[OpenAFS-devel] Re: developer takeaways from EAKC 2014

[Andrew Deason]

On Mon, 31 Mar 2014 10:48:52 -0400 (EDT)
Benjamin Kaduk <kaduk@MIT.EDU> wrote:

> I think there were a couple of sites that indicated they had tuned the
> client cache size and chunk size away from the defaults, and got
> noticably improved performance.  I seem to recall that our defaults
> are mostly unchanged from when they were created a long time ago, for
> hardware several generations removed from the current state of
> affairs.  Revisiting the defaults seems reasonable.

chunksize should default to 1 MiB if the cache size is reasonable. I
thought 1 MiB should be okay; I'd like to see what it was changed
from/to if that did make a difference.

Cache size itself is handled by packaging; so improving that may require
getting someone else to change it in some cases.

> CERN also has a patch that does per-volume throttling, in my notes.  I
> am less sure I accurately remember what this one was doing, but I
> think it was limiting the number of threads servicing requests for a
> given volume, so that other ("normal") volume accesses were not
> affected by a single user thrashing one volume.

This idea is generally the same as the existing per-host throttling, but
is just counting the number of threads per-volume instead of per-host.
In the past we've mentioned doing this per-volume and per-vnode.

-- 
Andrew Deason
adeason@sinenomine.net