Efficient magnetohydrodynamic simulations on distributed multi-GPU systems using a novel GPU Direct–MPI hybrid approach

Modern graphics processing units (GPUs) have been widely utilized in magnetohydrodynamic (MHD) simulations in recent years. Due to the limited memory of a single GPU, distributed multi-GPU systems are needed to be explored for large-scale MHD simulations. However, the data transfer between GPUs bottlenecks the efficiency of the simulations on such systems. In this paper we propose a novel GPU Direct–MPI hybrid approach to address this problem for overall performance enhancement. Our approach consists of two strategies: (1) We exploit GPU Direct 2.0 to speedup the data transfers between multiple GPUs in a single node and reduce the total number of message passing interface (MPI) communications; (2) We design Compute Unified Device Architecture (CUDA) kernels instead of using memory copy to speedup the fragmented data exchange in the three-dimensional (3D) decomposition. 3D decomposition is usually not preferable for distributed multi-GPU systems due to its low efficiency of the fragmented data exchange. Our approach has made a breakthrough to make 3D decomposition available on distributed multi-GPU systems. As a result, it can reduce the memory usage and computation time of each partition of the computational domain. Experiment results show twice the FLOPS comparing to common 2D decomposition MPI-only implementation method. The proposed approach has been developed in an efficient implementation for MHD simulations on distributed multi-GPU systems, called MGPU–MHD code. The code realizes the GPU parallelization of a total variation diminishing (TVD) algorithm for solving the multidimensional ideal MHD equations, extending our work from single GPU computation (Wong et al., 2011) to multiple GPUs. Numerical tests and performance measurements are conducted on the TSUBAME 2.0 supercomputer at the Tokyo Institute of Technology. Our code achieves 2 TFLOPS in double precision for the problem with 1200³ grid points using 216 GPUs.     

Efficient magnetohydrodynamic simulations on graphics processing units with CUDA

Magnetohydrodynamic (MHD) simulations based on the ideal MHD equations have become a powerful tool for modeling phenomena in a wide range of applications including laboratory, astrophysical, and space plasmas. In general, high-resolution methods for solving the ideal MHD equations are computationally expensive and Beowulf clusters or even supercomputers are often used to run the codes that implemented these methods. With the advent of the Compute Unified Device Architecture (CUDA), modern graphics processing units (GPUs) provide an alternative approach to parallel computing for scientific simulations. In this paper we present, to the best of the author?s knowledge, the first implementation of MHD simulations entirely on GPUs with CUDA, named GPU-MHD, to accelerate the simulation process. GPU-MHD supports both single and double precision computations. A series of numerical tests have been performed to validate the correctness of our code. Accuracy evaluation by comparing single and double precision computation results is also given. Performance measurements of both single and double precision are conducted on both the NVIDIA GeForce GTX 295 (GT200 architecture) and GTX 480 (Fermi architecture) graphics cards. These measurements show that our GPU-based implementation achieves between one and two orders of magnitude of improvement depending on the graphics card used, the problem size, and the precision when comparing to the original serial CPU MHD implementation. In addition, we extend GPU-MHD to support the visualization of the simulation results and thus the whole MHD simulation and visualization process can be performed entirely on GPUs.     

Accelerating incompressible flow computations with a Pthreads-CUDA implementation on small-footprint multi-GPU platforms

Graphics processor units (GPU) that are originally designed for graphics rendering have emerged as massively-parallel “co-processors” to the central processing unit (CPU). Small-footprint multi-GPU workstations with hundreds of processing elements can accelerate compute-intensive simulation science applications substantially. In this study, we describe the implementation of an incompressible flow Navier–Stokes solver for multi-GPU workstation platforms. A shared-memory parallel code with identical numerical methods is also developed for multi-core CPUs to provide a fair comparison between CPUs and GPUs. Specifically, we adopt NVIDIA’s Compute Unified Device Architecture (CUDA) programming model to implement the discretized form of the governing equations on a single GPU. Pthreads are then used to enable communication across multiple GPUs on a workstation. We use separate CUDA kernels to implement the projection algorithm to solve the incompressible fluid flow equations. Kernels are implemented on different memory spaces on the GPU depending on their arithmetic intensity. The memory hierarchy specific implementation produces significantly faster performance. We present a systematic analysis of speedup and scaling using two generations of NVIDIA GPU architectures and provide a comparison of single and double precision computational performance on the GPU. Using a quad-GPU platform for single precision computations, we observe two orders of magnitude speedup relative to a serial CPU implementation. Our results demonstrate that multi-GPU workstations can serve as a cost-effective small-footprint parallel computing platform to accelerate computational fluid dynamics (CFD) simulations substantially.     

A parallel lattice Boltzmann method for large eddy simulation on multiple GPUs

To improve the simulation efficiency of turbulent fluid flows at high Reynolds numbers with large eddy dynamics, a CUDA-based simulation solution of lattice Boltzmann method for large eddy simulation (LES) using multiple graphics processing units (GPUs) is proposed. Our solution adopts the “collision after propagation” lattice evolution way and puts the misaligned propagation phase at global memory read process. The latest GPU platform allows a single CPU thread to control up to four GPUs that run in parallel. In order to make use of multiple GPUs, the whole working set is evenly partitioned into sub-domains. We implement Smagorinsky model and Vreman model respectively to verify our multi-GPU solution. These two LES models have different relaxation time calculation behavior and lead to different CUDA implementation characteristics. The implementation based on Smagorinsky model achieves 190 times speedup over the sequential implementation on CPU, while the implementation based on Vreman model archives more than 90 times speedup. The experimental results show that the parallel performance of our multi-GPU solution scales very well on multiple GPUs. Therefore large-scale (up to 10,240 $\times $ 10,240 lattices) LES–LBM simulation becomes possible at a low cost, even using double-precision floating point calculation.     

Simulation of reaction diffusion processes over biologically relevant size and time scales using multi-GPU workstations

Simulation of in vivo cellular processes with the reaction–diffusion master equation (RDME) is a computationally expensive task. Our previous software enabled simulation of inhomogeneous biochemical systems for small bacteria over long time scales using the MPD-RDME method on a single GPU. Simulations of larger eukaryotic systems exceed the on-board memory capacity of individual GPUs, and long time simulations of modest-sized cells such as yeast are impractical on a single GPU. We present a new multi-GPU parallel implementation of the MPD-RDME method based on a spatial decomposition approach that supports dynamic load balancing for workstations containing GPUs of varying performance and memory capacity. We take advantage of high-performance features of CUDA for peer-to-peer GPU memory transfers and evaluate the performance of our algorithms on state-of-the-art GPU devices. We present parallel efficiency and performance results for simulations using multiple GPUs as system size, particle counts, and number of reactions grow. We also demonstrate multi-GPU performance in simulations of the Min protein system in E. coli. Moreover, our multi-GPU decomposition and load balancing approach can be generalized to other lattice-based problems.     

CUDA体系结构上的一层浅水系统的模拟

对于使用支持NVIDACUDA程序设计模型的GPU的二维一层浅水系统,给出了如何加速平衡性良好的有限体积模式的数值解,同时给出并实现了在单双浮点精度下使用CUDA模型利用潜在数据并行的算法。数值实验表明,CUDA体系结构的求解程序比CPU并行实现求解程序高效。     

Accelerating dissipative particle dynamics with multiple GPUs

Dissipative particle dynamics (DPD) simulation is implemented on multiple GPUs by using NVIDIA’s Compute Unified Device Architecture (CUDA) in this paper. Data communication between each GPU is executed based on the POSIX thread. Compared with the single-GPU implementation, this implementation can provide faster computation speed and more storage space to perform simulations on a significant larger system. In benchmark, the performance of GPUs is compared with that of Material Studio running on a single CPU core. We can achieve more than 90x speedup by using three C2050 GPUs to perform simulations on an 80∗80∗80 system. This implementation is applied to the study on the dispersancy of lubricant succinimide dispersants. A series of simulations are performed on lubricant–soot–dispersant systems to study the impact factors including concentration and interaction with lubricant on the dispersancy, and the simulation results are agreed with the study in our present work.     

High-performance cone beam reconstruction using CUDA compatible GPUs

Compute unified device architecture (CUDA) is a software development platform that allows us to run C-like programs on the nVIDIA graphics processing unit (GPU). This paper presents an acceleration method for cone beam reconstruction using CUDA compatible GPUs. The proposed method accelerates the Feldkamp, Davis, and Kress (FDK) algorithm using three techniques: (1) off-chip memory access reduction for saving the memory bandwidth; (2) loop unrolling for hiding the memory latency; and (3) multithreading for exploiting multiple GPUs. We describe how these techniques can be incorporated into the reconstruction code. We also show an analytical model to understand the reconstruction performance on multi-GPU environments. Experimental results show that the proposed method runs at 83% of the theoretical memory bandwidth, achieving a throughput of 64.3 projections per second (pps) for reconstruction of 512³-voxel volume from 360 512²-pixel projections. This performance is 41% higher than the previous CUDA-based method and is 24 times faster than a CPU-based method optimized by vector intrinsics. Some detailed analyses are also presented to understand how effectively the acceleration techniques increase the reconstruction performance of a naive method. We also demonstrate out-of-core reconstruction for large-scale datasets, up to 1024³-voxel volume.     

Exploiting hierarchy parallelism for molecular dynamics on a petascale heterogeneous system

Heterogeneous systems with nodes containing more than one type of computation units, e.g., central processing units (CPUs) and graphics processing units (GPUs), are becoming popular because of their low cost and high performance. In this paper, we have developed a Three-Level Parallelization Scheme (TLPS) for molecular dynamics (MD) simulation on heterogeneous systems. The scheme exploits multi-level parallelism combining (1) inter-node parallelism using spatial decomposition via message passing, (2) intra-node parallelism using spatial decomposition via dynamically scheduled multi-threading, and (3) intra-chip parallelism using multi-threading and short vector extension in CPUs, and employing multiple CUDA threads in GPUs. By using a hierarchy of parallelism with optimizations such as communication hiding intra-node, and memory optimizations in both CPUs and GPUs, we have implemented and evaluated a MD simulation on a petascale heterogeneous supercomputer TH-1A. The results show that MD simulations can be efficiently parallelized with our TLPS scheme and can benefit from the optimizations.     

Introducing and Implementing the Allpairs Skeleton for Programming Multi-GPU Systems

Algorithmic skeletons simplify software development: they abstract typical patterns of parallelism and provide their efficient implementations, allowing the application developer to focus on the structure of algorithms, rather than on implementation details. This becomes especially important for modern parallel systems with multiple graphics processing units (GPUs) whose programming is complex and error-prone, because state-of-the-art programming approaches like CUDA and OpenCL lack high-level abstractions. We define a new algorithmic skeleton for allpairs computations which occur in real-world applications, ranging from bioinformatics to physics. We develop the skeleton’s generic parallel implementation for multi-GPU Systems in OpenCL. To enable the automatic use of the fast GPU memory, we identify and implement an optimized version of the allpairs skeleton with a customizing function that follows a certain memory access pattern. We use matrix multiplication as an application study for the allpairs skeleton and its two implementations and demonstrate that the skeleton greatly simplifies programming, saving up to 90 % of lines of code as compared to OpenCL. The performance of our optimized implementation is up to 6.8 times higher as compared with the generic implementation and is competitive to the performance of a manually written optimized OpenCL code.     

Highly accelerated simulations of glassy dynamics using GPUs: Caveats on limited floating-point precision

Modern graphics processing units (GPUs) provide impressive computing resources, which can be accessed conveniently through the CUDA programming interface. We describe how GPUs can be used to considerably speed up molecular dynamics (MD) simulations for system sizes ranging up to about 1 million particles. Particular emphasis is put on the numerical long-time stability in terms of energy and momentum conservation, and caveats on limited floating-point precision are issued. Strict energy conservation over 10⁸ MD steps is obtained by double-single emulation of the floating-point arithmetic in accuracy-critical parts of the algorithm. For the slow dynamics of a supercooled binary Lennard-Jones mixture, we demonstrate that the use of single-floating point precision may result in quantitatively and even physically wrong results. For simulations of a Lennard-Jones fluid, the described implementation shows speedup factors of up to 80 compared to a serial implementation for the CPU, and a single GPU was found to compare with a parallelised MD simulation using 64 distributed cores.     

High‐speed Marching Cubes using HistoPyramids

We present an implementation approach for Marching Cubes (MC) on graphics hardware for OpenGL 2.0 or comparable graphics APIs. It currently outperforms all other known graphics processing units (GPU)‐based iso‐surface extraction algorithms in direct rendering for sparse or large volumes, even those using the recently introduced geometry shader (GS) capabilites. To achieve this, we outfit the Histogram Pyramid (HP) algorithm, previously only used in GPU data compaction, with the capability for arbitrary data expansion. After reformulation of MC as a data compaction and expansion process, the HP algorithm becomes the core of a highly efficient and interactive MC implementation. For graphics hardware lacking GSs, such as mobile GPUs, the concept of HP data expansion is easily generalized, opening new application domains in mobile visual computing. Further, to serve recent developments, we present how the HP can be implemented in the parallel programming language CUDA (compute unified device architecture), by using a novel 1D chunk/layer construction.     

An MPI-CUDA implementation of an improved Roe method for two-layer shallow water systems

The numerical solution of two-layer shallow water systems is required to simulate accurately stratified fluids, which are ubiquitous in nature: they appear in atmospheric flows, ocean currents, oil spills, etc. Moreover, the implementation of the numerical schemes to solve these models in realistic scenarios imposes huge demands of computing power. In this paper, we tackle the acceleration of these simulations in triangular meshes by exploiting the combined power of several CUDA-enabled GPUs in a GPU cluster. For that purpose, an improvement of a path conservative Roe-type finite volume scheme which is specially suitable for GPU implementation is presented, and a distributed implementation of this scheme which uses CUDA and MPI to exploit the potential of a GPU cluster is developed. This implementation overlaps MPI communication with CPU–GPU memory transfers and GPU computation to increase efficiency. Several numerical experiments, performed on a cluster of modern CUDA-enabled GPUs, show the efficiency of the distributed solver.     

A flexible Patch-based lattice Boltzmann parallelization approach for heterogeneous GPU-CPU clusters

Sustaining a large fraction of single GPU performance in parallel computations is considered to be the major problem of GPU-based clusters. We address this issue in the context of a lattice Boltzmann flow solver that is integrated in the WaLBerla software framework. Our multi-GPU implementation uses a block-structured MPI parallelization and is suitable for load balancing and heterogeneous computations on CPUs and GPUs. The overhead required for multi-GPU simulations is discussed in detail. It is demonstrated that a large fraction of the kernel performance can be sustained for weak scaling on InfiniBand clusters, leading to excellent parallel efficiency. However, in strong scaling scenarios using multiple GPUs is much less efficient than running CPU-only simulations on IBM BG/P and x86-based clusters. Hence, a cost analysis must determine the best course of action for a particular simulation task and hardware configuration. Finally we present weak scaling results of heterogeneous simulations conducted on CPUs and GPUs simultaneously, using clusters equipped with varying node configurations.     

SkelCL: a high-level extension of OpenCL for multi-GPU systems

Application development for modern high-performance systems with graphics processing units (GPUs) currently relies on low-level programming approaches like CUDA and OpenCL, which leads to complex, lengthy and error-prone programs. We present SkelCL—a high-level programming approach for systems with multiple GPUs and its implementation as a library on top of OpenCL. SkelCL makes three main enhancements to the OpenCL standard: (1) memory management is simplified using parallel container data types (vectors and matrices); (2) an automatic data (re)distribution mechanism allows for implicit data movements between GPUs and ensures scalability when using multiple GPUs; (3) computations are conveniently expressed using parallel algorithmic patterns (skeletons). We demonstrate how SkelCL is used to implement parallel applications, and we report experimental evaluation of our approach in terms of programming effort and performance.     

GPU近实时线性双目立体代价聚合

目的 近年来双目视觉领域的研究重点逐步转而关注其“实时化”策略的研究,而立体代价聚合是双目视觉中最为复杂且最为耗时的步骤,为此,提出一种基于GPU通用计算(GPGPU)技术的近实时双目立体代价聚合算法。方法 选用一种匹配精度接近于全局匹配算法的局部算法——线性立体匹配算法(linear stereo matching)作为代价聚合策略;结合线性代价聚合的原理,对其主要步骤(代价计算、均值滤波及系数求解等)的计算流程进行有针对性地并行优化。结果 对于相同的实验样本,用本文方法在NVIDA GTX780 实验平台上能在更短的时间计算出代价矩阵,与原有的CPU实现方法相比,代价聚合的效率平均有了数十倍的提升。结论 实时双目立体代价聚合方法,为在个人通用PC平台上实时获取高质量双目视觉深度信息提供了一个高效可靠的途径。     

Molecular dynamics simulations with many-body potentials on multiple GPUs—The implementation,package and performance

Molecular dynamics (MD) is an important research tool extensively applied in materials science. Running MD on a graphics processing unit (GPU) is an attractive new approach for accelerating MD simulations. Currently, GPU implementations of MD usually run in a one-host-process-one-GPU (OHPOG) scheme. This scheme may pose a limitation on the system size that an implementation can handle due to the small device memory relative to the host memory. In this paper, we present a one-host-process-multiple-GPU (OHPMG) implementation of MD with embedded-atom-model or semi-empirical tight-binding many-body potentials. Because more device memory is available in an OHPMG process, the system size that can be handled is increased to a few million or more atoms. In comparison with the serial CPU implementation, in which Newton’s third law is applied to improve the computational efficiency, our OHPMG implementation has achieved a 28.9x–86.0x speedup in double precision, depending on the system size, the cut-off ranges and the number of GPUs. The implementation can also handle a group of small simulation boxes in one run by combining the small boxes into a large box. This approach greatly improves the GPU computing efficiency when a large number of MD simulations for small boxes are needed for statistical purposes.     

Performance analysis and optimization strategies for a D3Q19 lattice Boltzmann kernel on nVIDIA GPUs using CUDA

This paper presents implementation strategies and optimization approaches for a D3Q19 lattice Boltzmann flow solver on nVIDIA graphics processing units (GPUs). Using the STREAM benchmarks we demonstrate the GPU parallelization approach and obtain an upper limit for the flow solver performance. We discuss the GPU-specific implementation of the solver with a focus on memory alignment and register shortage. The optimized code is up to an order of magnitude faster than standard two-socket x86 servers with AMD Barcelona or Intel Nehalem CPUs. We further analyze data transfer rates for the PCI-express bus to evaluate the potential benefits of multi-GPU parallelism in a cluster environment.     

Optimized OpenCL implementation of the Elastodynamic Finite Integration Technique for viscoelastic media

Development of parallel codes that are both scalable and portable for different processor architectures is a challenging task. To overcome this limitation we investigate the acceleration of the Elastodynamic Finite Integration Technique (EFIT) to model 2-D wave propagation in viscoelastic media by using modern parallel computing devices (PCDs), such as multi-core CPUs (central processing units) and GPUs (graphics processing units). For that purpose we choose the industry open standard Open Computing Language (OpenCL) and an open-source toolkit called PyOpenCL. The implementation is platform independent and can be used on AMD or NVIDIA GPUs as well as classical multi-core CPUs. The code is based on the Kelvin–Voigt mechanical model which has the gain of not requiring additional field variables. OpenCL performance can be in principle, improved once one can eliminate global memory access latency by using local memory. Our main contribution is the implementation of local memory and an analysis of performance of the local versus the global memory using eight different computing devices (including Kepler, one of the fastest and most efficient high performance computing technology) with various operating systems. The full implementation of the code is included.     

A Compute Unified System Architecture for Graphics Clusters Incorporating Data Locality

We present a development environment for distributed GPU computing targeted for multi-GPU systems, as well as graphics clusters. Our system is based on CUDA and logically extends its parallel programming model for graphics processors to higher levels of parallelism, namely, the PCI bus and network interconnects. While the extended API mimics the full function set of current graphics hardware—including the concept of global memory—on all distribution layers, the underlying communication mechanisms are handled transparently for the application developer. To allow for high scalability, in particular for network-interconnected environments, we introduce an automatic GPU-accelerated scheduling mechanism that is aware of data locality. This way, the overall amount of transmitted data can be heavily reduced, which leads to better GPU utilization and faster execution. We evaluate the performance and scalability of our system for bus and especially network-level parallelism on typical multi-GPU systems and graphics clusters.