Abstract: |
|
Reverse time migration (TRM) and full waveform inversion (FWI) are critical exploration techniques that compute high-resolution images of the subsurface from seismic data, by propagating waves in 3D models. Commonly used time domain wave equation solvers based on explicit finite differences are usually memory bandwidth bound due to low flop-per-byte ratio and non-contiguous memory access. State-of-the-art implementations exploit spatial and temporal cache blocking for better CPU balance during wave propagation. However, in imaging a forward-propagated source wavefield is combined at regular time steps with a back-propagated receiver wavefield. Synchronization of both wavefields can result in very large volume of IO, disrupting another balance on typical supercomputers. Multilayer checkpointing techniques, from DRAM to burst buffer to disk, potentially offer dramatic improvements. This talk presents an overview of large-scale seismic depth imaging implementation challenges, illustrated with field data applications. |
|