[OpenAFS] openafs fileservers in VMware ESX
Thu, 21 Apr 2005 16:28:57 +1200
The question is how much does the overhead of virtualisation (which with
afs is not much) actually matter with an AFS fileserver and the client
The data I have collected over time on our hardware server suggests that
our afs servers while containing a lot of data and user volumes are not
very heavily loaded and 95% of the time we are underutilizing the
hardware we had with it sitting at 1-2 % cpu, not using much memory, the
GB network links idling and the disk IO requirements not the heavy on
We also have had to deal with large budget cuts which meant we could not
purchase what we have been able to in the past nor afford to run it.
ESX allows us to share HBA's, FC ports for multiple hosts meaning less
FC overhead in setting it up. We are using 70% less power, need less
cooling and less space.
We now have less (higher quality) servers than we did before so harware
support cost has dropped. Also even with multiple AFS FS servers running
on a single ESX box we do not see any IO issues yet. If we ever do we
can just vmotion the problem server to another underulitlised ESX server
(without having to shutdown the FS server in about 1 minute).
We also have a three year lease cycle that means we have to swap servers
hardware on a regular basis, but with vmotion and shared sans luns we do
not need to use vos move.
If we have any problem volumes/apps we have a couple of hardware afs
servers for them. But so far apart from the issues with vos listvol
(which works but is slow) we have not had any issues.
We are working into it slowly with only about 1Tb of 4Tb of data in VM
afs servers and we can always back out if we run into big problems but
we had not seem any issues in testing because we did scale it up enougn.
Derek Atkins wrote:
> I've never seen any reason to virtualize an AFS server. Ever. The key is IO
> bandwith, which isn't increased by virtualization. You really want separate
> PHYSICAL servers for AFS servers. Virtualization does not give you any
> benefits due to hardware failure, power failure, or any other failure. It just
> adds overhead.