[PVE-User] Ceph Cache Tiering

Lindsay Mathieson lindsay.mathieson at gmail.com
Mon Oct 10 13:46:01 CEST 2016

On 10/10/2016 8:19 PM, Brian :: wrote:
> I think with clusters with VM type workload and at the scale that
> proxmox users tend to build < 20 OSD servers that cache tier is adding
> layer of complexity that isn't going to payback. If you want decent
> IOPS / throughput at this scale with Ceph no spinning rust allowed
> anywhere:)

I think you're right, for all the talk of small scale deployments and 
commodity hardware, ceph on the small business scale is a poor 
price/performance ratio. To get decent performance out of it you have to 
spend big on terrabyte SSD's, battery backed disk controllers etc.

Its a shame, the flexibility of it is really good and rock solid in the 
9.x range if you don't push it :) Its ability to quickly heal only dirty 
data is outsanding. I just tested a simple 3 node setup backed by our 
ZFS pools and was only getting 50MB/s seq writes. If I added SSD 
journals that probably would have got to the 100MB/s level, which would 
have been good enough, but there is one deal breaker for us and thats 
snapshots - they are incredibly slow to restore, 45 minutes for one that 
was only a few minutes old. It gets worse the more writes you add. qcow2 
snapshots on gluster only take a couple of minutes, and we use snapshots 
a lot for testing development and support.

> Additionally i think it is error prone.
> I ran into the problem, that a ssd stuck because it was full causing the complete storage to stuck.

It does seem to be very much a work in progress :)

Lindsay Mathieson

More information about the pve-user mailing list