This event has ended. Create your own event → Check it out
This event has ended. Create your own
View analytic
Wednesday, March 2 • 3:30pm - 3:50pm
Algorithms & Accelerators I: Performance of DGTD Finite Element Methods for the RTM Procedure on GPU Clusters

Sign up or log in to save this to your schedule and see who's attending!


Nodal discontinuous Galerkin time-domain (DGTD) methods exhibit attractive features for the large scale simulation of seismic waves in complex media. First, such methods provide accurate wavefield solutions for complicated geological structures thanks to the use of unstructured meshes and high-degree discontinuous basis functions. Additionally, the dense algebraic operations required per element and the weak element-to-element coupling of DGTD methods make them suitable schemes for efficient computations on modern clusters with massively parallelized many-core devices, such as GPUs. Both these aspects, accuracy and computational performance, are very important for seismic imaging in the Oil and Gas industry.

In collaboration with Shell, we have conceived a high-performance tool for seismic migration that can be run on clusters of GPUs [*]. This tool, named RiDG, includes reverse time migration (RTM) capabilities and multiple wave models. The model solver is based on a high-order DGTD method for first-order systems which uses unstructured meshes and multi-rate local time-stepping to efficiently deal with multi-scale solutions. Imaging conditions based on vertical characteristics provide improved RTM images.

We adopted the MPI+X approach for distributed programming together with OCCA, a unified framework to make use of major multi-threading languages (e.g. OpenMP, OpenCL and CUDA), offering a flexible approach to handling the multi-threading X. While the RTM procedure generally has extensive data storage requirements with slow I/O, low storage requirements for DGTD boundary data allows halo trace data to be stored in memory rather than relying on disk based check-pointing. The load balancing of our implementation reduces both device--host data movement and MPI node-to-node communication.

In this talk, we present the main features of our RTM implementation and recent results for GPU computing. In particular, the computational performance of the DGTD solver is analysed using the roofline model and compared with alternative strategies. The strong scalability of the implementation is tested using a three-dimensional RTM synthetic case on a GPU cluster. These results confirm the quality of RiDG implementation and the relevance of programming strategies.

[*] A. Modave, A. St-Cyr, W.A. Mulder, and T. Warburton. A nodal discontinuous Galerkin method for reverse-time migration on GPU clusters. Geophysical Journal International, 203(2):1419– 1435, 2015.


Axel Modave

Postdoctoral Associate, VirginiaTech
avatar for Tim Warburton

Tim Warburton

John K. Costain Chair & Professor of Mathematics, Virginia Tech

Attendees (4)