Skip to content

Latest commit

 

History

History
39 lines (25 loc) · 2.08 KB

hpc_notes.md

File metadata and controls

39 lines (25 loc) · 2.08 KB
Raw

#HPC = High Performance Computing

At USC, the new cluster is available to test at discovery.usc.edu.

Help email is [email protected].

Why HPC?

  • You need more than your laptop can give you. (computation time, memory, disk space, etc.)
  • You have trouble installing things on your laptop (the cluster is a linux system)

Parallelization

A cluster of nodes, each of which has many cores.

  • MPI: "message passing interface" -- Multiple copies of the same program are running. Can be either on multiple cores on the same node, or across multiple nodes.
    • Generically, you would run a program with mpi with something like mpirun python my_parallel_script.py (if you have mpi installed on your laptop, this kind of command would work)
    • There are different implementations of MPI, e.g. OpenMPI, Intel-MPI, etc.
  • OpenMP: shared-memory parallelization on a single node across multiple cores. Also might be called "multi-threading".

For, e.g., cosmological inference with cobaya/class, if a node has 24 cores, you might want to have 4 copies of the program running, each using 6 threads.

Terms of interest

  • ssh this is a program that you use to connect to another computer. This is the tool that we use to log on to the cluster.
    • Usually you have to type a password, but you can avoid this with "ssh keys" (google it).
    • .ssh/config is handy. We can talk about that later, too.
  • submitting a batch job: this is where you tell the cluster to do something for you.
  • SLURM: this is the batch-submission software that manages the cluster. You submit a job by writing a submission script and submitting it to the cluster with a program called sbatch.
  • scp: transfer files between machines. Uses ssh to do so.

For example, you would have a file called myjob.sh that contains information that SLURM needs to submit, as well as the program that you want to run. Then you would execute the following:

$ sbatch myjob.sh

Other useful SLURM commands: sinfo (shows cluster status) and squeue -u <username> shows status of your current jobs.