A portable HPC toolbox for the simulation and inversion of wavefields

The objective of this project is the development of an HPC toolbox with which acoustic/elastic waves can be simulated and inverted. Application areas are first of all in geophysics, but the results can be transferred directly to medical applications based on ultrasound and other areas. The simulation and inversion of the propagation of acoustic waves is a typical application area of HPC. Usually it is solved by means of explicit Finite Difference (FD) methods. In WAVE we extend and port established FD methods. To this end, we introduce an adaptive discretization, which allows a higher resolution necessary in complex model structures. At the same time, an efficient simulation and inversion of real-world problems is facilitated. Imbalances in the computational load will be addressed by load balancing techniqes that consider the heterogeneity of emerging HPC systems. By the intended extension of the LAMA library for numerical algorithms and data structures, an open-source HPC toolbox for scalable wave simulation and inversion is created. It shall be portable to different parallel architectures with robust performance. The practical evaluation of this HPC toolbox happens with a real-world use case on seismic exploration data.

The module for load balancing receives the data describing the simulation and the computer topology from the application or the LAMA library. Our load balancing methodology shall combine the best of two worlds. Primarily we want to employ space-filling curves due to their speed and scalability. In order to alleviate the deficiencies of space-filling curves in terms of solution quality, we will employ combinatorial methods additionally within a performance model. By using a local approach, this improvement step shall be scalable as well. Moreover, we will provide methods for repartitioning and mapping. They partially rely on the partitioning methods described above.

Meyerhenke, Henning Prof. Dr. (Details) (Modeling and Anlysis of Complex Systems)

Duration of Project
Start date: 02/2016
End date: 07/2019

Research Areas
Massively Parallel and Data-Intensive Systems

Research Areas

Eugenio Angriman, Alexander van der Grinten, Moritz von Looz, Henning Meyerhenke, Martin Nöllenburg, Maria Predari, Charilaos Tzovas (2019). Guidelines for Exp erimental Algorithmics: A Case Study in Network Analysis. Algorithms 12(7): 127 (2019).

Roland Glantz, Maria Predari, Henning Meyerhenke (2018). Topology-induced Enhancement of Mappings. In: Proc. 47th Intl. Conf. on Parallel Processing (ICPP): 9:1-9:10.

Moritz von Looz, Charilaos Tzovas, Henning Meyerhenke (2018). Balanced k-means for Parallel Geometric Partitioning. In: Proc. 47th Intl. Conf. on Parallel Processing (ICPP): 52:1-52:10.

David A. Bader, Andrea Kappes, Henning Meyerhenke, Peter Sanders, Christian Schulz, Dorothea Wagner (2018). Benchmarking for Graph Clustering and Partitioning. Encyclopedia of Social Network Analysis and Mining. 2nd ed., Springer-Verlag.

Last updated on 2021-04-01 at 17:47