perfomnace decease (scilab 5.4.0-beta3)

classic Classic list List threaded Threaded
6 messages Options
kjubo kjubo
Reply | Threaded
Open this post in threaded view
|

perfomnace decease (scilab 5.4.0-beta3)

hello,

im testing performance of scilab with my own file (heat transfer computation),
which consist of for cycles and if conditions. i found that scilab 5.4(beta3) runs 60% slower than other scilab version. have you any explanation to this behavior?
(im testing on the same hardware every time, windows OS)
here are my results, attached exec file.

thanks

Jakub Kopáč



results:
//SCILAB BENCHMARK
// heat transfer 2D
//scilab 4.1.2 272sek
//scilab 5.2.2 320sek
//scilab 5.3.0-beta1 295sec
//scilab 5.3.0-beta2 293sec
//scilab 5.3.0-beta3 296sec
//scilab 5.3.0-SSE3  286sec
//scilab 5.3.1       302sec
//scilab 5.3.2-64bit 265sec
//scilab 5.4.0-beta3 483sec

bench_prestup_tepla.sce
Mathieu Dubois Mathieu Dubois
Reply | Threaded
Open this post in threaded view
|

Re: perfomnace decease (scilab 5.4.0-beta3)

Hello,

Maybe this is related to the option used to compile the beta version
(like enabling debugging, less optimization, etc.).

You could try to compile it yourself from the source...

HTH,
Mathieu

Le 01/09/2012 14:18, kjubo a écrit :

> hello,
>
> im testing performance of scilab with my own file (heat transfer
> computation),
> which consist of for cycles and if conditions. i found that scilab
> 5.4(beta3) runs 60% slower than other scilab version. have you any
> explanation to this behavior?
> (im testing on the same hardware every time, windows OS)
> here are my results, attached exec file.
>
> thanks
>
> Jakub Kopáč
>
>
>
> results:
> //SCILAB BENCHMARK
> // heat transfer 2D
> //scilab 4.1.2 272sek
> //scilab 5.2.2 320sek
> //scilab 5.3.0-beta1 295sec
> //scilab 5.3.0-beta2 293sec
> //scilab 5.3.0-beta3 296sec
> //scilab 5.3.0-SSE3  286sec
> //scilab 5.3.1       302sec
> //scilab 5.3.2-64bit 265sec
> //scilab 5.4.0-beta3 483sec
>
> http://mailinglists.scilab.org/file/n4024765/bench_prestup_tepla.sce
> bench_prestup_tepla.sce
>
>
>
> --
> View this message in context: http://mailinglists.scilab.org/perfomnace-decease-scilab-5-4-0-beta3-tp4024765.html
> Sent from the Scilab users - Mailing Lists Archives mailing list archive at Nabble.com.
> _______________________________________________
> users mailing list
> [hidden email]
> http://lists.scilab.org/mailman/listinfo/users
>
_______________________________________________
users mailing list
[hidden email]
http://lists.scilab.org/mailman/listinfo/users
michael.baudin michael.baudin
Reply | Threaded
Open this post in threaded view
|

Re: perfomnace decease (scilab 5.4.0-beta3)

In reply to this post by kjubo
Hi,

My message does not explain the performance differences between
Scilab versions, but may solve your actual performance
problem.

It seems that your script has 3 nested loops to
perform the computation.
So, whatever the performance of Scilab, your script
will be slow.
And, indeed, 272 sec is quite slow...
What you may consider is to vectorize your code.

As an example of a fast 2D implementation for the
solution of a PDE, you may try the scibench
module from ATOMS:

http://atoms.scilab.org/toolboxes/scibench

Look at the scibench_poisson function, which
solves the Poisson equation.
You may find the code in the Source tab of :

http://forge.scilab.org/index.php/p/scibench/

The main trick here is to compute all the
matrix in just one statement, based on the
Kronecker product:

http://forge.scilab.org/index.php/p/scibench/source/tree/HEAD/macros/scibench_poissonA.sci

For iterative methods, we just need the A*u product, and
this is done here :

http://forge.scilab.org/index.php/p/scibench/source/tree/HEAD/macros/scibench_poissonAu.sci

The performance difference is huge, as shown here :

http://wiki.scilab.org/Solving%20Poisson%20PDE%20with%20Sparse%20Matrices

Best regards,

Michaël

Le 2012-09-01 14:18, kjubo a écrit :

> hello,
>
> im testing performance of scilab with my own file (heat transfer
> computation),
> which consist of for cycles and if conditions. i found that scilab
> 5.4(beta3) runs 60% slower than other scilab version. have you any
> explanation to this behavior?
> (im testing on the same hardware every time, windows OS)
> here are my results, attached exec file.
>
> thanks
>
> Jakub Kopáč
>
>
>
> results:
> //SCILAB BENCHMARK
> // heat transfer 2D
> //scilab 4.1.2 272sek
> //scilab 5.2.2 320sek
> //scilab 5.3.0-beta1 295sec
> //scilab 5.3.0-beta2 293sec
> //scilab 5.3.0-beta3 296sec
> //scilab 5.3.0-SSE3  286sec
> //scilab 5.3.1       302sec
> //scilab 5.3.2-64bit 265sec
> //scilab 5.4.0-beta3 483sec
>
> http://mailinglists.scilab.org/file/n4024765/bench_prestup_tepla.sce
> bench_prestup_tepla.sce
>
>
>
> --
> View this message in context:
>
> http://mailinglists.scilab.org/perfomnace-decease-scilab-5-4-0-beta3-tp4024765.html
> Sent from the Scilab users - Mailing Lists Archives mailing list
> archive at Nabble.com.
> _______________________________________________
> users mailing list
> [hidden email]
> http://lists.scilab.org/mailman/listinfo/users
_______________________________________________
users mailing list
[hidden email]
http://lists.scilab.org/mailman/listinfo/users
Sylvestre Ledru-4 Sylvestre Ledru-4
Reply | Threaded
Open this post in threaded view
|

Re: perfomnace decease (scilab 5.4.0-beta3)

In reply to this post by Mathieu Dubois
No, it is probably not the source of the issue.
You should report a bug about your problem.

Sylvestre


Le 02/09/2012 19:57, Mathieu Dubois a écrit :

> Hello,
>
> Maybe this is related to the option used to compile the beta version
> (like enabling debugging, less optimization, etc.).
>
> You could try to compile it yourself from the source...
>
> HTH,
> Mathieu
>
> Le 01/09/2012 14:18, kjubo a écrit :
>> hello,
>>
>> im testing performance of scilab with my own file (heat transfer
>> computation),
>> which consist of for cycles and if conditions. i found that scilab
>> 5.4(beta3) runs 60% slower than other scilab version. have you any
>> explanation to this behavior?
>> (im testing on the same hardware every time, windows OS)
>> here are my results, attached exec file.
>>
>> thanks
>>
>> Jakub Kopáč
>>
>>
>>
>> results:
>> //SCILAB BENCHMARK
>> // heat transfer 2D
>> //scilab 4.1.2 272sek
>> //scilab 5.2.2 320sek
>> //scilab 5.3.0-beta1 295sec
>> //scilab 5.3.0-beta2 293sec
>> //scilab 5.3.0-beta3 296sec
>> //scilab 5.3.0-SSE3  286sec
>> //scilab 5.3.1       302sec
>> //scilab 5.3.2-64bit 265sec
>> //scilab 5.4.0-beta3 483sec
>>
>> http://mailinglists.scilab.org/file/n4024765/bench_prestup_tepla.sce
>> bench_prestup_tepla.sce
>>
>>
>>
>> --
>> View this message in context:
>> http://mailinglists.scilab.org/perfomnace-decease-scilab-5-4-0-beta3-tp4024765.html
>> Sent from the Scilab users - Mailing Lists Archives mailing list
>> archive at Nabble.com.
>> _______________________________________________
>> users mailing list
>> [hidden email]
>> http://lists.scilab.org/mailman/listinfo/users
>>
> _______________________________________________
> users mailing list
> [hidden email]
> http://lists.scilab.org/mailman/listinfo/users


--
Sylvestre Ledru
Operations Manager / Community Manager
-----------------------------------------------------------
Scilab Enterprises
143bis rue Yves Le Coz - 78000 Versailles, France
http://www.scilab-enterprises.com


_______________________________________________
users mailing list
[hidden email]
http://lists.scilab.org/mailman/listinfo/users
kjubo kjubo
Reply | Threaded
Open this post in threaded view
|

Re: perfomnace decease (scilab 5.4.0-beta3)

In reply to this post by michael.baudin
This file is just for performance change measurement between different Scilab releases.
Is my simple benchmark, it not used to "serious" computations.
BTW: There are some algorithms, where you can not avoid using many for loops, for example: http://en.wikipedia.org/wiki/Non-Random_Two_Liquid_model , state General equations.

But thanks for your reply.
kjubo kjubo
Reply | Threaded
Open this post in threaded view
|

Re: perfomnace decease (scilab 5.4.0-beta3)

In reply to this post by Sylvestre Ledru-4
it is done.