Modeling and Information Processing

Parallel algorithms and parallelization of codes

Objective

Our activity is directed towards designing parallel algorithms and the parallelization of existing codes in all the numerical domains for large data sets: linear algebra, aerodynamics, structures, acoustics, electromagnetism, materials, etc.

The applications are programmed for any number of processors both on distributed and shared memory systems. In the case of the parallelization of an existing code, we have often found that we need to use or imagine a different algorithm or propose a combination of algorithms to increase the degree of parallelism.

Onera's contribution

We can mention some work in linear algebra for very large matrices:

• creation of an eigenvalue code, SPAI preconditionners, circulating matrix solver,
• library of parallel programs for resolving large linear systems.

And in the more specific fields, in collaboration with the specialists in the fields concerned:

• Programming scheme for explicit 3D finite difference problems.
• Theoretical and implementation study on a processor grid for a binary fluid mechanics method.
• Dynamic structure mechanics: structures subject to arbitrary forces and damping, with real time parallel coupling display.
• Electromagnetism: Wavelet based radar hyperimages, waveguides, Maxwell using finite differences (2D, 3D).
• Parallelization and implementation of particle trajectory simulation in an accelerator and real time display.
• Parallelization of a superalloy recrystallization code.
• Dynamic task balancing for distributed heterogeneous systems.

Currently we are interested in numerical coupling strategies for iterative solver based FEM-BEM codes.

Last Update: 11 October 2006 - © ONERA 2009 - Terms of use