open TELEMAC-MASCARET The mathematically superior suite of solvers

Hello.

I want to ask whether is it possible to use ParMETIS instead of METIS for parallel computing with TELEMAC2D package. I've tried to do that on a 64-bit system but with no success.

Thanks in advance.

Hi Budo

It seems to me that METIS is only used by PARTEL program for the domain decomposition.
If the problem comes from 64 bit version, you should find or generate a 64 bit METIS library.

As I know, ParMETIS will allow you to manage a domain decomposition in parallel but I think PARTEL is not up to date to do that ...

Regards

The METIS library are used by the PARTEL mesh partitioning program - included in the TELEMAC system - to split the mesh into subdomains. After that, the PARTEL program builds connectivity tables, computes interfaces nodes, applies boundary conditions for each subdomain...
The METIS computing time is usually less than 10% of the PARTEL global computing time.

In the V6P0 version, the PARTEL program is sequential. In a next version, the PARTEL program will be parallel but the domain decomposition will be still performed by METIS.
Why ? Because the quality of the domain decomposition is better with the METIS library than with the ParMETIS library.
In my opinion, there will be a need to use the ParMETIS library only to split a very huge mesh inducing a METIS failure. By the way, it would be better to use the PT-Scotch library which seems to be more powerfull than the ParMetis one ...

Dear Denis,

The ParMETIS was just one option for mesh partition. I was not aware of PT-Schotch library until yesterday when I dived into the Partel source code. The reason for my transition from METIS is its lack of internal support for 64bit architecture. There is an alpha version of 5^th revision but it has some issues when compiled under Windows.

Best regards.
Budo

Hello,

I would risk to add a few comments:

(1) Metis library inside of ParMetis 3.1.1 is not better than the "good old" Metis 4.0 library, from my particular point of view the only advantage it can deal with heterogeneous meshes, i.e. e.g. triangles and quadrangles as elements; but Telemac uses homogeneous triangles, of course, so this is no advantage here. And it tends to deliver discontinuous partitions, not bad, but ugly...

(2) Programs linked with ParMetis "as is" would be parallel MPI-Programs to be executed in the proper MPI environment even when they use one processor. Removing MPI from ParMetis would be quite a piece of work!

(3) ParMetis would be interesting when it comes to partition of really huge meshes - when the partitioner (Partel) is a parallel program itself... However, presently the most execution time is burnt elsewhere but in calling Metis, which is very performant...

(4) You might try to use the newer Metis 5.0pre2, but beware that some data structures in this new release are not flat simple C arrays/fields which can be easily called as routine parameters from Fortran; you might require proper ISO C-bindings Fortran-2003 style (C_PTR, C_INT, etc.) in irder to success in calling Metis C routines from Fortran - welcome to the new brave world of mixed language programming! (In Linux it goes also without C-bindings.) You might also require to modify the function interfaces in Metis for linking with Fortran - all that large/small letter sensitivities, undescores...

(5) Summary: Use "good old" Metis 4.0 or prepare for a battle!

Best Regards,
jaj

One quick fix, if you have problem with METIS (i.e. PARTEL) in Windows 64bit, is ...

Compile your PARTEL (and only this one, separately, as a 32bit windows executable, for which you can download the 32bit METIS library). Then compile the rest of your system as a 64bit windows.

That should work since PARTEL is called independently of TELEMAC.

I will try to compile METIS under windows with g95 and intel shortly - I just did it under Ubuntu gfortran.

Hope this helps.
Sebastien.