On Tue, Sep 19, 2006, Matus UHLAR - fantomas wrote: > On 19.09.06 10:11, Adrian Chadd wrote: > > The COSS code in Squid-2.6 has come quite far from its original design by > > Eric Stern. Steven Wilton has put an enormous amount of effort into the > > COSS design to fix the remaining bugs and dramatically improve its > > performance. > > > > I've assembled a quick webpage showing the drop in CPU usage and the > > negligible effect on hit-rate. Steven Wilton provided the statistics > > from two Squid caches he administers. > > > > You can find it here - http://www.squid-cache.org/~adrian/coss/. > > Great. Did you play with max-size option to try to find out at which size > the efficiency of aufs and COSS nears? I haven't really administered large-scale production caches in a few years but there's plenty of academic papers circa 1997 which show there's a clear win doing your own IO vs using a unix FS for object sizes under a couple hundred kilobytes. Modern UNIX FSes haven't gotten (much) better in that regard. COSS is still pretty new and there's not a lot of documentation on how to tune it. But even the defaults smoke using a normal unix filesystem. Just start at max-size at, say, 64k or 128k and go from there. Don't bother going above a few hundred kilobytes. > > Steven is running a recent snapshot of squid-2.6. The latest -STABLE > > release of Squid-2.6 doesn't incorporate all of the COSS bugfixes > > (and there's at least one really nasty bug!) so if you're interested > > in trying COSS out please grab the latest Squid-2.6 snapshot from > > the website. > > are you speaking about current release (STABLE3) or about stable squid in > general? Steven has committed a bunch of bugfixes to squid-2.6 after Squid-2.6-STABLE3 was released. So no, the current snapshots are fine but stable3 isn't. Adrian
