open TELEMAC-MASCARET The mathematically superior suite of solvers

an other link much better

mega.co.nz/#!DhkBFRKJ!urGWGB98bvuHealNhn...m1TqRL-Y8TCN5Q-f8hrY

Hello,
I have found that the problem occured when "CONVECTION DE U ET DE V" was enabled in my steering file. (telemac2d)

But i have also this kind of problem with tomawac.
I have this RAM problem for this two tomawac testing case :

• bottom_friction
• whirl_current

But its working well with the testing case "fetch_litmited".
If someone have some ideas.
Do you thinks its a steering file problem ? I keep it all from the cas test. But only change the mpi processes number .


Thanks for help.

Hello,

Well it is a problem that occurs in very few installations (never seen here at EDF, but seen by Artelia), linked with the use of characteristics (probably triggered when you add CONVECTION DE U ET V), with heavy use of MPI, but we do not know if it comes from Telemac side or MPI side. It is also according to some posts machine dependent (with the same compiler). Normally from Telemac side the maximum of memory must be allocated after one time step, unless we have somewhere a hidden automatic allocation (e.g. in a subroutine : DOUBLE PRECISION TOTO(NPOIN) with NPOIN an argument and TOTO not an argument, it requires then a somewhat bugged compiler that does not deallocates TOTO before returning to make a problem). Normally there is no automatic allocation in Telemac, for this very reason, unless a user does it in a personal Fortran.

This is where we are... any other hint welcome,

Regards,

Jean-Michel Hervouet

Hello JmHervouet,

May i know which kind of compiler and MPI do you use ?

Thanks.

Hello,
We found the incriminated parameter for tomawac in the steering file :

"PRISE EN COMPTE DE LA PROPAGATION"

When this parameter was enabled we have a memory leak.
Maybe its related with the same subroutine for CONVECTION DE U ET DE V ?

Thanks.

Hello,

No, it is always the same reason, this parameter triggers the call to characteristics, and it is when we have characteristics that this problem occurs.

Regards,

JMH

Hello Jean-Michel,

I'm working with Fabien Valentin.

As you say, we have two parts to check : the Telemac side and the MPI side.
About the Telemac side, which investigation could be appropriate to solve the issue? In other terms, which subroutines are specific to the use of characteristics? Is there a check of the memory allocation in this subroutines? It is not easy to know where and what looking for...

The solution to deactivate CONVECTION DE U ET DE V in Telemac2d and PRISE EN COMPTE DE LA PROPAGATION in Tomawac allows the computation but it is not satisfying about the physic of the computation...

Regards

Hello Nicolas,

All the characteristics subroutines that call MPI are in the module streamline in BIEF. If you look at the allocations in it, you will see that they are done once, unless the routines are called a second time for a larger problem, typically advection on all points of a 2D domain after advection of boundary points for Thompson boundary conditions. This is why after one time step there should not be any new allocation. This could be checked by changing the "call allocate(..." in streamline by a "call allocate2(..." where allocate2 calls allocate and prints a message. After one time step we should not see any message. Anyway this should not create a leak because we deallocate also. Then we have the automatic arrays like TOTO(NPOIN) where TOTO is not an argument. For this I do not know how to find them but officially we have none (maybe a program that analyses the Fortran more in depth, I'll ask if we have this here).

regards,

JMH

Hi Jean-Michel

The probably best tool for that should be VALGRIND.
I attended to install it but it's not really easy as you have to run in parallel. I don't have enough time to investigate this but maybe in a near future ...

Regards

Hello Christophe,

This one has been tested but says nothing on automatic arrays apparently (or we have really none...).

Regards,

JM