[BENCHMARK] 100Hz v 1000Hz with contest

[BENCHMARK] 100Hz v 1000Hz with contest

[Con Kolivas]

-----BEGIN PGP SIGNED MESSAGE-----
Hash: SHA1

I've attempted to answer the question does 1000Hz hurt responsiveness in 2.5 
as much as I've found in 2.4; since subjectively the difference wasn't there 
in 2.5. Using the same config with preempt enabled here are results from 
2.5.70-mm3 set at default 1000Hz and at 100Hz (mm31):

no_load:
Kernel         [runs]   Time    CPU%    Loads   LCPU%   Ratio
2.5.70-mm3          1   79      94.9    0.0     0.0     1.00
2.5.70-mm31         1   77      94.8    0.0     0.0     1.00
cacherun:
Kernel         [runs]   Time    CPU%    Loads   LCPU%   Ratio
2.5.70-mm3          1   76      97.4    0.0     0.0     0.96
2.5.70-mm31         1   74      98.6    0.0     0.0     0.96
process_load:
Kernel         [runs]   Time    CPU%    Loads   LCPU%   Ratio
2.5.70-mm3          2   108     68.5    64.5    28.7    1.37
2.5.70-mm31         2   107     69.2    67.0    29.0    1.39
ctar_load:
Kernel         [runs]   Time    CPU%    Loads   LCPU%   Ratio
2.5.70-mm3          3   114     70.2    1.0     5.3     1.44
2.5.70-mm31         3   105     73.3    0.7     3.8     1.36
xtar_load:
Kernel         [runs]   Time    CPU%    Loads   LCPU%   Ratio
2.5.70-mm3          3   123     62.6    2.3     5.7     1.56
2.5.70-mm31         3   122     61.5    2.0     4.9     1.58
io_load:
Kernel         [runs]   Time    CPU%    Loads   LCPU%   Ratio
2.5.70-mm3          4   116     66.4    40.6    18.8    1.47
2.5.70-mm31         4   114     65.8    41.0    19.3    1.48
io_other:
Kernel         [runs]   Time    CPU%    Loads   LCPU%   Ratio
2.5.70-mm3          2   116     66.4    50.0    22.2    1.47
2.5.70-mm31         2   112     67.9    46.1    21.4    1.45
read_load:
Kernel         [runs]   Time    CPU%    Loads   LCPU%   Ratio
2.5.70-mm3          2   104     75.0    8.2     5.8     1.32
2.5.70-mm31         2   100     76.0    7.5     7.0     1.30
list_load:
Kernel         [runs]   Time    CPU%    Loads   LCPU%   Ratio
2.5.70-mm3          2   95      80.0    0.0     7.4     1.20
2.5.70-mm31         2   92      82.6    0.0     5.4     1.19
mem_load:
Kernel         [runs]   Time    CPU%    Loads   LCPU%   Ratio
2.5.70-mm3          2   98      80.6    53.0    2.0     1.24
2.5.70-mm31         2   95      81.1    53.0    2.1     1.23
dbench_load:
Kernel         [runs]   Time    CPU%    Loads   LCPU%   Ratio
2.5.70-mm3          4   313     24.3    5.0     56.9    3.96
2.5.70-mm31         4   297     24.9    4.5     52.5    3.86

At first glance everything looks faster at 100Hz. However it is well known 
that it will take slightly longer even with no load at 1000Hz. Taking that 
into consideration and looking more at the final ratios than the absolute 
numbers it is apparent that the difference is statistically insignificant, 
except on ctar_load.

Previously I had benchmark results on 1000Hz which showed preempt improved the 
results in a few of the loads. For my next experiment I will compare 100Hz 
with preempt to 100Hz without.

Con
-----BEGIN PGP SIGNATURE-----
Version: GnuPG v1.2.1 (GNU/Linux)

iD8DBQE+3BRIF6dfvkL3i1gRAnEbAKCpaj/kajzKV3qVrWGRIhOh+Q8O8gCfZp6c
M3Iq1D/41t+4SB2jtNYQc48=
=NMfC
-----END PGP SIGNATURE-----

Re: [BENCHMARK] 100Hz v 1000Hz with contest

[Zwane Mwaikambo]

On Tue, 3 Jun 2003, Con Kolivas wrote:

> I've attempted to answer the question does 1000Hz hurt responsiveness in 2.5
> as much as I've found in 2.4; since subjectively the difference wasn't there
> in 2.5. Using the same config with preempt enabled here are results from
> 2.5.70-mm3 set at default 1000Hz and at 100Hz (mm31):

Thanks for carrying out these =)

> ctar_load:
> Kernel         [runs]   Time    CPU%    Loads   LCPU%   Ratio
> 2.5.70-mm3          3   114     70.2    1.0     5.3     1.44
> 2.5.70-mm31         3   105     73.3    0.7     3.8     1.36

> dbench_load:
> Kernel         [runs]   Time    CPU%    Loads   LCPU%   Ratio
> 2.5.70-mm3          4   313     24.3    5.0     56.9    3.96
> 2.5.70-mm31         4   297     24.9    4.5     52.5    3.86
>
> At first glance everything looks faster at 100Hz. However it is well known
> that it will take slightly longer even with no load at 1000Hz. Taking that
> into consideration and looking more at the final ratios than the absolute
> numbers it is apparent that the difference is statistically insignificant,
> except on ctar_load.

What about dbench_load?
 
> Previously I had benchmark results on 1000Hz which showed preempt improved the
> results in a few of the loads. For my next experiment I will compare 100Hz
> with preempt to 100Hz without.

Cheers,
	Zwane
-- 
function.linuxpower.ca

Re: [BENCHMARK] 100Hz v 1000Hz with contest

[Nick Piggin]

Well thats nice, AS holds up OK...

Con Kolivas wrote:

>-----BEGIN PGP SIGNED MESSAGE-----
>Hash: SHA1
>
>I've attempted to answer the question does 1000Hz hurt responsiveness in 2.5 
>as much as I've found in 2.4; since subjectively the difference wasn't there 
>in 2.5. Using the same config with preempt enabled here are results from 
>2.5.70-mm3 set at default 1000Hz and at 100Hz (mm31):
>
>no_load:
>Kernel         [runs]   Time    CPU%    Loads   LCPU%   Ratio
>2.5.70-mm3          1   79      94.9    0.0     0.0     1.00
>2.5.70-mm31         1   77      94.8    0.0     0.0     1.00
>cacherun:
>Kernel         [runs]   Time    CPU%    Loads   LCPU%   Ratio
>2.5.70-mm3          1   76      97.4    0.0     0.0     0.96
>2.5.70-mm31         1   74      98.6    0.0     0.0     0.96
>process_load:
>Kernel         [runs]   Time    CPU%    Loads   LCPU%   Ratio
>2.5.70-mm3          2   108     68.5    64.5    28.7    1.37
>2.5.70-mm31         2   107     69.2    67.0    29.0    1.39
>ctar_load:
>Kernel         [runs]   Time    CPU%    Loads   LCPU%   Ratio
>2.5.70-mm3          3   114     70.2    1.0     5.3     1.44
>2.5.70-mm31         3   105     73.3    0.7     3.8     1.36
>xtar_load:
>Kernel         [runs]   Time    CPU%    Loads   LCPU%   Ratio
>2.5.70-mm3          3   123     62.6    2.3     5.7     1.56
>2.5.70-mm31         3   122     61.5    2.0     4.9     1.58
>io_load:
>Kernel         [runs]   Time    CPU%    Loads   LCPU%   Ratio
>2.5.70-mm3          4   116     66.4    40.6    18.8    1.47
>2.5.70-mm31         4   114     65.8    41.0    19.3    1.48
>io_other:
>Kernel         [runs]   Time    CPU%    Loads   LCPU%   Ratio
>2.5.70-mm3          2   116     66.4    50.0    22.2    1.47
>2.5.70-mm31         2   112     67.9    46.1    21.4    1.45
>read_load:
>Kernel         [runs]   Time    CPU%    Loads   LCPU%   Ratio
>2.5.70-mm3          2   104     75.0    8.2     5.8     1.32
>2.5.70-mm31         2   100     76.0    7.5     7.0     1.30
>list_load:
>Kernel         [runs]   Time    CPU%    Loads   LCPU%   Ratio
>2.5.70-mm3          2   95      80.0    0.0     7.4     1.20
>2.5.70-mm31         2   92      82.6    0.0     5.4     1.19
>mem_load:
>Kernel         [runs]   Time    CPU%    Loads   LCPU%   Ratio
>2.5.70-mm3          2   98      80.6    53.0    2.0     1.24
>2.5.70-mm31         2   95      81.1    53.0    2.1     1.23
>dbench_load:
>Kernel         [runs]   Time    CPU%    Loads   LCPU%   Ratio
>2.5.70-mm3          4   313     24.3    5.0     56.9    3.96
>2.5.70-mm31         4   297     24.9    4.5     52.5    3.86
>
>At first glance everything looks faster at 100Hz. However it is well known 
>that it will take slightly longer even with no load at 1000Hz. Taking that 
>into consideration and looking more at the final ratios than the absolute 
>numbers it is apparent that the difference is statistically insignificant, 
>except on ctar_load.
>
>Previously I had benchmark results on 1000Hz which showed preempt improved the 
>results in a few of the loads. For my next experiment I will compare 100Hz 
>with preempt to 100Hz without.
>
>Con
>-----BEGIN PGP SIGNATURE-----
>Version: GnuPG v1.2.1 (GNU/Linux)
>
>iD8DBQE+3BRIF6dfvkL3i1gRAnEbAKCpaj/kajzKV3qVrWGRIhOh+Q8O8gCfZp6c
>M3Iq1D/41t+4SB2jtNYQc48=
>=NMfC
>-----END PGP SIGNATURE-----
>
>-
>To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
>the body of a message to majordomo@vger.kernel.org
>More majordomo info at  http://vger.kernel.org/majordomo-info.html
>Please read the FAQ at  http://www.tux.org/lkml/
>
>  
>

RE: [BENCHMARK] 100Hz v 1000Hz with contest

[Giuliano Pochini]

On 03-Jun-2003 Con Kolivas wrote:
> I've attempted to answer the question does 1000Hz hurt responsiveness in 2.5
> as much as I've found in 2.4; since subjectively the difference wasn't there
> in 2.5. Using the same config with preempt enabled here are results from
> 2.5.70-mm3 set at default 1000Hz and at 100Hz (mm31):

Is there any problem using a frequency other than 100 and 1000Hz ?


Bye.

Re: [BENCHMARK] 100Hz v 1000Hz with contest

[Con Kolivas]

On Tue, 3 Jun 2003 18:00, Giuliano Pochini wrote:
> On 03-Jun-2003 Con Kolivas wrote:
> > I've attempted to answer the question does 1000Hz hurt responsiveness in
> > 2.5 as much as I've found in 2.4; since subjectively the difference
> > wasn't there in 2.5. Using the same config with preempt enabled here are
> > results from 2.5.70-mm3 set at default 1000Hz and at 100Hz (mm31):
>
> Is there any problem using a frequency other than 100 and 1000Hz ?

Not at all. These were chosen because they were the default 2.4 (100) and 2.5 
(1000) frequencies. The large difference in Hz was postulated to increase the 
in-kernel overhead and the amount of time spent tearing down and building up 
the cpu cache again. 2.4 running at 1000Hz shows poor performance at high 
(>4) loads whereas 2.5 doesn't seem to do this. I originally thought it was 
cache thrashing/trashing responsible. However since 2.5 performance is almost 
comparable at 100/1000 it seems to be that the pure interrupt overhead in 2.5 
is lower?

Con

Re: [BENCHMARK] 100Hz v 1000Hz with contest

[William Lee Irwin III]

On Tue, 3 Jun 2003 18:00, Giuliano Pochini wrote:
>> Is there any problem using a frequency other than 100 and 1000Hz ?

On Tue, Jun 03, 2003 at 08:36:49PM +1000, Con Kolivas wrote:
> Not at all. These were chosen because they were the default 2.4 (100) and 2.5 
> (1000) frequencies. The large difference in Hz was postulated to increase the 
> in-kernel overhead and the amount of time spent tearing down and building up 
> the cpu cache again. 2.4 running at 1000Hz shows poor performance at high 
> (>4) loads whereas 2.5 doesn't seem to do this. I originally thought it was 
> cache thrashing/trashing responsible. However since 2.5 performance is almost 
> comparable at 100/1000 it seems to be that the pure interrupt overhead in 2.5 
> is lower?

You could try profiling cache misses etc.

I blame count_active_tasks(). =)


-- wli