The numerical simulation and HPCaprofessions have benefited significantly from technological advances in the field of processors, particularly multi-core architectures, in recent decades.

The parallelization methods employed to enhance the performance of numerical calculation are useful but their efficiency decreases as the degree increases. To overcome this limitation, IFPEN’s researchers have developed pragmatic approaches.

Two examples, implemented on IFPEN’s Grains3D and PeliGRIFF computational codes respectively, illustrate their approach:

• The first consisted in developing specific methods (for example, applied to domain decomposition techniquesb) aimed at optimizing the “efficiency” of HPC calculations, thereby avoiding the additional costs associated with inter-processor communications. For the biggest system studiedc, the performance obtained for 768 cores(1) (figure) was around 91% of the ideal case.

Science 33
Grains3Dd parallel computing performances in fluidized bed calculations.


• A second method was based on a multiscale strategy(2), consisting in modeling and solving physical problems on a small scale, before transferring data filtered using statistical approaches to larger scales via cascading.

This approach was successfully used in 2015 within the context of the ANR’s MORE4LESS collaborative project, dedicated to the multiscale modeling of reactive particle flows.

a - High Performance Computing.
b - Separation into coupled sub-problems, defined for smaller domains forming a division of the global domain.
c - Around 230 million fluidized particles.
d - Standardized for a complete knot of 16 cores.

(1) A. D. Rakotonirina, A. Wachs, Powder Technology, 2015, 154-172.
    DOI : 10.1016/j.powtec.2017.10.033

(2) A. Esteghamatian, F. Euzenat, A. Hammouti, M. Lance, A. Wachs, International Journal of Multiphase Flow.
    DOI :10.1016/j.ijmultiphaseflow.2017.11.003


Scientific contact: abdelkader.hammouti@ifpen.fr