two unit tests fail on Ubuntu 16.04

[pGarlick]

Hi,

I have just updated FEniCS and Oasis to 2016.2 and re-run the unit tests (cd tests; py.test). I have noticed that two of the tests that previously passed now fail.

The two failing tests are the ones that are skipped for Darwin installations; test_default_mpi_Coupled() and test_naive_mpi_Coupled(). FEniCS has been installed with HashDist and Oasis updated with git.

One difference between FEniCS 2016.1 and FEniCS 2016.2 that might be relevant is the set of linear solvers that is made available. In FEniCS 2016.1, installed with HashDist, MUMPS is included. In FEniCS 2016.2 it is not. Do the Oasis tests need MUMPS on Linux as well?

I have tried the following command to test:

$ mpirun -np 2 python NSCoupled.py problem=Cylinder testing=True

The process responds with:

Start Newton iterations flow
Iter 1, Error = 0.000415191268604

and then hangs, with both CPUs fully occupied.

In the solvers/NSCoupled/init.py file, where the solver is selected, I have changed line 46 to specify LUSolver('superlu_dist'). However, this appears to have no effect. Maybe superlu_dist is selected automatically by PETSc?

Also, along the way I noticed a typo in problems/Cylinder.py. In line 11 'Linus' should be 'Linux'.

Best,

Paul.

[mikaem]

Hi Paul

Looks like you don't have a working parallel LU solver? What does list_lu_solver_methods() tell you? superlu_dist should work in parallel.

Thanks for the typo, I'll fix it.

[pGarlick]

Hi Mikael,

Thank you.

I have checked the LU solver configuration. list_lu_solver_methods() gives:

LU method     |  Description
---------------------------------------------------------------------------
default       |  default LU solver
petsc         |  PETSc built in LU solver
superlu_dist  |  Parallel SuperLU
umfpack       |  UMFPACK (Unsymmetric MultiFrontal sparse LU factorization)

Looking at the dolfin code la/PETScLUSolver.cpp it appears that superlu_dist would be the chosen solver in a 2016.2 HashDist installation (with no MUMPS):

{
      #if PETSC_HAVE_MUMPS
      method = "mumps";
      #elif PETSC_HAVE_SUPERLU_DIST
      method = "superlu_dist";
      #elif PETSC_HAVE_PASTIX
      method = "pastix";
      #else
      dolfin_error("PETScLUSolver.cpp",
                   "solve linear system using PETSc LU solver",
                   "No suitable solver for parallel LU found. Consider configuring PETSc with MUMPS or SuperLU_dist");
      #endif
    }

Possibly the divergence seen in the unit tests is related to the version of superlu_dist? The version installed by HashDist is version 5.1.0.

One other test I have made was to try setting the "mesh_partitioner" parameter to "ParMETIS". However, this did not change the behaviour (divergence after the first iteration).

Paul.

[mikaem]

Ok,
Sorry I cannot reproduce. I'm using an Anaconda built Fenics and this case is working just fine on my Ubuntu 16.04. The working Fenics version is obtainable if you have Anaconda with conda install -c mikaem/labels/ubuntu1604-host-gcc fenics, and the scripts used to compile fenics are here: https://github.com/mikaem/docker-fenics-recipes. I'm not using hashdist anymore.