History File over 2GB

Alex Kiernan alexk at demon.net
Wed Sep 4 08:03:04 UTC 2002


"Jeffrey M. Vinocur" <jeff at litech.org> writes:

> On Tue, 3 Sep 2002, Russ Allbery wrote:
> 
> > bill davidsen <davidsen at tmr.com> writes:
> > > Russ Allbery  <rra at stanford.edu> wrote:
> > 
> > > | If you set nnrpdcheckart to true in inn.conf, nnrpd will
> > > | verify the existence of the articles at the time of LISTGROUP
> > > | or XOVER and only send lines for those articles that are
> > > | actually present in the spool.
> > 
> > > What does that do to performance?
> > 
> > Depends on your storage backend, but it can cause some minor
> > delays (particularly for very large groups).
> 
> Actually, I don't think minor describes it accurately.  I've seen
> large groups (tens of thousands of articles) stored in tradspool on
> a single disk take several minutes to do xover, and it appears to be
> related to this.
> 

Or on CNFS, where checkart absolutely massacres your performance as
you end up with lots of random, synchronous page-ins of the bitmap.

> I suspect there's something else going on here, because it has all
> of the delay at once and then sends all of the data, which doesn't
> match my mental model of what should happen (send response header,
> and then send multiline response one article at a time as they are
> discovered).  But there's definitely a potential downside to
> checkart as things stand now.
> 

But that sounds weird and is nothing like what I've seen. Is it just
that monster directories (assuming tradspool) are upsetting your
kernel's cacheing?

-- 
Alex Kiernan, Principal Engineer, Development, THUS plc


More information about the inn-workers mailing list