MUMPS, solve and possible memory leaks

General Discussion
1

Dear all,

I am trying to parallelize a solution of a nonlinear time-dependent problem using FreeFEM++ 4.2.1 with the MUMPS package. At every time step I use “solve” in this way

load “MUMPS”
…
for (int m = 0; m<mmax; m++){
f=u1old^2+u2old^2;
solve PDE([u1,u2], [v1,v2], solver=sparsesolver,master=-1)=
int3d(th,mpirank)(u1 * v1/dt + u2 * v2/dt - u1old * v1/dt - u2old * v2/dt +f * u1 * v1+…)
…;
u1old=u1;
u2old=u2;
}

It works fine in the sense that the results are the same as for a non-parallel solver and there is a reasonable speed-up, but I have encountered memory leaks - something stays in the memory and after several loops, my laptop runs out of memory.

I think I use “solve” in the wrong way and the matrix or the solution stays in the memory. Is it so? Could you point out please, where I can find out about the master parameter?

I have tried to use “problem” in the same way and the results are the same (in the sense both the precision and memory leaks).

Thank you!

2

hi,
Have you solved the memory leak problem？ I meet the same trouble when I parallelize a linear problem with MUMPS package, and don’t know how to solve. If you have solved the problem, could you give me some advices what should I do .If you haven’t , maybe we could discuss together.

3

You can solve this problem by using the PETSc interface to MUMPS. See this post, replace -pc_type hypre by -pc_type lu (though for simple problems such as heat equations and such, you should definitely use hypre and not MUMPS).