Ceph - Distributed File System in Proxmox
Ceph is massively scalable distributed file system which runs on Linux. I primarily use Ceph within a Proxmox cluster. If you haven’t checked out Proxmox as a virtual machine hypervisor then I strongly suggest you do so!
Clearing Crash Notifications
From the console, simply run:
ceph crash archive-all
Removing Ceph Completely
If you need to completely remove Ceph from a server as you no longer need it or you’d like to install cleanly then you can follow these steps.
apt purge ceph-mon ceph-osd ceph-mgr ceph-mds
rm -rf /var/lib/ceph/mon/ /var/lib/ceph/mgr/ /var/lib/ceph/mds/
rm -r /var/lib/ceph
rm /etc/pve/priv/ceph -r
Improve Recovery / Rebalance Performance
By default Ceph prioritises client traffic over recovery and rebalancing. If you need to give Ceph recovery a higher priority (for instanc; when recovering from a disk replacement), you can change the settings on each OSD to optimise the cluster behaviour.
ceph tell 'osd.*' injectargs '--osd-max-backfills 32'
ceph tell 'osd.*' injectargs '--osd-recovery-max-active 32'
Don’t forget to reset afterwards:
ceph tell 'osd.*' injectargs '--osd-max-backfills 1'
ceph tell 'osd.*' injectargs '--osd-recovery-max-active 0'
Additional performance tuning options you can experiment with:
ceph tell 'osd.*' config set osd_min_pg_log_entries 10
ceph tell 'osd.*' config set osd_max_pg_log_entries 10
ceph tell 'osd.*' config set osd_pg_log_dups_tracked 10
ceph tell 'osd.*' config set osd_pg_log_trim_min 10
Destroying an unused disk reference
Destroying an old (unused) disk that is still being held onto as a reference in Proxmox.
ceph-volume lvm zap /dev/sdX --destroy
Changing the Ceph Public Network
Changing the Ceph Public Network once the cluster is up and running can be challenging to figure out but is quite straightforward.
- Edit ceph config file from shell (/etc/ceph/ceph.conf) and set new IP range
- One by one destroy and recreate monitors and managers from the UI, they should come up with an IP on the new public network
- Check the cluster health at each step
- Reboot each server one by one
It’s often helpful to measure the performance of your Ceph setup. For me this has helped verify the network configuration is setup correctly, as well as ensuring the disk performance settings are optimal.
Create a dedicated Ceph pool to benchmark in:
ceph osd pool create scbench 128 128
ceph osd pool application enable scbench rdb
Run benchmarks using rados:
rados bench -p scbench 10 write --no-cleanup
rados bench -p scbench 10 seq
rados bench -p scbench 10 rand
Cleanup once you’re finished:
rados -p scbench cleanup