Intensive Computational Cost Research Articles

A unified spatio-temporal parallelization framework for accelerated Monte Carlo radiobiological modeling of electron tracks and subsequent radiation chemistry

Monte Carlo (MC) nano-scale modeling of the cellular damage is desirable but most times is prohibitive for large scaled systems due to their intensive computational cost. In this study a parallelized computational framework is presented, for accelerated MC simulations of both particle propagation and subsequent radiation chemistry at the subcellular level. Given the inherent parallelism of the electron tracks, the physical stage was “embarrassingly parallelized” into a number of independent tasks. For the chemical stage, the diffusion–reaction of the radical species was simulated with a time-driven kinetic Monte Carlo algorithm (KMC) based on the Smoluchowski formalism and the parallelization was realized by employing a spatio-temporal linked-list cell method based on a spatial subdivision with a uniform grid. The evaluation of our method was established on two metrics: speedup and efficiency. The results indicated a linear speedup ratio for the physical stage and a linear latency for shared- versus a distributed-memory system with a maximum of 3.6⋅10−3% per electron track. For the chemical stage, a series of simulations were performed to show how the execution time per step was scaling with respect to the number of radical species and a 5.7× speedup was achieved when a larger number of reactants were simulated and eight processors were employed. The simulations were deployed on the Amazon EC2 infrastructure. It is also elucidated how the overhead started becoming significant as the number of reactant species decrease relative to the number of processors. The method reported here lays the methodological foundations for accelerated MC simulations and allows envisaging a future use for large-scale radiobiological modeling of multi-cellular systems involved into a clinical scenario.

The issues of prestack reverse time migration and solutions with Graphic Processing Unit implementation

ABSTRACTPrestack reverse time migration (RTM) is a very useful tool for seismic imaging but has mainly three bottlenecks: highly intensive computation cost, low‐frequency band imaging noise and massive memory demand. Traditionally, PC‐clusters with thousands of computation nodes are used to perform RTM but it is too expensive for small companies and oilfields. In this article, we use Graphic Processing Unit (GPU) architecture, which is cheaper and faster to implement RTM and we obtain an order of magnitude higher speedup ratio to solve the problem of intensive computation cost. Aiming at the massive memory demand, we adopt the pseudo random boundary condition that sacrifices the computation cost but reduces the memory demand. For rugged topography RTM, it is difficult to deal with the rugged free boundary condition with the finite difference method. We employ a simplified boundary condition that avoids the abundant logical judgment to make the GPU implementation possible and does not induce any sacrifice on efficiency. Besides, we have also done some tests on multi‐GPU implementation for wide azimuth geometries using the latest GPU cards and drivers. Finally, we discuss the challenges of anisotropy RTM and GPU solutions. All the jobs stated above are based on GPU and the synthetic data examples will show the efficiency of the algorithm and solutions.

Denoising and Storage in Seismic Reverse Time Migration

AbstractPre‐stack reverse time migration (RTM) is a very useful tool for seismic imaging. It has, however, some problems such as highly intensive computation cost, low‐frequency imaging noise and massy memory demand. The problem of time consuming for calculation can be solved by GPU/CPU collaborative computation. This paper focuses on suppressing noise and data storage. First, we analyze the generation mechanism of low‐frequency imaging noise and suggest that the phase and frequency spectra of seismic data are modified before migration and a Laplacian filter is used to remove the low wavenumber noise efficiently. Aiming at the problem of massy memory demand, we adopt the random boundary condition which sacrifices the computation cost but reduces the memory demand. The implementation can be performed on GPU and the communication between CPU and GPU can be saved to reduce the computation cost in another way. The tests on synthetic data examples illustrate that the difference between the migration result from the method of random boundary condition and that from the traditional method of saving the footprint of the wavefield can be overlooked.

The Algorithm of High Order Finite Difference Pre‐Stack Reverse Time Migration and GPU Implementation

AbstractPre‐stack reverse time migration (RTM) is a very useful tool for seismic imaging. But it has not been widely used because of the highly intensive computation cost, imaging noise and massy memory demand. In this paper, we present the implementation process of RTM and analyze the stability condition and dispersion relation of the finite difference (FD) method. For the problem of intensive computation cost, we use the Graphic Processing Unit (GPU) architecture to realize RTM and get an order of magnitude higher speedup ratio compared to the traditional CPU architecture. Compared to the one way wave equation migration methods, RTM does not have the imaging dip limit and the imaging effect is significantly improved. The tests on vertical fault and BP models prove this conclusion. The problems of imaging noise removal and massy memory demand will be discussed in other papers.

Practical issues of reverse time migration: true amplitude gathers, noise removal and harmonic-source encoding

Recently, reverse-time migration (RTM) has drawn a lot of attention in the industry. Unlike one-way wave equation migration, RTM does not need to deal with the theory of singular pseudo-differential operators. A straightforward implementation of RTM correctly handles complex velocities and produces a complete set of acoustic waves (reflections, refractions, diffractions, multiples, evanescent waves, etc.). The RTM propagator also carries the correct propagation amplitude and imposes no dip limitations on the image. In the past, the strong migration artifacts and the intensive computational cost have been the two major problems that prevented RTM from being used in production. In this paper, we first formulate RTM based on inversion theory and then we address some solutions to suppress the low frequency migration artifacts. At the end, we propose harmonic-source migration as a way to improve the efficiency of delayed-shot RTM.

Far-Field Decoupled Basis for the Method of Moments-2d cAse - Abstract

When solving electromagnetic problems, integral equation methods, such as the Method of Moments, are known to yield to dense matrix systems whose solution usually supposes an intensive computational cost. Continuous efforts have been developed to overcome this drawback. In this paper, the goal is to generate a set of basis functions defined over the entire domain and characterized by a null radiated power coupling in the far-field region. This behavior permits to diminish the density of unknowns per wavelength, considering only a reduced set of the new basis. Furthermore, the resulting matrix has a sparse structure. All these factors amounts to considerable savings in terms of computer storage and CPU requirements.

Far-Field Decoupled Basis for the Method of Moments-2D Case

When solving electromagnetic problems, integral equation methods, such as the Method of Moments, are known to yield to dense matrix systems whose solution usually supposes an intensive computational cost. Continuous efforts have been developed to overcome this drawback. In this paper, the goal is to generate a set of basis functions defined over the entire domain and characterized by a null radiated power coupling in the far-field region. This behavior permits to diminish the density of unknowns per wavelength, considering only a reduced set of the new basis. Furthermore, the resulting matrix has a sparse structure. All these factors amounts to considerable savings in terms of computer storage and CPU requirements.

A Bayesian probabilistic approach for structure damage detection

A Bayesian probabilistic approach is presented for the damage detection of multistorey frame structures. In this paper, a Bayesian probabilistic approach is applied to identify multiple damage locations using estimated modal parameters when (1) the measurement data are potentially corrupted with noise, (2) only a small number of degrees of freedom are measured, and (3) a few fundamental modes are estimated. To reduce the potentially intensive computational cost of the proposed method, a branch-and-bound search scheme is proposed and a simplified approach for the modelling of multistorey frame structures is employed. A six-storey shear frame example and two multistorey frame examples, with multiple damage locations, are presented to illustrate the applicability of the proposed approach. © 1997 John Wiley & Sons, Ltd.