DOWNLOAD PRESENTATIONWATCH VIDEOReverse time migration (RTM) is the method of first choice in seismic imaging. It fully respects the two-way wave equation. It provides high quality imaging also in regions with complex geological structures, e.g. regions composed of steep flanks and/or high velocity contrasts. This comes at a price. RTM has high demands on the underlying compute resources. Being of pure academical interest for a long time, progress in hardware development has made RTM feasible also on the industrial scale. Complex physical modeling, large target output domains, large migration apperture and/or high frequency content still require efficient parallelization on the algorithmic side. There, the memory footprint might be too large for the computation of shot results on a single device. RTM benefits from high throughput accelerators like e.g. GPUs. In order to deal with the heterogeneity at the hardware level, RTM needs a high level of parallelism and improved load balancing features in order to fully exploit the underlying hardware resources. Furthermore, with an ever increasing floating point throughput, I/O should be avoided as much as possible to preserve scalability.
These requirements also arise in the context of interactive velocity model building based on RTM, which comes into the realms of possibility nowadays. Here, time to solution matters and efficient internode parallelization is required to achieve good scalability.
We propose to introduce the concept of asynchronous constraint execution complemented by random velocity perturbation boundary conditions to achieve good scalability. We show the parallel efficiency achieved for forward propagation in an acoustic isotropic medium in a strong scalability set up on a 1024 cube grid. For 64 processes running on a Intel(R) Xeon(R) CPU E5-2680 using 10 cores each, the parallelization efficiency is 97%.
