Communication method and system for CPU and coprocessor

Abstract

The invention relates to a communication method and a communication system for a CPU and a coprocessor. A wave field at T-1 moment is transmitted while the coprocessor computes the wave field at T moment; communication between the coprocessor and the CPU is achieved in an asynchronous mode, namely transmission is only to transmit a command of starting transmission, wave field continuation computation of the following coprocessor can be continued, with no need to wait for the ending of transmission. The method provided by the invention is an efficient parallel algorithm of which 40% operation time and machine-hour fee are saved compared with a conventional reverse time migration algorithm, and is a good method which is reduced in service cost and improved in exploration efficiency. The computation result is completely the same as that of the conventional method; when being applied to seismic wave imaging of actual data of an oil exploration field, the conventional method usually needs to estimate per month to complete the whole reverse time migration; the method provided by the invention can save a lot of time and machine-hour fee, and is the good method which is reduced in service cost and improved in exploration efficiency.

Description

technical field [0001] The invention belongs to the field of artificial seismic reflection wave imaging in geological surveys such as oil exploration and mineral exploration, and relates to a communication method and system for a CPU and a coprocessor. Background technique [0002] The current seismic wave imaging method uses reverse time migration as the most accurate algorithm, but its characteristics are: large amount of calculation, large storage capacity, large memory requirement, etc. These practical bottlenecks restrict the scope of application. At present, the main points of existing technologies to solve these bottlenecks include: multi-node parallelism to improve computing efficiency, GPU implementation of wave field finite difference continuation to improve computing efficiency of some codes, optimization of checkpoint methods to reduce storage, and use of compression algorithms to reduce memory requirements, etc. . Although the existing technology can be applied...

AI Technical Summary

Problems solved by technology

Although the existing technology can be applied to actual data, it is still an algorithm with the longest application period, and it needs to further improve the efficiency

[0003] It is a research hotspot in recent years that the RTM algorithm is implemented by using a CPU processor plus a high-performance coprocessor (such as GPU, MAC, FPGA, etc.). According to the requirements of the time offset algorithm, multiple coprocessors are usually required to complete the processing of one shot of data, but the data communication between coprocessors is transmitted through the PCI-E transmission bus, and its transmission efficiency is compared with that of CPU and memory. The transmission between is much lower, so the bottleneck of communication efficiency has always existed in the conventional algorithm

[0004] In 2013, Liu Shouwei, Tang Xianggong, etc. in the literature [1] (Liu Shouwei, Wang Huazhong, Chen Shengchang, etc. Research on the realization scheme of 3D reverse time migration GPU / CPU cluster. Acta Geophysics, 2013, 56(10): 3487-3496.) and literature [2] (Tang Xianggong, Kuang Bin, Du Jixiu, etc. Research and application of multi-GPU cooperative 3D pre-stack reverse time migration method. Petroleum Geophysical Exploration, 2013, 48(6): 910-914), using wave field reconstruction The algorithm design of GPU-RTM in a large-scale cluster environment is carried out in the way of GPU-RTM. The core idea is to use the reversibility of wave field propagation to simulate the source wave field for two rounds. Compared with the algorithm on the conventional CPU, this method does not require waves. Field storage, which solves the problem of large RTM storage capacity, but one more simulation of the source wave field, so the calculation amount increases by 30%, the memory requirement of the GPU also doubles, and the throughput also doubles

In practical applications, it can be foreseen that the GPU memory is far from meeting the needs of this type of algorithm. When the scale increases to a certain extent, there are still many problems in computing and communication efficiency.

In addition, if the wave field reconstruction method adopts PML absorption boundary, a large number of branch calculations will be required, which will greatly reduce the calculation efficiency of the GPU.

In 2013, Liu Shouwei mentioned in the literature [1] that the unified wave equation can be used to avoid the judgment of boundary conditions, but the calculation amount of the unified wave equation of the boundary is more than doubled compared with the conventional wave equation, and the memory requirement is also increased by one. times, this is also the Achilles' heel of TRON reconstruction

It is worth noting that there are at most two consecutive wave fields at a time, and wave fields at other moments cannot be stored in memory

Therefore, the wave field calculation of CPU processor transmission and coprocessor can only be executed serially

Images