[OpenAFS-devel] Re: [OpenAFS] namei interface lockf buggy on Solaris (and probably HP-UX and AIX)

[Rainer Toebbicke]

Ken Hornstein wrote:

> 
> I interpreted Tom's comments about "batch updates" as not wanting to
> suppress calls to fsync() (well, I imagine that's part of it), but to
> actually batch up calls to write().  If I'm wrong, then my apologies.
> I agree that turning off fsync for volume operations should be relatively
> straightforward.
> 

Well it is. The patch is in CVS since some time. It does the job on 
the link table quite nicely.

However, it only solves the "volser"  problem. Not the fileserver 
problem which should not be underestimated. It order to create 10000 
files you'll be doing 10000 or 20000 sync()s unless I am mistaken. As 
each create is in its own RPC, batching them up in some logical 
fashion is complicated, batching them up on time intervals is again easy.

On a vanilla native system you can create some 2000-4000 
file/directories per second, ext3 doesn't sync in between. Doing 100 
over AFS is a challenge.

Just by converting the syncing into some periodic task as to at least 
limit the effects of a crash speeds this up by a factor 5-10. Still 
far from 2000 or 4000, but then the local disk cache probably also 
runs into trouble finding and ftruncating cache files.

-- 
=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=
Rainer Toebbicke
European Laboratory for Particle Physics(CERN) - Geneva, Switzerland
Phone: +41 22 767 8985       Fax: +41 22 767 7155