Speaker
Ondřej Meca
(IT4Innovations)
Description
A wide range of engineers from the industrial practice use numerical methods to simulate physical processes. A process of simulation can be divided into two parts -- creation of a numerical model (finite element mesh, definition of boundary conditions, contact interfaces, etc.) and its solution. Since the solution quality is dependent on the numerical model, it is necessary to use an efficient and robust tool for its creation. A typical example are problems in the structural mechanics, where the creation of a high-quality numerical model is one of the most complicated (and important) parts during the simulation process (especially for complex geometries).
In general, high-quality models are usually provided by (commercial) tools that produce sequential mesh databases. Hence, some parallel solvers provide a converter of external (sequential) database files into their (parallel) database structure. However, the converter is usually only sequential. It greatly slows down or inhibits the connection between tools that generate sequential data and parallel solvers. It is one of the factors of a low utilization of HPC by the mainstream engineering community.
Our motivation is to create a library enabling the connection between tools for the creation of complex engineering models (such as ANSYS, HyperMesh, ANSA, ABAQUS, etc.) along with open source parallel solvers to allow broader usage of HPC by the engineering community. By the straight connection, we get the robust preprocessing together with possibility to connect various highly parallel solvers that are able to solve non-standard problems.
We propose an algorithm that is able to load and decompose a sequentially stored mesh database. Thus, it can be used as a parallel converter for various formats. Moreover, scalability tests show that the algorithm can be also used as a direct loader and preprocessor of massively parallel solvers (e.g ESPRESO). The algorithm is composed of several commonly used approaches that together lead to a robust and fast solution.
The workflow is composed from several steps. At the beginning a mesh database is read and parsed. Since the parsed data can be randomly scattered among MPI processes, we apply parallel sorting to get a known data distribution. Then, we are ready to assign regions and make a spatial clusterization. This step assures scalability of the next processing that prepares a mesh for a parallel solver.
References
ESPRESO (espreso.it4i.cz)
Summary
We will present the workflow for parallel loading of sequentially stored mesh databases. It can be used as the connection between tools for the creation of complex engineering models along with parallel solvers to allow broader usage of HPC by the engineering community. The scalability tests show that the proposed algorithm is able to prepare a mesh with hundreds of millions nodes/elements in several second. The workflow is implemented into in-house library ESPRESO.
Primary author
Ondřej Meca
(IT4Innovations)
Co-authors
Lubomír Říha
(IT4Innovations)
Tomas Brzobohaty
(IT4Innovations)
