[RFC] shipping kernels with default modules?
Ivan Voras
ivoras at freebsd.org
Sat Jun 18 15:21:51 UTC 2011
On 11/06/2011 22:49, Doug Barton wrote:
> And no one has ever been surprised that assumptions proved invalid in
> the light of actual testing? :) Theorizing on this point is of
> less-than-zero utility. Who is going to volunteer to do the actual
> benchmarking?
For what it's worth, I've done the comparison with GEOM_ZERO and
"diskinfo -vt". The results are surprising. I've done about 7 iterations
of the diskinfo benchmark in each case (module and compiled-in) and
analyzed all their outputs collectively with ministat - as there is no
actually IO performed, all operations are basically NOPs no matter what
seeks are issued.
Here's the ministat analysis of "seek times" (x is as module, + is as
directly compiled; GENERIC is used in both cases, AMD64 8.2-RELEASE):
x 70 0.021 0.031 0.028 0.027585714 0.0016637034
+ 98 0.025 0.033 0.031 0.02972449 0.0020548189
Difference at 95.0% confidence
0.00213878 +/- 0.000583402
7.7532% +/- 2.11487%
(Student's t, pooled s = 0.00190204)
Here's the analysis for bandwidth part of the benchmark (same symbols):
N Min Max Median Avg Stddev
x 30 1785465 2096170 1909806 1915060.4 77626.781
+ 42 1777623 2083800 1819215 1833298.4 62484.689
Difference at 95.0% confidence
-81761.9 +/- 32966.2
-4.26942% +/- 1.72142%
(Student's t, pooled s = 69161.2)
It looks like for both benchmarks, the results are better for the module
case (lower seek times, higher bandwidth).
As the results were surprising, I did ABAB interleaving of the two cases
(module vs built-in).
More information about the freebsd-arch
mailing list