A Hybrid Multi-GPU/CPU Computational Framework for Rotorcraft Flows on Unstructured Overset Grids

[Abstract] A method of unstructured dynamic overset grids is developed for the simulation of rotor forward flight flow and unsteady rotor-fuselage interaction aerodynamics. Two overset grids are formed by dividing the initial integral grid into two parts, namely, rotational grid and static grid, and the rotational grid rotates with the rotor while the static grid keep static during the calculating process. Also the hole fringes are formed and donor elements for hole fringe points are founded at the same time when the overset grids are formed. An appropriate overlapped area can be ensured during the unsteady calculation process, and the hole fringe needn’t to be formed repeatedly in this method. A fast searching method is used to update the donor elements, which can reduce the computational effort. The linear spring analogy is used in the rotational grid to calculate the flapping and pitching movement. The unsteady Euler equations are solved using duel time stepping to simulate a two-bladed rotor forward flight flow and Georgia Tech configuration rotor-fuselage interaction flowfield and the calculation results agree well with Experimental data.

A parallel Navier-Stokes solver based on dynamic overset unstructured grids method is presented to simulate the unsteady turbulent flow field around helicopter in forward flight. The grid method has the advantages of unstructured grid and Chimera grid and is suitable to deal with multiple bodies in relatively moving. Unsteady Navier-Stokes equations are solved on overset unstructured grids by an explicit dual time-stepping, finite volume method. Preconditioning method applied to inner iteration of the dual-time stepping is used to speed up the convergence of numerical simulation. The Spalart-Allmaras one-equation turbulence model is used to evaluate the turbulent viscosity. Parallel computation is based on the dynamic domain decomposition method in overset unstructured grids system at each physical time step. A generic helicopter Robin with a four-blade rotor in forward flight is considered to validate the method presented in this paper. Numerical simulation results show that the parallel dynamic overset unstructured grids method is very efficient for the simulation of helicopter flow field and the results are reliable.

SummaryA robust and efficient dynamic grid strategy based on an overset grid coupled with mesh deformation technique is proposed for simulating unsteady flow of flapping wings undergoing large geometrical displacement. The dynamic grid method was implemented using a hierarchical unstructured overset grid locally coupled with a fast radial basis function (RBF)‐based mapping approach. The hierarchically organized overset grid allows transferring the grid resolution for multiple blocks and overlapping/embedding the meshes. The RBF‐based mapping approach is particularly highlighted in this paper in view of its considerable computational efficiency compared with conventional RBF evaluation. The performance of the proposed dynamic mesh strategy is demonstrated by three typical unsteady cases, including a rotating rectangular block in a fixed domain, a relative movement between self‐propelled fishes and the X‐wing type flapping‐wing micro air vehicle DelFly, which displays the clap‐and‐fling wing‐interaction phenomenon on both sides of the fuselage. Results show that the proposed method can be applied to the simulation of flapping wings with satisfactory efficiency and robustness. Copyright © 2016 John Wiley & Sons, Ltd.

An effort to develop and integrate grid-generation and flow-solver tools capable of computing the flow over a commercial subsonic aircraft in a high-lift configuration is described. An overset-grid approach is utilized. Surface grids are generated using the Boeing Aero-Grid-and-Paneling-System plus the NASA Chimera-Grid-Tools package. Volume grids are efficiently generated using new enhancements to the HYPGEN hyperbolic-grid program and a new localelliptic grid generator. The overset grids are linked together using the PEGASUS software and automatically generated hole-cutting surfaces, which are necessary for dealing with the small gaps between high-lift elements. The grid-generation system is applied to a Boeing 747PD high-lift configuration, and the flow field is computed using the OVERFLOW program. The computed lift-coefficient result for an angle-ofattack of 8 deg is in agreement with an experimental value.

Compared to traditional body-fitted moving grids and overset grids, adaptive Cartesian grids offer significant advantages, including no need for grid deformation, highly automated generation process, natural adaptability to complex geometries, and ease of implementing solution-based adaptive mesh refinement. However, in high Reynolds number simulations involving large motions, these grids face the issue of “grid catastrophe” due to the extreme grid resolution required for turbulent boundary layer simulation. To address this, we employ the immersed boundary-dynamic wall function model combined with an implicit dual-time-stepping method, effectively resolving near-wall flow structures without significantly increasing grid count. To handle the assignment of newly emerged Eulerian cells during motion, a three-point interpolation method is used for flow field initialization. This computational framework has been validated through a series of steady-state benchmark cases. Through simulations of typical dive processes and typical pitching airfoil benchmarks, we demonstrate the high efficiency and fidelity of our method in solving turbulent flow fields under dynamic boundary conditions at subsonic and critical sonic speeds. This method provides efficient and reliable numerical simulation support for the aerodynamic design and optimization of various new aircraft.

General Purpose Graphics Processing Units (GPGPU) have been used in high performance computing platforms to accelerate the performance of scientific applications such as simulations. With the increased computing resources required for large-scale network simulation, one GPU device may not have enough memory and computation capacities. It is therefore necessary to enhance the system scalability by introducing multiple GPU devices. It is also attractive to investigate the performance scalability of Multi-GPU simulations. This paper describes the simulation of information propagation on multiple GPU devices, including the optimized network simulation algorithms, the network partitioning and replication strategy, and the data synchronization scheme. The experimental results for scalable random networks show that the number of simulation steps, computation time, synchronization time, and data transfer time all affect the overall simulation performance. In order to compare with random networks, we also conduct simulations of scale-free networks. We can observe that the node replication ratio in scale-free networks is smaller than that in random networks and therefore the cost of data transfer and synchronization is significantly reduced. This indicates that the network structure is also an important factor that influences the simulation performance in a Multi-GPU system.

Recently, General Purpose Graphical Processing Units (GP-GPUs) have been identified as an intriguing technology to accelerate numerous data-parallel algorithms. Several GPU architectures and programming models are beginning to emerge and establish their niche in the High-Performance Computing (HPC) community. New massively parallel architectures such as the Nvidia’s Fermi and AMD/ATi’s Radeon pack tremendous computing power in their large number of multiprocessors. Their performance is unleashed using one of the two GP-GPU programming models: Compute Unified Device Architecture (CUDA) and Open Computing Language (OpenCL). Both of them offer constructs and features that have direct bearing on the application runtime performance. In this paper, we compare the two GP-GPU architectures and the two programming models using a two-level character recognition network. The two-level network is developed using four different Spiking Neural Network (SNN) models, each with different ratios of computation-to-communication requirements. To compare the architectures, we have chosen the two extremes of the SNN models for implementation of the aforementioned two-level network. An architectural performance comparison of the SNN application running on Nvidia’s Fermi and AMD/ATi’s Radeon is done using the OpenCL programming model exhausting all of the optimization strategies plausible for the two architectures. To compare the programming models, we implement the two-level network on Nvidia’s Tesla C2050 based on the Fermi architecture. We present a hierarchy of implementations, where we successively add optimization techniques associated with the two programming models. We then compare the two programming models at these different levels of implementation and also present the effect of the network size (problem size) on the performance. We report significant application speed-up, as high as 1095× for the most computation intensive SNN neuron model, against a serial implementation on the Intel Core 2 Quad host. A comprehensive study presented in this paper establishes connections between programming models, architectures and applications.

Multi-GPU-based Swendsen–Wang multi-cluster algorithm for the simulation of two-dimensional [formula omitted]-state Potts model

Significant efforts have been made in accelerating computer vision and machine learning algorithms by utilizing parallel processors such as multi-core CPUs and GPUs. Although the suitability of GPU is well-known for computer graphics and image processing applications which require massively parallel floating-point computations, recent research movement towards general purpose computing on-GPU (GPGPU) makes it possible to take advantage of parallel processors to accelerate text processing applications as well. However, how to fully leverage different types of parallel processor architectures to obtain optimal performance (especially with text) without making specific efforts to each platform still remains a great challenge. We applied performance and accuracy enhancements to Naive Bayes algorithm to develop a practically sound implementation of text classification. A platform-aware dynamic configuration support automation flow is also proposed to support the seamless execution of our work across platforms. Experiments on various (integrated graphics, dedicated multiple GPUs) platforms demonstrate that our proposed approach improves both accuracy and performance of text classification.

The propeller slipstream is numerically simulated by solving three-dimensional unsteady Euler equations. The method of unstructured overset grids is applied to the simulation of relative motion between the propeller and airframe. The rotationalmotion of the propeller is simulated by establishing a rotational subzone, and the airframe is contained in the stationary subzone. During the unsteady simulation with the rotational subzone rotating, there is always an overlapping area between the rotational and stationary subzones, and the overlapping area is used to convect the flow variables of the two subzones. The governing equations are solved by using the dual-time method and standard explicit four-stage Runge-Kutta method. The unsteady flowfields of different working conditions are calculated, and the influence of propeller slipstream is analyzed. Some calculating results are compared with experimental data, and they have a good agreement with each other. It is demonstrated that the present method is efficient and robust for the simulation of an unsteady propeller slipstream flow.

This work presents a numerical simulation of the flow field in a liquid propellant rocket engine chamber and exit nozzle using techniques to allow the results to be taken as starting points for designing those propulsive systems. This was done using a Finite Volume method simulating the different flow regimes which usually take place in those systems. As the flow field has regions ranging from the low subsonic to the supersonic regimes, the numerical code used, initially developed for compressible flows only, was modified to work proficiently in the whole velocity range. It is well known that codes have been developed in CFD, for either compressible or incompressible flows, the joint treatment of both together being complex even today, given the small number of references available in this area. Here an existing code for compressible flow was used and primitive variables, the pressure, the Cartesian components of the velocity and the temperature instead of the conserved variables were introduced in the Euler and Navier-Stokes equations. This was done to permit the treatment at any Mach number. Unstructured meshes with adaptive refinements were employed here. The convective terms were treated with upwind first and second order methods. The numerical stability was kept with artificial dissipation and in the spatial coverage one used a five stage Runge-Kutta scheme for the Fluid Mechanics and the VODE (Value of Ordinary Differential Equations) scheme along with the Chemkin II in the chemical reacting solution. During the development of this code simulating the flow in a rocket engine, comparison tests were made with several different types of internal and external flows, at different velocities, seeking to establish the confidence level of the techniques being used. These comparisons were done with existing theoretical results and with other codes already validated and well accepted by the CFD community.

This study proposes a method for modelling the flow interaction of multiple moving objects where the flow field variables are communicated between multiple separate moving computational domains. Instead of using the conventional approach with a single fixed computational domain covering the whole flow field, this method advances the moving computational domain (MCD) method in which the computational domain itself moves in line with the motions of an object inside. The computational domains created around each object move independently, and the flow fields of each domain interact where the flows cross. This eliminates the spatial restriction for simulating multiple moving objects. Firstly, a shock tube test verifies that the overset implementation and grid movement do not adversely affect the results and that there is communication between the grids. A second test case is conducted in which two spheres are crossed, and the forces exerted on one object due to the other’s crossing at a short distance are calculated. The results verify the reliability of this method and show that it is applicable to the flow interaction of multiple moving objects.

This study presents a parallel acceleration framework for computational fluid dynamics (CFD) simulations in OpenFOAM through the integration of domestic general-purpose GPU (GPGPU) accelerator cards with two novel matrix coloring algorithms based on the LDU matrix format. Focusing on preconditioning techniques optimization, we specifically address the parallel implementation challenges of Diagonal Incomplete LU (DILU) decomposition and Gauss-Seidel smoother algorithms. To resolve critical data race issues inherent in parallel computations, we develop two innovative coloring strategies: (1) a hierarchical parallelization method utilizing the matrix's directed acyclic graph (DAG) topology for dependency sorting, and (2) an enhanced multicolor ordering algorithm derived from red-black Gauss-Seidel principles. Experimental validation using both motorBike benchmark and 3D lid-driven cavity cases demonstrates the proposed methods' computational efficacy. Comparative analysis reveals that our GPU-accelerated implementation achieves 5× speedup for standalone DILU/DIC preconditioners and 4-10× acceleration for hybrid DILU-Gauss-Seidel configurations compared to conventional CPU-based approaches. The developed framework maintains numerical accuracy while significantly enhancing solution efficiency, providing a robust acceleration paradigm for large-scale CFD simulations in heterogeneous computing environments.

In this paper, we demonstrate a machine learning algorithm with multiple GPU processing in hematocytes detection. Object detection in compound microscopy images presents a specific task. Microscopy image import directly to a specimen slide under the compound microscope by an image sensor device. We propose Faster R-CNN with a customizing model applying cutting-edge object detection systems. The MicrosisDCN (Microbes Diagnosis with Deep Convolutional Neural Network) deploys Faster R-CNN in python script configures to access multiple GPUs computation with 7168 CUDA cores of dual GPUs with Linux command options: worker_replicas and num_clones equal to several graphics processing units. The datasets consist of hematocytes extracted from raw slides under a microscope. These images separate the cells of interest into three groups: red blood cell (RBCs), white blood cell (WBCs), and platelets. The training dataset consists of 80 percent of the 40,000 images. And the testing dataset consists of 20 percent of the 40,000 images. Our algorithm also provides the result of mean average precision (mAP) and enables multiple GPU training models in Tensorflow and OpenCV. The mAP is the average of average precision (AP) with intersection-over-union (IoU) in measuring the score of object detection accuracy. If the mAP score approaches 1.0, it indicates well accuracy. Our 9,000 steps valuation algorithm model by the python script estimates the mAP in 3 groups: RBCs, WBCs, and platelets as about 0.9147, 0.9664, and 0.9548. The optimum would be at nearly 12,000 steps because the algorithm estimates the mAP model as 0.96 in all types. The experiment aims to verify the neural network model using a compound microscope.

Hybrid parallel iterative sparse linear solver framework for reservoir geomechanical and flow simulation