lock contention etc.

[Julian Elischer]

Andrew Gallatin wrote:
> Julian Elischer writes:
>  > I'm guessing that the extra system time represents a lot of
>  > contention somewhere.
>  > 
>  > What is the best way to get actual numbers on this?
> 
> The Dtrace lockstat tool is designed for this, and I've used it to
> track down contention like this on Solaris.  It seems to be on the
> TODO list for the FreeBSD Dtrace port at
> http://people.freebsd.org/~jb/dtrace/todo.html.  I wonder how soon it
> will arrive..  It is one of the things that I'm most looking forward
> to.
> 
> Drew

I just realised that since this is an HTT machine the massive
increase in user and system times between -j2 and -j4 is probably
due to every thing running slower but over twice as many cpus..
The increase between -j1 and -j2 is more relevant and
interesting I think.

make -j 2 buildworld
     2465.10 real      2958.94 user      1918.64 sys
     2463.97 real      2968.43 user      1910.65 sys
     2472.13 real      2978.64 user      1912.80 sys
     2582.83 real      2965.51 user      2047.19 sys


make -j 4 buildworld
     2262.88 real      4322.31 user      3333.58 sys
     2255.88 real      4330.32 user      3319.48 sys
     2256.09 real      4318.84 user      3336.57 sys
     2257.63 real      4338.34 user      3313.07 sys