HEADS UP: new cpuset and scheduler topology code.
Aryeh M. Friedman
aryeh.friedman at gmail.com
Mon Mar 3 08:21:01 UTC 2008
Jeff Roberson wrote:
> On Sun, 2 Mar 2008, Aryeh M. Friedman wrote:
>>> Many programs should run faster on dual and quadcore machines but
>>> there is still some tuning to do. If you notice a workload is now
>>> significantly faster or slower please report it to me.
>> I know this is not scientific but so far my experiences are:
>> HW Spec: e6850 w/ 4GB on AMD64
>> Kernel Options: All debugging removed and IPI_PREEMPTION enabled
>> Build/install world/kernel (-j3): 44 mins (15 min. reduction)
>> done wile also running the following in the background:
> Can you back up your sched_ule.c by one version and try again to
> verify these results? That would be quite a fantastic reduction.
Sorry was not able to do it since I was not able to figure out how to
revert only the cpuset stuff (i.e. the only way I know to revert is cvs
update -D "..." on the whole source tree and attempting to only do
sched_ule.c lead to all kinds of nasty problems and panics)... if you
would be willing to send me a patch to do it I will test it.
That being said the savings seems to be genuine in that the port
reinstall (almost all ports not rebuild caused a panic) is roughly 25%
faster (the same improvenment reported in the world/kernel test)
More information about the freebsd-current