Optimizing Forward Wavefield Storage Leveraging High-Speed Storage Media

Full-Waveform Inversion (FWI) is a commonly used adjoint-based optimization technique in geosciences. Similar to other adjoint-based optimizations, FWI involves gradient calculations that impose a substantial memory footprint. This memory footprint is often considered the primary bottleneck in this FWI process and has been the subject of extensive research. Previous studies have explored various approaches to mitigate this issue, including optimal checkpointing strategies, lossy compression, subsampling in time, and other methods aimed at reducing the memory requirements. These investigations have typically operated under the assumption that writing the data to disk would result in significant and unacceptable performance degradation, leading to compromises in precision or increased computational costs to conserve memory resources. In this work, we challenge this assumption in light of recent advancements in high-bandwidth storage, particularly the utilization of Non-Volatile Memory Express solid-state drives (NVMe SSDs). We provide empirical evidence that, with specific combinations of modern hardware, it is indeed feasible to write temporary data to disk during the forward propagation phase and subsequently read it back during the adjoint computation without experiencing substantial performance degradation.