Home Directory Storage Quota#
There are strict quotas for each home directory (i.e., /home/<username>), and staying within the quota is vital for preventing issues on the HPC. This page provides some best practices for keeping within the quota. For more information about data storage on the HPC, see Data Storage Options.
Important
All commands on this page should be run from a compute node because they are CPU-intensive. You can find more information on getting a job on a compute node from Interactive Jobs: srun Command.
Analyze Disk Usage#
From a compute node, run the following command from your /home/<username> directory:
du -shc .[^.]* ~/*
This command will output the size of each file, directory, and hidden directory in your /home/<username> space, with the total of your /home directory being the last line of the output. After identifying the large files and directories, you can move them to the appropriate location (e.g., /work for research) or back up and delete them if they are no longer required. An example output would look like:
[<username>@<host> directory]$ du -shc .[^.]* ~/*
39M .git
106M discovery-examples
41K README.md
3.3M software-installation
147M total
Utilize /work and /scratch#
Use /work for long-term storage. PIs can request a folder in /work via New Storage Space Request and additional storage via Storage Space Extension Request. Utilize /scratch/<username> for temporary or intermediate files. Then, move files from /scratch to /work for persistent storage (i.e., the recommended workflow).
Note
Please be mindful of the /scratch purge policy, which can be found on the Research Computing Policy Page. See Data Storage Options for information on /work and /scratch.
Cleaning Directories#
Local#
We advise against using ‘pip install’ to install packages outside of a conda environment or python virtual environment (for example, while in a JupyterLab Notebook or interactive python session). These installations are placed in your .local directory, adding to your /home quota. Additionaly, the presence of different packages in .local can have a negative impact on the function of applications on the OOD. Please ensure all the packages you need are installed in a conda or virtual python environment.
If there are no activly running processes the entire .local directory can be moved to .local-off or individual packages can be removed usually from within: /home/username/.local/lib/pythonXX/site-packages
You can check for running processes via:
squeue -u <username>
To move your .local to .local-off
mv /home/username/.local /home/username/.local-off
Conda#
Note
Conda environments are part of your research and should be stored in your PI’s /work directory.
Here are some suggestions to reduce the storage size of the environments for those using the /home/<username>/.conda directory.
Remove unused packages and clear caches of Conda by loading an Anaconda module and running the following:
source activate <your environment>
conda clean --all
This will only delete unused packages in your ~/.conda/pkgs directory.
To remove any unused conda environments, run:
conda env list
conda env remove --name <your environment>
Singularity#
If you have pulled any containers to the HPC using Singularity, you can clean your container cache in your /home/<username> directory by running the following command from a compute node:
module load singularity/3.5.3
singularity cache clean all
To avoid your ~/.singularity directory filling up, you can set a temporary directory for when you pull a container to store the cache in that location; an example of this procedure (where <project> is your PI’s /work directory) is the following:
mkdir /work/<project>/singularity_tmp
export SINGULARITY_TMPDIR=/work/<project>/singularity_tmp
Then, pull the container using Singularity as usual.
Cache#
The ~/.cache directory can become large with the general use of HPC and Open OnDemand. Make sure you are not running any processes or jobs at the time by running the following:
squeue -u <username>
which prints a table with JOBID, PARTITION, NAME, USER ST, TIME, NODES, and NODELIST (REASON), which is empty when no jobs are running (i.e., it is safe to remove ~/.cache when no jobs are running).
Storing research environments#
Conda environments#
Use conda environments for Python on HPC. To create an environment in /work, use the --prefix flag as follows: (where <project> is your PI’s /work directory and <my conda env> is an empty directory to store your Conda environment):
conda create --prefix=/work/<project>/<my conda env>
Utilize the same conda environment to save storage space and time (i.e., avoid duplicate conda environments). Hence, shared environments can be easily done for a project accessing the same /work directory.
Singularity containers#
Containers pulled, built, and maintained for research work should be stored in your PI’s /work directory, not your /home/<username> directory.