Overlapping communication and computation of GPU/CPU heterogeneous parallel spatial domain decomposition MOC method

2020 ◽  
Vol 135 ◽  
pp. 106988 ◽  
Author(s):  
Liang Liang ◽  
Qian Zhang ◽  
Peitao Song ◽  
Zhijian Zhang ◽  
Qiang Zhao ◽  
...  
Author(s):  
Moresh J. Wankhede ◽  
Neil W. Bressloff ◽  
Andy J. Keane

Computational fluid dynamics (CFD) simulations to predict and visualize the reacting flow dynamics inside a combustor require fine resolution over the spatial and temporal domain, making them computationally very expensive. The traditional time-serial approach for setting up a parallel combustor CFD simulation is to divide the spatial domain between computing nodes and treat the temporal domain sequentially. However, it is well known that spatial domain decomposition techniques are not very efficient especially when the spatial dimension (or mesh count) of the problem is small and a large number of nodes are used, as the communication costs due to data parallelism becomes significant per iteration. Hence, temporal domain decomposition has some attraction for unsteady simulations, particularly on relatively coarse spatial meshes. The purpose of this study is two-fold: (i), to develop a time-parallel CFD simulation method and apply it to solve the transient reactive flow-field in a combustor using an unsteady Reynolds-averaged Navier Stokes (URANS) formulation in the commercial CFD code FLUENT™ and (ii) to investigate its benefits relative to a time-serial approach and its potential use for combustor design optimization. The results show that the time-parallel simulation method correctly captures the unsteady combustor flow evolution but, with the applied time-parallel formulation, a clear speed-up advantage, in terms of wall-clock time, is not obtained relative to the time-serial approach. However, it is clear that the time-parallel simulation method provides multiple stages of transient combustor flow-field solution data whilst converging towards a final converged state. The availability of this resulting data could be used to seed multiple levels of fidelity within the framework of a multi-fidelity co-Kriging based design optimization strategy. Also, only a single simulation would need to be setup from which multiple fidelities are available.


2014 ◽  
Vol 14 (1) ◽  
pp. 151-162 ◽  
Author(s):  
Paul Ellinghaus ◽  
Josef Weinbub ◽  
Mihail Nedjalkov ◽  
Siegfried Selberherr ◽  
Ivan Dimov

2018 ◽  
Author(s):  
Ruonan Shi ◽  
Jae-Hun Jung ◽  
Ferdinand Schweser

AbstractThe MRI image is obtained in the spatial domain from the given Fourier coefficients in the frequency domain. It is costly to obtain the high resolution image because it requires higher frequency Fourier data while the lower frequency Fourier data is less costly and effective if the image is smooth. However, the Gibbs ringing, if existent, prevails with the lower frequency Fourier data. We propose an efficient and accurate local reconstruction method with the lower frequency Fourier data that yields sharp image profile near the local edge. The proposed method utilizes only the small number of image data in the local area. Thus the method is efficient. Furthermore the method is accurate because it minimizes the global effects on the reconstruction near the weak edges shown in many other global methods for which all the image data is used for the reconstruction. To utilize the Fourier method locally based on the local non-periodic data, the proposed method is based on the Fourier continuation method. This work is an extension of our previous 1D Fourier domain decomposition method to 2D Fourier data. The proposed method first divides the MRI image in the spatial domain into many subdomains and applies the Fourier continuation method for the smooth periodic extension of the subdomain of interest. Then the proposed method reconstructs the local image based on L2 minimization regularized by the L1 norm of edge sparsity to sharpen the image near edges. Our numerical results suggest that the proposed method should be utilized in dimension-by-dimension manner instead of in a global manner for both the quality of the reconstruction and computational efficiency. The numerical results show that the proposed method is effective when the local reconstruction is sought and that the solution is free of Gibbs oscillations.


Sign in / Sign up

Export Citation Format

Share Document