Integrando checkpointing e compressão de dados de um Solver FWI com OpenMP em multi-GPUS

Abstract

High-performance computing (HPC) plays a crucial role in various scientific fields, and the use of GPUs for parallel computing has proven to be an efficient solution for large-scale problems, such as Full Waveform Inversion (FWI) in seismic simulations. This work proposes the implementation of domain decomposition, checkpointing, and data compression techniques in an FWI application executed on multiple GPUs, aiming to optimize computational performance and reduce memory usage. The research was conducted in three main stages: familiarization with the existing codebase, study of checkpointing and data compression techniques, and implementing domain decomposition in a multi-GPU environment. The domain decomposition strategy was applied by subdividing the three-dimensional grid into "slices," allowing distributed execution across multiple devices. Explicit control over data allocation and transfer between the CPU and GPUs, using OpenMP and nvcomp, enabled greater efficiency in memory management and data transfers. Experimental results, performed on a node with 4 NVidia V100 GPUs connected via NVLink, showed that the use of multiple GPUs offers significant performance gains for large-scale problems. For smaller problems, the overhead associated with data transfer between devices neutralizes the benefits of parallelism. The combination of checkpointing and data compression resulted in improvements in memory usage and communication between devices, particularly in larger domains and with more iterations. This work contributes to the advancement of FWI applications in distributed computing environments, providing a scalable and efficient solution for large-scale scientific problems.