Nodes must be able to be halted, destroyed and then re-integrated, possibly with a "light provision" or "no provision" alternative about hpc-collab HOT 5 OPEN

transcription of timing data...

some timings: Vagrant 2.2.9, Virtualbox 6.1.6, underlying platform; Fedora30: 5.6.13-100
VirtualBox VMs file system: xfs over luks, underlying SSD/nvme
tarballs file system: xfs, underlying SSD/nvme

vcfs: 9m, ingest/createrepo repos.tgz, yum install/timeouts, selinux homedirs
vcsvc: 3m, yum install/timeouts, rsyslog selinux
vcbuild: 11m, yum (epel/priorities) install/timeouts, slurm build - if triggered, prereq timeouts
vcdb: 7m, yum (epel/priorities) install/timeouts, mysql (community mysql repo?) install, prereq timeouts
vcaltdb: 9m, yum (epel/priorities?) install/timeouts, mysql (community mysql repo?) install, prereq timeouts
vcsched: 6m, yum (epel/priorities?) install/timeouts, prereq timeouts
vc1: 13m, yum (epel/priorities?) install/timeouts, prereq timeouts
vc2: 4m, yum (epel/priorities?) install timeouts, prereq timeouts
vclogin: 22m, slurm test jobs, slurm db/config, yum (epel/priorities?) install timeouts
vxsched: 3m, yum install (epel/priorities?) timeouts
vx1: 5m, yum install (epel/priorities?) timeouts
vx2: 4m, yum install (epel/priorities?) timeouts
vxlogin: 10m, slurm test jobs, slurm db/config, yum (epel/priorities?) install timetouts

muon/centos underlying similar, slower: VirtualBox VMs & tarballs: zfs, raidz1 over HDD

So, first-cut:

1. cache more from epel into local repositories, attempting to remove it from most/all nodes,
2. find another mechanism besides yum-plugin-prorities to force local repo first, then remote fallback,
3. tune prereq timeouts
4. all else, such as vclogin, slurm-test-jobs, selinux

from hpc-collab.

VMtouch, to lock memory pages in the host: https://github.com/hoytech/vmtouch
User-level NFS server, replacing vboxsf

preliminary timings don't indicate much of a gain, implying the problem isn't host I/O, but rather host to VM I/O or in-VM I/O.
Cpu load is not high when this is occurring, implying that it is not trashing on the host system. (in either the host or guests)

from hpc-collab.

Contacting the epel repository is a source of delay in node provisioning. This is a wait for network I/O, rather than any real productive work. This may be best addressed by a snapshot-checkpoint-suspend/resume cycle.

This has been addressed by careful caching of RPMs to avoid epel and yum search timeouts, especially in the early RPM installation step.

from hpc-collab.

compute nodes may be unprovisioned and reprovisioned safely, separate from other nodes.

from hpc-collab.

estale monitoring code added, allowing most nodes (except vcdb) to be reprovisioned. In particular, vcfs may disappear and reappear, much like a traditional NFS server

from hpc-collab.