Accelerating Seismic Imaging on GPUs

We discuss the performance of Petrobras production Kirchhoff prestack seismic migration on a cluster of 72 GPUs and 288 CPU cores with CPU and GPU cooperation. Porting and optimization of the application hot spot (contribution loop) to a single GPU reduces total execution time by a factor of 20 per node. Instead of porting the next hot spot (trace filtering) to the GPU, we also show that cooperation of CPU (trace filtering) and GPU (contribution loop) is optimal, since the CPU workload is totally overlapped by GPU workload.