
  abi:     MPICH
  libmpi:  libmpi_cray
  mpiexec: mpiexec

Package versions
  MPI.jl:             0.20.11
  MPIPreferences.jl:  0.1.8

Library information:
  libmpi:  libmpi_cray
  MPI version:  3.1.0
  Library version:
    MPI VERSION    : CRAY MPICH version (ANL base 3.4a2)
    MPI BUILD INFO : Mon Apr 18 12:05 2022 (git hash 4f56723)
When running on the login node, switch back to the default provided MPI binaries in MPI_jll.jl by removing the LocalPreferences.toml file.

GPU Support

NVIDIA GPU support is provided through the CUDA.jl package.

julia> ] add CUDA
In case you want write portable GPU kernels we highly recommend the KernelAbstractions.jl package. It provides a high-level abstraction for writing GPU kernels that can be compiled for different GPU backends.
julia> ] add KernelAbstractions
By loading either oneAPI.jl, AMDGPU.jl, or CUDA.jl (see quickstart guide below).

HDF5 Support

Parallel HDF5 support is provided by

module load cray-hdf5-parallel
After setting export JULIA_HDF5_PATH=$HDF5_DIR we can install the HDF5.jl package.
julia> ] add HDF5

Quickstart Guide

The following example shows how to use MPI.jl, CUDA.jl, and HDF5.jl to write a parallel program that computes the sum of two vectors on the GPU and writes the result to an HDF5 file. A repository with an example code computing an approximation of pi can be found at Polaris.jl. In this repository, you will also find a script that will build the HDF5.jl and MPI.jl package for the system libraries. The dependencies are installed with the following commands:

julia --project

julia> ] up
using CUDA
using HDF5
using MPI
using Printf
using Random

function pi_kernel(x, y, d, n)
    idx = (blockIdx().x-1) * blockDim().x + threadIdx().x
    if idx <= n
        d[idx] = (x[idx] - 0.5)^2 + (y[idx] - 0.5)^2 <= 0.25 ? 1 : 0
    return nothing

function approximate_pi_gpu(n::Integer)
    x = CUDA.rand(Float64, n)
    y = CUDA.rand(Float64, n)
    d = CUDA.zeros(Float64, n)

    nblocks = ceil(Int64, n/32)

    @cuda threads=32 blocks=nblocks pi_kernel(x,y,d,n)

    return sum(d)

function main()
    n = 100000  # Number of points to generate per rank
    Random.seed!(1234)  # Set a fixed random seed for reproducibility

    dsum = MPI.Allreduce(approximate_pi_gpu(n), MPI.SUM, MPI.COMM_WORLD)

    pi_approx = (4 * dsum) / (n * MPI.Comm_size(MPI.COMM_WORLD))

    if MPI.Comm_rank(MPI.COMM_WORLD) == 0
        @printf "Approximation of π using Monte Carlo method: %.10f\n" pi_approx
        @printf "Error: %.10f\n" abs(pi_approx - π)
    return pi_approx

if !isinteractive()
    pi_approx = main()
    h5open("pi.h5", "w") do file
        write(file, "pi", pi_approx)

Job submission script

This example can be run on Polaris with the following job submission script:

#!/bin/bash -l
#PBS -l select=1:system=polaris
#PBS -l place=scatter
#PBS -l walltime=0:30:00
#PBS -l filesystems=home:grand
#PBS -q debug


# MPI example w/ 4 MPI ranks per node spread evenly across cores
module load cray-hdf5-parallel
# Put in your Julia depot path
# Path to Julia executable. When using juliaup, it's in your julia_depot folder

mpiexec -n ${NTOTRANKS} --ppn ${NRANKS_PER_NODE} --depth=${NDEPTH} --cpu-bind depth julia --check-bounds=no --project pi.jl
Verify that JULIA_DEPOT_PATH is set to the correct path and JULIA_PATH points to the Julia executable. When using juliaup, the Julia executable is located in the juliaup folder of your JULIA_DEPOT_PATH.

Advanced features

CUDA-aware MPI

MPI.jl supports CUDA-aware MPI. This is enabled by setting the following environment variables

export JULIA_MPI_PATH=$PATH_TO_CUDA_MPI # /opt/cray/pe/mpich/8.1.16/ofi/nvidia/20.7

Note that MPI.jl needs to be rebuilt for the changes to take effect.

julia --project -e 'using Pkg;"MPI"; verbose=true)'

Large-scale parallelism

CUDA.jl uses the nvcc compiler to compile GPU kernels. This will create object files in the TEMP filesystem. Per default, the tempdir is a global directory that can lead to name clashes of the compiled kernel object files. To avoid this, we recommend setting the tempdir to a local directory on the compute node.

export TMPDIR=/local/scratch