Parallel solvers using petsc for nonlinear conduction problems in higher dimensions
Loading...
Date
2018-05
Authors
Muhammad Hisyamuddin Rosli
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
The Portable, Extensible, Toolkit for Scientific Computation (PETSc) library
package is a popular computational tool among researchers that provides the
capability of implementing parallel schemes in solving various kinds of finite element
analysis through the use of its many routines and libraries that can be manipulated
through runtime options. PETSc, as is claimed, allows a migration of existing
scientific code from a sequential structure into a scalable, parallel paradigm structure
in a convenient manner.
The main focus of this project is to evaluate the parallel performance of finite
element code through the use of PETSc. Performance comparisons between different
types of parallel solver are also done in this study. Different levels of parallel
optimisation are also studied. The test platform consists of an Intel Xeon X5650
processor which has 6 physical cores (12 logical cores) and 24GB RAM and the FEM
code is built with GCC compiler version 4.4 and OpenMPI 3.0.
The results from this project show a considerably good parallel performance
compared to the serial performance. For instance, most of the parallel computation
simulations could gain at least 3 times speedup compared to the serial computation.
Apart from this, the implementation of build-time optimisation option does not
necessarily increase the parallel performance of the code. In this study, the use of
different parallel solvers, in this case, the incomplete LU (ILU) and multigrid (MG)
solvers, also shows difference in the parallel performance.
In brief, this work shows that the performance of the parallel computation is
considerably better than serial computation performance. However, there is much
room for improvement of the scalability of the parallel computation speedups since
the most speedups start to scale poorly when more number of cores used. Moreover,
this work hints that some tunings of the compiler should be taken to enhance the
optimised versions of the parallel solver.