proof-of-concept patch for cpu accounting speedup
Ivan Voras
ivoras at fer.hr
Wed Nov 30 02:13:15 GMT 2005
Poul-Henning Kamp wrote:
> I am very interested to hear feedback and benchmarks of this patch.
So, here are some :) I've applied your patch to 6-stable with only minor
difficulties.
Here are the results of unixbench's context1 benchmark for "normal"
(unpatched) kernel with acpi-fast, tsc and i8254 timecounter hardware
and then with a patched kernel (timecounter hardware didn't make a
noticable difference here):
x ct1-6-stable-tsc
+ ct1-6-stable-tscphk
* ct1-6-stable-acpifast
% ct1-6-stable-i8254
+--------------------------------------------------------------------------+
|% * |
|% * |
|% * x +|
|% * xx ++|
|% * xx ++|
|A A A| |A|
+--------------------------------------------------------------------------+
N Min Max Median Avg Stddev
x 5 2331645 2353665 2342346 2342657.8 9825.4048
+ 5 2585053 2605310 2602141 2596283.8 10010.176
Difference at 95.0% confidence
253626 +/- 14465.2
10.8264% +/- 0.617468%
(Student's t, pooled s = 9918.22)
* 5 1719124 1727693 1725060 1724709.8 3464.6507
Difference at 95.0% confidence
-617948 +/- 10744.2
-26.3781% +/- 0.458633%
(Student's t, pooled s = 7366.9)
% 5 582065 584233 582435 582720 883.62832
Difference at 95.0% confidence
-1.75994e+06 +/- 10173.6
-75.1257% +/- 0.434275%
(Student's t, pooled s = 6975.65)
Also, I did complete runs of unixbench and here are the cumulative indexes:
unpatched-acpi-fast: 326.7
unpatched-tsc: 336.3
patched+tsc: 349.1
By this benchmark, it seems that overall performance could improve by
~7% by using your patch vs the default acpi-fast timecounter on
unpatched kernel. Of course, this result probably won't hold in real life :)
All results are available at:
http://ivoras.sharanet.org/stuff/timebench.tgz
More information about the freebsd-current
mailing list