singularity support

[tbugfinder]

DL benchmarks could also get executed within HPC environments. Often singularity is the container engine used within HPC therefore singularity should be supported.

https://sylabs.io/singularity/

[qhaas]

Singularity support would be quite useful to more than just HPC users.

For example, many of our production multi-user systems have Singularity, but not docker. Our IT policy doesn't permit docker on some systems due to the reasons outlined here. I'd like to use a containerized dlbs to compare pytorch our x86-64 and Power9 systems equipped with Tesla V100s, but our DGX-2's policy only allows for Singularity.

[sergey-serebryakov]

@qhaas Do you build Singularity containers by yourself or do you pull them from some kind of repository? We do have experience running DLBS + TensorFlow + Singularity but that never ended up in the github repo. I'll talk to people who actually ran that config. It will be super useful if you could provide example Singularity recipe. Thanks, Sergey.

[qhaas]

We normally just create singularity friendly docker images based on the upstream cuda images and build singularity images from the docker images. We would use the ngc pytorch image as a base, but we need ppc64le support for our IBM AC922 HPC.

[tbugfinder]

For getting started quickly we just pull images from dockerhub and those are converted to simg automatically. In that case singularity is almost an inline replacement to docker.

[tbugfinder]

@sergey-serebryakov Could the singularity test code be shared maybe in its own branch for now?

[xuagu37]

Any updates on singularity support?