matthew green writes:
> 
> hi folks.
> 
> 
> i've been benchmarking builds on a (dual) athlon64 x2 running
> amd64 mode.   i recently noticed that my builds were slower
> and i have tracked down the slowness to the x86pmap merge.  a
> bit of binary searching led me to believe it was that, and 
> backing out the change on -current gives me similar performance
> to the previous kernel.
> 
> it's a fairly severe degradation - i see almost 10% more runtime
> to build a kernel!
> 
> here are some numbers to illustrate:
> 
> plain -current:
> 
> 	256.8u 92.5s 2:56.98 197.4% 0+0k 0+7755io 0pf+0w
> 	256.0u 93.6s 2:56.94 197.6% 0+0k 0+7751io 0pf+0w
> 
> old amd64/pmap.c with -current:
> 
> 	241.7u 77.5s 2:41.72 197.4% 0+0k 0+7742io 0pf+0w
> 	242.7u 76.4s 2:41.71 197.4% 0+0k 0+7747io 0pf+0w
> 
> 1 month old kernel:
> 
> 	241.3u 73.2s 2:41.65 194.5% 0+0k 33+7748io 125pf+0w
> 	242.1u 72.6s 2:39.51 197.3% 0+0k 0+7745io 0pf+0w
> 
> 
> (so there may infact be more slowness since a month ago, but
> i have not looked into that at all -- just noticed it now.)
> 
> so that is adding over 15 seconds to a 160 second build..
> ouch.

In the pre/post x86pmap benchmarking I did, I couldn't find any 
noticable difference on a -j 16 build.sh on a quad opteron box 
with 3 builds with a kernel from just before the merge and with
3 builds with a kernel from just after..  All the builds came in
at about the 54 minute mark.  (Unfortunately, it seems I've since 
toasted the actual build results.. likely because they didn't seem 
overly interesting at the time :-} )

Later...

Greg Oster