Alderaan Overview

Configuration

  • 32 compute nodes with 64 cores and 512GB memory each, 30 available.
  • 2 nodes with 64 cores, 2TB memory, and 2 NVIDIA A100 GPUs each,
  • 800TB storage (1PB raw).

Quick basics

  • From the CU Denver | Anschutz network, connect by using your university username and password:
  • in a web browser to Jupyterhub to https://math-alderaan.ucdenver.pvt, or
  • by Remote Desktop (use Windows App on macOS) to math-alderaan, or

  • on the command line by ssh username@math-alderaan.ucdenver.pvt

  • Use the SLURM scheduler:

  • Write a job script, with the commands to execute and special comments how many cores and which kind of node it needs (=partition name)
  • Submit the script to the SLURM scheduler. See examples for details.
  • The scheduler finds a compute node with enough free cores and memory, marks the resources you requested as allocated, and starts executing the job script on the node.

Quick Tips

  • If you need software that is not installed on Alderaan, you can
  • check if it may be available in one of our singularity containers
  • check if it may be available in a module
  • install it in your account using Conda

  • ask for it to be installed, which may be typically done in a new singularity container or module

  • In your SLURM job,

  • request an estimated number of cores your code needs, if you do not know exactly how many (e.g., some codes in R)

  • if you run software that needs modules loaded first: add module load module_name in your job script.

  • If you need to run on compute nodes interactively, not as a script, you can:

  • use Jupyterhub, which provides multiple terminal and notebook windows with dedicated resources

  • run interactively on an alderaan node by entering the magic line
    srun --pty -N 1 -n 1 -t 1:00:00 -p math-alderaan /bin/bash -l
    on math-alderaan. This example will start an interactive shell job on one alderaan core for one hour max and teleport your session to that shell.

  • If you need to run Python with custom environments, use Conda.

  • If you have an MPI code (to use multiple cores and nodes as a single large computer), ask for enough cores and run your executable as an argument to the mpirun command. Make sure you use the same MPI version (GNU or Intel) to build and run the code. See https://github.com/ccmucdenver/templates for examples.