The de-facto standard algorithm in reflection seismology to construct images of the Earth’s subsurface from measured seismic data on the surface is Reverse Time Migration (RTM). It is computationally complex and typically takes several days to complete on large supercomputers. It is therefore paramount to use a highly optimised implementation that can utilise the available hardware to the best advantage.
This white paper gives practical tips on how modern CPU features such as wide vector units, cache hierarchies, NUMA memory hierarchies, and the latest instruction sets can be used to achieve large performance gains for RTM. It further details how to tap into the compute power of Nvidia GPUs and align the available processing cores and memory hierarchy with the RTM algorithm. As Intel’s Xeon Phi becomes increasingly popular, the paper explores its suitability for the RTM algorithm. Practical implementations need to combine accelerators and CPUs in a compute cluster, and we demonstrate how to best hide the necessary data transfer latencies.