Recent progress and challenges in exploiting graphics processors in computational fluid dynamics

Niemeyer, Kyle E.; Sung, Chih-Jen

doi:10.1007/s11227-013-1015-7

Cited by 85 publications

(50 citation statements)

References 111 publications

(194 reference statements)

Supporting

Mentioning

Contrasting

Order By: Relevance

“…Using graphics processing unit (GPU) hardware as the primary computation device or accelerator in a heterogeneous system is a viable, expedient option for high-performance cluster design that helps mitigate these problems. For this reason, GPUs and other emerging coprocessor architectures are increasingly used to accelerate CFD simulations [6].GPU technology has improved rapidly in recent years; in particular, NVIDIA GPUs have progressed from Kepler to Pascal architecture in four years. This development doubled and tripled peak single-and double-precision performance, respectively [7].…”

mentioning

confidence: 99%

Accelerating solutions of one-dimensional unsteady PDEs with GPU-based swept time–space decomposition

Magee

Niemeyer

2018

Journal of Computational Physics

Self Cite

View full text Add to dashboard Cite

The expedient design of precision components in aerospace and other high-tech industries requires simulations of physical phenomena often described by partial differential equations (PDEs) without exact solutions. Modern design problems require simulations with a level of resolution difficult to achieve in reasonable amounts of time-even in effectively parallelized solvers. Though the scale of the problem relative to available computing power is the greatest impediment to accelerating these applications, significant performance gains can be achieved through careful attention to the details of memory communication and access. The swept time-space decomposition rule reduces communication between subdomains by exhausting the domain of influence before communicating boundary values. Here we present a GPU implementation of the swept rule, which modifies the algorithm for improved performance on this processing architecture by prioritizing use of private (shared) memory, avoiding interblock communication, and overwriting unnecessary values. It shows significant improvement in the execution time of finite-difference solvers for one-dimensional unsteady PDEs, producing speedups of 2-9 × for a range of problem sizes, respectively, compared with simple GPU versions and 7-300 × compared with parallel CPU versions. However, for a more sophisticated one-dimensional system of equations discretized with a second-order finite-volume scheme, the swept rule performs 1.2-1.9 × worse than a standard implementation for all problem sizes.execution-simulation at the speed of nature-in accordance with the highperformance computing development goals set out in the CFD Vision 2030 report [1]. Classic approaches to domain decomposition for parallelized, explicit, time-stepping partial differential equation (PDE) solutions incur substantial computational performance costs from the communication between nodes required every timestep. This communication cost consists of two parts: latency and bandwidth, where latency is the fixed cost of each communication event and bandwidth is the variable cost that depends on the amount of data transferred. Latency in inter-node communication is a fundamental barrier to this goal, and advancements to network latency have historically been slower than improvements in other computing performance barriers such as bandwidth and computational power [2]. Performance may be improved by avoiding external node communication until exhausting the domain of dependence, allowing the calculation to advance multiple timesteps while requiring a smaller number of communication events. This idea is the basis of swept time-space decomposition [3,4].Extreme-scale computing clusters have recently been used to solve the compressible Navier-Stokes equations on over 1.97 million CPU cores [5]. The monetary cost, power consumption, and size of such a cluster impedes the realization of widespread peta-and exa-scale computing required for real-time, high-fidelity, CFD simulations. While these are significant challenges, they also pr...

show abstract

mentioning

confidence: 99%

Accelerating solutions of one-dimensional unsteady PDEs with GPU-based swept time–space decomposition

Magee

Niemeyer

2018

Journal of Computational Physics

Self Cite

View full text Add to dashboard Cite

show abstract

“…The vectorized implementation of influence matrix assembly function "V-0" offers 5.81× speed-up compared to the version "S-0". The attained performance was 10.54 GFLOP/s (31.4 % of peak floating point performance of the Bulldozer module). -Subsequent optimization consisted in inserting #pragma _CRI pipeline, which enables software-based vector pipelining optimization in addition to hardwarebased vector pipelining.…”

Section: Optimization Of the Cpu Implementationmentioning

confidence: 98%

“…We consider that factor to be an upper bound on attainable code speed-ups between the two architectures, provided both implementations make efficient use of available resources (see [10] for a comprehensive review of challenges in exploiting GPUs in computational fluid dynamics).…”

Section: Introductionmentioning

confidence: 99%

Accelerating low-fidelity aerodynamic codes on multi- and many-core architectures

Chrust

Laurendeau

Ostiguy³

2015

J Supercomput

View full text Add to dashboard Cite

Vortex lattice and panel methods belong to a broad family of aerodynamic codes based on potential flow theory. They are used in preliminary aerodynamic studies in early stages of aircraft design where hundreds of thousands candidate configurations are analyzed. In this paper, we describe their efficient implementation on modern multiand many-core architectures. We show how to bridge the 'ninja gap', defined as the performance gap between an unoptimized C/C++ code and best optimized CPU code. We port the Vortex Lattice Method to a Graphics Processing Unit using the OpenACC standard. An elegant solution for implementation of data movements for C++ classes is also presented.

show abstract

“…Historical development of the PI method, starting from a formula derived by Onsager and Machlup [16] for the Ornstein-Uhlenbeck process to the numerical implementation of the PI method [7,17] can be found in [18]. Despite the lack of rigorous mathematical substantiation, unlike the well defined Wiener path integral [19], the PI method has been widely used in various areas of engineering, physics and finance [20][21][22][23][24]. In stochastic dynamics the PI method has been successfully adapted for Markov processes along with the Chapman-Kolmogorov equation for finding a response PDF as well as reliability characteristics of a system [25][26][27].…”

Section: Introductionmentioning

confidence: 99%

GPU computing for accelerating the numerical Path Integration approach

Alevras

Yurchenko

2016

Computers & Structures

View full text Add to dashboard Cite

The paper discusses a novel approach of accelerating the numerical Path Integration method, used for generating a stationary joint response probability density function of a dynamic system subjected to a random excitation, by the GPU computing. The paper proposes the parallelization of nested loops technique and demonstrates the advantages of GPU computing. Two, three and four dimensional in space problems are investigated as a part of the pilot project and the achieved maximum accelerations are reported. Three degree-of-freedom system (6D) is approached by the Path Integration technique for the first time. The application of the proposed GPU methodology for problems of stochastic dynamics and reliability are discussed.

show abstract

Recent progress and challenges in exploiting graphics processors in computational fluid dynamics

Cited by 85 publications

References 111 publications

Accelerating solutions of one-dimensional unsteady PDEs with GPU-based swept time–space decomposition

Accelerating solutions of one-dimensional unsteady PDEs with GPU-based swept time–space decomposition

Accelerating low-fidelity aerodynamic codes on multi- and many-core architectures

GPU computing for accelerating the numerical Path Integration approach

Contact Info

Product

Resources

About