Singularity Container System for HPC

At work I had a task to port the WRFv4 atmospheric model that we currently run in the office, from a Docker container, up to the Lonestar5 (LS5) supercomputer at the Texas Advanced Computing Center (TACC). My first impulse was to simply replicate the Dockerfile steps that we use in a simple bash script to natively build the required WRF executables.

Since I always forget the module load commands, I opened up the LS5 documentation. In doing so, I noticed that a recent update system update provides support for “Singularity” containers. Somehow I missed the rising popularity and support for Singularity, but was intrigued because it attempts to solve several of the problems we deal with running parallel numerical models (MPI, GPU and queue/batch systems) with Docker containers. And so I went down the rabbit hole of researching and installing Singularity with the hopes of replicating our Docker-like container workflows in a way that will run on LS5.

Running WRF Singularity Container on TACC Systems

So prior the the recent TACC training on containers I was trying to get our WRF Docker images to run up on the TACC HPC infrastructure using Singularity.

I checked the version of Singularity on the Lonestar5 cluster and it was behind the most recent versions a bit, but luckily there are Docker images for Singularity so you can play with whatever version might be required. (The version on Lonestar5 may already be updated by now too, recent version on Stampede2 was 3.4.3.)