Automatic intra-node CPU-GPU load-balancing of domain-decomposed seismic applications
Load balancing; Wave propagation; Seismic modeling; Collaborative processing; Collaborative execution; Heterogeneous architecture
The seismic modeling process, applied to subsurface imaging techniques, re-quires high-performance computational solutions to extract more accurate information on the subsoil geological formations properties. In the last years, the integration between CPU and GPU has become very attractive thanks to the excellent combination of computational power and energy efficiency. However, traditional approaches that dele-gate workloads massively parallel to GPUs can leave CPUs idle, which leads to a waste of the system’s overall performance potential. In order to efficiently use all available processing units, this work aims to investigate and develop methods of load balancing in heterogeneous systems composed of multiple CPUs and GPUs. An acoustic wave modeling code was implemented as an experiment and then executed in two different computational architectures. The results show that performing collaborative processing between CPU and GPU can be more efficient than CPU-only and GPU-only execution. Besides, the results show that, depending on the computational resources configuration, different proportions of workload partitioning must be distributed to optimize the use of these resources and reduce the execution time.