Python is supported on ARCHER2 both for running intensive parallel jobs and also as an analysis tool. This section describes how to use Python in either of these scenarios.
The Python installations on ARCHER2 contain some of the most commonly
used modules. If you wish to install additional Python modules, we
recommend that you use the
pip command after loading the
cray-python module. This is described in more detail below.
When you log onto ARCHER2, no Python module is loaded by default. You
will generally need to load the
cray-python module to access the
functionality described below. Running
python without loading a module
first will result in your using the operating system default Python
which is likely not what you intend.
HPE Cray Python distribution
The recommended way to use Python on ARCHER2 is to use the HPE Cray Python distribution.
The HPE Cray distribution provides Python 3 along with some of the most common packages used for scientific computation and data analysis. These include:
- numpy and scipy - built using GCC against HPE Cray LibSci
- mpi4py - built using GCC against HPE Cray MPICH
The HPE Cray Python distribution can be loaded (either on the front-end or in a submission script) using:
module load cray-python
The HPE Cray Python distribution provides Python 3. There is no Python 2 version as Python 2 is now deprecated.
The HPE Cray Python distribution is built using GCC compilers. If you wish
to compile your own Python, C/C++ or Fortran code to use with HPE Cray
Python, you should ensure that you compile using
PrgEnv-gnu to make sure
they are compatible.
Adding your own packages
If the packages you require are not included in the HPE Cray Python
distribution, further packages can be added using
pip. However, as the
/home file systems are not available on the compute nodes, you will need
to modify the default install location that
pip uses to point to a
location on the /work file systems (by default,
pip installs into
$HOME/.local). To do this, you set the
variable to point to a location on /work, for example:
You will also need to ensure that:
- the location of commands installed by
pipare available on the command line by modifying the
- any packages you install are available to Python by modifying the
You can do this in the following way (once you have set
PYTHONUSERBASE as described
export PATH=$PYTHONUSERBASE/bin:$PATH export PYTHONPATH=$PYTHONUSERBASE/lib/python3.8/site-packages:$PYTHONPATH
We would recommend adding all three of these commands to your
file to ensure they are set by default when you log in to ARCHER2.
Once, you have done this, you can use
pip to add packages on top of the HPE
Cray Python environment. This can be done using:
module load cray-python pip install --user <package_name>
This uses the
--user flag to ensure the packages are installed in
the directory specified by
Setting up virtual environments
We recommend that you use the
virtualenv packages to
manage your Python environments. A summary of how to get a virtual environment set up is contained in the below, but for further information, see:
Sometimes, you may need several different versions of the same package installed, for example, due to dependency issues. Virtual environments allow you to manage these conflicting requirements. The first step is to run the commands contained in the above section so that you can install the
virtualenv package which will manage your environments. To install
pip install --user virtualenv # The --user flag indicates this should be installed in the user's package folder
virtualenvto set this folder up for storing virtual environment 'stuff'. This is done by running the command
mkdir /work/t01/t01/auser/<<name of your virtual environment>> # Create the folder virtualenv -p /opt/cray/pe/python/188.8.131.52/bin/python /work/t01/t01/asuser/<<name of your virtual environment>> # -p flag means use this python interpreter
activateyour environment. This is done by running
source /work/t01/t01/auser/<<name of your virtual environment>>/bin/activate
pip install <<package name>>. These packages will only be available within this environment. When running code that requires these packages you must activate the environment, by adding the above
source ... activateline of code to any submission scripts.
Running Python on the compute nodes
In this section we provide example Python job submission scripts for a variety of scenarios of using Python on the ARCHER2 compute nodes.
Example serial Python submission script
#!/bin/bash --login #SBATCH --job-name=python_test #SBATCH --nodes=1 #SBATCH --tasks-per-node=1 #SBATCH --cpus-per-task=1 #SBATCH --time=00:10:00 # Replace [budget code] below with your project code (e.g. t01) #SBATCH --account=[budget code] #SBATCH --partition=standard #SBATCH --qos=standard # Load the Python module module load cray-python # If using a virtual environment source <<path to virtual environment>>/bin/activate # Run your Python progamme python python_test.py
#!/bin/bash --login #SBATCH --job-name=python_test #SBATCH --nodes=1 #SBATCH --tasks-per-node=1 #SBATCH --cpus-per-task=1 #SBATCH --time=00:10:00 # Replace [budget code] below with your project code (e.g. t01) #SBATCH --account=[budget code] #SBATCH --partition=standard #SBATCH --qos=standard # Setup the batch environment module load epcc-job-env # Load the Python module module load cray-python # If using a virtual environment source <<path to virtual environment>>/bin/activate # Run your Python progamme python python_test.py
If you have installed your own packages you will need to set
PYTHONPATH as described above within your job submission script in order
to accesss the commands and packages you have installed.
Example mpi4py job submission script
Programs that have been parallelised with mpi4py can be run on
multiple processors on ARCHER2. A sample submission script is given
below. The primary difference from the Python submission script in the
previous section is that we must run the program using
srun python my_prog.py instead of
python my_prog,py. Failing to do so will
cause a segmentation fault in your program when it reaches the line
from mpi4py import MPI".
#!/bin/bash --login # Slurm job options (job-name, compute nodes, job time) #SBATCH --job-name=mpi4py_test #SBATCH --nodes=1 #SBATCH --tasks-per-node=2 #SBATCH --cpus-per-task=1 #SBATCH --time=0:10:0 # Replace [budget code] below with your budget code (e.g. t01) #SBATCH --account=[budget code] #SBATCH --partition=standard #SBATCH --qos=standard # Load the Python module module load cray-python # Run your Python programme # Note that srun MUST be used to wrap the call to python, otherwise an error # will occur srun python mpi4py_test.py
#!/bin/bash --login # Slurm job options (job-name, compute nodes, job time) #SBATCH --job-name=mpi4py_test #SBATCH --nodes=1 #SBATCH --tasks-per-node=2 #SBATCH --cpus-per-task=1 #SBATCH --time=0:10:0 # Replace [budget code] below with your budget code (e.g. t01) #SBATCH --account=[budget code] #SBATCH --partition=standard #SBATCH --qos=standard # Setup the batch environment module load epcc-job-env # Load the Python module module load cray-python # Run your Python programme # Note that srun MUST be used to wrap the call to python, otherwise an error # will occur srun python mpi4py_test.py
Using JupyterLab on ARCHER2
It is possible to view and run Jupyter notebooks from both login nodes and compute nodes on ARCHER2.
You can test such notebooks on the login nodes, but please do not attempt to run any computationally intensive work. Jobs may get killed once they hit a CPU limit on login nodes.
Please follow these steps.
Install JupyterLab in your work directory.
module load cray-python export PYTHONUSERBASE=/work/t01/t01/auser/.local export PATH=$PYTHONUSERBASE/bin:$PATH # source <<path to virtual environment>>/bin/activate # If using a virtualenvironment uncomment this line and remove the --user flag from the next pip install --user jupyterlab
If you want to test JupyterLab on the login node please go straight to step 3. To run your Jupyter notebook on a compute node, you first need to run an interactive session.Your prompt will change to something like below.
srun --nodes=1 --exclusive --time=00:20:00 --account=<your_budget> \ --partition=standard --qos=short --reservation=shortqos \ --pty /bin/bashIn this case, the node id is
nid001015. Now execute the following on the compute node.
cd /work/t01/t01/auser # Update the path to your work directory export PYTHONUSERBASE=$(pwd)/.local export PATH=$PYTHONUSERBASE/bin:$PATH export HOME=$(pwd) module load cray-python # source <<path to virtual environment>>/bin/activate # If using a virtualenvironment uncomment this line
Run the JupyterLab server.Once it's started, you will see a URL printed in the terminal window of the form
export JUPYTER_RUNTIME_DIR=$(pwd) jupyter lab --ip=0.0.0.0 --no-browser
http://127.0.0.1:<port_number>/lab?token=<string>; we'll need this URL for step 6.
Please skip this step if you are connecting from a machine running Windows. Open a new terminal window on your laptop and run the following command.where
ssh <username>@login.archer2.ac.uk -L<port_number>:<node_id>:<port_number>
<username>is your username, and
<node_id>is the id of the node you're currently on (for a login node, this will be
uan01, or similar; on a compute node, it will be a mix of numbers and letters). In our example,
nid001015. Note, please use the same port number as that shown in the URL of step 3. This number may vary, likely values are 8888 or 8889.
Please skip this step if you are connecting from Linux or macOS. If you are connecting from Windows, you should use MobaXterm to configure an SSH tunnel as follows.
- Click on the
Tunnellingbutton above the MobaXterm terminal. Create a new tunnel by clicking on
New SSH tunnelin the window that opens.
- In the new window that opens, make sure the
Local port forwardingradio button is selected.
- In the
forwarded porttext box on the left under
My computer with MobaXterm, enter the port number indicated in the JupyterLab server output (e.g., 8888 or 8890).
- In the three text boxes on the bottom right under
login.archer2.ac.uk, your ARCHER2 username and then
- At the top right, under
Remote server, enter the id of the login or compute node running the JupyterLab server and the associated port number.
- Click on the
- In the tunnelling window, you will now see a new row for the settings you just entered. If you like, you can give a name to the tunnel in the leftmost column to identify it.
- Click on the small key icon close to the right for the new connection to tell MobaXterm which SSH private key to use when connecting to ARCHER2. You should tell it to use the same
.ppkprivate key that you normally use when connecting to ARCHER2.
- The tunnel should now be configured. Click on the small start button (like a play '>' icon) for the new tunnel to open it. You'll be asked to enter your ARCHER2 account password -- please do so.
- Click on the
Now, if you open a browser window locally, you should be able to navigate to the URL from step 3, and this should display the JupyterLab server. If JupyterLab is running on a compute node, the notebook will be available for the length of the interactive session you have requested.
Please do not use the other http address given by the JupyterLab output,
the one formatted
http://<node_id>:<port_number>/lab?token=<string>. Your local
browser will not recognise the
<node_id> part of the address.
You willl of course need to use a different hostname for the SSH tunnel if you're running on the
ARCHER2 4-cabinet system,