Tips and tricks for Finite difference and i/o-less FWI
Tips and tricks for Finite difference and i/o-less FWI
复制标题
DOI：
10.1190/1.3627855
复制
发表时间：
2011
期刊：
影响因子：
--
通讯作者：
David Imbert;Khadija Imadoueddine;P. Thierry;H. Chauris;L. Borges
中科院分区：
文献类型：
--
作者：
David Imbert;Khadija Imadoueddine;P. Thierry;H. Chauris;L. Borges
由于近十年来计算机能力的发展，FD 方法在 3D 地震中重新出现，以模拟全双向波动方程，以及 3D 逆时偏移（RTM）和全波形反演（FWI）。除了 3D 模板的优化实现之外，逆时偏移很大程度上依赖于互连和 I/O（输入/输出）子系统来支持域分解通信以及前向和后向波场互相关所需的快照。在本文中，我们总结了可用于加速模板计算的不同级别的优化。我们还定义了一个模型来估计此类算法的实际效率。为了优化 FWI 算法，我们重新审视基于 Griewank (1992) 的检查点算法，该算法已在计算流体动力学和地震 (Symes, 2007) 中使用，以避免或限制 I/O。这些技术是使用主存储器和/或磁盘作为检查点设备来实现的，但代价是对前向传播进行一些重新计算。关键是要使用快速的 I/O 设备。我们还表明，使用 MPI 和 OpenMP 的专用混合并行实现有助于将该模型扩展到更大的工作负载，并将几乎所有数据保留在内存中，仍然实现良好的可扩展性：基于 SSD 技术的本地快速 I/O 设备可以保证额外的 I/O 需求，而不是使用大型并行文件系统。
Thanks to the computer capability grow in the last decade, the FD method reappeared in 3D seismic to simulate the full twoway wave equation, together with 3D reverse time migration (RTM) and full waveform inversion (FWI). In addition to the optimized implementation of 3D stencils, the reverse time migration heavily relies on interconnect and i/o (input/output) subsystem to support domain decomposition communication as well as the necessary snapshots for cross correlation of the forward and backward wave-fields. In this paper, we summarize the different levels of optimization available to speedup stencil computations. We also define a model to estimate the actual efficiency of such algorithm. To optimize the FWI algorithm we revisit check-pointing algorithms based on Griewank (1992) that were already used in Computational fluid dynamics and in seismic (Symes, 2007) to avoid or limit the i/o. Such techniques are implemented using main memory and/or disk as a checkpoint device at the cost of some recomputation of the forward propagation. The key point is to use a fast i/o device. We also show that a dedicated hybrid parallel implementation using MPI and OpenMP can help to extend this model to larger workload and to keep almost all data in memory, still achieving good scalability: Instead of using large parallel filesystem, local fast i/o devices based on SSD technology can ensure the extra i/o needs.