mbuf autotuning effect

[Ian Lepore]

On Sat, 2013-09-07 at 12:21 -0700, hiren panchasara wrote:
> On Sep 6, 2013 8:26 PM, "Warner Losh" <imp at bsdimp.com> wrote:
> >
> >
> > On Sep 6, 2013, at 7:11 PM, Adrian Chadd wrote:
> >
> > > Yeah, why is VM_KMEM_SIZE only 12mbyte for MIPS? That's a little
> low
> for a
> > > platform that has a direct map that's slightly larger than 12mb :)
> > >
> > > Warner? Juli?
> >
> > All architectures have it at 12MB, except sparc64 where it is 16MB.
> This
> can be changed with the options VM_KMEM_SIZE=xxxxx in the config file.
> 
> Right. Does that mean for any platform, if we do not have nmbclusters
> pre-set in kmeminit() than we will always have pretty low value of
> vm_kmem_size. And because of that, if maxmbufmem is not pre-set (via
> loader.conf) inside tunable_mbinit() , we will have very low value for
> maxmbufmem too.
> 
> I hope (partially believe) that my understanding is not entirely
> correct.
> Because if its correct, we arw depending on loader.conf instead of
> actually
> auto tuning.
> 
I think the part of this that strikes me as strange is calling 20% of
physical memory used for network buffers a "very low value".  It seems
outrageously high to me.   I'd be pissed if that much memory got wasted
on network buffers on one of our $work platforms with so little memory.

So the fact that you think it's crazy-low and I think it's crazy-high
may be a sign that it's auto-tuned to a reasonable compromise, and in
both our cases the right fix would be to use the available knobs to tune
things for our particular uses.

-- Ian