Parallel Deflated CG Method to Model
Groundwater Flow in a Layered Grid
Raju Ram
(COSSE student, double degree with
Friedrich-Alexander
University of Erlangen- Nurnberg)
Supervisor: Kees Vuik
Site of the project:
Deltares
Princetonlaan 6
3584 CB Utrecht
Supervisor Deltares: Jarno Verkaik
start of the project: October 2016
In March 2017 the
Interim Thesis
has appeared
and a
presentation
has been given.
The Master project has been finished in August 2017
by the completion of the
Masters Thesis
and a final
presentation
has been given.
For working address etc. we refer to our
alumnipage.
Summary of the master project:
At Deltares, we are developing large groundwater models for our national
and international customers (e.g. water boards, drinking water
companies, municipalities, etc.) in order to simulate effects of changes
by climate, such as drought, or effects induces by water managers, e.g.
increasing pumping wells for drinkwater. The simulation code we use is
MODFLOW, and this code is the worldwide standard for groundwater
computation.
MODFLOW is a
FORTRAN-77 code with extensions to FORTRAN-90 that originally dates from
1988 and is being developed by the United States Geological Survey
(USGS). The current core version is MODFLOW-2005. It solves the
groundwater equation for Darcy flow on a finite volume grid: if the
problem is non-linear (e.g. by Cauchy boundary conditions such as rivers
or drains) within each outer iteration a linear system is being solved.
Typically the outer iteration is done by Picard iteration and the inner
iteration by Conjugate Gradient. Currently, we are developing a new
MODFLOW package (module), called Parallel Krylov Solver package together
with the USGS and Utrecht University. Last year, we implemented PKS in
the new MODFLOW-USG code, that is a spin-off code from MODFLOW that is
being developed for unstructured grids. Subdomain partitioning is done
with the
METIS library in a
straightforward manner. Furthermore, MODFLOW-USG can also run for
structured grids and partitioning can be chosen as uniform in lateral
direction or with the Recursive Bisection Algorithm. The PKS currently
supports parallel MPI/OpenMP CG, BiCGSTAB and we are now working on a
parallel version of GMRES. It uses an Restrictive Additive Schwarz
overlapping parallel preconditioner, with a (preconditioning only)
ILU/ICC subdomain solve.
More details on the results for 2015 can be found in:
Parallel
Solver Report.
For this year we plan to solve issues with the BICGSTAB solver and the
rounding errors in the residual we are now encountering. We will also
focus more on overlapping computations. Instead of the Indonesia model,
we are also working on a parallel version of the global PCRGLOB model,
that will consist of 58 subdomains, each representing a hydrologic
stream region.
In the project the following steps can be done
-
Optimize the interface transmission conditions between the subdomains
for largely diagonal dominant groundwater flow, e.g. with optimized
Schwarz. Now we have a Dirichlet transmission.
-
Add a multi-level parallel preconditioner, e.g. coarse grid / deflation
-
Add an iterative subdomain solve, e.g. a stationary method or Krylov
subspace
-
Optimize METIS of alternative (parallel) graph partitioning algorithm
for partitioning of extremely large MODFLOW grids (> 1 billion cells)
-
Optimize the CG algorithms to minimize (global) MPI communication;
reordering updates
Contact information:
Kees
Vuik
Back to the
home page
or the
Master students page of Kees Vuik