While the basic BC Installation provides for a standard compute image, it's often the case that HPC systems rely on heterogeneous architectures, which can require different node images. Some sites may also wish to provide separate visualization or login nodes.
Additional roles in the BC exist for building two other predefined node types: Login, and GPU nodes. These each may require site-specific configuration, but act as guides for setting up these more customized appliances.
Where to install software?¶
It's important to balance the size of the compute node image with the ease of installing software. In many cases it is better to install to the shared filesystem, rather than adding additional bloat to the compute image, which slows down the boot/build process, and takes up space on the headnode root partition (depending on your filesystem configuration) and in the Warewulf database.
Step-by-Step Image Building (Warewulf)¶
This guide provides a minimal set of guidelines for building a custom node image. The pieces that are not generalizable are left as an exercise for the reader, based on your local requirements.
To initialize a new chroot environment, based on the same warewulf template as the original compute nodes:
wwmkchroot compute-nodes /opt/ohpc/admin/images/new-image #new-image can be whatever... as can this whole path!
It's important to be aware of the amount of disk space available in /opt when adding new images - these can balloon quite rapidly!
It's easiest to install new packages into the new image chroot via yum - from outside the image:
yum -y --installroot=/opt/ohpc/admin/images/new-image install chrony kernel lmod-ohpc grub2 freeipmi ipmitool
yum will not function correctly inside the chroot environment without various changes, which could lead to breaking the headnode.
If you need to install software that isn't available from yum into your new image, copy the source into the appropriate
place, chroot into the environment, and build away:
cp /root/package.tar.gz /opt/ohpc/admin/images/new-image/root/
tar xfvz ./package.tar.gz
It is preferable, however, to install software in a shared filesystem rather than inside of the image, since large image result in slower boot times and more RAM used for the tmpfs.
If you build development tools or some Python tools into the compute image, be aware that not all pieces of the chroot
are included in the final image! Check in
for lines like
which exclude /usr/lib from the final image. Comment these out as needed for your packges to function correctly!
If you need to build additional drivers for something like a GPU, it may be necessary to mount additional directories into the chroot environment.
Make sure you'll be able to ssh over to the new image:
cp /root/.ssh/authorized_keys /opt/ohpc/admin/images/new-image/root/.ssh/authorized_keys
Also, copy over your default /etc/fstab for your environment:
cp /opt/ohpc/admin/images/centos7-compute/etc/fstab /opt/ohpc/admin/images/new-image/root/.ssh/authorized_keys
You may also need to enable any important services in the chroot environment, so that they will start during
the stateless boot. This include things like slurmd, chrony, and munge in particular.
systemctl enable slurmd chronyd munge
Once you've installed all the customized software you'll need, you can create the new VNFS via:
wwvnfs -y --chroot /opt/ohpc/admin/images/new-image
That will add a vnfs image with the name
new-image to the database (there's a --name flag if you want a different name than the last directory in the path-to-chroot)
Confirm the new vnfs via
wwsh vnfs list
If you'd like to boot an existing node with the new image, you can modify the node entry in the WW database, via:
wwsh provision set compute-0 --wwvnfs=new-image
Before doing this, of course, be sure the node will not be scheduled by slurm:
scontrol update nodename=compute-0 state=down reason=testing-new-image #reason is necessary but no specific list of flags need be used