Podman and NFS
In my previous Podman in HPC environments article I introduced how Podman can be used to run containers under the control of Open MPI. In this article I want to extend my HPC environment to use a shared NFS home directory.
The following examples are running on CentOS 7.7 and are configuring Podman for rootless usage based on the official documentation.
The user in my examples is named centos.
$ sudo sh -c "echo 'user.max_user_namespaces=28633' > /etc/sysctl.d/userns.conf"
$ sudo sysctl -p /etc/sysctl.d/userns.conf
$ sudo sh -c "echo 'centos:200000:65536' >> /etc/subuid"
$ sudo sh -c "echo 'centos:200000:65536' >> /etc/subgid"
With this the system should be ready to run rootless containers. As I am focussing on containers running under Open MPI's control I am using Podman with --net=host, as mentioned in my previous article.
During system setup I am also configuring Podman to be ready to run on a NFS based home directory because, as far as I know, it is not possible for Podman to correctly setup the necessary user namespaces when the storage backend is running on NFS.
The following commands are necessary on my system to tell Podman to use /tmp/centos/containers as the storage backend:
$ podman info
$ sed -e "s,graphroot.*$,graphroot = \"/tmp/centos/containers\",g" -i .config/containers/storage.conf
$ rm -f ./.local/share/containers/storage/libpod/bolt_state.db ./.local/share/containers/cache/blob-info-cache-v1.boltdb
The first command lets Podman create an initial configuration for the current system. As the home directory is on a NFS mounted directory it is necessary to tell Podman to use a non NFS directory for backend storage (/tmp/centos/containers in this example). As this happens during initial system (or user) configuration and no container has yet been run by Podman I can easily delete Podman's local database which contains reference to the home directory as the storage backend. With these 3 steps Podman is ready to be used on a NFS based home directory once the user logs in for the first time.
I am now running the same Open MPI based container example as in my previous article.
$ mpirun --hostfile hostfile \
--mca orte_tmpdir_base /tmp/podman-mpirun \
podman run --env-host \
-v /tmp/podman-mpirun:/tmp/podman-mpirun \
--net=host --pid=host --ipc=host \
Rank 2 has cleared MPI_Init
Rank 2 has completed ring
Rank 2 has completed MPI_Barrier
Rank 3 has cleared MPI_Init
Rank 3 has completed ring
Rank 3 has completed MPI_Barrier
Rank 1 has cleared MPI_Init
Rank 1 has completed ring
Rank 1 has completed MPI_Barrier
Rank 0 has cleared MPI_Init
Rank 0 has completed ring
Rank 0 has completed MPI_Barrier
The difference to the previous article is that my home directory is now NFS based. Podman will now go to the specified registry (quay.io) to download for each host involved in the MPI job the specified container to /tmp/centos/containers.
This enables me to use Podman in a even more HPC like environment where shared home directories are very common to share input and output data.