On Tue, Sep 04, 2001 at 11:50:18AM +1000, Luke Mewburn wrote:
> > 
> > We should actually benchmark filesystems with many *less* cylinder groups
> > (32K filesystems) against those with about 50 cylinder groups on modern
> > disks to see which way we handle real workloads better.  However, I think
> > there's plenty of evidence to support switching to 16K blocks and as many
> > cpg as we can get, given the geometry (about 300 for most new disks) right
> > now.
> 
> So, I did a bit more testing, this time with the smallest number of
> cylinder groups I could get away with, and filesystems with 16K/2K
> and 32K/4K configuration.  I decided not to bother with olddirpref
> for the 16K and 32K file systems, because we know they suck. :)

So, in the 80,000 vnode case, it looks like 8K/1K is just barely better than
16K/2K which is about 15% better than 32K/4K.

However, I have a hunch we now have a _different_ bad interaction going on
here.  I think you need to scale the number of directories to allocate in
one place by the relative CG size, or at least by a pretty good factor to
compensate for the much larger CGs -- the penalty for a CG switch, after
all, got a *lot* larger when you made the CGs that much bigger.

Would you mind giving it a whirl?

Thor