DPM-HPC: Enhanced Design for High Performance Parallel Execution
This research proposes an extension of the previous work on Domain Decomposition which showed the benefit of parallel execution for the Dynamics module. The new design will allow parallel and distributed execution of coupled simulation modules of XDEM. The targeted execution platforms are clusters of multiprocessor computers.
The primary goal of this enhanced design is to provide:
- Parallel execution of coupled modules: the new design supports internal XDEM modules (like Dynamics and Conversion simulations) and external modules as well (like OpenFoam and Diffpack).
- Flexibility and Maintainability: Current simulation modules can be selected at runtime and the integration of new modules in the framework is made easy. A common codepath is used for parallel and sequential execution: the sequential execution is just seen as a special case of the parallel execution.
- High Performance distributed execution: Code and data structures are re-organized in order to provide the best performance. The MPI library offers efficient communications, portability and the ability to use high speed networks like InfiniBand.
In a second step, further improvements will be investigated. We will consider, among others, offloading part of the simulation onto GPUs, and using new dynamics load-balancing algorithms.
Orthogonal Recursive Bisection (ORB) partitioning according to different axes
- Prof. Pascal. Bouvry, Parallel Computing & Optimisation Group, University of Luxembourg
- Dr. Sébastien Varrette, Parallel Computing & Optimisation Group, University of Luxembourg