NERSC (Perlmutter, Cori)

This page describes setup for running calculations at NERSC, including on the Perlmutter and Cori systems.

To get started with calculations at NERSC:

  1. Ask Kristin about whether you will be running at NERSC and, if so, under what account / repository to charge.

  2. Request a NERSC account through the NERSC homepage (Google “NERSC account request”).

  3. A NERSC Liason or PI Proxy will validate your account and assign you an initial allocation of computing hours.

  4. At this point, you should be able to log in, check CPU-hour balances, etc. through “NERSC NIM” and “My NERSC” portals

  5. In order to log in and run jobs on the various machines at NERSC, review the NERSC documentation.

  6. In order to load and submit scripts for various codes (VASP, ABINIT, Quantum Espresso), NERSC has lots of information to help. Try Google, e.g. “NERSC VASP”.

    ... * Note that for commercial codes such as VASP, there is an online form that allows you to enter your VASP license, which NERSC will confirm and then allow you access to. Log in to, select "Open Request", and fill out the "VASP License Confirmation Request" form.

  7. Please make a folder inside your project directory and submit all your jobs there, as your home folder has only about 40GB of space. For example, for matgen project, your work folder path should be something like the following:


  8. You can also request a mongo database for your project to be hosted on NERSC. Google “MongoDB on NERSC” for instructions. Donny Winston or Patrick Huck can also help you get set up and provide you with a preconfigured database suited for running Materials Project style workflows.

Running Jobs on NERSC

This tutorial provides a brief overview of setting yourself up to run jobs on NERSC. If any information is unclear or missing, feel free to edit this document or contact Kara Fong.

Setting up a NERSC account:

Contact the group’s NERSC Liaison (currently Howard Li, see Group Jobs list). They will help you create an account and allocate you computational resources. You will then receive an email with instructions to fill out the Appropriate Use Policy form, set up your password, etc.

Once your account is set up, you can manage it at the NERSC Information Management (NIM) website.

Logging on (Setup):

You must use the SSH protocol to connect to NERSC. Make sure you have SSH installed on your local computer (you can check this by typing which ssh). You will also need to set up multi-factor authentication with NERSC. This will allow you to generate "one time passwords" (OTPs). You will need append a OTP to the end of your NIM password each time you log on to a NERSC cluster.

We also advise you to configure the NERSC sshproxy script so that you only have to log into NERSC with a OTP once per 24 hours (helpful if you are scp-ing things). To do this:

  1. Download the script from NERSC to your home folder

  2. At the terminal type ./ -u <nersc_username>

  3. Enter your password and OTP

You should now be able to log in without authenticating for 24 hours!

Logging on to Cori:

Log on to Cori, for example, by submitting the following command in the terminal:


You will be prompted to enter your passphrase, which is your NIM password+OTP (e.g. <your_password><OTP> without any spaces). This will log you onto the cluster and take you to your home directory. You may also find it useful to set up an alias for signing on to HPC resources. To do this, add the following line to your bash_profile:

alias cori="ssh <your_username>"

Now you will be able to initialize a SSH connection to cori just by typing cori in the command line and pressing enter.

Logging on to Perlmutter

Similar to Cori, you can set up an alias for Perlmutter, or you can ssh into Perlmutter by running the following command in the terminal

alias perlmutter="ssh <your_username>"


Transferring files to/from NERSC:

For small files, you can use SCP (secure copy). To get a file from NERSC, use:

scp /local/path

To send a file to NERSC, use:

scp /local/path/myfile.txt

To move a larger quantity of data using a friendlier interface, use Globus Online.

Mounting NERSC's file system locally:

If you regularly work with files from NERSC and you find it annoying to scp them, you could mount your NERSC home folder locally. First you need to install sshfs by downloading the stable releases of macFUSE and SSHFS from this website. If the installation doesn't work, then try getting sshfs from this website. Next, make a folder to mount onto:

mkdir ~sshfs
mkdir ~sshfs/cori

Next, add this to your bash profile, making sure to add your own username to replace NERSC_USERNAME:

alias sshfs-cori="sshfs -o 'IdentityFile=~/.ssh/nersc,reconnect,ServerAliveInterval=15,ServerAliveCountMax=3,auto_cache,Compression=no,follow_symlinks' ~/sshfs/cori"
function unmount { 
    pkill -9 sshfs
    sudo umount -f $HOME/sshfs/cori

Source your bash profile or restart your terminal. Try it out by running


Then you should have access to your cori home folder within the ~sshfs/cori/ folder. There are 4 data transfer nodes, so if the command is hanging or timing out, try a different data transfer node.

To terminate the connection run


Running and monitoring jobs:

The following instructions are for running on Cori.

Most jobs are run in batch mode, in which you prepare a shell script telling the batch system how to run the job (number of nodes, time the job will run, etc.). NERSC’s batch system software is called SLURM. Below is a simple batch script example, copied from the NERSC website:

#!/bin/bash -l

#SBATCH -N 2          #Use 2 nodes
#SBATCH -t 00:30:00   #Set 30 minute time limit
#SBATCH -q regular    #Submit to the regular QOS
#SBATCH -L scratch    #Job requires $SCRATCH file system
#SBATCH -C haswell    #Use Haswell nodes

srun -n 32 -c 4 ./my_executable

Here, the first line specifies which shell to use (in this case bash). The keyword #SBATCH is used to start directive lines (click here for a full description of the sbatch options you can specify). The word “srun” starts execution of the code.

To submit your batch script, use sbatch in the directory containing the script file.

Below are some useful commands to control and monitor your jobs:

sqs -u username (Lists jobs for your account)
scancel job_id     (Cancels a job from the queue)

For Perlmutter, the job scripts will look similar, but you need to specify whether you are running on CPU or GPU.

#!/bin/bash -l

#SBATCH -N 2          #Use 2 nodes
#SBATCH -t 00:30:00   #Set 30 minute time limit
#SBATCH -q regular    #Submit to the regular QOS
#SBATCH -C GPU        #Use GPU

srun -n 8 -c 32 --cpu-bind=cores -G 8 ./my_executable

For more options in the executable, please refer to NERSC documentation. To work with the high-throughput infrastructure, please refer to "Fireworks & Atomate" in this handbook.

Choosing a QOS (quality of service):

You specify which queue to use in your batch file. Use the debug queue for small, short test runs, the regular queue for production runs, and the premium queue for high-priority jobs.

Choosing a node type (haswell vs knl):

You may also specify the resource type you would like your job to run on, witnin your batch file. When running on Cori, there are two CPU architectures available, Haswell, and Knights Landing (known as KNL). Running on a Haswell node will afford you high individual core performance with up to 32 cores per node (or 64 threads per node). A KNL node provides a large core count (68 cores per node or 272 threads per node) which is suitable for programs capable of effectively utilizing multithreading. On Cori, there are 2388 Haswell nodes and 9688 KNL nodes.

Automatic job submission on NERSC: crontab

In order to automatically manage job submission at NERSC, you can use crontab. You can submit jobs periodically even when you are not signed in to any NERSC systems and perhaps reduce the queue time from 5-10 days to a few hours. This is possible because of the way jobs are managed in atomate/fireworks. Please make sure you feel comfortable submitting individual jobs via atomate before reading this section.

In atomate, by using --maxloop 3 for example when setting rocket_launch in your my_qadapter.yaml, after 3 trials in each minute if there are still no READY jobs available in your Launchpad Fireworks would stop the running job on NERSC to avoid wasting computing resources. On the other hand, if you have Fireworks available with the READY state and you have been using crontab for a few days, even if the jobs you submitted a few days ago start running on NERSC, they would pull any READY Fireworks and start RUNNING them reducing the turnaround from a few days to a few hours! So how to setup crontab? Please follow the instructions here: 1. ssh to the node where you want to setup the crontab; try one that is easy to remember such as cori01 or edison01; for logging in to a specific node just do for example “ssh cori01” after you log in to the system (Cori in this example).

  1. Type and enter: crontab -e

  2. Now you can setup the following command in the opened vi editor. What it does is basically running the file every 120 minutes of every day of every week of every month of every year (or simply /120 *):

    */120 * * * * /bin/bash -l >> PATH_TO_LOGFILE
  3. Setup your like the following: (as a suggestion, you can simply put this file and the log file which keeps a log of submission states in your home folder):

    qlaunch --fill_mode rapidfire -m 1000 --nlaunches 1

    The last line of this 3-line file is really what submitting your job inside your production folder with the settings that you set in FW_config.yaml file. See atomate documentation for more info.

  4. Please make sure to set your PRODUCTION_FOLDER under /global/project/projectdirs/ that has much more space than your home folder and it is also backed up. Make sure to keep an eye on how close you are to disk space and file number limitations by checking periodically.

Running Jupyter Notebooks on Cori

Jupyter notebooks are quickly becoming an indispensable tool for doing computational science. In some cases, you might want to (or need to) harness NERSC computing power inside of a jupyter notebook. To do this, you can use NERSC’s new Jupyterhub system at These notebooks are run on a large memory node of Cori and can also submit jobs to the batch queues (see for details). All of your files and the project directory will be accessible from the Jupyterhub, but your conda envs won’t be available before you do some configuration.

To set up a conda environment so it is accessible from the Jupyterhub, activate the environment and setup an ipython kernel. To do this, run the command “pip install ipykernel”. More info can be found at

Automatic Job Packing with FireWorks

DISCLAIMER: Only use job packing if you have trouble with typical job submission. The following tip is not 100% guaranteed to work., and is based on limited, subjective experience on Cori. Talk to Alex Dunn ( for help if you have trouble.

The Cori queue system can be unreasonably slow when submitting many (e.g., hundreds, thousands) of small (e.g., single node or 2 nodes) jobs with qos-normal priority on Haswell. In practice, we have found that the Cori job scheduler will give your jobs low throughput if you have many jobs in queue, and you will often only be able to run 5-30 jobs at a time, while the rest wait in queue for far longer than originally expected (e.g., weeks). While there is no easy way to increase your queue submission rate (AFAIK), you can use FireWorks job-packing to “trick” Cori’s SLURM scheduler into running many jobs in serial on many parallel compute nodes with a single queue submission, vastly increasing throughput.

You can use job packing with the “multi” option to rlaunch. This command launches N parallel python processes on the Cori scheduling node, each which runs a job using M compute nodes.

The steps to job packing are: 1. Edit your my_qadapter.yaml file to reserve N * M nodes for each submission. For example, if each of your jobs takes M = 2 nodes, and you want a N = 10 x speedup, reserve 20 nodes per queue submission. 2. Change your rlaunch command to:

rlaunch -c /your/config multi N

To have each FireWorks process run as many jobs as possible in serial before the walltime, use the --nlaunches 0 option. To prevent FireWorks from submitting jobs with little walltime left (causing jobs to frequently get stuck as “RUNNING”), set the --timeout option. Make sure --timeout is set so that even a long running job submitted at the end of your allocation will not run over your walltime limit. Your my_qadapter.yaml should then have something similar to the following lines:

rocket_launch: rlaunch -c /your/config multi 10 --nlaunches 0 --timeout 169200
nodes: 20

Typically, setting N <= 10 will give you a good N-times speedup with no problems. There are no guarantees, however, when N > 10-20. Use N > 50 at your own risk!

Jobs on Premium QOS

By default, premium QOS access is turned off for everyone in the group. When there is a scientific emergency (for example, you need to complete a calculation ASAP for a meeting with collaborators the next day), premium QOS can be utilized. In such cases, please contact Howard ( or on Slack) to request premium QOS access. The access will be then turned off automatically after three weeks or after the emergency has been dealt with.

Last updated