automatic installation of Metis

[Chebli Rezki]

Hello,

I have got a problem with the automatic installation with the Metis.

Part of the install_saturne.log file is attached.

I have got this error message:  [ Error during compile stage of Metis. ]

Thank you for your help.

Rezki

[Alexandre Douce]

Hello,

Metis is not mandatory, so I advice you to not use this prerequisite. Code_Saturne embedded an internal partitioner.

[Chebli Rezki]

Hi,

Thank's for your Help

But the installation is stopped and I cannot what make to pass this step because I used the automatic installation ?

Regards.

Rezki

[Chebli Rezki]

Hi,

I attach here the last version of my log file.

Regards.

Rezki

[Claus Andersen]

I have a solution to the metis problem, but you're not gonna like it or bother with it, because it will be much simpler to disable metis in the setup file where you have the installer script extracted:

Change this:

metis      None    yes   yes

to this:

metis      None    no   no

And heres the real solution, conveniently (for you!) in French

http://www.code-aster.org/forum2/viewto ... ?pid=19214

[Chebli Rezki]

It works.

Thank you very much Claus.

Regards

Rezki

[Laurent Malod-Panisset]

Hi
 
I would like to know what's the difference between using the embedded partitioner and other like Metis or Scotch.
Is it better? Do you have more control over it?
 
Regards

[Yvan Fournier]

Hello,

The embedded partitioner is based on the Morton space-filling curve and leads to slightly lower quality partitions than SCOTCH or METIS, and we have only run a few benchmarks, but it seems that calculations run with the embedded partitioner are usually 15 to 20% slower than using METIS or SCOTCH. Comparisons with parallel partitioning on a tetrahedral mesh which will be published in an upcoming article seem to place the embedded partitioner at the same level as ParMETIS (whose partition quality drops with the number of processors, while that of the embedded partitioner is deterministic and independent of the number of processors used to build it), below serial SCOTCH and METIS, with PT-SCOTCH in between.

In the future, I would like to switch the embedded partitioner from the Morton to the Peano-Hilbert curve, in which case we could expect partitioning quality closer to that of METIS, but I have not had the time to do this yet.

In any case, if the 15 to 20% performance difference is important to you, do not forget to set linear system convergence criteria from the default 10-8 to 10-5 (which will make a bigger difference), and I would recommend SCOTCH over METIS: the project is maintained (while METIS 5 is still in beta since 2007, and both METIS 4 and 5 need patching to compile on modern Linux distributions), and it uses a true free software licence, while METIS's licence allows you to build Code_Saturne with METIS for your own use, but does not allow us to redistribute it (we only provide a link). The METIS version from Code_Aster which Claus pointed to may be redistributed (at least by the Code_Aster team, as they have requested permission from the METIS developpers, though I did not check on the details to see if this also encompasses other EDF codes), but in any case, METIS is not maintained much, and its installation is not worth the bother when other equivalent (SCOTCH) or simpler (embedded partitioner) options are available.

Best regards,

  Yvan

[Laurent Malod-Panisset]

Thanks for the quick answer.
It is clear now.