Integration method pre-stack time migration method based on Shenwei architecture

Abstract

The invention provides an integral method pre-stack time migration method based on a Shenwei architecture. The method comprises the steps of 1, preparing seismic data and migration speed required by integral method pre-stack time migration calculation; 2, establishing a plurality of calculation groups, and distributing the seismic data; step 3, calculating seismic data; 4, re-sampling and storing the seismic data; step 5, storing a pre-stack time migration calculation result of the resampled seismic data; 6, when calculation of all seismic data channels is completed, a calculation result is stipulated and stored; 7, when all the calculation groups finish calculation, the calculation result is stipulated and stored; and step 8, performing resampling to obtain a regular sampling interval result. According to the integral method pre-stack time migration method based on the Shenwei architecture, the pressure on read-write and network communication during super-large-scale parallel is reduced, and meanwhile, the operation speed of the slave core is increased by reducing the number of access times between the main memory and the local memory.

Description

Technical Field

[0001] The present invention relates to the technical field of petroleum geophysical exploration, and in particular to an integral method prestack time migration method based on a Shenwei architecture. Background Art

[0002] Seismic exploration is the most important and effective method for solving oil and gas exploration problems in geophysical exploration. Seismic data processing is in the middle of seismic exploration and can provide a basis for subsequent geological researchers to find oil and gas reservoirs. Among them, migration imaging technology can convert seismic reflection signals into position images of underground geological bodies, which is one of the core technologies of seismic data processing. Among many migration methods, the integral method prestack time migration algorithm is the most widely used in the industry due to its high computational efficiency, flexible parallel mode, and good processing effect. With the increasing amount of seismic data processing, the amount of calculation for migration processing has also increased, and supercomputers are needed for parallel accelerated calculations. The Shenwei supercomputer was independently developed by my country and is equipped with a purely domestically produced Shenwei 26010 multi-core processor. The use of a two-level parallel structure of "master core + slave core" can give full play to the computing performance of the Shenwei processor.

[0003] The current parallel strategies of the integral method prestack time migration algorithm are based on CPU, GPU, etc. Among them, the CPU-based parallel strategy uses the network to transmit data and messages between nodes, and the CPU is used inside the node for parallel computing; the heterogeneous parallel strategy based on CPU and GPU uses the network to transmit data and messages between nodes, and the CPU and GPU inside the node cooperate to perform parallel computing, and the computing program is assigned to the CPU or GPU according to the specific characteristics of the calculation. When the Shenwei supercomputer performs multi-node parallel computing, it also uses the network to transmit data and messages between nodes. Unlike conventional CPUs, it has a master core and a slave core. The number of master cores is relatively small, with only four. They are responsible for inter-node communication, main memory access, simple calculations, sending computing tasks to slave cores, and recovering slave core calculation results. Each master core manages a slave core group, and each core group contains 64 slave cores. The slave core comes with 64KB of high-speed local memory. The slave core directly accesses the local memory to improve the reading speed and reduce the access delay. The use of master core + slave core group can achieve parallel acceleration.

[0004] The Sunway supercomputer uses technologies such as flattened object storage and high-speed data transmission based on domestic networks to support large-scale concurrent I / O operations and data transmission, but it still encounters bottlenecks when operating in ultra-large-scale parallelism, resulting in a significant decrease in parallel efficiency. The Sunway supercomputer relies on slave cores for parallel accelerated computing. The slave cores read quickly from their own local memory, but the local memory capacity is small. The main memory capacity of the host is large. In order to meet the storage needs of computing, the slave cores inevitably need to frequently read and write main memory data during computing, but the slave cores read and write main memory slower than local memory. Frequent reading and writing of main memory will reduce overall computing performance.

[0005] In the Chinese patent application with application number: CN202110883510.8, a three-dimensional Kirchhoff integral method pre-stack time migration fast imaging method is involved, including: step 1, collecting velocity models and seismic data for migration imaging; step 2, completing the calculation of travel time of scattering points in different layers on different threads; step 3, based on the instruction set, parallelly realizing the calculation of the travel of 4 longitudinal sampling points in each channel, and storing it in the SSE register; step 4, for each scattering point, performing geometric diffusion compensation of amplitude energy, completing scanning and stacking of diffraction energy, and realizing migration imaging processing based on scattering points; step 5, performing migration imaging within the migration aperture range, and completing multi-threaded fast imaging processing in a single machine. The three-dimensional Kirchhoff integral method pre-stack time migration fast imaging method provides a faster and more attractive pre-stack time migration technology, which provides a basis and guarantee for subsequent field data collection on-site quality monitoring.

[0006] In the Chinese patent application with application number: CN201910397720.9, a data optimization method and an integral method prestack depth migration method are involved, which involve the field of seismic exploration technology. The optimization method includes obtaining a target matrix to be optimized; generating a first sequence according to the target matrix; thinning the first sequence according to a preset grid density to obtain the value position of each element of the second sequence, and obtaining the value of each element of the second sequence based on the least squares principle; interpolating the second sequence to obtain a third sequence; calculating the target matrix corresponding to the third sequence; calculating the error between the target matrix to be optimized and the target matrix corresponding to the third sequence; when the error is less than the first error threshold, recording the target matrix corresponding to the second sequence as the optimized target matrix of the target matrix to be optimized. The data optimization method provided by the embodiment of the invention can greatly improve the efficiency of parallel calculation of the Kirchhoff integral method prestack depth migration technology using GPU to compensate for medium absorption.

[0007] In the Chinese patent application with application number: CN201410367547.5, a large-scale parallel Kirchhoff prestack depth migration method and device are involved, wherein the method includes: storing the data to be migrated as multiple offset files, the offsets of the data to be migrated in the same offset file are the same; dividing all nodes participating in the Kirchhoff prestack depth migration calculation into multiple node groups; triggering each node group to perform Kirchhoff prestack depth migration calculation on different offset files respectively, triggering the leader node of the node group to read an offset file and broadcast it to the member nodes of the node group, triggering each node in the node group to receive an imaging task of the offset file, and obtaining the common imaging point travel time table corresponding to the imaging task, and independently completing the migration calculation of each imaging task according to the obtained common imaging point travel time table. This solution solves the problem of intensive network transmission of seismic data and avoids the local I / O problem caused by reading the travel time table multiple times.

[0008] In the Chinese patent application with application number: CN201210367672.7, a parallel method for Kirchhoff prestack time migration suitable for large-scale and non-acceleration bottlenecks is involved in the processing of oil geophysical exploration seismic data. The output space is subdivided into four dimensions, such as line number and shot offset, and the number of shot offsets is calculated according to the minimum shot offset, maximum shot offset, and shot offset increment parameters. All nodes involved in the calculation are grouped, and the number of nodes in each group is uniform. The nodes in the same group are continuous in the order of node names, and each group of nodes calculates the prestack time migration of one shot offset component at a time. The invention realizes the minimum amount of network transmission of seismic data, completely solves the network transmission intensive problem of Kirchhoff prestack time migration, and can expand the scale of computing hardware and shorten the calculation cycle of prestack time migration.

[0009] The above existing technologies are all significantly different from the present invention and fail to solve the technical problem we want to solve. Therefore, we have invented a new integral method pre-stack time migration method based on the Shenwei architecture. Summary of the invention

[0010] The purpose of the present invention is to provide an integral method prestack time migration method based on the Shenwei architecture, which can reduce the reading and writing and network communication pressure in ultra-large-scale parallelism, reduce the number of accesses between main memory and local memory, and improve the parallel computing speed.

[0011] The object of the present invention can be achieved by the following technical measures: an integral method prestack time migration method based on the Shenwei architecture, the integral method prestack time migration method based on the Shenwei architecture comprises:

[0012] Step 1, prepare the seismic data and migration velocity required for prestack time migration calculation using the integral method;

[0013] Step 2, establish multiple calculation groups and distribute the seismic data;

[0014] Step 3, calculating seismic data;

[0015] Step 4, resampling and storing the seismic data;

[0016] Step 5, storing the prestack time migration calculation results of the resampled seismic data;

[0017] Step 6, when all seismic data channels are calculated, the calculation results are simplified and stored;

[0018] Step 7: When all calculation groups have completed the calculation, the calculation results are simplified and stored;

[0019] Step 8, resampling is performed to obtain results with regular sampling intervals.

[0020] The purpose of the present invention can also be achieved by the following technical measures:

[0021] In step 1, the seismic data is actual seismic data or forward modeling seismic data that has been preprocessed by denoising, deconvolution, and static correction.

[0022] In step 1, the migration velocity is the time domain root mean square velocity that matches the seismic data and is suitable for prestack time migration using the integral method. The velocity can be obtained by velocity analysis and appropriate smoothing.

[0023] In step 1, if the seismic data is forward modeling data, the migration velocity can be converted from the depth-domain layer velocity used in the forward modeling and appropriately smoothed to obtain the time-domain root mean square velocity.

[0024] In step 2, multiple computing groups are established, each of which contains a certain number of Shenwei computing core groups; the seismic data is divided into multiple parts with the seismic channel as the smallest unit, and each idle computing group is assigned to calculate one of the parts.

[0025] In step 3, the calculation group reads the seismic data to be calculated into the main memory and distributes it to the Shenwei calculation core group within the group for calculation, with each core group calculating one line.

[0026] In step 4, the main core of the core group copies a seismic data into the local memory, uses the slave core to resample the seismic data according to the frequency requirement, stores the result in the local memory, and releases the local memory occupied by the original seismic data; the main core of the core group reads the corresponding velocity data according to the position of the seismic track and copies it into the local memory.

[0027] In step 5, each core group applies for memory space in the main memory to store the pre-stack time migration calculation results of the resampled seismic data; each slave core of the core group calculates a resampled seismic data sample point, caches the numerical value of the calculation result and the main memory address corresponding to the value in the local memory, and accumulates and merges the results with the same memory address in the local memory. When the cached results accumulate to a certain scale, they are imported into the main memory in batches, and the data is updated according to the memory address.

[0028] In step 6, when not all seismic data channels have been calculated, after a core group in the calculation group has completed calculation, an uncalculated channel is selected from the seismic data that the calculation group is responsible for calculating and assigned to the core group for calculation, and the process returns to step 4.

[0029] In step 6, when there is no seismic data to be allocated, after the calculation group completes the calculation, the main core is used to reduce the calculation results within the group and store them in the main memory.

[0030] In step 7, when not all calculation groups have completed the calculation, and a calculation group is idle after completing the calculation, a piece of unallocated seismic data is allocated to the group for calculation, and the process returns to step 3.

[0031] In step 7, after all computing groups have completed the calculation and result reduction, the main core is used to reduce the reduction results of all computing groups and store them in the main memory.

[0032] In step 8, the reduced results are resampled according to the sampling interval of the original seismic data or the sampling interval specified by the parameters to obtain the results of the regular sampling interval, and output to the disk, and the calculation is now completed.

[0033] The purpose of the present invention can also be achieved through the following technical measures: an integral method pre-stack time migration system based on the Shenwei architecture, which uses an integral method pre-stack time migration method based on the Shenwei architecture to optimize the parallel computing of the algorithm and improve the parallel computing speed.

[0034] The integral method prestack time migration method based on the Shenwei architecture in the present invention optimizes the parallel computing implementation method of the algorithm according to the characteristics of the Shenwei architecture and the integral method prestack time migration algorithm, which can reduce the reading and writing and network communication pressure during ultra-large-scale parallelization, reduce the number of accesses between the main memory and the local memory, and improve the parallel computing speed.

[0035] The core of the integral prestack time migration method based on the Shenwei architecture is to reduce the reading, writing and network communication pressure in ultra-large-scale parallelism through parallel grouping, and use methods such as discrete update and hot memory resident cache to reduce the number of accesses between main memory and local memory, thereby improving the running speed of the slave core.

[0036] The present invention provides an integral method prestack time migration method based on the Shenwei architecture, which can reduce the pressure on reading, writing and network communication in ultra-large-scale parallelism through a unique parallel grouping method, and at the same time improve the operating speed of the slave core by reducing the number of accesses between the main memory and the local memory. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] Figure 1 A schematic diagram of a parallel grouping structure and computing task allocation in a specific embodiment 1 of the present invention;

[0038] Figure 2 This is a schematic diagram of implementing data compression using non-uniform sampling in a specific embodiment 1 of the present invention;

[0039] Figure 3 It is a schematic diagram of discrete update in a specific embodiment 1 of the present invention;

[0040] Figure 4 This is a comparison diagram of the average time consumption per core group per time step before and after optimization in a specific embodiment 2 of the present invention;

[0041] Figure 5 The flowchart is a specific embodiment of the integral method pre-stack time migration method based on the Shenwei architecture of the present invention. DETAILED DESCRIPTION

[0042] It should be noted that the following detailed descriptions are exemplary and are intended to provide further explanation of the present invention. Unless otherwise specified, all technical and scientific terms used herein have the same meanings as those commonly understood by those skilled in the art to which the present invention belongs.

[0043] It should be noted that the terms used herein are only for describing specific embodiments and are not intended to limit exemplary embodiments according to the present invention. As used herein, unless the context clearly indicates otherwise, the singular form is also intended to include the plural form. In addition, it should be understood that when the terms "comprising" and / or "including" are used in this specification, it indicates the presence of features, steps, operations and / or combinations thereof.

[0044] The integral method prestack time migration method based on the Shenwei architecture of the present invention includes reducing the pressure of reading, writing and network communication in ultra-large-scale parallelism through parallel grouping, and using discrete updates and hotspot memory resident cache and other methods to improve the running speed of the slave core. Figure 5 As shown, Figure 5 The flowchart of the integral method prestack time migration method based on the Shenwei architecture of the present invention includes:

[0045] 101. Prepare the seismic data and migration velocity required for pre-stack time migration calculation using the integral method. The seismic data is the actual seismic data or forward modeling seismic data that has been pre-processed by denoising, deconvolution, and static correction.

[0046] The migration velocity is the time domain root mean square velocity that matches the seismic data and is suitable for integral method prestack time migration. This velocity can be obtained by velocity analysis and appropriate smoothing. If the seismic data is forward simulation data, the migration velocity can be converted from the depth domain layer velocity used in the forward simulation and appropriately smoothed to obtain the time domain root mean square velocity.

[0047] 102. Establish multiple calculation groups, each of which contains a certain number of Shenwei calculation core groups; take the seismic channel as the smallest unit, divide the seismic data into multiple parts, and specify each idle calculation group to calculate one of the parts.

[0048] 103. The calculation group reads the earthquake data to be calculated into the main memory and distributes it to the Shenwei calculation core group in the group for calculation, with each core group calculating one line.

[0049] 104. The main core of the core group copies a seismic data into the local memory, uses the slave core to resample the seismic data according to the frequency requirement, stores the result in the local memory, and releases the local memory occupied by the original seismic data. The main core of the core group reads the corresponding velocity data according to the position of the seismic trace and copies it into the local memory.

[0050] 105. Each core group applies for memory space in the main memory to store the pre-stack time migration calculation results of the resampled seismic data. Each slave core of the core group calculates a resampled seismic data sample point, caches the value of the calculation result and the main memory address corresponding to the value in the local memory, and accumulates and merges the results with the same memory address in the local memory. When the cached results accumulate to a certain scale, they are imported into the main memory in batches, and the data is updated according to the memory address.

[0051] 106. When a core group in the calculation group has completed calculation, select an uncalculated line from the seismic data that the calculation group is responsible for calculating and assign it to the core group for calculation, and jump to step 104.

[0052] 107. When a calculation group is free after completing calculation, a copy of unallocated seismic data is allocated to the group for calculation, and the process jumps to step 103.

[0053] 108. When there is no seismic data to be allocated, after the calculation group completes the calculation, the main core is used to reduce the calculation results within the group and store them in the main memory; when all calculation groups complete the calculation and result reduction, the main core is used to reduce the reduced results of all calculation groups and store them in the main memory; the reduced results are resampled according to the original seismic data sampling interval or the sampling interval specified by the parameters to obtain the results of the regular sampling interval, and output to the disk. The calculation is now complete.

[0054] The following are several specific embodiments of the present invention.

[0055] Example 1

[0056] like Figure 1 As shown in the figure, (1+M*(N+1)) processes are started on the Sunway computer, including 1 global main process and M computing groups. Each computing group contains N computing processes and 1 master process within the group. The global master process is responsible for allocating computing tasks between process groups and reducing the results of all computing groups. The master process within the group communicates with the global master process to apply for computing tasks, reads the data required for computing at one time according to the computing tasks and forwards it to the computing processes within the group. After the computing within the group is completed, the results are reduced within the group. The computing process receives the computing data of the master process within the group for computing. When each core group is computing, according to Figure 2 As shown in the figure, non-uniform resampling is performed and stored in local memory, and calculations are performed on the non-uniform resampled data. Before the calculation results are output, resampling is performed to restore the uniform sampling interval. When using the slave core for calculation within the core group, according to Figure 3 As shown, the numerical value of the calculation result and the main memory address corresponding to the value are cached in the local memory, and the results of the same memory address are accumulated and merged in the local memory. When the cached results accumulate to a certain scale, they are imported into the main memory in batches and the data is updated according to the memory address.

[0057] Example 2

[0058] In a specific embodiment 2 of the present invention, the input data is 500,000 channels in total, totaling 5.8Gb, using 800 core groups, each of which uses 64 slave cores. Start (1+10*(80+1))=811 processes for parallel computing, where 1 global master process distributes the computing tasks as evenly as possible to the master processes in the 10 process groups. Each master process in the group reads the data required for the calculation into the main memory at one time according to the assigned tasks and broadcasts it to the computing processes in the group, and distributes the computing tasks of the group as evenly as possible to the 80 computing processes in the group. The computing processes in the group perform non-uniform resampling on the data, and use slave cores for parallel computing. The calculation results are cached in the local memory and merged according to the memory address. When the cached results accumulate to a certain scale, they are imported into the main memory in batches, and the data is updated according to the memory address. When the calculation in the group is completed, the results are reduced by the computing process in the group to the master process in the group. After the group master process is completed, it continues to receive the computing tasks assigned by the master process and performs the group calculations in the above manner. The new calculation results are merged with the old calculation results of the group master process. When all process groups have completed all the computing tasks, the calculation results of all the group master process specifications are reduced to the global master process and resampled at a uniform sampling interval, and then the global master process outputs the results.

[0059] By using the parallel optimization method, the average time consumed by each core group per time step in the above application example is as follows: Figure 4 As shown, the time is shortened from 6.35s to 1.32s, with a speedup ratio of 4.81.

[0060] Example 3

[0061] In the specific embodiment 3 of the present invention, the input data totals 87,000,000 channels, 16,000 core groups are used, and each core group uses 64 slave cores. Start (1+200*(80+1))=16201 processes for parallel computing, among which 1 global master process distributes the computing tasks as evenly as possible to the master processes in the 200 process groups. Each master process in the group reads the data required for the calculation into the main memory at one time according to the assigned tasks and broadcasts it to the computing processes in the group, and distributes the computing tasks of the group as evenly as possible to the 80 computing processes in the group. The computing processes in the group perform non-uniform resampling on the data, and use slave cores for parallel computing. The calculation results are cached in the local memory and merged according to the memory address. When the cached results accumulate to a certain scale, they are imported into the main memory in batches, and the data is updated according to the memory address. When the calculation in the group is completed, the results are reduced by the computing process in the group to the master process in the group. After the group master process is completed, it continues to receive the computing tasks assigned by the master process and performs the group calculations in the above manner. The new calculation results are merged with the old calculation results of the group master process. When all process groups have completed all the computing tasks, the calculation results of all the group master process specifications are reduced to the global master process and resampled at a uniform sampling interval, and then the global master process outputs the results.

[0062] Finally, it should be noted that the above is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art can still modify the technical solutions recorded in the aforementioned embodiments or replace some of the technical features therein with equivalents. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

[0063] Except for the technical features described in the specification, all other technical features are known technologies to those skilled in the art.

Claims

1. Integral prestack time migration method based on the Shenwei architecture, It is characterized in that The integral prestack time migration method based on the Shenwei architecture includes: Step 1, prepare the seismic data and migration velocity required for prestack time migration calculation using the integral method; Step 2, establish multiple calculation groups and distribute the seismic data; Step 3, calculating seismic data; Step 4, resampling and storing the seismic data; Step 5, storing the prestack time migration calculation results of the resampled seismic data; Step 6, when all seismic data channels are calculated, the calculation results are simplified and stored; Step 7: When all calculation groups have completed the calculation, the calculation results are simplified and stored; Step 8, resampling is performed to obtain results with regular sampling intervals.

2. The integral method prestack time migration method based on the Shenwei architecture according to claim 1, It is characterized in that In step 1, the seismic data is actual seismic data or forward modeling seismic data that has been preprocessed by denoising, deconvolution, and static correction.

3. The integral method prestack time migration method based on the Shenwei architecture according to claim 1, It is characterized in that In step 1, the migration velocity is the time domain root mean square velocity that matches the seismic data and is suitable for prestack time migration using the integral method. The velocity can be obtained by velocity analysis and appropriate smoothing.

4. The integral method prestack time migration method based on the Shenwei architecture according to claim 3, It is characterized in that In step 1, if the seismic data is forward modeling data, the migration velocity can be converted from the depth-domain layer velocity used in the forward modeling and appropriately smoothed to obtain the time-domain root mean square velocity.

5. The integral method prestack time migration method based on the Shenwei architecture according to claim 1, It is characterized in that In step 2, multiple computing groups are established, each of which contains a certain number of Shenwei computing core groups; the seismic data is divided into multiple parts with the seismic channel as the smallest unit, and each idle computing group is assigned to calculate one of the parts.

6. The integral method prestack time migration method based on the Shenwei architecture according to claim 5, It is characterized in that In step 3, the calculation group reads the seismic data to be calculated into the main memory and distributes it to the Shenwei calculation core group within the group for calculation, with each core group calculating one line.

7. The integral method prestack time migration method based on the Shenwei architecture according to claim 5, It is characterized in that In step 4, the main core of the core group copies a seismic data into the local memory, uses the slave core to resample the seismic data according to the frequency requirement, stores the result in the local memory, and releases the local memory occupied by the original seismic data; the main core of the core group reads the corresponding velocity data according to the position of the seismic track and copies it into the local memory.

8. The integral method prestack time migration method based on the Shenwei architecture according to claim 5, It is characterized in that In step 5, each core group applies for memory space in the main memory to store the pre-stack time migration calculation results of the resampled seismic data; each slave core of the core group calculates a resampled seismic data sample point, caches the numerical value of the calculation result and the main memory address corresponding to the value in the local memory, and accumulates and merges the results with the same memory address in the local memory. When the cached results accumulate to a certain scale, they are imported into the main memory in batches, and the data is updated according to the memory address.

9. The integral method prestack time migration method based on the Shenwei architecture according to claim 5, It is characterized in that In step 6, when not all seismic data channels have been calculated, after a core group in the calculation group has completed calculation, an uncalculated channel is selected from the seismic data that the calculation group is responsible for calculating and assigned to the core group for calculation, and the process returns to step 4.

10. The integral method prestack time migration method based on the Shenwei architecture according to claim 5, It is characterized in that In step 6, when there is no seismic data to be allocated, after the calculation group completes the calculation, the main core is used to reduce the calculation results within the group and store them in the main memory.

11. The integral method prestack time migration method based on the Shenwei architecture according to claim 5, It is characterized in that In step 7, when not all calculation groups have completed the calculation, and a calculation group is idle after completing the calculation, a piece of unallocated seismic data is allocated to the group for calculation, and the process returns to step 3.

12. The integral method prestack time migration method based on the Shenwei architecture according to claim 5, It is characterized in that In step 7, after all computing groups have completed the calculation and result reduction, the main core is used to reduce the reduction results of all computing groups and store them in the main memory.

13. The integral method prestack time migration method based on the Shenwei architecture according to claim 1, It is characterized in that In step 8, the reduced results are resampled according to the sampling interval of the original seismic data or the sampling interval specified by the parameters to obtain the results of the regular sampling interval, and output to the disk, and the calculation is now completed.

14. Integral method pre-stack time migration system based on Shenwei architecture, It is characterized in that The integral method pre-stack time migration system based on the Shenwei architecture adopts the integral method pre-stack time migration method based on the Shenwei architecture described in any one of claims 1 to 13 to optimize the parallel computing of the algorithm and improve the parallel computing speed.

Images