基于MPI和CUDA的转换波Kirchhoff叠前时间偏移并行计算
|
摘要
转换波Kirchhoff叠前时间偏移可以实现全空间三维转换波资料的准确成像。但转换波叠前偏移数据量巨大,而且需偏移迭代多次来寻找匹配的偏移速度模型,导致偏移处理周期长、效率低,限制了转换波偏移技术在生产上的应用规模。目前解决海量运算问题的方法主要是应用CPU集群来提高计算效率,但集群存在成本高、功耗大、占用空间大、维护成本高等缺点。给出了一种基于MPI(Message Passing Interface)和CUDA的转换波Kirchhoff叠前时间偏移并行算法,将细粒度线程级的GPU(Graphic Processing Unit)并行计算融入粗粒度进程级MPI并行编程模型。利用实际转换波数据分别在CPU(单核)、GPU(单卡)、MPI和GPU(2个节点)测试平台上对算法性能进行了测试,结果表明,MPI和GPU(2个节点)的计算速度是CPU(单核)的近400倍,可以大幅度提高转换波Kirchhoff叠前时间偏移的计算效率,降低计算成本。
Converted-wave Kirchhoff prestack time migration can image accurately in 3-D space.However,the prestack time migration generates huge computation amount and need migration for several times to discover matched migration velocity model,which leads to long processing period and low efficiency,and restricts the application scale in production.Nowadays,CPU-cluster computation is applied widely to solve the problem,but the CPU-cluster has some disadvantages such as high cost,large power consumption,large space occupation and high maintenance cost.In order to improve the time-consuming and save the computing cost,we proposed a parallel algorithm of converted-wave Kirchhoff prestack time migration based on MPI+CUDA,which merged the GPU parallel computation of fine-grained thread-level into the MPI parallel programming model of coarse-grained process-level.We tested the performance of the parallel algorithm on CPU(single core),GPU(single card),MPI+GPU(2 nodes).The testing results indicate that the computation speed of the MPI+GPU(2 nodes) platform is nearly 400 times of the CPU's(single core),largely improves the computation efficiency of converted-wave Kirchhoff prestack time migration and decreases the computation cost.
Converted-wave Kirchhoff prestack time migration can image accurately in 3-D space.However,the prestack time migration generates huge computation amount and need migration for several times to discover matched migration velocity model,which leads to long processing period and low efficiency,and restricts the application scale in production.Nowadays,CPU-cluster computation is applied widely to solve the problem,but the CPU-cluster has some disadvantages such as high cost,large power consumption,large space occupation and high maintenance cost.In order to improve the time-consuming and save the computing cost,we proposed a parallel algorithm of converted-wave Kirchhoff prestack time migration based on MPI+CUDA,which merged the GPU parallel computation of fine-grained thread-level into the MPI parallel programming model of coarse-grained process-level.We tested the performance of the parallel algorithm on CPU(single core),GPU(single card),MPI+GPU(2 nodes).The testing results indicate that the computation speed of the MPI+GPU(2 nodes) platform is nearly 400 times of the CPU's(single core),largely improves the computation efficiency of converted-wave Kirchhoff prestack time migration and decreases the computation cost.
引文
[1]马昭军,唐建明.叠前时间偏移在三维转换波资料处理中的应用[J].石油物探,2007,46(2):174-180 Ma Z J,Tang J M.The application of prestack time migration in3Dconverted-wave data processing[J].Geophysical Prospecting for Petroleum,2007,46(2):174-180
[2]张兵,赵改善,黄俊,等.地震叠前深度偏移在CU-DA平台上的实现[J].勘探地球物理进展,2008,31(6):427-432 Zhang B,Zhao G S,Huang J.Implement prestack depth migration on CUDA platform[J].Progress in Exploration Geophysics,2008,31(6):427-432
[3]刘国峰,刘洪,王秀闽,等.Kirchhoff积分叠前时间偏移的两种走时计算及并行算法[J].地球物理学进展,2009,24(1):131-136 Liu G F,Liu H,Wang X M,et al.Two kinds of trav-eling time computation and parallel computing meth-ods of Kirchhoff migration[J].Progress in Geophys-ics,2009,24(1):131-136
[4]李肯立,彭俊杰,周仕勇.基于CUDA的Kirchhoff叠前时间偏移算法设计与实现[J].计算机应用研究,2009,26(12):4474-4477 Li K L,Peng J J,Zhou S Y.Implement Kirchhoff prestack time migration algorithm on CUDA archi-tecture[J].Application Research of Computers,2009,26(12):4474-4477
[5]刘国峰,刘钦,李博,等.油气勘探地震资料处理GPU/CPU协同并行计算[J].地球物理学进展,2009,24(5):1671-1678 Liu G F,Liu Q,Li B,et al.GPU/CPU co-processingparallel computation for seismic data processing in oil and gas exploration[J].Progress in Geophysical(in Chinese),2009,24(5):1671-1678
[6]Li X Y.Converted-wave moveout analysis revisited:the search for a standard approach[J].Expanded Abstracts of73 rd Annual Internat SEG Mtg,2003,805-808
[7]Li X Y,Dai H C,Mancini F.Converted-wave ima-ging in anisotropic media:theory and case studies[R].Edinburgh:Edinburgh Anisotropy Project,2006
[8]Li X Y,Yuan J X.Converted wave traveltime equations in layered anisortropic media:an overview[R].Edinburgh:EAP Annual Research Report,2001:3-32
[9]Dai H C,Li X Y,Conway P.3Dpre-stack Kichhoff time migration of PS-waves and migration velocity model building[J].Expanded Abstracts of 74 th Annual Internat SEG Mtg,2004,1115-1118
[10]John D O.A surey of general-purpose computation on graphics hardware,eurographics[EB/OL].(2005-06-23)[2012-03-15]http://www.realworldtech.com/page.cfm?ArticleID=RWT0908
[11]Nvidia.Nvidia CUDA compute unified device archi-tecture programming guide version4.0[EB/OL].[2011-12-9]http://developer.download.Nvidia.com/compute/cuda/4_0/docs/CudaReferenceManu-al_4.0.pdf
[12]张舒,褚艳丽,赵开勇,等.GPU高性能计算之CU-DA[M].北京:中国水利电力出版社,2009:25-30 Zhang S,Chu Y L,Zhao K Y,et al.GPU high per-formance computing on CUDA[M].Beijing:China Publishing Company of Water Conservancy and E-lectric Power,2009:25-30
[13]Budruk R,Anderson D,Shanley T.PCI express sys-tem architecture[EB/OL].[2012-03-15]http://www.docin.com/touch/detail.do?id=501507982
[14]都志辉.高性能计算之并行编程技术-MPI并行程序设计[EB/OL].(2007-10-9)[2012-03-15]http://wenku.baidu.com/view/00bacc204b35eefdc8d33375.html Du Z H.The parallel programme technolegy of high performance computing-MPI parallel programming[EB/OL].(2007-10-9)[2012-03-16]http://wenku.baidu.com/view/00bacc204b35eefdc8d33375.html
[15]John D O,Mike H,David L,et al.GPU computing[J].Proceedings of the IEEE,2008,96(5):879-897
[2]张兵,赵改善,黄俊,等.地震叠前深度偏移在CU-DA平台上的实现[J].勘探地球物理进展,2008,31(6):427-432 Zhang B,Zhao G S,Huang J.Implement prestack depth migration on CUDA platform[J].Progress in Exploration Geophysics,2008,31(6):427-432
[3]刘国峰,刘洪,王秀闽,等.Kirchhoff积分叠前时间偏移的两种走时计算及并行算法[J].地球物理学进展,2009,24(1):131-136 Liu G F,Liu H,Wang X M,et al.Two kinds of trav-eling time computation and parallel computing meth-ods of Kirchhoff migration[J].Progress in Geophys-ics,2009,24(1):131-136
[4]李肯立,彭俊杰,周仕勇.基于CUDA的Kirchhoff叠前时间偏移算法设计与实现[J].计算机应用研究,2009,26(12):4474-4477 Li K L,Peng J J,Zhou S Y.Implement Kirchhoff prestack time migration algorithm on CUDA archi-tecture[J].Application Research of Computers,2009,26(12):4474-4477
[5]刘国峰,刘钦,李博,等.油气勘探地震资料处理GPU/CPU协同并行计算[J].地球物理学进展,2009,24(5):1671-1678 Liu G F,Liu Q,Li B,et al.GPU/CPU co-processingparallel computation for seismic data processing in oil and gas exploration[J].Progress in Geophysical(in Chinese),2009,24(5):1671-1678
[6]Li X Y.Converted-wave moveout analysis revisited:the search for a standard approach[J].Expanded Abstracts of73 rd Annual Internat SEG Mtg,2003,805-808
[7]Li X Y,Dai H C,Mancini F.Converted-wave ima-ging in anisotropic media:theory and case studies[R].Edinburgh:Edinburgh Anisotropy Project,2006
[8]Li X Y,Yuan J X.Converted wave traveltime equations in layered anisortropic media:an overview[R].Edinburgh:EAP Annual Research Report,2001:3-32
[9]Dai H C,Li X Y,Conway P.3Dpre-stack Kichhoff time migration of PS-waves and migration velocity model building[J].Expanded Abstracts of 74 th Annual Internat SEG Mtg,2004,1115-1118
[10]John D O.A surey of general-purpose computation on graphics hardware,eurographics[EB/OL].(2005-06-23)[2012-03-15]http://www.realworldtech.com/page.cfm?ArticleID=RWT0908
[11]Nvidia.Nvidia CUDA compute unified device archi-tecture programming guide version4.0[EB/OL].[2011-12-9]http://developer.download.Nvidia.com/compute/cuda/4_0/docs/CudaReferenceManu-al_4.0.pdf
[12]张舒,褚艳丽,赵开勇,等.GPU高性能计算之CU-DA[M].北京:中国水利电力出版社,2009:25-30 Zhang S,Chu Y L,Zhao K Y,et al.GPU high per-formance computing on CUDA[M].Beijing:China Publishing Company of Water Conservancy and E-lectric Power,2009:25-30
[13]Budruk R,Anderson D,Shanley T.PCI express sys-tem architecture[EB/OL].[2012-03-15]http://www.docin.com/touch/detail.do?id=501507982
[14]都志辉.高性能计算之并行编程技术-MPI并行程序设计[EB/OL].(2007-10-9)[2012-03-15]http://wenku.baidu.com/view/00bacc204b35eefdc8d33375.html Du Z H.The parallel programme technolegy of high performance computing-MPI parallel programming[EB/OL].(2007-10-9)[2012-03-16]http://wenku.baidu.com/view/00bacc204b35eefdc8d33375.html
[15]John D O,Mike H,David L,et al.GPU computing[J].Proceedings of the IEEE,2008,96(5):879-897
版权所有：© 2023 中国地质图书馆 中国地质调查局地学文献中心