extraordinary stoc/ thrash ...

Michael Meeks michael.meeks at suse.com
Tue Apr 3 02:12:27 PDT 2012


Hi Enrico,

On Tue, 2012-04-03 at 05:15 +0200, Enrico Weigelt wrote:
> > 	What would be ultimately nice would be to rid ourselves of this
> > XRegistry 'stuff' and have a simple XML file, read into some simple
> > structures, and added to a couple of hashes for the common lookup
> > patterns we want. Is there any chance of doing that separately, so we
> > can sever the back-compatible .rdb reading for LibreOffice 4 mostly
> > by code removal ?
> 
> Maybe that XML file could even be mmap'ed, just having the index on
> heap.

	mmapping is the least of our worries wrt. memory use and load
performance (read Stephan's nice write-up) :-) if we're truly worried
about the page-cache we could f/madvise DONTNEED after parsing; only if
we wanted to deal better with many concurrent processes touching the
same data would we mmap it; and then we'd need to switch to UTF-8
strings internally - and of course be rather vulnerable to SIGBUS'

>  Depending on the actual structure and access patterns, the
> whole thing could also be splitted into several files, and just use
> the filesystem hierachy as index.

	Sure - but then we loose wrt. seek time on cold starts :-)

> Do you already have some ideas how that XML file would look like ?

	Checkout services.rdb in the install; again do you actively want to
work on this ? or are you just offering the benefit of your insight ?

	Regards,

		Michael.

-- 
michael.meeks at suse.com  <><, Pseudo Engineer, itinerant idiot



More information about the LibreOffice mailing list