References of the abstract :
The Extended Discrete Element Method (XDEM) is a new and innovative numerical simulation technique that extends the classical Discrete Element Method (DEM) (which simulates the motion of granular material), by additional properties such as the chemical composition, thermodynamic state, and stress/strain for each particle. It has been successfully applied to numerous industries involving the processing of granular materials such as sand, rock, wood or coke. In this context, computational simulation with (X)DEM has become an increasingly essential tool for researchers and scientific engineers to set up and explore their experimental processes. However, increasing the size or accuracy of a model requires the use of High-Performance Computing (HPC) platforms with a parallelized implementation to accommodate
the growing needs in terms of memory and computation time.
At a large scale, load-balancing has a critical impact on performance. The load-balancing returned by the partitioning libraries can only be as good as the workload estimated on the load-balancing unit (i.e. the weight representing the computation cost on each node of the graph). For applications with complex algorithms and/or non-homogeneous computation distribution, this workload estimation can be difficult to do.
With this work, we have the following objectives:
- Optimise the implementation of the MPI communications in XDEM;
- Study the relationship between the workload estimation on each load-balancing unit (i.e., DEM or CFD cell);
- Study the quality of the load-balancing (e.g. returned by libraries like METIS, SCOTCH, Zoltan, ...) in relation to the actual performance;
- Provide better workload estimation with the final objective to improve the global performance.
In practice, XDEM simulations with millions of particles require significant computation time. It makes it difficult for engineers and industrial partners because they require a long simulated time. With this work, we want to break this barrier by addressing the main issues limiting the performance of XDEM on many thousands of cores: the performance of the MPI communication, in particular the packing/unpacking of data, and the quality of the load balancing.