[pbs-devel] Scheduler causing connectivity issues?

[Thomas Lamprecht]

Am 13/07/2022 um 12:41 schrieb Mark Schouten:
> Requested files sent offlist.

Thanks!

You have 30% of runnable process getting stalled due waiting for IO, that
naturally should not cause the request accept future to get starved but is
the reason for why it happened with the current (or better old)
architecture. Increasing available memory, so that the page cache can hold
more entries, could already relieve that system a bit.

We improved on the reproducer we got locally by simulating a higher latency
disk using dm-delay on a small single core VM.

For one we made the libpve-storage-perl do more efficient list-snapshot
requests if they can be filtered by VMID, and on the PBS side we moved most
operations that cause IO (and are related to backup groups/snapshots) to a
separate thread pool so that the main thread should be less
congested/blocked.

The results got packaged and uploaded to our test repositories and are
available with:

- proxmox-backup-server version 2.2.4-1
- libpve-storage-perl version 7.2-7

It'd be great if you could try out those and report back if they actually
helped in your setup(s) too.

cheers,
Thomas