| 三维逆时偏移GPU/CPU机群实现方案研究 |
| |
| 引用本文： | 刘守伟,王华忠,陈生昌,孔祥宁.三维逆时偏移GPU/CPU机群实现方案研究[J].地球物理学报,2013,56(10):3487-3496. |
| |
| 作者姓名： | 刘守伟 王华忠 陈生昌 孔祥宁 |
| |
| 作者单位： | 1. 浙江大学地球科学系, 杭州 310027;
2. 同济大学海洋与地球科学学院波现象与反演成像研究组, 上海 200092;
3. 中国石化物探技术研究院, 南京 211103 |
| |
| 基金项目： | “973”项目(CB201002);重大专项项目(2011ZX05005-005-008HZ,2011ZX05006-002,2011ZX05003-003)资助 |
| |
| 摘 要： | 叠前逆时偏移是当前最为准确的地震成像方法,由于计算量大、存储量大等原因需要合适的实现策略和高效的计算平台.本文以高阶有限差分逆时偏移为基础,重点讨论了在GPU上实现需要解决的显存不足问题和人工边界问题.利用区域分解技术可以在当前GPU上高效地实现任意生产规模的三维逆时偏移成像,不会受到GPU显存规模的制约.常规最佳匹配层边界条件边界区域控制方程与内部区域差异较大,不适于GPU高速运算.本文在GPU上实现近似最佳匹配层(NPML)边界条件,使得高阶有限差分计算不需要分支判断,边界区域辅助波场的存储量也较低,保证了在GPU上进行波场传播的高效性.三维理论数据和实际资料成像结果表明了本文方法的正确性.
|
| 关 键 词： | 逆时偏移 图形处理器 区域分解 近似最佳匹配层 |
| 收稿时间： | 2012-07-31 |
Implementation strategy of 3D reverse time migration on GPU/CPU clusters |
| |
| LIU Shou-Wei,WANG Hua-Zhong,CHEN Sheng-Chang,KONG Xiang-Ning.Implementation strategy of 3D reverse time migration on GPU/CPU clusters[J].Chinese Journal of Geophysics,2013,56(10):3487-3496. |
| |
| Authors: | LIU Shou-Wei WANG Hua-Zhong CHEN Sheng-Chang KONG Xiang-Ning |
| |
| Institution: | 1. Department of Earth Science, Zhejiang University, Hangzhou 310027, China;
2. Wave Phenomena and Inversion Imaging Research Group (WPI), School of Ocean & Earth Science, Tongji University, Shanghai 200092, China;
3. Sinopec Geophysical Research Institute (SGRI), Nanjing 211103, China |
| |
| Abstract: | Prestack reverse time migration(RTM) is currently the most precise seismic imaging method. It needs appropriate implementation strategies and efficient computation platforms, because of the needs for large amount of calculation and large storage. Based on high order finite difference RTM, this paper focuses on solving the questions of the lack of video memory and artificial boundary, which must be solved for Graphic Processing Unit (GPU) RTM. Using the domain decomposition technique, we can efficiently realize three dimensional (3D) RTM of any production scale. It eliminates the limitation of GPU memory size. The control equation of the traditional perfectly matched layer (PML) boundary condition is different to that of the inner area. It is not suitable for GPU calculation. This paper realizes the Nearly PML (NPML) boundary condition on GPU. It makes high order finite difference calculation eliminate the needs of conditional branches. And the amount of auxiliary wavefields in boundary area is small. Therefore, wavefields propagation using NPML is efficient on GPU. 3D theoretical data and real field data RTM results prove the correctness of this paper's methods. |
| |
| Keywords: | Reverse time migration(RTM) Graphic processing unit (GPU) Domain decomposition Nearly perfectly matched layer(NPML) |
| 本文献已被 CNKI 等数据库收录！ |
| 点击此处可从《地球物理学报》浏览原始摘要信息 |
| 点击此处可从《地球物理学报》下载免费的PDF全文 |
|
