Hi!
I'm implementing a generalized eigenvalue problem in libmesh with SLEPc,
and noticed a problem when running with certain numbers of MPI nodes. I
would get some erroneous eigenvalues in those cases. This seems to be
independent of PETSc/SLEPc settings, because I have reproduced the
problem with several different solver settings on two different
computers, as well as by exporting the eigenproblem system matrices to
Matlab format and used Matlab's eig().
After simplifying to a minimal test case, I have ended up with something
which essentially is the Poisson eigenvalue problem in a 3D box with
homogeneous Dirichlet boundary conditions. So the question is: why does
this program return different eigenvalues when the number of MPI nodes N
is 4 and 8?:
Here are the seven lowest eigenvalues when I have run the program on
different number of MPI nodes (N):
N=1: 32.9925 73.0811 73.0811 73.0811 120.25 120.25 120.25
N=2: 32.9925 73.0811 73.0811 73.0811 120.25 120.25 120.25
N=3: 32.9925 73.0811 73.0811 73.0811 120.25 120.25 120.25
N=4: 32.2965 66.4538 73.0811 73.0811 106.261 120.25 120.25 <----
Different values, and broken degeneracy among the values at 73.0811
N=5: 32.9925 73.0811 73.0811 73.0811 120.25 120.25 120.25
N=6: 32.9925 73.0811 73.0811 73.0811 120.25 120.25 120.25
N=7: 32.9925 73.0811 73.0811 73.0811 120.25 120.25 120.25
N=8: 30.8978 60.2595 67.7427 72.9946 96.2636 106.388 119.182 <----
Different values, and broken degeneracy again
The degeneracy I'm referring to should be there as a consequence of the
symmetry in the libmesh mesh generation on the 3D cube. I have attached
the program that produces these results when run as follows (example
with 3 MPI nodes):
mpiexec -n 4 ./bug.exe
I'm using a new GIT version of LibMesh, with complex numbers enabled.
PETSc and SLEPc are the newest versions from the maint-branch, built
with complex number support.
Best regards,
Torquil Sørensen
------------------------------------------------------------------------------
Transform Data into Opportunity.
Accelerate data analysis in your applications with
Intel Data Analytics Acceleration Library.
Click to learn more.
http://pubads.g.doubleclick.net/gampad/clk?id=278785231&iu=/4140
_______________________________________________
Libmesh-users mailing list
Libmesh-users@lists.sourceforge.net
https://lists.sourceforge.net/lists/listinfo/libmesh-users
