Developing an Exascale Trajectory; Leveraging GPU-based Seismic Imaging Experiences
Ty McKercher · 72nd EAGE Conference and Exhibition - Workshops and Fieldtrips · 2010
The lessons learned from implementing GPU-based computing solutions in large-scale seismic imaging production environments will set a trajectory for transparent scalability of future generation many-core architectures. GPUs offer an additional level of parallelism to existing distributed memory parallel environments. As hardware evolves, it is important to consider: memory architecture, execution control, control flow efficiency, and unified address schemes. At the core of NVIDIA's strategy is the requirement to complement existing compute infrastructure with industry-standard COTS components that leverage existing software knowledge. A mandatory requirement is to provide extensible APIs that allow augmentation of modern software tools/practices, thereby offering developers freedom to choose the best GPU-based tool/method for a given problem. But achieving optimal performance from many-core architectures also requires improving computational thinking skills. Experience has shown that it is imperative to focus on domain decomposition, truly understanding data access patterns during optimization investigations for distributed many-core systems. Academic institutions have adopted CUDA-based techniques to help build the foundation for computational thinking because the API supports the proper parallel constructs. Also, as the costs for seismic acquisition and processing spiral upward, modeling plays an ever increasing role in oil/gas work flows. Mesh-based modeling approaches that exploit GPU parallelism will continue to demonstrate excellent scalability and help solve some of the inverse challenges. The seismic work flow is melding, and large-scale Visual Computing will play an expanding role. Centralized, secure delivery of seismic applications from HPC data centers will improve crossdiscipline collaboration to shorten cycle time, reduce risk, and improve exploration/production success rates. In this presentation, we will share customer experiences that relate to the forces shaping the future of seismic imaging. This is an exciting time for innovations in GPU run-time environments, and as programming tools evolve, we can leverage our experience to help set a trajectory toward meeting the exascale demands that loom on the horizon.