Getting back to that topic I would like to share some more results. This
time I was testing Core(TM) i7 870 @ 2.93GHz. It has 8 logical cores and
bigger allowed TurboBoost effect. I was testing real time of net/mpd5
port building, using single CPU. I was testing it with HZ=1000 with
different C-states allowed and with/without kern.eventtimer.idletick
sysctl enabled (supported by the latest event timer code in HEAD). This
sysctl, when disabled, allows to avoid most of timer interrupts on idle
cores, allowing them to sleep deeper.
Here is results:
x C1i
+ C2i
* C3i
% C1
# C2
@ C3
+-----------------------------------------------------------------+
| @ |
| # O ** + % xx |
|#...@o@ + ** + %%% xx |
|#...@o@ + ** +++ %%% xx x|
| |_A_| |
| |___A__M_| |
| A| |
| |A |
||A_| |
| |A| |
+-----------------------------------------------------------------+
N Min Max Median Avg Stddev
x 8 12.01 12.14 12.02 12.03125 0.04421942
+ 8 11.45 11.68 11.67 11.61125 0.097897541
Difference at 95.0% confidence
-0.42 +/- 0.0814651
-3.49091% +/- 0.677113%
(Student's t, pooled s = 0.0759582)
* 8 11.51 11.54 11.52 11.52375 0.011877349
Difference at 95.0% confidence
-0.5075 +/- 0.0347234
-4.21818% +/- 0.28861%
(Student's t, pooled s = 0.0323761)
% 8 11.9 11.93 11.92 11.915 0.011952286
Difference at 95.0% confidence
-0.11625 +/- 0.0347382
-0.966234% +/- 0.288733%
(Student's t, pooled s = 0.0323899)
# 8 10.72 10.79 10.73 10.74875 0.031819805
Difference at 95.0% confidence
-1.2825 +/- 0.0413146
-10.6597% +/- 0.343394%
(Student's t, pooled s = 0.0385218)
@ 8 10.76 10.8 10.79 10.785 0.014142136
Difference at 95.0% confidence
-1.24625 +/- 0.035208
-10.3584% +/- 0.292638%
(Student's t, pooled s = 0.032828)
As you may see, with full timer interrupt rate TurboBoost effect (part
of it, that enabled by some number of idle cores) is about 3-4%. CPUs
are not sleeping long enough. But without extra interrupts on idle cores
effect increasing to more then 10%!
--
Alexander Motin
_______________________________________________
freebsd-performance@freebsd.org mailing list
http://lists.freebsd.org/mailman/listinfo/freebsd-performance
To unsubscribe, send any mail to "freebsd-performance-unsubscr...@freebsd.org"
