Workflow for Parallel Processing of Sequential Mesh Databases

The work presents a novel approach for parallel loading of unstructured meshes. The key challenge is that the database file stores elements in a spatially scattered fashion. The reason is that the file was not stored with the expectation of being read in parallel.

Features

Library for scalable data processing of the sequential database files for connection of the massively parallel open source solvers.

Sequential input

Create data in external tools like ANSYS, HyperMesh, ABAQUS...

Parallel loading

MPI I/O & DATA PARSING Loading and understanding data

Parallel sorting

Balance distributed data

Spatial locality

CLUSTERIZATION
Space Filling Curve decomposition

Fast Dual Graph

The dual graph is suitable for decomposition of complex geometries with various parameters.

Hybrid Domain Decomposition

Multilevel decomposition suitable for massively parallel solvers.

Output for HPC

Efficient connection engineering comunity to HPC

Massivelly parallel solver connection

Hybrid multilevel FETI solver for complex engineering problems

image

Jet Engine

This figure present a evaluation of the entire workflow on the large dataset, which contains about 800 million nodes and 500 million elements. These dataset can be processed in 6.9s using 6144 MPI processes on 512 nodes of the Salomon supercomputer. An additional 6.8s is needed for MPI I/O to load and distribute the 142GB (21 GB/s throughput) file from the Lustre storage to compute nodes memory.

Contact Us

IT4Innovations National Supercomputing Center
VSB – Technical University of Ostrava
Studentsk√° 6231/1B
708 33 Ostrava-Poruba
Czech Republic