Re: [boost] performance of a linear algebra/matrix library

Subject: Re: [boost] performance of a linear algebra/matrix library
From: Karsten Ahnert (karsten.ahnert_at_[hidden])
Date: 2010-05-08 09:19:13

On 05/07/2010 03:06 PM, DE wrote:
> on 06.05.2010 at 22:48
>  Kenny Riddile wrote :
>> If you mean a language-imposed, inherent abstraction penalty, then I
>> doubt it (certainly not 33%).  You can find out for yourself though:
> 
>> http://www.stepanovpapers.com/AbstractionPenaltyBenchmark.cpp
> 
> i've run the code for my everyday compiler (msvc80) and it greported
> no abstraction penalty
> 
> i've also tried icc11 and it showed minor abstraction penalty
> (mean = 1.17)
> 
> 
> accidentaly i tried to compile my code with icc11 and -- it's a
> miracle! -- the test showed no abstraction penalty
> 
> i believe the implementation in fact works just as i expected
> so i guess if one wants speed he should get a better compiler
> 
> for curious ones here are some results:
> 
> n       2       4       8       16      32      64      128     256     512     1024    2048
> msvc80  1.77    1.57    1.34    1.12    1.05    1.12    1.32    1.33    1.23    1.23    1.23
> icc11   1.53    1.32    1.26    1.55    2.17    1.07    1.02    0.98    1.00    1.03    1.02
> 
> n is matrices' dimensions
> numbers in the latter 2 lines are ratios of runtime of C++ code to
> that of C code (time_cpp/time_c)
> 'msvc80' denotes time ratio for msvc version
> 'icc11' denotes the one for icc version

Hi,

maybe this is a stupid question: Do you average the performance results?
When I run AbstractionPenaltyBenchmark.cpp the results differ from run
to run by ca. 30%, such that the one single result is not significant. I
usually average them over 100-1000 runs until they have converged.

Karsten