Spack¶

Terminology¶

Basic Procedure for the Public Instance¶

Use the procedure in this section when using software that has already been built by the system.

Load the Spack Environment¶

If you use bash or zsh, run the following after logging in.

. /shared/software/spack-1.2.0/share/spack/setup-env.sh

If you use csh or tcsh, run the following.

source /shared/software/spack-1.2.0/share/spack/setup-env.csh

When using software in a batch job, include the same setup command in the job script.

Confirm That Spack Is Available¶

After loading the environment, run the following.

spack --version
which spack

If spack: command not found is displayed, Load the Spack Environment has not been done. In the same shell, load setup-env.sh again.

Check Available Software¶

Check the prebuilt software explicitly provided by the system.

spack find -x

To also display hashes, use the following.

spack find -lx

Common check commands are as follows.

If you run spack find without -x, packages installed as dependencies are also shown. For normal use, start with spack find -x.

Load Software¶

For example, to use cp2k, run the following.

spack load cp2k

Loading a package sets environment variables such as PATH, making the application available in that shell or job.

Check that the executable is visible.

which COMMAND

For some applications, the package name and executable command name are different. For example, Quantum ESPRESSO uses application-specific commands such as pw.x. If you do not know the executable command, check the application's official manual or administrator guidance.

You can check currently loaded packages with the following.

spack find --loaded

Unload Software When Finished¶

To remove a package from the current shell after use, run the following.

spack unload cp2k

When trying another package with the same name but different build conditions, unload the old package before loading the new one.

Using Software in Batch Jobs¶

Do not run heavy computations on login nodes. Run actual computations on compute nodes through batch jobs or interactive jobs.

Basic GPU Job Form¶

When using GPU-enabled software, loading the software with Spack alone does not allocate GPUs. Request GPU resources in the Slurm job options.

#!/bin/bash
#SBATCH --job-name=gpu-app-test # Job name
#SBATCH --time=00:10:00         # Time limit
#SBATCH --gpus=8                # Number of GPUs
#SBATCH --ntasks=8              # Number of processes

. /shared/software/spack-1.2.0/share/spack/setup-env.sh
spack load PACKAGE_NAME

# Examples for checking that GPUs are allocated
srun nvidia-smi
srun bash -c 'echo CUDA_VISIBLE_DEVICES=$CUDA_VISIBLE_DEVICES'

srun COMMAND INPUT_FILE

Main GPU-Enabled Software Provided¶

The following software is currently provided as GPU-enabled builds.

• petsc
• lammps
• quantum-espresso
• gromacs
• kokkos

The provided software may be updated. Check the latest list with the following.

spack find -x

Example Quantum ESPRESSO Job Script¶

The following is an example using quantum-espresso.

#!/bin/bash
#SBATCH --job-name=qe-test # Job name
#SBATCH --time=00:30:00    # Time limit
#SBATCH --gpus=8           # Number of GPUs
#SBATCH --ntasks=8         # Number of processes

. /shared/software/spack-1.2.0/share/spack/setup-env.sh
spack load quantum-espresso

srun pw.x -in qe.in

qe.in is a Quantum ESPRESSO input file. For actual calculations, prepare the input file and required pseudopotential files in UPF format in advance.

Using hpcx-mpi¶

NVIDIA HPC-X MPI is available on this system. HPC-X MPI is an MPI implementation based on Open MPI and uses communication libraries such as UCX.

However, not all MPI applications provided in the public instance are built with hpcx-mpi support. Before running an MPI application, confirm that the target application was built with an hpcx-mpi configuration.

Main Packages Provided with hpcx-mpi Support¶

The following applications and libraries are currently provided with hpcx-mpi support.

• quantum-espresso
• gromacs

Check Whether a Package Supports hpcx-mpi¶

First, check the hash of the target package.

spack find -lv PACKAGE_NAME

Example:

spack find -lv quantum-espresso

After finding the hash, use it to check the build configuration.

spack spec /HASH

If the dependency tree contains the following entry, that configuration uses hpcx-mpi.

^hpcx-mpi

When multiple packages have the same name, checking only by package name, such as spack spec quantum-espresso, may show a configuration different from the installed package you intended. We recommend checking the hash with spack find -lv PACKAGE_NAME and then confirming the configuration with spack spec /HASH.

MPI Runtime Considerations¶

MPI applications must use the same MPI implementation expected by the loaded application at runtime. When using MPI applications provided in this system's public instance, generally load the target application with spack load and run it with Slurm srun.

. /shared/software/spack-1.2.0/share/spack/setup-env.sh
spack load PACKAGE_NAME
srun COMMAND INPUT_FILE

If an MPI communication error occurs, see Troubleshooting.

Finding Provided Software¶

The provided software may be updated. Always check the latest list on the system.

spack find -x

Main Provided Software¶

Example Output of spack find -lx¶

The following is example output. Hashes, versions, and package counts may change when the system is updated.

spack find -lx

Example output:

-- linux-ubuntu24.04-neoverse_v2 / %c,cxx,fortran=gcc@13.3.0 ----
g6hpeea cp2k@2026.1            qxqrq4o parallel-netcdf@1.14.1
gmcipc4 darshan-runtime@3.5.0  vephnns paraview@6.1.1
q6ezzfa frontistr@5.3          e7q773o petsc@3.25.2
dtkhf7f julia@1.12.6           kg4hkb6 py-scipy@1.17.1

-- linux-ubuntu24.04-neoverse_v2 / %c,cxx,fortran=nvhpc@26.3 ----
efwm4pc quantum-espresso@7.5

-- linux-ubuntu24.04-neoverse_v2 / %c,cxx=gcc@13.3.0 ------------
icb2hpj ffmpeg@8.1       j25jgtt openfoam-org@12
g4gyqaz ffmpeg@8.1       snjqbss povray@3.7.0.10
kiylntv gnuplot@6.0.0    l2zajbr py-matplotlib@3.11.0
teveiql grads@2.2.3      wypy6fy py-mpi4py@4.1.1
b2zyy2l gromacs@2026.1   qpo72vu py-pandas@3.0.3
sjduy44 lammps@20260211  esjclj6 py-scikit-learn@1.9.0
dcheirs openbabel@3.2.0  usvgll2 rust@1.96.0
lrbixw3 openfoam@2512

-- linux-ubuntu24.04-neoverse_v2 / %c,fortran=gcc@13.3.0 --------
53j2h54 cpmd@4.3              t3je6ga salmon-tddft@2.0.0
djkxpao genesis@1.6.0         4hnwin4 scale@5.4.4
7phdib7 netcdf-fortran@4.6.2  vjefr4h wrf@4.7.1
6trnbya openmx@3.9

-- linux-ubuntu24.04-neoverse_v2 / %c=gcc@13.3.0 ----------------
hb4jb3t gsl@2.8         djy67le py-netcdf4@1.7.2
i7qisgo netcdf-c@4.9.2  fxzltob tmux@3.6a

-- linux-ubuntu24.04-neoverse_v2 / %cxx=gcc@13.3.0 --------------
y4boldd autodock-vina@1.2.6  kup5bkr kokkos@5.1.1

-- linux-ubuntu24.04-neoverse_v2 / no compilers -----------------
52c5kr6 py-ase@3.28.0   us4lyh6 py-xarray@2026.4.0
kn3r4xs py-toml@0.10.2
==> 40 installed packages

When Multiple Packages Have the Same Name¶

Spack can manage multiple builds of the same software at the same time, with different versions, compilers, MPI implementations, GPU support, dependencies, and other settings. For this reason, the same package name may appear more than once.

Typical Error¶

For example, suppose you run the following when multiple builds of fftw are installed.

spack load fftw

If there are multiple candidates, an error like the following is shown.

==> Error: fftw matches multiple packages.
  Matching packages:
    erk4i5v fftw@3.3.11 platform=linux os=ubuntu24.04 target=neoverse_v2 %c,fortran=gcc@13.3.0
    5rny4xu fftw@3.3.11 platform=linux os=ubuntu24.04 target=neoverse_v2 %c,fortran=gcc@13.3.0
    nkvjfgj fftw@3.3.11 platform=linux os=ubuntu24.04 target=neoverse_v2 %c,fortran=nvhpc@26.3

Recommended: Specify by Hash¶

When multiple packages have the same name, first check the short hashes.

spack find -lx fftw

Then load the build you want by using its short hash.

spack load /5rny4xu

For the candidates in the example output, you can specify them as follows.

spack load /erk4i5v
spack load /5rny4xu
spack load /nkvjfgj

Hashes may change when the environment is updated. Do not memorize the hashes on this page as fixed values. Check them at runtime with spack find -lx PACKAGE_NAME.

Specify by Version or Compiler¶

You can also specify a version number.

spack load fftw@3.3.11

However, if multiple builds have the same version, this alone cannot distinguish them. In that case, include the compiler.

spack load fftw%gcc
spack load fftw%nvhpc

For more detailed specification, write the compiler version as follows.

spack load fftw%gcc@13.3.0
spack load fftw%nvhpc@26.3

If multiple candidates still remain, specify the package by hash.

Using a Private Instance¶

This section is for users who want to build and use OSS themselves. If you only use prebuilt software provided by the public instance, you do not need to perform the steps in this section.

When a Private Instance Is Needed¶

Use a private instance in the following cases.

• You want to use OSS that is not in the public instance.
• You want to use a different version from the provided one.
• You want to build with custom build options, dependencies, or compiler settings.
• You want to manage custom packages within a research group.

Create a Spack Instance¶

The following example creates a personal Spack instance under your home directory.

cd $HOME
git clone SPACK_REPO_URL spack
cd spack
git checkout BRANCH_NAME

Follow administrator guidance for the repository URL and branch name to use.

Load the Private Instance Environment¶

. $HOME/spack/share/spack/setup-env.sh

If you load both the public instance and private instance setup-env.sh files in the same shell, it becomes difficult to tell which Spack is being used. When using a private instance, we recommend opening a new shell and loading the private-side setup-env.sh.

Configure Chaining with the Public Instance¶

In a private instance, you can refer to prebuilt packages in the public instance by configuring Spack's upstreams.yaml. This reduces the cost of building dependencies each time.

A standard configuration example is as follows.

mkdir -p ~/.spack
cat > ~/.spack/upstreams.yaml <<'EOF'
upstreams:
  gb200-public:
    install_tree: /shared/software/spack-1.2.0/opt/spack
EOF

The install_tree path depends on the actual public instance configuration. If administrators announce a different path, use that path.

After configuring it, check whether packages from the public instance are visible.

spack find -lx

Search for Packages¶

Search for package names available in Spack.

spack list
spack list mpi

Check package versions and build options.

spack info openmpi

Install Packages¶

For example, to install openmpi, run the following.

spack install openmpi

You can also specify a version.

spack install openmpi@4.1.1

After installation, check it with the following.

spack find -lx openmpi

Uninstall Packages¶

spack uninstall openmpi

If multiple packages have the same name, prevent accidental deletion by checking the target by hash before running the uninstall command.

spack find -lx openmpi
spack uninstall /HASH

Using the ml-linux-aarch64 Stack¶

In addition to the standard software set, the system provides the ml-linux-aarch64 stack, a set of machine learning, AI, and data science packages for the Arm environment that are verified and provided together.

Load the Dedicated Environment¶

To use the ml-linux-aarch64 stack, load the environment setup script provided in a dedicated path separate from the regular public instance.

If you use bash or zsh, run the following.

. /shared/software/ml-linux-aarch64/env.sh

Check Provided Packages¶

To check the list of packages provided by this stack and their build configurations, run the following with the environment loaded.

spack find -lx

Example output:

-- linux-ubuntu24.04-aarch64 / %c,cxx=clang@18.1.8 --------------
rvdltfy py-jaxlib@0.10.1  35firrg py-tensorflow@2.20.0

-- linux-ubuntu24.04-aarch64 / %c,cxx=gcc@13.3.0 ----------------
26dq6yh py-scikit-learn@1.9.0  dvrmjfz py-tensorflow-metadata@1.17.2  rujxraz py-torch@2.12.0  kjorvey py-torch-nvidia-apex@24.04.01  rcjt7fg py-torchaudio@2.11.0  62r7jr7 py-torchvision@0.27.0

-- linux-ubuntu24.04-aarch64 / %cxx=gcc@13.3.0 ------------------
vfsb26o py-transformers@4.57.0

-- linux-ubuntu24.04-aarch64 / no compilers ---------------------
5qmndhd py-botorch@0.8.4  2omptwj py-kornia@0.8.3                       yjzhzb4 py-tensorboard@2.20.0             nbebl2h py-timm@1.0.24            zmw5kt2 py-vector-quantize-pytorch@0.3.9
4z67epj py-gpytorch@1.13  2hcpwjx py-lightning@2.6.5                    jiiecon py-tensorboardx@2.6.2.2           aglujkh py-torch-geometric@2.5.3
gf4b375 py-jax@0.10.1     ucsotkz py-pytorch-lightning@2.6.1            j45akk5 py-tensorflow-datasets@4.4.0      stxjtil py-torchgeo@0.9.0
ootk7uq py-keras@3.14.1   7ce5zk3 py-segmentation-models-pytorch@0.5.0  vbl5kb5 py-tensorflow-probability@0.25.0  aowxp5a py-torchmetrics@1.9.0

-- linux-ubuntu24.04-neoverse_v2 / %c,cxx=gcc@13.3.0 ------------
b7qoi5i llvm@18.1.8
==> 27 installed packages

Run Applications¶

Loading the environment setup script automatically sets the paths to the required packages (such as Python). Run your program directly from job submission commands such as srun or sbatch.

Example:

srun --gpus=1 --time=00:10:00 python -c "import torch; print(torch.cuda.is_available())"

Troubleshooting¶

Information to Check Before Contacting Support¶

When contacting administrators, include the following information if possible.

hostname
date
echo $SHELL
which spack
spack --version
spack find --loaded
spack find -lx PACKAGE_NAME

If the problem occurs in a job, also include the job ID, job script, standard output, and standard error.