You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
I have observed errors triggered by the combined use of GridapPETSc and GridapGmsh as it is done in the expansion driver. Eliminating the line using GridapGmsh and the line include("Expansion.jl") in GridapMHD.jl, these errors do not arise (none of them).
The way the errors manifest depend on how GridapPETSc is installed. I considered two options
choosing a 32 bit PETSc installation (configured without --with-64-bit-indices)
choosing a 64 bit PETSc installation (configured with --with-64-bit-indices)
When using GridapGmsh a dependency on Metis, which is compiled with 32 bits (see here) is introduced. It is possible to bypass the call to Metis from MUMPS, which can be done through the option "-mat_mumps_icntl_7 0" in PETSc arguments that selects using AMD ordering. Therefore, I consider the following commands
julia --project=. -e ' using GridapMHD; GridapMHD.hunt(nc=(10,10),np=(1,1),backend=:sequential,solver=:petsc,petsc_options="-snes_monitor -ksp_error_if_not_converged true -ksp_converged_reason -ksp_type preonly -pc_type lu -pc_factor_mat_solver_type mumps")'
julia --project=. -e ' using GridapMHD; GridapMHD.hunt(nc=(10,10),np=(1,1),backend=:sequential,solver=:petsc,petsc_options="-snes_monitor -ksp_error_if_not_converged true -ksp_converged_reason -ksp_type preonly -pc_type lu -pc_factor_mat_solver_type mumps -mat_mumps_icntl_7 0")'
mpiexec -n 1 julia --project=. -e 'using GridapMHD; GridapMHD.hunt(np=(1,1),nc=(10,10),backend=:mpi,solver=:petsc,petsc_options="-snes_monitor -ksp_error_if_not_converged true -ksp_converged_reason -ksp_type preonly -pc_type lu -pc_factor_mat_solver_type mumps")'
mpiexec -n 1 julia --project=. -e 'using GridapMHD; GridapMHD.hunt(np=(1,1),nc=(10,10),backend=:mpi,solver=:petsc,petsc_options="-snes_monitor -ksp_error_if_not_converged true -ksp_converged_reason -ksp_type preonly -pc_type lu -pc_factor_mat_solver_type mumps -mat_mumps_icntl_7 0")'
The behavior of these four commands under the three installations mentioned above is summarized in the following table, that points to the errors detailed below (which are very similar, always crashing at mumps.c:1686 with different info values though).
I have observed errors triggered by the combined use of GridapPETSc and GridapGmsh as it is done in the expansion driver. Eliminating the line
using GridapGmsh
and the lineinclude("Expansion.jl")
in GridapMHD.jl, these errors do not arise (none of them).The way the errors manifest depend on how GridapPETSc is installed. I considered two options
--with-64-bit-indices
)--with-64-bit-indices
)When using GridapGmsh a dependency on Metis, which is compiled with 32 bits (see here) is introduced. It is possible to bypass the call to Metis from MUMPS, which can be done through the option "-mat_mumps_icntl_7 0" in PETSc arguments that selects using AMD ordering. Therefore, I consider the following commands
julia --project=. -e ' using GridapMHD; GridapMHD.hunt(nc=(10,10),np=(1,1),backend=:sequential,solver=:petsc,petsc_options="-snes_monitor -ksp_error_if_not_converged true -ksp_converged_reason -ksp_type preonly -pc_type lu -pc_factor_mat_solver_type mumps")'
julia --project=. -e ' using GridapMHD; GridapMHD.hunt(nc=(10,10),np=(1,1),backend=:sequential,solver=:petsc,petsc_options="-snes_monitor -ksp_error_if_not_converged true -ksp_converged_reason -ksp_type preonly -pc_type lu -pc_factor_mat_solver_type mumps -mat_mumps_icntl_7 0")'
mpiexec -n 1 julia --project=. -e 'using GridapMHD; GridapMHD.hunt(np=(1,1),nc=(10,10),backend=:mpi,solver=:petsc,petsc_options="-snes_monitor -ksp_error_if_not_converged true -ksp_converged_reason -ksp_type preonly -pc_type lu -pc_factor_mat_solver_type mumps")'
mpiexec -n 1 julia --project=. -e 'using GridapMHD; GridapMHD.hunt(np=(1,1),nc=(10,10),backend=:mpi,solver=:petsc,petsc_options="-snes_monitor -ksp_error_if_not_converged true -ksp_converged_reason -ksp_type preonly -pc_type lu -pc_factor_mat_solver_type mumps -mat_mumps_icntl_7 0")'
The behavior of these four commands under the three installations mentioned above is summarized in the following table, that points to the errors detailed below (which are very similar, always crashing at mumps.c:1686 with different info values though).
E1
E2
E3
The text was updated successfully, but these errors were encountered: