Performance of parallel MUMPS solver

General
1

I have done some benchmarking of different MUMPS versions available on debian

Comparison between libmumps-5.7 and libmumps-ptscotch-5.7 packages, configuring getfem build respectively with

LIBS='-lmumps_common -lpord' --enable-paralevel=2 '--with-mumps=-lsmumps -ldmumps -lcmumps -lzmumps' -disable-openmp

and

LIBS='-lmumps_common_ptscotch -lpord_ptscotch' --enable-paralevel=2 '--with-mumps=-lsmumps_ptscotch -ldmumps_ptscotch -lcmumps_ptscotch -lzmumps_ptscotch' -disable-openmp

The produced python interface libraries link to the following shared libraries (right side with ptscotch)

image
image1337×892 179 KB

Running a benchmark model with approx 250000 dofs gives the following timings (right side with ptscotch is much faster)

image
image1113×687 86.4 KB

2

Does the version of Mumps without the Scotch library use Metis or some internal reordering algorithm?

3

my post refers to testing of debian mumps packages. The exact build options for these packages. with alternative ordering algorithms can be found here:

Still, at runtime, one can change blas/lapack version in debian with the update-alternatives --config ... command, which is another factor to take intro account when benchmarking on debian.