Efficient Implementation of 3D Surface-related Multiple Elimination for Large-scale Compute Clusters

Suppression of multiple reflections is an important step in processing the offshore seismic data, since such reflections seriously complicate the data interpretation. SRME is often the method of choice for eliminating multiple reflections related to the free surface. Application of the method to the 3D data demands for a huge amount of computations and data traffic, and scalability for a large number of cluster nodes is a serious problem. In the present work, we suggest an efficient 3D SRME algorithm providing high scalability of computations on Linux clusters with thousands of compute nodes. The main idea is that of organizing a quasi-sequential stream of data from the storage to compute nodes. Implementation details for both CPU-only and hybrid CPU+GPU systems are considered. Performance measurements taken on a high-end supercomputer are presented and discussed.