open TELEMAC-MASCARET The mathematically superior suite of solvers

Hello,

I am checking the correct installation of telemac-mascaret v6p3 on a cluster using gfortran, openMPI-1.8.4 and METIS 5.0.2 (both installed locally).

I precise I have both the v6p3 scalar and parallel modes successfully installed on a separate PC offline. I am therefore considering the results given by the PC as a reference (of course, results in scalar and parallel modes are perfectly consistent with each another on the PC).

On the cluster, the scalar mode seems to be ok as the results of T2D example cases perfectly match the reference. Now the parallel mode - which compiled successfully - seems to run normally but produces velocity fields which are slightly different from the reference (see attached picture for Telemac2d example case 'pildepon') so that scalar and parallel modes are not fully consistent with each other.

Why do I observe this difference?...
How could I fix the problem?...

Config file is also given in attachment (codes are launched interactively on the cluster at the moment for quick dev purposes, later QSUB commands for job queueing shall be added in the .cfg file).

Thank you for your help!
JBS.

Hello

This is normal to have slight differences between scalar and parallel results. this is due to the truncation error and to the non reproducible operations.
We are working the improvement of the reproducability of the Telemac system, however this is a tremendous task and will take very long time to be achieved.

with my best regards

Riadh

Thank you Riadh.

What do you mean exactly by "non reproducible" operations?
Is it meant that the discrepancies get worse by increasing the number of processors/subdomains?

Thank you again,
JBS.

Hello

No this is not that simple. There are some algebraic operations that are not always associative with respect to parallelisme. For example a+(b+c) and (a+b)+c will give different values when a,b and c are of very different magnitudes). This has a direct influence on the parallel assembly that we use in Telemac.
The difference is not linear with the number of CPUs, but this difference is in a great majority of cases negligeable. However, for some cases this differences can be very important (such as the van Karma vortex behind a bridge pier).

I hope that this helps

with my best regards

Riadh