Evaluation of Different Parallelization Strategies on a Cluster for Parametric Sweep in EC_NDT

L.Santandrea 1
1Laboratoire de Génie ́ Electrique et ́ Electronique de Paris, CNRS, Gif sur Yvette, France
发布日期2023

As part of its research, the Group of Electrical Engineering Paris (GeePs) uses the COMSOL Multiphysics® software for applications in an Electrical Engineering context. At present, a multitude of problems, such as artificial intelligence and the database generation that goes with it, metamodel construction, uncertainty quantification or inverse methods, require numerical simulations with parametric sweeps. These involve running the same physical model with different sets of parameters, which may be physical or geometric. Our study concerns the identification of 2D mechanical stresses in ferromagnetic materials using an eddy current non-destructive testing (EC-NDT) method. The inversion problem consists in identifying the results of impedance measurements on maps obtained by numerical simulation. COMSOL Multiphysics® will enable us to generate these impedance maps as a function of the stress state along x and y and the orientation of the sensor in this plane. It has been established that for ferromagnetic materials, the magnetic permeability of the magnetic field also depends on the stress state. To take account of this magneto-elastic behavior, a multi-scale model is used. For each state of mechanical stress in XY, this model returns a permeability tensor for the material, which is then entered into the software. Our numerical model consists in a 3D magnetodynamic frequency (AC/DC Module - MF model). For its resolution, we have the option of using the Ruche cluster in the “Mésocentre” computing center of Paris-Saclay, on a local level, which will enable us to speed up our calculations. It's not a question of parallelizing problems with millions of degrees of freedom or numerous iterations in time, but rather of simulating the same numerical model hundreds of times. The "mesocentre" provides access to several compute nodes, each of which has several cores. COMSOL® is used in batch mode via a script and the SLURM job manager. It's easy to imagine that the use of such computational resources could reduce computing time. But things aren't that simple, and efficiency depends not only on the size of the problem itself, but also on the strategy for distributing simulations over the different cluster resources. The possibilities for parallelization at both Comsol and Cluster level (using nodes, cores, MPI or OPENMP processes) are multiple. The aim of this work is to evaluate with a EC-NDT typical problem in order to determine the best parallelization strategy for this type of problem. In addition, each of our models can be spread over several cores in shared memory or over several nodes in distributed memory. Here again, we'll be trying to assess scalability and the most efficient strategies for this kind of small but repetitive problem. A summary of the various possible approaches will be presented and discussed at the conference. (Optimal number of cores for solving a single problem, evolution of this optimal number as a function of the number of degrees of freedom, use of compute nodes or cores of the same node, reproducibility of results, Influence of problem size etc.).

Baidu
map