Heterogeneous CPU/GPU co-execution of CFD simulations on the POWER9 architecture: Application to airplane aerodynamics

Borrell, R.; Dosimont, Damien; García-Gasulla, Marta; Houzeaux, Guillaume; Lehmkuhl, O.; Mehta, V.J.; Owen, Herbert; Vázquez, Mariano; Oyarzun, Guillermo

doi:10.1016/j.future.2020.01.045

Cited by 31 publications

(14 citation statements)

References 27 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…Author details 1 School of Computer Science and Engineering, Sun Yat-sen University, Guangzhou, China. 2 Hartree Centre, STFC Daresbury Laboratory, Warrington, UK.…”

Section: Fundingmentioning

confidence: 99%

See 1 more Smart Citation

Effects of mesh loop modes on performance of unstructured finite volume GPU simulations

et al. 2021

View full text Add to dashboard Cite

In unstructured finite volume method, loop on different mesh components such as cells, faces, nodes, etc is used widely for the traversal of data. Mesh loop results in direct or indirect data access that affects data locality significantly. By loop on mesh, many threads accessing the same data lead to data dependence. Both data locality and data dependence play an important part in the performance of GPU simulations. For optimizing a GPU-accelerated unstructured finite volume Computational Fluid Dynamics (CFD) program, the performance of hot spots under different loops on cells, faces, and nodes is evaluated on Nvidia Tesla V100 and K80. Numerical tests under different mesh scales show that the effects of mesh loop modes are different on data locality and data dependence. Specifically, face loop makes the best data locality, so long as access to face data exists in kernels. Cell loop brings the smallest overheads due to non-coalescing data access, when both cell and node data are used in computing without face data. Cell loop owns the best performance in the condition that only indirect access of cell data exists in kernels. Atomic operations reduced the performance of kernels largely in K80, which is not obvious on V100. With the suitable mesh loop mode in all kernels, the overall performance of GPU simulations can be increased by 15%-20%. Finally, the program on a single GPU V100 can achieve maximum 21.7 and average 14.1 speed up compared with 28 MPI tasks on two Intel CPUs Xeon Gold 6132.

show abstract

“…Author details 1 School of Computer Science and Engineering, Sun Yat-sen University, Guangzhou, China. 2 Hartree Centre, STFC Daresbury Laboratory, Warrington, UK.…”

Section: Fundingmentioning

confidence: 99%

“…CFD plays an important part in the design of aircraft [1], which is applied in many scenarios including predicting aerodynamics of wing [2], analyzing combustion in engine [3], and so on. More and more computing resources are required, for high precision analysis.…”

Section: Introductionmentioning

confidence: 99%

Effects of mesh loop modes on performance of unstructured finite volume GPU simulations

et al. 2021

View full text Add to dashboard Cite

show abstract

“…The problem considered in this work is the simulation of an airplane, extensively described in Borrell et al (2020). The mesh is hybrid and comprises 31.5M elements: 15.4M tetrahedra, 7K pyramids and 16.1M prisms.…”

Section: Physical Problemmentioning

confidence: 99%

“…Avoiding duplicated work on these cells, provides a certain advantage in the assembly phase for the Finite-Element method, where the scalability only depends upon the control of the load balance. Although we have treated the load imbalance due to the different types of elements in previous papers Borrell et al 2020), here, no specific strategy has been employed.…”

Section: Parallelisationmentioning

confidence: 99%

A Generic Performance Analysis Technique Applied to Different CFD Methods for HPC

García-Gasulla

Banchelli

Peiro

et al. 2020

International Journal of Computational Fluid Dynamics

View full text Add to dashboard Cite

For complex engineering and scientific applications, Computational Fluid Dynamics (CFD) simulations require a huge amount of computational power. As such, it is of paramount importance to carefully assess the performance of CFD codes and to study them in depth for enabling optimisation and portability. In this paper, we study three complex CFD codes, OpenFOAM, Alya and CHORUS representing two numerical methods, namely the finite volume and finite-element methods, on both structured and unstructured meshes. To all codes, we apply a generic performance analysis method based on a set of metrics helping the code developer in spotting the critical points that can potentially limit the scalability of a parallel application. We show the root cause of the performance bottlenecks studying the three applications on the MareNostrum4 supercomputer. We conclude providing hints for improving the performance and the scalability of each application.

show abstract

“…In contrast, the schemes used in the present work are much cheaper computationally, so the algorithm is clearly memory-bound. Another example of a CUDAbased supercomputing code for aerodynamics problems can be found in [5]. Co-execution of scale-resolving simulations on CPUs and GPUs is demonstrated there.…”

Section: Introductionmentioning

confidence: 99%

Adapting Complex and Clumsy CFD Code to Rapidly Changing Supercomputing Realities

Gorobets¹

2021

14th WCCM-ECCOMAS Congress

View full text Add to dashboard Cite

This work is devoted to acceleration and upgrade of the CFD code NOISEtte for scale-resolving simulations of compressible turbulent flows using edge-based high-accuracy methods on unstructured hybrid meshes. Attempts to extend the baseline multilevel MPI+OpenMP parallelization towards GPU-based hybrid systems have faced the problem: the code is too complex. It is an in-house research code with plenty of numerical methods, schemes, models, most of which are experimental and are not used in practical simulations. This chaotic zoo leads to excessive conditional branches, switches, redundant functional calls that slow down computations. Although the parallel algorithm is fully adapted to the stream processing paradigm, such an immense amount of code is too difficult to port efficiently to OpenCL or CUDA and maintain it in consistency with the CPU version. An approach to survive in the process of adaptation to hybrid systems has been elaborated. It consists of various components, such as creation of a simplified configurations, combining different stages of the algorithm in order to reduce memory traffic, collapsing multiple functions in one function without branches and switches, mixing single and double precision, etc. As a result, the upgraded code is about twice as fast on CPUs and can use GPUs from different manufacturers-AMD, NVIDIA, Intel through the OpenCL standard.

show abstract

Heterogeneous CPU/GPU co-execution of CFD simulations on the POWER9 architecture: Application to airplane aerodynamics

Cited by 31 publications

References 27 publications

Effects of mesh loop modes on performance of unstructured finite volume GPU simulations

Effects of mesh loop modes on performance of unstructured finite volume GPU simulations

A Generic Performance Analysis Technique Applied to Different CFD Methods for HPC

Adapting Complex and Clumsy CFD Code to Rapidly Changing Supercomputing Realities

Contact Info

Product

Resources

About