[OpenAFS] Current Limits of OpenAfs ?

Steven Jenkins steven.jenkins@gmail.com
Tue, 13 Jan 2009 18:18:50 -0500

On Sat, Jan 10, 2009 at 1:08 PM, Matt Benjamin <matt@linuxbox.com> wrote:
> Hash: SHA256
> Adding to that, I'd suggest reading the slides from Alistair Ferguson's
> keynote at the 2008 AFS workshop.  The current client/server ratio
> (2000/1, going to 5000/1?) and configured callback's per file server (4
> million?) in one Morgan Stanley enviornment are mentioned, IIRC.

It should be emphasized that their highest client/server ratios occur
in what they call 'read-only' cells, which only contain read-only, not
read-write data.  Callbacks for those two different types of data are
handled very differently (e.g., file level for read-write data vs
volume level for read-only data); thus I suggest carefully
benchmarking with actual access patterns to determine how far you want
to scale.

I would be further cautious taking numbers from that presentation and
applying them elsewhere.  Slide 20 points out they currently have 1.5K
hosts per cell (750 per fileserver) now, and their goal is for OpenAFS
to scale to handle 10K hosts per cell (5K per server) down the road.
Those numbers are further muddled because they do not specify how many
read-only copies are available total for data (or even if those ratios
are for read-only or read-write cells), how (un)balanced access
patterns are, how much headroom for growth those ratios include,
hardware configurations, etc.

Fileservers aside, and without knowing more about what Gary is trying
to accomplish, the ~10 million user number is quite interesting.  I do
not know of any OpenAFS installation (or IBM AFS) installation that
has tried to scale beyond 100K or so users.  I know that we have done
tests in a lab with very large numbers of users, but having that scale
in production would likely present some new challenges.

Steven Jenkins
End Point Corporation