ChattyGraph:面向异构多协处理器的高可扩展图计算系统

现阶段,随着数据规模扩大化和结构多样化的趋势日益凸现,如何利用现代链路内链的异构多协处理器为大规模数据处理提供实时、可靠的并行运行时环境,已经成为高性能以及数据库领域的研究热点.利用多协处理器(GPU)设备的现代服务器(multi-GPU server)硬件架构环境,已经成为分析大规模、非规则性图数据的首选高性能平台.现有研究工作基于Multi-GPU服务器架构设计的图计算系统和算法(如广度优先遍历和最短路径算法),整体性能已显著优于多核CPU计算环境.然而,这类图计算系统中,多GPU协处理器间的图分块数据传输性能受限于PCI-E总线带宽和局部延迟,导致通过增加GPU设备数量无法达到整体系统性能的类线性增长趋势,甚至会出现严重的时延抖动,进而已无法满足大规模图并行计算系统的高可扩展性要求.经过一系列基准实验验证发现,现有系统存在如下两类缺陷:(1)现代GPU设备间数据通路的硬件架构发展日益更新(如NVLink-V1,NVLink-V2),其链路带宽和延迟得到大幅改进,然而现有系统受限于PCI-E总线进行数据分块通信,无法充分利用现代GPU链路资源(包括链路拓扑、连通性和路由);(2)在...     

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.     

多GPU系统虚实地址转换架构研究

近年来,随着大数据的发展,GPU应用的数据集规模急剧增加,这对GPU的处理能力提出了挑战.由于摩尔定律即将达到极限,提升单一GPU的性能变得越发困难,而多GPU系统通过提升GPU处理器级的并行性,成为应对该挑战的一种解决方案.GPU制造商对内存虚拟化的支持进一步简化了多GPU系统的编程,提升了资源利用率.内存虚拟化需要...     

Equalizer: A Scalable Parallel Rendering Framework

Continuing improvements in CPU and GPU performances as well as increasing multi-core processor and cluster-based parallelism demand for flexible and scalable parallel rendering solutions that can exploit multipipe hardware accelerated graphics. In fact, to achieve interactive visualization, scalable rendering systems are essential to cope with the rapid growth of data sets. However, parallel rendering systems are non-trivial to develop and often only application specific implementations have been proposed. The task of developing a scalable parallel rendering framework is even more difficult if it should be generic to support various types of data and visualization applications, and at the same time work efficiently on a cluster with distributed graphics cards. In this paper we introduce a novel system called Equalizer, a toolkit for scalable parallel rendering based on OpenGL which provides an application programming interface (API) to develop scalable graphics applications for a wide range of systems ranging from large distributed visualization clusters and multi-processor multipipe graphics systems to single-processor single-pipe desktop machines. We describe the system architecture, the basic API, discuss its advantadges over previous approaches, present example configurations and usage scenarios as well as scalability results.     

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.     

Data-intensive document clustering on graphics processing unit (GPU) clusters

Document clustering is a central method to mine massive amounts of data. Due to the explosion of raw documents generated on the Internet and the necessity to analyze them efficiently in various intelligent information systems, clustering techniques have reached their limitations on single processors. Instead of single processors, general-purpose multi-core chips are increasingly deployed in response to diminishing returns in single-processor speedup due to the frequency wall, but multi-core benefits only provide linear speedups while the number of documents in the Internet is growing exponentially. Accelerating hardware devices represent a novel promise for improving the performance for data-intensive problems such as document clustering. They offer more radical designs with a higher level of parallelism but adaptation to novel programming environments.In this paper, we assess the benefits of exploiting the computational power of graphics processing units (GPUs) to study two fundamental problems in document mining, namely to calculate the term frequency-inverse document frequency (TF-IDF) and cluster a large set of documents. We transform traditional algorithms into accelerated parallel counterparts that can be efficiently executed on many-core GPU architectures. We assess our implementations on various platforms, ranging from stand-alone GPU desktops to Beowulf-like clusters equipped with contemporary GPU cards. We observe at least one order of magnitude speedups over CPU-only desktops and clusters. This demonstrates the potential of exploiting GPU clusters to efficiently solve massive document mining problems. Such speedups combined with the scalability potential and accelerator-based parallelization are unique in the domain of document-based data mining, to the best of our knowledge.     

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.     

Zippy: A Framework for Computation and Visualization on a GPU Cluster

Due to its high performance/cost ratio, a GPU cluster is an attractive platform for large scale general‐purpose computation and visualization applications. However, the programming model for high performance general‐purpose computation on GPU clusters remains a complex problem. In this paper, we introduce the Zippy frame‐work, a general and scalable solution to this problem. It abstracts the GPU cluster programming with a two‐level parallelism hierarchy and a non‐uniform memory access (NUMA) model. Zippy preserves the advantages of both message passing and shared‐memory models. It employs global arrays (GA) to simplify the communication, synchronization, and collaboration among multiple GPUs. Moreover, it exposes data locality to the programmer for optimal performance and scalability. We present three example applications developed with Zippy: sort‐last volume rendering, Marching Cubes isosurface extraction and rendering, and lattice Boltzmann flow simulation with online visualization. They demonstrate that Zippy can ease the development and integration of parallel visualization, graphics, and computation modules on a GPU cluster.     

Work stealing for GPU‐accelerated parallel programs in a global address space framework

Task parallelism is an attractive approach to automatically load balance the computation in a parallel system and adapt to dynamism exhibited by parallel systems. Exploiting task parallelism through work stealing has been extensively studied in shared and distributed‐memory contexts. In this paper, we study the design of a system that uses work stealing for dynamic load balancing of task‐parallel programs executed on hybrid distributed‐memory CPU‐graphics processing unit (GPU) systems in a global‐address space framework. We take into account the unique nature of the accelerator model employed by GPUs, the significant performance difference between GPU and CPU execution as a function of problem size, and the distinct CPU and GPU memory domains. We consider various alternatives in designing a distributed work stealing algorithm for CPU‐GPU systems, while taking into account the impact of task distribution and data movement overheads. These strategies are evaluated using microbenchmarks that capture various execution configurations as well as the state‐of‐the‐art CCSD(T) application module from the computational chemistry domain. Copyright © 2016 John Wiley & Sons, Ltd.     

面向分布式机器学习的大消息广播设计

MPI (Message Passing Interface)专为节点密集型大规模计算集群设计,然而,随着MPI+CUDA (Compute Unified Device Architecture)应用程序以及计算节点拥有GPU的计算机集群的出现,类似于MPI的传统通信库已无法满足.而在机器学习领域,也面临着同样的挑战,如Caff以及CNTK (Microsoft CognitiveToolkit)的深度学习框架,由于训练过程中, GPU会缓存庞大的数据量,而大部分机器学习训练的优化算法具有迭代性特点,导致GPU间的通信数据量大,通信频率高,这些已成为限制深度学习训练性能提升的主要因素之一,虽然推出了像NCCL(Nvidia Collective multi-GPU Communication Library)这种解决深度学习通信问题的集合通信库,但也存在不兼容MPI等问题.因此,设计一种更加高效、符合当前新趋势的通信加速机制便显得尤为重要,为解决上述新形势下的挑战,本文提出了两种新型通信广播机制:(1)一种基于MPI_Bcast的管道链PC (Pipelined Chain)通信机制:为GPU缓存提供高效的节点内外通信.(2)一种适用于多GPU集群系统的基于拓扑感知的管道链TA-PC (TopologyAware Pipelined Chain)通信机制:充分利用多GPU节点间的可用PCIe链路.为了验证提出的新型广播设计,分别在三种配置多样化的GPU集群上进行了实验:GPU密集型集群RX1、节点密集型集群RX2、均衡型集群RX3.实验中,将新的设计与MPI+NCCL1 MPI_Bcast进行对比实验,对于节点内通信和节点间的通信,分别取得了14倍和16.6倍左右的性能提升;与NCCL2的对比试验中,小中型消息取得10倍左右的性能提升,大型消息取得与其相当的性能水平,同时TA-PC设计相比于PC设计,在64GPU集群上实现50%左右的性能提升.实验结果充分说明,提出的解决方案在可移植性以及性能方面有较大的优势.     

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.     

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.     

一种适应GPU的混合访问缓存索引框架

散列表（hash table）作为一类根据关键码值（key value）提供高效数据访问的数据索引结构,其广泛应用于各类计算机应用中,尤其是在对性能要求极高的系统软件、数据库以及高性能计算领域.在网络、云计算和物联网服务方面,以散列表为核心结构已经成为缓存系统的重要系统组件.然而,随着大规模数据量的大幅度增加,以多核CPU为核心设计散列表结构的系统已经逐渐出现性能瓶颈,亟需进一步改进散列表的高性能和可扩展性.随着通用图形处理器（graphic processing unit,简称GPU）的日益普及以及硬件计算能力和并发性能的大幅度提升,各类以并行计算为核心的系统软件任务在GPU上进行了优化设计并得到可观的性能提升.由于存在稀疏性和随机性,采用现有散列表的并行结构直接在GPU上应用势必会带来高频次的内存访问和频繁的总线数据传输,影响了散列表在GPU上的性能发挥.重点分析了缓存系统中散列表索引的内存访问、命中率与索引开销,提出并设计了一种适应GPU的混合访问缓存索引框架CCHT（cache cuckoo hash table）,提供了两种适应不同命中率和索引开销要求的缓存策略,允许写入与查询操作并发执行,最大程度地利用了GPU硬件的计算性能与并发特性,减少了内存访问与总线传输.通过在GPU硬件上的实现与实验验证,CCHT在保证缓存命中率的同时,性能优于其他用于缓存索引的散列表.     

Concurrent warp execution: improving performance of GPU-likely SIMD architecture by increasing resource utilization

Hardware parallelism should be exploited to improve the performance of computing systems. Single instruction multiple data (SIMD) architecture has been widely used to maximize the throughput of computing systems by exploiting hardware parallelism. Unfortunately, branch divergence due to branch instructions causes underutilization of computational resources, resulting in performance degradation of SIMD architecture. Graphics processing unit (GPU) is a representative parallel architecture based on SIMD architecture. In recent computing systems, GPUs can process general-purpose applications as well as graphics applications with the help of convenient APIs. However, contrary to graphics applications, general-purpose applications include many branch instructions, resulting in serious performance degradation of GPU due to branch divergence. In this paper, we propose concurrent warp execution (CWE) technique to reduce the performance degradation of GPU in executing general-purpose applications by increasing resource utilization. The proposed CWE enables selecting co-warps to activate more threads in the warp, leading to concurrent execution of combined warps. According to our simulation results, the proposed architecture provides a significant performance improvement (5.85 % over PDOM, 91 % over DWF) with little hardware overhead.     

Parallel processing of the Building-Cube Method on a GPU platform

The Building-Cube Method (BCM) based on equally-spaced Cartesian meshes has been proposed as a next generation CFD method. Due to the equally-spaced meshes, it is well suited for highly parallel computation. This paper proposes a parallel implementation scheme of BCM on a GPU cluster system, which needs efficient hierarchical parallel processing to exploit the potential of the cluster system. The proposed scheme employs the Red-Black SOR method for the pressure calculations, which is the most time-consuming part of BCM, to obtain massive data parallelism of BCM. By exploiting the coarse-grain and fine-grain parallelism of BCM, the proposed scheme hierarchically assigns equally-divided tasks into the GPU cluster system. Furthermore, to exploit the computational power of GPUs in the cluster system, the proposed scheme employs an efficient data management such as coalesced data transfer and reusing data on an on-chip memory. Experimental results show that the single GPU implementation can achieve about three times higher performance than the single CPU one. Moreover, the multiple GPU implementation can achieve an almost ideal scalability. Finally, the possibility of further acceleration of not only the pressure calculation but also the whole BCM is discussed.     

A scalable approach to solving dense linear algebra problems on hybrid CPU‐GPU systems

Aiming to fully exploit the computing power of all CPUs and all graphics processing units (GPUs) on hybrid CPU‐GPU systems to solve dense linear algebra problems, we design a class of heterogeneous tile algorithms to maximize the degree of parallelism, to minimize the communication volume, and to accommodate the heterogeneity between CPUs and GPUs. The new heterogeneous tile algorithms are executed upon our decentralized dynamic scheduling runtime system, which schedules a task graph dynamically and transfers data between compute nodes automatically. The runtime system uses a new distributed task assignment protocol to solve data dependencies between tasks without any coordination between processing units. By overlapping computation and communication through dynamic scheduling, we are able to attain scalable performance for the double‐precision Cholesky factorization and QR factorization. Our approach demonstrates a performance comparable to Intel MKL on shared‐memory multicore systems and better performance than both vendor (e.g., Intel MKL) and open source libraries (e.g., StarPU) in the following three environments: heterogeneous clusters with GPUs, conventional clusters without GPUs, and shared‐memory systems with multiple GPUs. Copyright © 2014 John Wiley & Sons, Ltd.     

Multi-Kepler GPU vs. multi-Intel MIC for spin systems simulations

We present and compare the performances of two many-core architectures: the Nvidia Kepler and the Intel MIC both in a single system and in cluster configuration for the simulation of spin systems. As a benchmark we consider the time required to update a single spin of the 3D Heisenberg spin glass model by using the Over-relaxation algorithm. We present data also for a traditional high-end multi-core architecture: the Intel Sandy Bridge. The results show that although on the two Intel architectures it is possible to use basically the same code, the performances of a Intel MIC change dramatically depending on (apparently) minor details. Another issue is that to obtain a reasonable scalability with the Intel Phi coprocessor (Phi is the coprocessor that implements the MIC architecture) in a cluster configuration it is necessary to use the so-called offload mode which reduces the performances of the single system. As to the GPU, the Kepler architecture offers a clear advantage with respect to the previous Fermi architecture maintaining exactly the same source code. Scalability of the multi-GPU implementation remains very good by using the CPU as a communication co-processor of the GPU. All source codes are provided for inspection and for double-checking the results.     

量子线路模拟器QuEST在多GPU平台上的性能优化

在当前量子计算的研究中,量子线路模拟器作为重要的研究工具,一直受到研究者们的高度重视.QuEST是一款开源的通用量子线路模拟器,能在单个CPU结点、多个CPU结点和单个GPU等多种测试平台上灵活运行.量子线路模拟固有的并行性使其非常适合在GPU上运行,并能获得较大的性能加速.但是其缺点在于所消耗的内存空间巨大,单个GP...     

Simulation of shallow-water systems using graphics processing units

This paper addresses the speedup of the numerical solution of shallow-water systems in 2D domains by using modern graphics processing units (GPUs). A first order well-balanced finite volume numerical scheme for 2D shallow-water systems is considered. The potential data parallelism of this method is identified and the scheme is efficiently implemented on GPUs for one-layer shallow-water systems. Numerical experiments performed on several GPUs show the high efficiency of the GPU solver in comparison with a highly optimized implementation of a CPU solver.     

A multigrain Delaunay mesh generation method for multicore SMT-based architectures

Given the proliferation of layered, multicore- and SMT-based architectures, it is imperative to deploy and evaluate important, multi-level, scientific computing codes, such as meshing algorithms, on these systems. We focus on Parallel Constrained Delaunay Mesh (PCDM) generation. We exploit coarse-grain parallelism at the subdomain level, medium-grain at the cavity level and fine-grain at the element level. This multi-grain data parallel approach targets clusters built from commercially available SMTs and multicore processors. The exploitation of the coarser degree of granularity facilitates scalability both in terms of execution time and problem size on loosely-coupled clusters. The exploitation of medium-grain parallelism allows performance improvement at the single node level. Our experimental evaluation shows that the first generation of SMT cores is not capable of taking advantage of fine-grain parallelism in PCDM. Many of our experimental findings with PCDM extend to other adaptive and irregular multigrain parallel algorithms as well.