国产异构系统上的HPCG并行算法及高效实现

HPCG基准测试程序是一种新的超级计算机排名度量标准.该测试基准主要用于衡量超级计算机解决大规模稀疏线性系统的能力,更贴近实际应用,近年来广受关注.基于国产超级计算机研究异构众核并行HPCG软件具有非常重要的意义,其不仅可以提升国产超级计算机HPCG的排名,还对很多应用提供了并行算法、优化技术等方面的参考.面向某国产复杂异构超级计算机开展研究,首先采用了分块图着色算法对HPCG进行并行,并提出一种适用于结构化网格的图着色算法.该算法并行性能高于传统的JPL、CC等算法,且着色质量高,运用于HPCG后,迭代次数减少了3次,整体性能提升了6%.分析了复杂异构系统各个部件传输的开销,提出一套更适用于HPCG的任务划分方法,并从稀疏矩阵存储格式、稀疏矩阵重排、访存等角度开展了细粒度的优化.在多进程计算时,还采用内外区划分算法将核心函数SpMV、SymGS中的邻居通信操作进行了隐藏.最终整机测试时,性能达到了国产超级计算机峰值性能的1.67%,与单节点相比,整机弱可扩展性并行效率达到了92%.     

一类国产复杂异构系统上的HPCG并行算法及高效实现研究

HPCG基准测试程序是一种新的超级计算机排名度量标准.该测试基准主要用于衡量超级计算机解决大规模稀疏线性系统的能力,更贴近实际应用,近年来广受关注.基于国产超级计算机研究异构众核并行HPCG软件具有非常重要的意义,其不仅可以提升国产超级计算机HPCG的排名,还对很多应用提供了并行算法、优化技术等方面的参考.本文面向某国产复杂异构超级计算机开展研究,首先采用了分块图着色算法对HPCG进行并行,并提出一种适用于结构化网格的图着色算法,该算法并行性能高于传统的JPL、CC等算法,且着色质量高,运用于HPCG后,迭代次数减少了3次,整体性能提升了6%.本文还分析了复杂异构系统各个部件传输的开销,提出一套更适用于HPCG的任务划分方法,并从稀疏矩阵存储格式、稀疏矩阵重排、访存等角度开展了细粒度的优化.另外在多进程计算时,还采用了内外区划分算法将核心函数SpMV、SymGS中的邻居通信操作进行了隐藏.最终整机测试时,性能达到国产超级计算机峰值性能的1.67%,相比单节点,整机弱可扩展性并行效率达到了92%.     

Scientific Computing Kernels on the Cell Processor

In this work, we examine the potential of using the recently-released STI Cell processor as a building block for future high-end scientific computing systems. Our work contains several novel contributions. First, we introduce a performance model for Cell and apply it to several key numerical kernels: dense matrix multiply, sparse matrix vector multiply, stencil computations, and 1D/2D FFTs. Next, we validate our model by comparing results against published hardware data, as well as our own Cell blade implementations. Additionally, we compare Cell performance to benchmarks run on leading superscalar (AMD Opteron), VLIW (Intel Itanium2), and vector (Cray X1E) architectures. Our work also explores several different kernel implementations and demonstrates a simple and effective programming model for Cell’s unique architecture. Finally, we propose modest microarchitectural modifications that could significantly increase the efficiency of double-precision calculations. Overall results demonstrate the tremendous potential of the Cell architecture for scientific computations in terms of both raw performance and power efficiency.     

基于深度学习的稀疏矩阵向量乘运算性能预测模型

稀疏矩阵向量乘（SpMV）是求解稀疏线性方程组的计算核心,被广泛应用在经济学模型、信号处理等科学计算和工程应用中,对于SpMV及其调优技术的研究有助于提升解决相关领域问题的运算效率。传统SpMV自动调优方法基于硬件平台的体系结构参数设置来提升SpMV性能,但巨大的参数设置量导致搜索空间变大且自动调优耗时大幅增加。采用深度学习技术,基于卷积神经网络,构建由双通道稀疏矩阵特征融合以及稀疏矩阵特征与体系结构特征融合组成的SpMV运算性能预测模型,实现快速自动调优。为提高SpMV运算时间的预测精度,选取特征数据并利用箱形图统计SpMV时间信息,同时在佛罗里达稀疏矩阵数据集上进行实验设计与验证,结果表明,该模型的SpMV运算时间预测准确率达到80%以上,并且具有较强的泛化能力。     

大规模稀疏矩阵的主特征向量计算优化方法

矩阵主特征向量(principal eigenvectors computing,PEC)的求解是科学与工程计算中的一个重要问题。随着图形处理单元通用计算(general-purpose computing on graphics pro cessing unit,GPGPU)的兴起,利用GPU来优化大规模稀疏矩阵的图形处理单元求解得到了广泛关注。分别从应用特征和GPU体系结构特征两方面分析了PEC运算的性能瓶颈,提出了一种面向GPU的稀疏矩阵存储格式——GPU-ELL和一个针对GPU的线程优化映射策略,并设计了相应的PEC优化执行算法。在ATI HD Radeon5850上的实验结果表明,相对于传统CPU,该方案获得了最多200倍左右的加速,相对于已有GPU上的实现,也获得了2倍的加速。     

基于RISC-V向量指令的稀疏矩阵向量乘法实现与优化

开源指令集架构RISC-V具有高性能、模块化、简易性和易拓展等优势,在物联网、云计算等领域的应用日渐广泛,其向量拓展部分V模块更是很好地支持了矩阵数值计算.稀疏矩阵向量乘法SpM V作为矩阵数值计算的一个重要组成部分,具有深刻的研究意义与价值.利用RISC-V指令集的向量可配置性和寻址特性,分别对基于CSR、ELLPA...     

Memory bandwidth optimization of SpMV on GPGPUs

It is an important task to improve performance for sparse matrix vector multiplication (SpMV), and it is a difficult task because of its irregular memory access. General purpose GPU (GPGPU) provides high computing ability and substantial bandwidth that cannot be fully exploited by SpMV due to its irregularity. In this paper, we propose two novel methods to optimize the memory bandwidth for SpMV on GPGPU. First, a new storage format is proposed to exploit memory bandwidth of GPU architecture more efficiently. The new storage format can ensure that there are as many non-zeros as possible in the format which is suitable to exploit the memory bandwidth of the GPU. Second, we propose a cache blocking method to improve the performance of SpMV on GPU architecture. The sparse matrix is partitioned into sub-blocks that are stored in CSR format. With the blocking method, the corresponding part of vector x can be reused in the GPU cache, so the time to access the global memory for vector x is reduced heavily. Experiments are carried out on three GPU platforms, GeForce 9800 GX2, GeForce GTX 480, and Tesla K40. Experimental results show that both new methods can efficiently improve the utilization of GPU memory bandwidth and the performance of the GPU.     

GPU-based iterative transmission reconstruction in 3D ultrasound computer tomography

As today’s standard screening methods frequently fail to detect breast cancer before metastases have developed, early diagnosis is still a major challenge. With the promise of high-quality volume images, three-dimensional ultrasound computer tomography is likely to improve this situation, but has high computational needs. In this work, we investigate the acceleration of the ray-based transmission reconstruction by a GPU-based implementation of the iterative numerical optimization algorithm TVAL3. We identified the regular and transposed sparse-matrix–vector multiply as the performance limiting operations. For accelerated reconstruction we propose two different concepts and devise a hybrid scheme as optimal configuration. In addition we investigate multi-GPU scalability and derive the optimal number of devices for our two primary use-cases: a fast preview mode and a high-resolution mode. In order to achieve a fair estimation of the speedup, we compare our implementation to an optimized CPU version of the algorithm. Using our accelerated implementation we reconstructed a preview 3D volume with 24,576 unknowns, a voxel size of (8 mm)³ and approximately 200,000 equations in 0.5 s. A high-resolution volume with 1,572,864 unknowns, a voxel size of (2mm)³ and approximately 1.6 million equations was reconstructed in 23 s. This constitutes an acceleration of over one order of magnitude in comparison to the optimized CPU version.     

RAM（h）模型下SpMV存储访问复杂度的分析

稀疏矩阵向量乘(SpMV)采取压缩行存储格式的算法性能非常差,而寄存器分块算法可以使得数据尽量在靠近处理器的存储层次中访问而提高性能.利用RAM(h)模型进行分析和比较不同算法形式的存储访问复杂度,可以比较两种算法的优劣.通过RAM(h)分析SpMV两种实现形式的存储访问复杂度,同时在奔腾四平台上,测试了7个稀疏矩阵的SpMV性能,并统计了这两种算法中L1,L2,和TLB的缺失率,实验结果与模型分析的数据一致.     

面向国产申威26010众核处理器的SpMV实现与优化

世界首台峰值性能超过100P的超级计算机——神威太湖之光已经研制完成,该超级计算机采用了国产申威异构众核处理器,该处理器不同于现有的纯CPU,CPU-MIC,CPU-GPU架构,采用了主-从核架构,单处理器峰值计算能力为3TFlops/s,访存带宽为130GB/s.稀疏矩阵向量乘SpMV（sparse matrix-vector multiplication）是科学与工程计算中的一个非常重要的核心函数,众所周知,其是带宽受限型的,且存在间接访存操作.国产申威处理器给稀疏矩阵向量乘的高效实现带来了很大的挑战.针对申威处理器提出了一种CSR格式SpMV操作的通用异构众核并行算法,该算法从任务划分、LDM空间划分方面进行精细设计,提出了一套动静态buffer的缓存机制以提升向量x的访存命中率,提出了一套动静态的任务调度方法以实现负载均衡.另外还分析了该算法中影响SpMV性能的几个关键因素,并开展了自适应优化,进一步提升了性能.采用Matrix Market矩阵集中具有代表性的16个稀疏矩阵进行了测试,相比主核版最高有10倍左右的加速,平均加速比为6.51.通过采用主核版CSR格式SpMV的访存量进行分析,测试矩阵最高可达该处理器实测带宽的86%,平均可达到47%.