CPU-GPU Heterogeneous Platform Research Articles

Parallelization with load balancing of the weather scheme WSM7 for heterogeneous CPU-GPU platforms

This article provides an enhanced parallelization of the WSM7 microphysics scheme for the Weather Research and Forecasting Model (WRF). The parallelization is designed to maximize the utilization of a heterogeneous computing system consisting of CPUs, GPUs or both. Therefore the reference implementation of the WSM7 scheme is re-implemented for the heterogeneous execution model. For each time step, a dynamic load distribution is introduced which balances the computational load between the two components aiming for an overall minimum execution time. The evaluation of the parallelized implementation is done for a specific weather situation. Specifically, the precipitation of the low-pressure zone “Bernd” from July 2021 is simulated using an Intel Core i7-7700 CPU and a NVIDIA GTX 1070 GPU. The results show a speedup of up to 28.51 for the GPU version in comparison with the reference implementation. The heterogeneous dynamic load balancing increases the speedup achieved even further by introducing a distribution factor that is updated for each time step.

Development of a CPU-GPU heterogeneous platform based on a nonlinear parallel algorithm

Abstract In order to seek a refined model analysis software platform that can balance both the computational accuracy and computational efficiency, a CPU-GPU heterogeneous platform based on a nonlinear parallel algorithm is developed. The modular design method is adopted to complete the architecture construction of structural nonlinear analysis software, clarify the basic analysis steps of nonlinear finite element problems, so as to determine the structure of the software system, conduct module division, and clarify the function, interface, and call relationship of each module. The results show that when the number of model layers is 10, the GPU is 210.5/s and the CPU is 1073.2/s, and the computational time of the GPU is significantly better, with an acceleration ratio of 5.1. For all the models, the GPU calculation time is much less than that of the CPU, and when the number of model degrees of freedom increases, the acceleration effect of the GPU becomes more obvious. Therefore, the CPU-GPU heterogeneous platform can more accurately describe the nonlinear behavior in the complex stress states of the shear walls, and is computationally efficient.

A heterogeneous accelerated simulation framework for wind field dynamic model

This paper proposes an efficient and extensible simulation framework for mid-fidelity wind field model based on heterogeneous architecture. The wind field model consists of 3-D wind inflow, wake dynamics, and wind turbine models. The CPU-GPU heterogeneous platform accelerates the 3-D wind inflow model simulation by performing large-scale parallel computing based on GPU. The CPU is utilized to dispatch the simulation and perform serial computing. Compared with the conventional simulation tool, the simulation speed of the 3-D wind inflow model is highly improved, which fulfils the requirement of the real-time wind field simulation. The hardware architecture of the framework is designed to be highly extensible by multiple computing nodes for 3-D wind inflow, wake dynamic, and wind turbine model. The extendability enables the ability for large-scale wind field simulation with high efficiency. A 5 km $\rm km$ × 2 km $\rm km$ × 0.35 km $\rm km$ wind field consisting of three wind turbines is used as the case. The case study was conducted and analyzed, proving that the framework can capture the detailed wake dynamic, wind turbine dynamic, and wind inflow dynamic. Meanwhile, the calculation speed of disturbed wind speed increased by 264 times.

Intelligent vehicle visual pose estimation algorithm based on deep learning and parallel computing for dynamic scenes

In view of the disadvantages of the existing pose estimation algorithm, which has low real-time performance and the positioning accuracy will be greatly reduced in dynamic scene, a compound deep learning and parallel computing algorithm (DP-PE) is proposed. The detection algorithm based on deep learning is used to detect dynamic objects in the environment, and the dynamic feature points are removed before the matching of feature points to reduce the impact of dynamic objects on the positioning accuracy; A method for distinguishing “pseudo-dynamic objects” is proposed to solve the problem that the stationary vehicles and pedestrians in the environment are regarded as dynamic objects. The parallel computing framework for feature point extraction and matching is established on CPU-GPU heterogeneous platform to speed up DP-PE; In the localization part of DP-PE, we propose a 3D interior point detection strategy to achieve parallel search of map points, and the saturated linear kernel function is used to act on reprojection error to realize the parallelization of pose optimization. We verify the algorithm on KITTI dataset, the experimental results show that average speedup ratio of feature point extraction and matching is 6.5 times, and the overall computational efficiency of DP-PE is about 7 times higher than that before acceleration, which can realize high precision and efficient pose estimation in dynamic scene.

High performance computing of stiff bubble collapse on CPU-GPU heterogeneous platform

SCB is an efficient fluid solver developed for computing two-phase compressible flows involving strong shocks and expansion waves. It solves a four-equation diffuse-interface model, which is derived from the five-equation model proposed by Kapila et al. The governing equations are discretized by a finite volume method with explicit time stepping. SCB uses a fully parallel environment via Message Passing Interfaces (MPI). With the fast growing number of heterogeneous computing platforms including disparate hardware architectures, it becomes nowadays necessary to develop hybrid parallelization strategies with a special care to portability. In this context, we present an heterogeneous computing framework based on MPI library and OpenACC. The choice of OpenACC is discussed. Performances, scalability and adaptability are illustrated through a series of tests on an heterogeneous architecture. Validations are proposed on various bubble collapses, in free-field or near a rigid wall. Comparisons are done with existing results and analytical solutions. Furthermore a stiff shock-induced bubble collapse demonstrates the capabilities and the high potential of the code.

HCGrid: a convolution-based gridding framework for radio astronomy in hybrid computing environments

ABSTRACT Gridding operation, which is to map non-uniform data samples on to a uniformly distributed grid, is one of the key steps in radio astronomical data reduction process. One of the main bottlenecks of gridding is the poor computing performance, and a typical solution for such performance issue is the implementation of multicore CPU platforms. Although such a method could usually achieve good results, in many cases, the performance of gridding is still restricted to an extent due to the limitations of CPU, since the main workload of gridding is a combination of a large number of single instruction, multidata stream operations, which is more suitable for GPU, rather than CPU implementations. To meet the challenge of massive data gridding for the modern large single-dish radio telescopes, e.g. the Five-hundred-meter Aperture Spherical radio Telescope, inspired by existing multicore CPU gridding algorithms such as Cygrid, here we present an easy-to-install, high-performance, and open-source convolutional gridding framework, HCGrid, in CPU-GPU heterogeneous platforms. It optimizes data search by employing multithreading on CPU, and accelerates the convolution process by utilizing massive parallelization of GPU. In order to make HCGrid a more adaptive solution, we also propose the strategies of thread organization and coarsening, as well as optimal parameter settings under various GPU architectures. A thorough analysis of computing time and performance gain with several GPU parallel optimization strategies show that it can lead to excellent performance in hybrid computing environments.

The Immersed Boundary-Lattice Boltzmann Method Parallel Model for Fluid-Structure Interaction on Heterogeneous Platforms

Immersed boundary-lattice Boltzmann method (IB-LBM) has become a popular method for studying fluid-structure interaction (FSI) problems. However, the performance issues of the IB-LBM have to be considered when simulating the practical problems. The Graphics Processing Units (GPUs) from NVIDIA offer a possible solution for the parallel computing, while the CPU is a multicore processor that can also improve the parallel performance. This paper proposes a parallel algorithm for IB-LBM on a CPU-GPU heterogeneous platform, in which the CPU not only controls the launch of the kernel function but also performs calculations. According to the relatively local calculation characteristics of IB-LBM and the features of the heterogeneous platform, the flow field is divided into two parts: GPU computing domain and CPU computing domain. CUDA and OpenMP are used for parallel computing on the two computing domains, respectively. Since the calculation time is less than the data transmission time, a buffer is set at the junction of two computational domains. The size of the buffer determines the number of the evolution of the flow field before the data exchange. Therefore, the number of communications can be reduced by increasing buffer size. The performance of the method was investigated and analyzed using the traditional metric MFLUPS. The new algorithm is applied to the computational simulation of red blood cells (RBCs) in Poiseuille flow and through a microchannel.

Analysis of energy efficiency of a parallel AES algorithm for CPU-GPU heterogeneous platforms

Encryption plays an important role in protecting data, especially data transferred on the Internet. However, encryption is computationally expensive and this leads to high energy costs. Parallel encryption solutions using more CPU/GPU cores can achieve high performance. If we consider energy efficiency to be cost effective using parallel encryption solutions at the same time, this problem can be alleviated effectively. Because many CPU/GPU cores and encryption are pervasive currently, saving energy cost by parallel encrypting has become an unavoidable problem. In this paper, we propose an energy-efficient parallel Advance Encryption Standard (AES) algorithm for CPU-GPU heterogeneous platforms. These platforms, such as the Green 500 computers, are popular in both high performance and general computing. Parallelizing AES algorithm, using both GPUs and CPUs, balances the workload between CPUs and GPUs based on their computing capacities. This approach also uses the Nvidia Management Library (NVML) to adjust GPU frequencies, overlaps data transfers and computation, and fully utilizes GPU computing resources to reduce energy consumption as much as possible. Experiments conducted on a platform with one K20M GPU and two Xeon E5-2640 v2 CPUs show that this approach can reduce energy consumption by 74% compared to CPU-only parallel AES algorithm and 21% compared to GPU-only parallel AES algorithm on the same platform. Its energy efficiency is 4.66 MB/Joule on average higher than both CPU-only parallel AES algorithm (1.15 MB/Joule) and GPU-only parallel AES algorithm (3.65 MB/Joule). As an energy-efficient parallel AES algorithm solution, it can be used to encrypt data on heterogeneous platforms to save energy, especially for the computers with thousands of heterogeneous nodes.

Sparse matrix partitioning for optimizing SpMV on CPU-GPU heterogeneous platforms

Sparse matrix–vector multiplication (SpMV) kernel dominates the computing cost in numerous applications. Most of the existing studies dedicated to improving this kernel have been targeting just one type of processing units, mainly multicore CPUs or graphics processing units (GPUs), and have not explored the potential of the recent, rapidly emerging, CPU-GPU heterogeneous platforms. To take full advantage of these heterogeneous systems, the input sparse matrix has to be partitioned on different available processing units. The partitioning problem is more challenging with the existence of many sparse formats whose performances depend both on the sparsity of the input matrix and the used hardware. Thus, the best performance does not only depend on how to partition the input sparse matrix but also on which sparse format to use for each partition. To address this challenge, we propose in this article a new CPU-GPU heterogeneous method for computing the SpMV kernel that combines between different sparse formats to achieve better performance and better utilization of CPU-GPU heterogeneous platforms. The proposed solution horizontally partitions the input matrix into multiple block-rows and predicts their best sparse formats using machine learning-based performance models. A mapping algorithm is then used to assign the block-rows to the CPU and GPU(s) available in the system. Our experimental results using real-world large unstructured sparse matrices on two different machines show a noticeable performance improvement.

Efficient adaptive load balancing approach for compressive background subtraction algorithm on heterogeneous CPU–GPU platforms

Mixture of Gaussians (MoG) and compressive sensing (CS) are two common approaches in many image and audio processing systems. The combination of these algorithms is recently used for the compressive background subtraction task. Nevertheless, the result of this combination has not been exploited to take advantage of the evolution of parallel computing architectures. This paper proposes an efficient strategy to implement CS-MoG on heterogeneous CPU–GPU computing platforms. This is achieved through two elements. The first one is ensuring the better acceleration and accuracy that can be achieved for this algorithm on both CPU and GPU processors: The obtained results of the improved CS-MoG are more accurate and performant than other published MoG implementations. The second contribution is the proposition of the Optimal Data Distribution Cursor ODDC, a novel adaptive data partitioning approach to exploit simultaneously the heterogeneous processors on any given platform. It aims to ensure an automatic workload balancing by estimating the optimal data chunk size that must be assigned to each processor, with taking into consideration its computing capacity. Furthermore, our method ensures an update of the partitioning at runtime to take into account any influence of data content irregularity. The experimental results, on different platforms and data sets, show that the combination of these two contributions allows reaching 98% of the maximal possible performance of targeted platforms.

On the Performance and Energy Consumption of Molecular Dynamics Applications for Heterogeneous CPU-GPU Platforms Based on Gromacs

Abstract High Performance Computing (HPC) accelerates life science discoveries by enabling scientists to analyze large data sets, to develop detailed models of entire biological systems and to simulate complex biological processes. As computational experiments, molecular dynamics simulations are widely used in life sciences to evaluate the equilibrium nature of classical many-body systems The modelling and molecular dynamics study of surfactant, polymer solutions and the stability of proteins and nucleic acids under different conditions, as well as deoxyribonucleic acid proteins are studied. The study aims to understand the scaling behavior of Gromacs (Groningen machine for chemical simulations) on various platforms, and the maximum performance in the prospect of energy consumption that can be accomplished by tuning the hardware and software parameters. Different system sizes (48K, 64K, and 272K) from scientific investigations have been studied show that the GPU (Graphics Processing Unit) scales rather beneficial than other resources, i.e., with GPU support. We track 2-3 times speedup compared to the latest multi-core CPUs. However, the so-called “threading effect” leads to the better results.

Parallel GMRES solver for fast analysis of large linear dynamic systems on GPU platforms

In this paper, we propose an efficient parallel dynamic linear solver, called GPU-GMRES, for transient analysis of large linear dynamic systems such as large power grid networks. The new method is based on the preconditioned generalized minimum residual (GMRES) iterative method implemented on heterogeneous CPU–GPU platforms. The new solver is very robust and can be applied to power grids with different structures as well as for general analysis problems for large linear dynamic systems with asymmetric matrices. The proposed GPU-GMRES solver adopts the very general and robust incomplete LU based preconditioner. We show that by properly selecting the right amount of fill-ins in the incomplete LU factors, a good trade-off between GPU efficiency and convergence rate can be achieved for the best overall performance. Such tunable feature can make this algorithm very adaptive to different problems. GPU-GMRES solver properly partitions the major computing tasks in GMRES solver to minimize the data traffic between CPU and GPUs to enhance performance of the proposed method. Furthermore, we propose a new fast parallel sparse matrix–vector (SpMV) multiplication algorithm to further accelerate the GPU-GMRES solver. The new algorithm, called segSpMV, can enjoy full coalesced memory access compared to existing approaches. To further improve the scalability and efficiency, segSpMV method is further extended to multi-GPU platforms, which leads to more scalable and faster multi-GPU GMRES solver. Experimental results on the set of the published IBM benchmark circuits and mesh-structured power grid networks show that the GPU-GMRES solver can deliver order of magnitudes speedup over the direct LU solver, UMFPACK. The resulting multi-GPU-GMRES can also deliver 3–12×speedup over the CPU implementation of the same GMRES method on transient analysis.

Optimization of minimum volume constrained hyperspectral image unmixing on CPU–GPU heterogeneous platform

Hyperspectral unmixing is essential for efficient hyperspectral image processing. Nonnegative matrix factorization based on minimum volume constraint (MVC-NMF) is one of the most widely used methods for unsupervised unmixing for hyperspectral image without the pure-pixel assumption. But the model of MVC-NMF is unstable, and the traditional solution based on projected gradient algorithm (PG-MVC-NMF) converges slowly with low accuracy. In this paper, a novel parallel method is proposed for minimum volume constrained hyperspectral image unmixing on CPU–GPU Heterogeneous Platform. First, a optimized unmixing model of minimum logarithmic volume regularized NMF is introduced and solved based on the second-order approximation of function and alternating direction method of multipliers (SO-MVC-NMF). Then, the parallel algorithm for optimized MVC-NMF (PO-MVC-NMF) is proposed based on the CPU–GPU heterogeneous platform, taking advantage of the parallel processing capabilities of GPUs and logic control abilities of CPUs. Experimental results based on both simulated and real hyperspectral images indicate that the proposed algorithm is more accurate and robust than the traditional PG-MVC-NMF, and the total speedup of PO-MVC-NMF compared to PG-MVC-NMF is over 50 times.

Optimized FFT computations on heterogeneous platforms with application to the Poisson equation

We develop optimized multi-dimensional FFT implementations on CPU–GPU heterogeneous platforms for the case when the input is too large to fit on the GPU global memory, and use the resulting techniques to develop a fast Poisson solver. The solver involves memory bound computations for which the large 3D data may have to be transferred over the PCIe bus several times during the computation. We develop a new strategy to decompose and allocate the computation between the GPU and the CPU such that the 3D data is transferred only once to the device memory, and the executions of the GPU kernels are almost completely overlapped with the PCI data transfer. We were able to achieve significantly better performance than what has been reported in previous related work, including over 145 GFLOPS for the three periodic boundary conditions (single precision version), and over 105 GFLOPS for the two periodic, one Neumann boundary conditions (single precision version). The effective bidirectional PCIe bus bandwidth achieved is 9–10 GB/s, which is close to the best possible on our platform. For all the cases tested, the single 3D data PCIe transfer time, which constitutes a lower bound on what is possible on our platform, takes almost 70% of the total execution time of the Poisson solver.