Parallel calculations
Zebulon has an iterative sub-domain parallel solver which have been
proven effective and scalable for non-linear problems. Sub-domains are
used in Zebulon, each having a small inverted matrix, and their joint
contributions are solved using a conditioned iterative FETI (Finite
Element Tearing and Interconnecting) approach. Even on single processor
machines, this method has value because it is a good compromise between
a full implicit solution with stiffness inverse, and a pure iterative
solver (which is likely to diverge in non-linear solutions).
- PVM or MPI interface The parallel solver in Zebulon uses
both standard interfaces for interprocess communication. These
interfaces abstract the hardware or machine dependent issues so the
calcul ation can run in particularly inhomogeneous environments. This
means that a single processor machine can act like a
multiprocessor machine, and a network of dissimilar computers (even
remotely located) can be used as a parallel computer.
- Advanced solver The solver uses a preconditioned
multi-domain FETI solver to handle the parallel matrix inversion. Tools
are provided to automatically divide a mesh into well shaped,
uniform size domains. This solver has been proven effective for strongly
non-linear problems in viscoplasticity with thermal loading cycles.
- Scalability The nature of the domain method with a
FETI solver gives reduced message passing requirements between
processes, which is normally the limiting factor for large scale
parallelism. The method therefore remains efficient up to several
dozen processors.
- Local level parallelism For problems where the material
law integration is more costly than the global solution (small meshes
with a very large number of state variables) parallel calculations
can be applied to the local integration. This has allowed interesting
results using even separate finite element models for each Gauss
point in a structure (FE inside FE)!
|
Up: NW Numerics Software
Back: User additions
Next: Input files