Speeding-up FWI by One Order of Magnitude

Etienne, V.; Tonellot, Thierry; Thierry, Philippe; Berthoumieux, Vincent; Andreolli, C.

doi:10.3997/2214-4609.20141905

Cited by 5 publications

(2 citation statements)

References 0 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…This problem is becoming more tractable due to the evolution of computing systems, and FWI is now being applied to obtain subsurface parameters for both short-and long-offset acquisitions (Virieux & Operto, 2009;Vigh et al, 2014). Recent works have shown the capabilities of FWI using traditional computer architectures (Bunks et al, 1995;Thierry, Donno & Noble, 2014;Etienne et al, 2014;Cao & Liao, 2014). Furthermore, as a consequence of the development of parallel computer architectures such as Graphics Processing Units (GPUs), some studies have targeted to implementing FWI in time domain using GPUs, achieving speedups in the computation time in comparison to its serial implementation counterpart (Wang et al, 2011;Mao, Wu & Wang, 2012;Kim, Shin & Calandra, 2012;Weiss & Shragge, 2013;Zhang et al, 2014;Yang, Gao & Wang, 2015).…”

Section: Introductionmentioning

confidence: 99%

A practical implementation of acoustic full waveform inversion on graphical processing units

Abreo-Carrillo

Ramirez

Reyes

et al. 2015

CT&F Cienc. Tecnol. Futuro

View full text Add to dashboard Cite

R ecently, Full Waveform Inversion (FWI) has gained more attention in the exploration geophysics community as a data fitting method that provides high-resolution seismic velocity models. Some of FWI essential components are a cost function to measure the misfit between observed and modeled data, a wave propagator to compute the modeled data and an initial velocity model that is iteratively updated until an acceptable decrease of the cost function is reached.Since FWI is a wave equation based method, the computational costs are elevated. In this paper, it is presented a fast Graphical Processing Unit (GPU) FWI implementation that uses a 2D acoustic wave propagator in time and updates the model using the gradient of the cost function, which is efficiently computed with the adjoint state method. The proposed parallel implementation is tested using the Marmousi velocity model. The performance of the proposed implementation is evaluated using the NVIDIA GeForce GTX 860 GPU and compared to a serial Central Processing Unit (CPU) implementation, in terms of execution time. We also evaluate the GPU occupancy and analyze the memory requirements. Our tests show that the GPU implementation can achieve a speed-up of 26.89 times when compared to its serial CPU implementation. ABSTRACT 6R ecentemente, a inversão de onda completa (FWI, sigla em inglês) ganhou maior atenção na comunidade de exploração geofísica como método de ajuste de dados, que fornece modelos de velocidades sísmicas de alta resolução. Alguns dos componentes essenciais do FWI são uma função de custo para estimar a diferença entre os dados observados e os dados modelados, um propagador do campo de ondas acústicas para os dados modelados e um modelo de velocidade inicial, que é atualizada de forma iterativa.Como o FWI está baseado no método da equação da onda, as exigências computacionais de execução são altas. Neste artigo apresentamos uma implementação rápida do FWI acústico 2D em tempo em uma unidade de processamento gráfico (GPU, sigla em inglês). Esta implementação utiliza um propagador da equação de onda e atualiza o modelo de velocidade, utilizando o gradiente da função objetivo, que é calculada de forma eficiente usando o método do estado adjunto. Proposta de implementação paralela é testada utilizando o modelo de velocidade Marmousi. O desempenho da implementação proposta é avaliada usando uma GeForce GTX 860 e comparada com uma aplicação de série em, um único processador, em termos de tempo de execução. Avaliamos também a quantidade de recursos utilizados pela GPU e analisamos os requisitos de memória. Os testes mostram que a implementação em GPU pode conseguir uma taxa de aceleração de 26.89 vezes quando comparada com uma implementação serial do processador. R ecientemente, la inversión de onda completa (FWI, por sus siglas en inglés) ha ganado una mayor atención en la comunidad de exploración geofísica como un método de ajuste de datos que provee modelos de velocidades sísmicas de gran resolución. Algunos de los componentes esenciales de la FWI cor...

show abstract

Section: Introductionmentioning

confidence: 99%

A practical implementation of acoustic full waveform inversion on graphical processing units

Abreo-Carrillo

Ramirez

Reyes

et al. 2015

CT&F Cienc. Tecnol. Futuro

View full text Add to dashboard Cite

show abstract

“…In order to mitigate the computational burden, previous work has mostly focused on improving the efficiency of the computational kernel [11,13,21], which takes most of the FWI execution time. The computational kernel is responsible for simulating seismic wavefields with the purpose of measuring data misfits, generating gradients and obtaining approximations to the Hessian.…”

Section: Introductionmentioning

confidence: 99%

Acceleration strategies for elastic full waveform inversion workflows in 2D and 3D

et al. 2016

View full text Add to dashboard Cite

Full waveform inversion (FWI) is one of the most challenging procedures to obtain quantitative information of the subsurface. For elastic inversions, when both compressional and shear velocities have to be inverted, the algorithmic issue becomes also a computational challenge due to the high cost related to modelling elastic rather than acoustic waves. This shortcoming has been moderately mitigated by using high-performance computing to accelerate 3D elastic FWI kernels. Nevertheless, there is room in the FWI workflows for obtaining large speedups at the cost of proper grid pre-processing and data decimation techniques. In the present work, we show how by making full use of frequency-adapted grids, composite shot lists and a novel dynamic offset control strategy, we can reduce by several orders of magnitude the compute time while improving the convergence of the method in the studied cases, regardless of the forward and adjoint compute kernels used.

show abstract