[OpenAFS] Re: OpenAFS failing rmdir performance

Andrew Deason adeason@sinenomine.net
Fri, 8 Oct 2010 17:11:04 -0500

On Fri, 8 Oct 2010 17:43:44 -0400
Simon Wilkinson <sxw@inf.ed.ac.uk> wrote:

> On 8 Oct 2010, at 17:24, Andrew Deason wrote:
> > This is probably the throttling behavior of the fileserver, where
> > the fileserver throttles client activity if it sends too many RPCs
> > that return error codes in a certain amount of time.

Actually, this is a per-connection (and per-call) count, IIRC; it's not
within a certain amount of time and it never goes down unless you get a
new conn.

> I do wonder if the abort threshold is too aggressive when applied to
> authenticated clients. Whilst a denial of service attack is possible
> from authenticated clients, it's also more likely that there will be
> administrative means available to track down and punish offenders.

Yeah, I agree with the general idea. I'm not sure if just increasing the
threshold is good enough, since once you hit it the performance just
goes straight to N req/s (right now N=1, I think) effectively forever as
long as you're constantly using the conn. Maybe we should also change
the throttled rate, or just reset the count after a certain amount of
time (which would effectively do the same thing amortized).

Although, just a separate thresh is a lot easier, and being able to
configure '-aborttreshold 0' just for auth'd conns would be nice.

> Maybe we need two different thresholds depending on the class of the
> client.

Or maybe N thresholds, for any of the existing or future security class
indices? (This sounds annoying to tweak via command line arguments; when
are we getting a config file again? ;)

Andrew Deason