无约束最优化问题的BFGS松弛异步并行算法

为解决大规模非线性最优化问题的串行求解速度慢的问题,提出应用松弛异步并行算法求解无约束最优化问题。根据无约束最优化问题的BFGS串行算法,在PC机群环境下将其并行化。利用CHOLESKY方法分解系数为对称正定矩阵的线性方程组,运用无序松弛异步并行方法求解解向量和Wolfe-Powell非线性搜索步长,并行求解BFGS修正公式,构建BFGS松弛异步并行算法,并对算法的时间复杂性、加速比进行分析。在PC机群的实验结果表明,该算法提高了无约束最优化问题的求解速度且负载均衡,算法具有线性加速比。     

并行演化算法在MEMS继电器参数优化中的应用

使用演化算法求解MEMS继电器参数优化主要瓶颈在于算法运行时间过长,而算法运行时间过长主要由于电磁仿真软件进行建模和分析需要耗费大量的计算时间。针对该问题,采用主从并行模式,对演化算法个体适应值计算阶段并行化处理。在充分考虑计算机资源的使用效率与负载均衡等因素下,使服务器尽量少地参与任务计算及减少与客户机的通信以增强并行模式的分布能力,并且增加了客户端掉线处理,任务重分配等操作以增强并行模式的容错能力。经过测试,该并行演化算法在MEMS微波继电器参数优化上加速比接近线速,具有良好的并行效率且容错性较高。     

单颗粒重构软件EMAN算法分析与高效并行实现

单颗粒重构技术是确定大分子三维结构的重要手段之一.近年来,由于其本身独有的一些优点,单颗粒重构技术受到越来越广泛的关注.然而其处理过程极其耗时,并且缺少高效的并行实现,极大地限制了该技术的应用.对当今使用最广泛的单颗粒重构软件EMAN进行了性能优化及并行加速.通过分析各部分的具体算法,发现其中最核心的问题是如何在低通信开销的前提下实现负载平衡.针对这一问题,提出了自适应动态调度算法.该算法不仅适合于EMAN,同样适合于其他类似的独立任务调度问题.实际运行结果表明,经过优化的串行程序运行时间减少11. 50%.由于采用了自适应动态调度算法,提供的并行实现比EMAN自带的实现具有更高的加速比,其中最耗时的分类操作加速比接近线性.在16个处理器核上的总体并行效率比EMAN自带的并行实现高29. 8%.因此,提供的并行实现可有效利用计算资源,显著缩短单颗粒重构所需时间.     

一种使用GPU加速地震叠前时间偏移的方法

应用GPU通用高性能编程技术实现一种加速地震叠前时间偏移的新方法.该技术是地震勘探处理的常规流程,其核心算法具有计算密集、数据独立性强、并行性高等特点.通过性能剖析获得其计算热点,通过CUDA技术对其进行并行化改造,并利用CUDA的流技术实现CPU到GPU的异步传输.通过集群环境下的性能测试,应用GPU并行化的PSTM程序可明显缩短运行时间.     

一种基于多处理器任务复制的分簇调度算法

任务调度的优劣是决定并行分布式计算机系统性能好坏的重要因素之一。为优化任务调度,基于一些典型算法(如LG、PPA算法等),提出了一种新的任务调度算法。该算法一方面复制满足条件的前驱任务来缩短调度长度;另一方面合理地复制其他前驱任务和合并冗余簇来减少所需处理器的数目。实验表明,该算法在调度长度和所需处理器的数目上优于以上典型算法,并具有更小的时间复杂度,对并行计算机系统性能的提升具有一定的意义。     

分布存储系统上一种新的并行调度算法

在一般的分布存储系统上各个处理器可能不同且资源共享,导致了并行任务在各个处理器上的执行时间具有很大的随机性,主要根据系统及并行任务特性等引进特征参数,采用计算与通信重叠等方法设计出了一种新的并行调度算法,即使在多用户环境下应用此算法不仅能达到极高的负载平衡,充分利用系统资源而且能有效地提高并行效率及加速比。实验结果表明,提出的新的并行调度算法与已有的类似调度算法相比能更加有效地利用系统资源及提高并行效率。     

基于神威·太湖之光的非结构网格计算加速算法

在国产异构众核平台神威·太湖之光上的非结构网格计算具有稀疏存储、离散访存、数据依赖等特点,严重制约了众核处理器的性能发挥。为解决稀疏存储和离散访存问题,提出一种N阶对角染色算法,以有效平衡主从核计算并利用从核将全局访存转化为LDM访问。针对数据依赖造成的计算竞争问题,采用自适应和无依赖的任务划分方法,避免并行计算时的数据冲突。为对处理器架构和非结构网格计算进行优化,采用主核与从核异步并行的方式,差异化使用主从核以充分利用硬件资源,同时,取消处理器提供的寄存器通信机制,降低从核阵列的同步开销同时便于扩展到新一代神威平台。此外,使用计算访存异步重叠技术来充分隐藏访存延迟。利用SpMV、Integration、calcLudsFcc算子进行实验,结果表明,相比主核实现,组合加速算法在不同算例规模下平均取得了10倍的加速效果,加速比最高可达24倍,N阶对角染色算法相比非染色分块算法取得了超过5.8倍的性能加速,有效提升了数据局部性和计算并行度。该算法对有依赖关系的计算冲突算子同样具有良好的加速性能,验证了自适应和无依赖任务划分方法的有效性。     

一种基于动态分组策略和多线程并行IO的并行镶嵌算法优化

遥感图像的镶嵌处理具有数据量大,流程复杂,算法处理耗时巨大的特点,并行计算是加速镶嵌处理过程速度的有效手段。但是,传统的并行镶嵌算法由于任务分配采用静态策略,导致计算节点负载不均衡,并行效率不高。同时,由于传统并行镶嵌算法中存在大量非常耗时的数据存取操作,并且在重采样和匀色过程中存在不合理的流程配置,使得并行效率降低,难以得到比较线性的加速比。本文提出的基于动态任务分配和多线程并行I/O的并行镶嵌算法,较好地解决了上述问题,通过对比分析和实验表明,本算法对大规模图像的镶嵌处理,具有较好的并行处理速度,以及理想的线性并行加速比曲线,节点扩展能力较强。     

基于并行约束规划的最大团识别研究

为提高大数据平台下大规模图例的最大团问题求解效率,提出一种基于并行约束规划的最大团识别算法.通过BMT图划分策略将一个复杂图例分割为若干个可独立计算的子图,并将其分配给Spark集群中的计算节点,每个计算节点采用约束规划方法对分割产生的子问题分别进行建模和求解,实现最大团问题的并行化处理.引入时间预测模型,设计基于任务运行时间预测模型的并行图划分方法,从而有效解决计算节点的负载均衡问题.实验结果表明,与基于BMC图划分策略的最大团并行识别算法相比,该算法具有更高的求解效率,可取得近似线性的加速比.     

基于图形处理器的模糊C均值聚类分割算法

针对模糊C均值聚类图像分割算法运算量大、难于实时处理的问题,提出了一种基于图形处理器的加速算法。通过分析模糊C均值聚类算法各阶段可以并行处理的运算部分,利用计算统一设备架构软硬件结构,分别将隶属度矩阵计算、聚类中心计算和像素按隶属度归类3个部分改造成适合图形处理器硬件并行运行的形式。实验结果表明,相对于CPU串行算法,基于图形处理器的加速算法效率提升明显。鉴于大多数图像处理算法均具有可并行处理的部分,利用图形处理器进行加速具有普适性。     

Parallelization of structured,hierarchical adaptive mesh refinement algorithms

We describe an approach to parallelization of structured adaptive mesh refinement algorithms. This type of adaptive methodology is based on the use of local grids superimposed on a coarse grid to achieve sufficient resolution in the solution. The key elements of the approach to parallelization are a dynamic load-balancing technique to distribute work to processors and a software methodology for managing data distribution and communications. The methodology is based on a message-passing model that exploits the coarse-grained parallelism inherent in the algorithms. The approach is illustrated for an adaptive algorithm for hyperbolic systems of conservation laws in three space dimensions. A numerical example computing the interaction of a shock with a helium bubble is presented. We give timings to illustrate the performance of the method. Received: 28 April 1999 / Accepted: 25 November 1999     

Parallel genetic algorithms: a survey and problem state of the art

In relation with development of computer capabilities and the appearance of multicore processors, parallel computing made it possible to reduce the time for solution of optimization problems. At present of interest are methods of parallel computing for genetic algorithms using the evolutionary model of development in which the main component is the population of species (set of alternative solutions to the problem). In this case, the algorithm efficiency increases due to parallel development of several populations. The survey of basic parallelization strategies and the most interesting models of their implementation are presented. Theoretical ideas on improvement of existing parallelization mechanisms for genetic algorithms are described. A modified model of parallel genetic algorithm is developed. Since genetic algorithms are used for solution of optimization problems, the proposed model was studied for the problem of optimization of a multicriteria function. The algorithm capabilities of getting out of local optima and the influence of algorithm parameters on the deep extremum search dynamics were studied. The conclusion on efficiency of application of dynamic connections of processes, rather than static connections, is made. New mechanisms for implementation and analysis of efficiency of dynamic connections for distributed computing in genetic algorithms are necessary.     

Parallel Algorithms for Dynamic Shortest Path Problems

The development of intelligent transportation systems (ITS) and the resulting need for the solution of a variety of dynamic traffic network models and management problems require faster‐than‐real‐time computation of shortest path problems in dynamic networks. Recently, a sequential algorithm was developed to compute shortest paths in discrete time dynamic networks from all nodes and all departure times to one destination node. The algorithm is known as algorithm DOT and has an optimal worst‐case running‐time complexity. This implies that no algorithm with a better worst‐case computational complexity can be discovered. Consequently, in order to derive algorithms to solve all‐to‐one shortest path problems in dynamic networks, one would need to explore avenues other than the design of sequential solution algorithms only. The use of commercially‐available high‐performance computing platforms to develop parallel implementations of sequential algorithms is an example of such avenue. This paper reports on the design, implementation, and computational testing of parallel dynamic shortest path algorithms. We develop two shared‐memory and two message‐passing dynamic shortest path algorithm implementations, which are derived from algorithm DOT using the following parallelization strategies: decomposition by destination and decomposition by transportation network topology. The algorithms are coded using two types of parallel computing environments: a message‐passing environment based on the parallel virtual machine (PVM) library and a multi‐threading environment based on the SUN Microsystems Multi‐Threads (MT) library. We also develop a time‐based parallel version of algorithm DOT for the case of minimum time paths in FIFO networks, and a theoretical parallelization of algorithm DOT on an ‘ideal’ theoretical parallel machine. Performances of the implementations are analyzed and evaluated using large transportation networks, and two types of parallel computing platforms: a distributed network of Unix workstations and a SUN shared‐memory machine containing eight processors. Satisfactory speed‐ups in the running time of sequential algorithms are achieved, in particular for shared‐memory machines. Numerical results indicate that shared‐memory computers constitute the most appropriate type of parallel computing platforms for the computation of dynamic shortest paths for real‐time ITS applications.     

Parallelization of Pipelined Algorithms for Sets of Linear Banded Systems

Solution of independent sets of linear banded systems is a core part of implicit numerical algorithms. In this study we propose a novel pipelined Thomas algorithm with low parallelization penalty. We introduce two-step pipelined algorithms (PAs) formally and show that the idle processor time is invariant with respect to the order of backward and forward steps. Therefore, the parallelization efficiency of the PA cannot be improved directly. However, the processor idle time can be used if some lines have been computed by the time processors become idle. We develop the immediate backward pipelined Thomas algorithm (IB-PTA). The backward step is computed immediately after the forward step has been completed for the first portion of lines. The advantage of the IB-PTA over the basic PTA is the presence of solved lines, which are available for other computations, by the time processors become idle. Implementation of the IB-PTA is based on a proposed static processor schedule that switches between forward and backward computations and controls communication between processors. Computations are performed on the Cray T3E MIMD computer. Combination of the proposed IB-PTA with the “burn from two ends” algorithm shows low parallelization penalty.     

基于多核阵列体系结构的嵌套循环并行优化

多核处理器已广泛应用于高性能计算领域,如何有效地将传统串行程序转换为并行代码并减少程序中嵌套循环所占用时间仍是该领域的挑战性问题。本文首先基于多面体模型对嵌套循环进行依赖特征分析并实现瓦片分割,据此自动生成粗粒度并行代码。针对多核阵列处理器的结构特点,采用遗传算法生成通信优化的瓦片任务序列,在此基础上建立了有效的任务调度模型。最后将上述方法应用于LU分解,结果表明该方法与传统调度算法相比,在增加数据局部性、实现负载平衡方面具有更好效果。     

A method for speeding up beam-tracing simulation using thread-level parallelization

In recent years, the computational power of modern processors has been increasing mainly because of the increase in the number of processor cores. Computationally intensive applications can gain from this trend only if they employ parallelism, such as thread-level parallelization. Geometric simulations can employ thread-level parallelization because the main part of a geometric simulation can be divided into a subset of mutually independent tasks. This approach is especially interesting for acoustic beam tracing because it is an intensive computing task. This paper presents the parallelization of an existing beam-tracing simulation composed of three algorithms. Two of them are iterative algorithms, and they are parallelized with an already known technique. The most novel method is the parallelization of the third algorithm, the recursive octree generation. To check the performance of the multi-threaded parallelization, several tests are performed using three different computer platforms. On all of the platforms, the multi-threaded octree generation algorithm shows a significant speedup, which is linear when all of the threads are executed on the same processor.     

An Optimal Parallel Co-Connectivity Algorithm

In this paper we consider the problem of computing the connected components of the complement of a given graph. We describe a simple sequential algorithm for this problem, which works on the input graph and not on its complement, and which for a graph on n vertices and m edges runs in optimal O(n+m) time. Moreover, unlike previous linear co-connectivity algorithms, this algorithm admits efficient parallelization, leading to an optimal O(log n)-time and O((n+m)log n)-processor algorithm on the EREW PRAM model of computation. It is worth noting that, for the related problem of computing the connected components of a graph, no optimal deterministic parallel algorithm is currently available. The co-connectivity algorithms find applications in a number of problems. In fact, we also include a parallel recognition algorithm for weakly triangulated graphs, which takes advantage of the parallel co-connectivity algorithm and achieves an O(log² n) time complexity using O((n+m²) log n) processors on the EREW PRAM model of computation.     

Parallelization of the Ising model and its performance evaluation

We report on the parallelization of two widely used algorithms in computational physics: The Monte Carlo simulation of the Ising model and a cluster identification algorithm which is used for percolation or percolation-like problems. Both parallel algorithms were tested on a multi-transputer system using up to 128 processors. The results show that the algorithms can perform with a linear speedup. We propose a scaling law for the speedup and show that the speedup for both algorithms satisfies this scaling.     

On the Performance of Randomized Embedding of Reproduction Trees in Static Networks

High performance computing requires high quality load distribution of processes of a parallel application over processors in a parallel computer at runtime such that both the maximum load and dilation are minimized. The performance of a simple randomized tree embedding algorithm that dynamically supports tree-structured parallel computations on arbitrary static networks is analyzed in this paper. The algorithm spreads newly created tree nodes to neighboring processors, which actually provides randomized dilation-1 tree embedding in static networks. We develop a linear system of equations that characterizes expected loads on all processors under the reproduction tree model, which can generate trees of arbitrary size and shape. It is shown that as the tree size becomes large, the asymptotic performance ratio of the randomized tree embedding algorithm is the ratio of the maximum processor degree to the average processor degree. This implies that the simple randomized tree embedding algorithm is able to generate high quality load distributions on virtually all static networks commonly employed in parallel and distributed computing.     

Intermediate-level feature extraction in novel parallel environments

Abstract. Parallel systems provide an approach to robust computing. The motivation for this work arises from using modern parallel environments in intermediate-level feature extraction. This study presents parallel algorithms for the Hough transform (HT) and the randomized Hough transform (RHT). The algorithms are analyzed in two parallel environments: multiprocessor computers and workstation networks. The results suggest that both environments are suitable for the parallelization of HT. Because scalability of the parallel RHT is weaker than with HT, only the multiprocessor environment is suitable. The limited scalability forces us to use adaptive techniques to obtain good results regardless of the number of processors. Despite the fact that the speedups with HT are greater than with RHT, in terms of total computation time, the new parallel RHT algorithm outperforms the parallel HT. Received: 8 December 2001 / Accepted: 5 June 2002 Correspondence to: V. Kyrki