Abstract :
[en] Numerical models are commonly used to simulate or model physical processes such as weather forecasts, fluid action, rocket trajectory, building designs, or biomass combustion. These simulations are immensely complex and require a hefty amount of time and computation, making it impossible to run on a standard modern laptop in a reasonable and fair period. This research work targets large-scale and parallel simulations of DEM and DEM-CFD couplings using high-performance computing techniques and optimizations. This thesis aims to analyze, contribute, and apply the DEM approach using the XDEM multi-Physics toolbox to physical processes that have been reluctant to be used due to their required computational resources and time.
The first step of this work is to analyze and investigate the performance bottlenecks of the XDEM software. Therefore, the latter has been profiled and some critical parts as the contact detection were identified as the main bottlenecks of the software. A benchmark has also been set up to assess each bottleneck part’s performance using a baseline case. This step is crucial as it defines the general guidelines to follow in optimizing any application in general.
A complete framework has been developed from scratch and aims to test and compare several contact detection algorithms and implementations. The framework, which also has a parallel version, has been used to select an appropriate algorithm and implementation for the XDEM software. The link-cell approach, combined with a new Verlet list concept, proved to be the best option for significantly reducing the contact detection part’s computational time. The Verlet buffer concept developed during this thesis takes the particle flow regime into account when selecting the skin margin to enhance the algorithm’s efficiency further.
In order to target the high-performance computers for large-scale simulations, a full hybrid distributed-shared memory parallelization has been introduced by adding a fine-grain OpenMP implementation layer to the existing MPI approach. A shared memory parallelization allows taking full advantage of personal workstations with modern CPU architecture. On the other hand, a hybrid approach is one of the best ways to fully exploit the computing node capacities of our modern CPU clusters that mainly have a NUMA architecture. Macro-benchmarking performance analysis showed that we could entirely exploit 80% (speed-up) of 85 computing nodes representing 2380 cores on the ULHPC supercomputer.
Finally, a life-size biomass combustion furnace is developed and used as an application test to demonstrate the complex and heavy cases that the XDEM software can accommodate at this time. The furnace is the combustion chamber of a 16 MW geothermal steam super-heater, partoftheEnelGreenPower"Cornia2." powerplant located in Italy. It proves that DEM, in general, and XDEM in particular, can be used for real case applications that discourage users due to their complexity and especially the time required to deliver the outcome results.
