Memory leak in PETSc Hypre PILUT preconditioner

[mikekryjak]

I'm running Hermes-3 in 2D and trying to optimise PETSc preconditioners, focusing on the ILU methods Euclid and PILUT. Ignoring the fact both are now deprecated in Hypre, I found that PILUT is giving me a significant memory leak where the more solver steps there are, the more the memory increases. I've had a 10 hour simulation grow from 1GB to 25GB. I have tested Euclid and CVODE for the same case and they don't grow their memory usage.

Here is my input deck:

[solver]
type = beuler # Backward Euler steady-state solver
snes_type = newtonls # Nonlinear solver
ksp_type = gmres # Linear solver
max_nonlinear_iterations = 10
pc_type = hypre # Preconditioner type
pc_hypre_type = pilut # Hypre preconditioner type
lag_jacobian = 500 # Iterations between jacobian recalculations
atol = 1e-12 # Absolute tolerance
rtol = 1e-8 # Relative tolerance

[petsc]
pc_hypre_pilut_tol = 1e-7 # Default: 1e-4
options_view # All picked up options
options_left # Unused/unrecognised options
log_view # Performance info

[johnomotani]

Does the same thing happen with the logging disabled (i.e. remove log_view option)? Just trying to rule out one possible cause.

[mikekryjak]

@johnomotani yes, I just did the test to make sure.