[OpenAFS-devel] fileserver profiling

[Troy Benjegerdes]

On Sun, Mar 13, 2005 at 02:47:06PM -0500, Kyle Moffett wrote:
> On Mar 13, 2005, at 14:32, Tom Keiser wrote:
> >ok.  I've been operating under the assumption that implementing a
> >purely userspace gettimeofday would be too costly, but apparently that
> >assumption is false.  If you're only storing seconds since the epoch,
> >your value is atomic, so you don't need sanity checking to prevent
> >time apparently going backwards.
> 
> Agreed (Although that's somewhat architecture and compiler dependent).
> 
> >I agree that if you're going to store usecs too, then you need the
> >read barriers.  When I said we could coalesce ops and reduce the
> >number of barriers, I was thinking about a technique similar to RCU
> >that Herlihy discussed back in the late 80's
> 
> Interesting.  In any case, it's very efficient to do both time() and
> gettimeofday() in userspace without using much CPU time.
> 
> >>My goal for this implementation
> 
> Hrm, I appear to have inadvertently deleted a paragraph or two from
> my response in that email, oops!

Have you tried implementing this in the fileserver yet, so we've got
some real numbers to look at?

I'm kinda stumped as to why I can get 50MB/sec via NFS to the /vicepa
partition, but through AFS, I'm only getting about 5. (I've tried both
openafs and the 2.6.11 in-kernel AFS).

This is a single 4.2GB file I'm testing with, FYI.