Niklas Edmundsson <[EMAIL PROTECTED]> writes:
> Package: openafs-fileserver
> Version: 1.4.1-2
> We had various performance-issues with 1.4.0, and during that debugging
> we found out that the problem was made even worse by the fact that the
> openafs fileserver is started without any tuning options. Since our
> fileserver was initially installed using an older package, this might be
> fixed now, and in that case I appologise.
> The interesting line in /etc/openafs/BosConfig should be modified to
> include:
> parm /usr/lib/openafs/fileserver -p 23 -busyat 600 -rxpck 400 -s 1200 -l 1200
> -cb 65535 -b 240 -vc 1200
> The above tuning is recommended according to Harald Barth, for the
> following reasons:
> * It's unclear what the defaults are unless you dig into the source.
> * Modern hardware have no problem with increasing buffers and threads
> a bit.
> * Having too few threads will cause your fileserver to stall if too
> many threads ends up waiting for something (breaking callbacks and
> timing out, whatever).
My ideal would be to get the defaults fixed, of course, but in the
meantime I can see the argument that relying on the current defaults is a
bad idea. I think this only affects the afs-newcell command and the
documentation, though; in general, AFS leaves it to you to set up your
fileserver command however you want it, and only the Debian afs-newcell
automated cell creation script encodes parameters (or lack thereof).
We use:
-L -p 125 -rxpck 400 -busyat 200 -cb 65535
but I think that may actually be too many threads. I'm curious about the
differences. Ideally, before picking defaults to put into the
documentation, I'd like this to be discussed on openafs-info towards some
sort of consensus (which is where the current client cache parameters came
from). If you have a chance and are willing to start that conversation,
that would be great; if not, I'll try to do so in the not too distant
future.
--
Russ Allbery ([EMAIL PROTECTED]) <http://www.eyrie.org/~eagle/>
--
To UNSUBSCRIBE, email to [EMAIL PROTECTED]
with a subject of "unsubscribe". Trouble? Contact [EMAIL PROTECTED]
