<<< Date Index >>>     <<< Thread Index >>>

Re: header_cache broken in cvs



Thomas Glanzmann wrote:  [Mon Mar 13 2006, 03:17:42PM EST]
> > exception that I've been testing with gdbm instead of db4.  Just to
> 
> I use qdbm not db4. Because it is the fastet of the three supported and
> it is the only who does support compression which reduces header cache
> size to 1/5 of the original header cache in my usage scenario. gdbm is
> faster than db4.

There's some confusion between the configuration you posted and your
statement here.  The configuration that you showed in your previous
email showed --without-qdbm --without-gdbm.  Additionally you were
setting header_cache_pagesize which only exists in init.h for
HAVE_GDBM || HAVE_DB4.  So I assume you are referring to your usual
configuration instead of your testing configuration?

> > The second run took 5 seconds, third run took 0.5 seconds, clearly
> > using the cache.
> 
> at seems so but it could be also the buffer cache of the filesystem. But
> killing the cache short after creation and do it again will give you the
> strong prove that you use it. I also had a patch which gathers hit/miss
> statics. Whatever.

I'm sorry, but I don't understand what you're saying in light of what
I stated.  My testing involves three runs.  The first warms the
filesystem buffer cache so that it is not a factor.  After the first
run, I remove the mutt header_cache and perform two additional runs.

You're right that timing is a poor method of determining whether the
header_cache is working, but the discrepancy in times when caching is
used is significant.

Aron

Attachment: pgpOxZELNkRNU.pgp
Description: PGP signature