On Mon, Mar 21, 2005 at 01:58:12PM +1100, Simon Burge wrote:
> Chuck Silvers wrote:
> 
> > On Thu, Mar 10, 2005 at 10:24:22AM +1100, Simon Burge wrote:
> > > Just for kicks, here's some times with a -current kernel from sources
> > > around March 7th.  This is a trimmed down kernel with no IPV6,
> > > filesystems I don't use, etc (same as the above 2.0F benchmarks).  When
> > > the kernel boots, I see:
> > > 
> > > 	NetBSD 2.99.16 (PC532) #65: Thu Mar 10 08:36:15 EST 2005
> > > 	total memory = 8192 KB
> > > 	avail memory = 6992 KB
> > > 
> > > time (real)                  vm.bufmem_hiwater vm.{anon,exec,file}{min,max}
> > > 7:44.74 43.3%   0+0k 101+39io 44309pf+1w  524288  {{10,95},{ 5,30},{10,50}}
> > > 4:43.54 43.3%   0+0k   5+26io 17627pf+1w  524288  {{10,95},{ 5,10},{ 5,15}}
> > > 2:53.01 53.1%   0+0k   4+24io  7883pf+1w  131072  {{10,95},{ 5,10},{ 5,15}}
> > > 2:14.59 60.2%   0+0k   4+24io  4552pf+1w  131072  {{10,95},{ 5,10},{ 0, 5}}
> > > 1:56.30 65.7%   0+0k   4+24io  3404pf+1w  131072  {{10,95},{ 5,10},{ 0, 1}}
> > > 1:54.14 66.5%   0+0k   4+24io  3261pf+1w  131072  {{10,95},{ 0,10},{ 0, 1}}
> > > 1:51.39 67.9%   0+0k   4+23io  3209pf+1w  131072  {{10,80},{ 0,10},{ 0, 1}}
> > > 
> > > Note that the user time component of was a relatively constant 60 to 61
> > > seconds, and system time was 139 seconds in the first down to 15 seconds
> > > in the last test.
> > > 
> > > Compare those times with a trimmed down kernel against a GENERIC kernel
> > > (6284 KB free at boot) for the vm tunables with the fastest time above.
> > > 
> > > time (real)                  vm.bufmem_hiwater vm.{anon,exec,file}{min,max}
> > > 8:55.57 44.2%  0+0k 3+24io 52780pf+1w  131072  {{10,80},{ 0,10},{ 0, 1}}
> > > 
> > > And then compare that with a 1.5X GENERIC kernel (6220KB free):
> > > 
> > > 55.189u  20.468s  2:06.78 59.6% 0+0k 71+63io  5364pf+0w
> > 
> > so basically, the extra ~700k makes a huge difference.  this benchmark just
> > barely fits in memory.  I doubt there are any sysctl settings that would
> > allow a GENERIC -current kernel to get a 2 minute run.
> 
> I keep meaning to play with more tests, but I just haven't had the time
> recently.
> 
> The main sticking point from my POV is that my pc532 with a 1.5X GENERIC
> without any sysctl knobs and reporting "avail mem 6220 KB" at boot
> builds machdep.o in around 2mins.  An untuned -current GENERIC gets
> something silly like 40+ minutes.  The only way I can get close to 2min
> with -current is playing around with trimmed down kernels that report
> nearly 7MB free then playing around with sysctl knobs until I find some
> combination that "works".

oh, I misread your earlier message.  I was thinking that the need to
strip down the kernel config was because it had grown since 1.5X,
but that's not what you were saying.

one thing that changed shortly after 1.5X was that we started counting
shared library pages as "exec" (in addition to pages for base executables),
that's probably what made the difference.  not that knowing that helps much.


> Note too that userland has been constant since I first starting doing
> these tests, primarily because it's too hard to build a recent userland
> :-/

-Chuck