[OpenAFS] calculating memory

David Boyes dboyes@sinenomine.net
Mon, 31 Jan 2011 12:17:37 -0600

>I also strongly second the comments made about VM vs large quantities of
>IOPs. We're debating virtualization of servers here (umich) and the jury
>is still out. Thomas Kula has more to say on that topic in his note, but
>in general I have strong doubts about any file server that runs in
>anything except a one-server-per-phyisical-host model.

I'd modify that comment to include the words "on Intel hardware." Largely,
we're discussing the defects in current Intel virtualization. I would have
zero hesitation to deploy file servers on virtual machines on Power or
s390x virtual machines -- this is really a case of right tool, right job;
not all virtualization solutions are created equal.

A suggestion: Even on Intel, if you do go VM-based, start at half the RAM
size as on physical hardware and measure performance, then increase the
size of the VM by 32=10M and retest until you reach a steady-state with 1-2=
swapping inside the VM at a maximum. Unless your OS supports controlling
the size of the disk cache, as the working set of the guest OS increases
and more of the guest RAM is consumed with I/O buffer cache, it becomes
harder to schedule the guest, which hurts the performance of the guest.
Often a 256M guest outperforms a 2G guest in a VM environment -- it's less
difficult to schedule a 256M guest than bigger guests.