应用程序属性感知的Yarn资源调度模型研究

Hadoop应用程序存在计算密集属性和调度时间属性,但是Hadoop大数据平台集成的第二代资源管理器Yarn内置的三种资源调度器无法将相同属性的应用程序均衡分配到计算节点上,导致部分节点负载过高,出现严重的计算任务长尾效应。文中提出了一种应用程序属性感知的Yarn负载均衡调度模型——APB Scheduler。APB Scheduler自动感知应用程序属性,将相同属性应用程序的Container按照动态资源计划均衡分配到集群计算节点上,并使用NSGA-Ⅲ算法完成最优分配方案计算。实验结果表明,APB Scheduler解决了相同属性应用程序的Container分配倾斜问题,大幅提升了集群的性能和稳定性。

基于YARN集群的计算加速部件扩展支持

以GPU和Intel MIC为代表的计算加速部件已在科学计算、图形图像处理等领域得到了广泛的应用,其在基于云平台的高性能计算及大数据处理等方向也具有广泛的应用前景.YARN是新一代Hadoop分布式计算框架,其对计算资源的分配调度主要针对CPU,缺少对计算加速部件的支持.在YARN中添加计算加速部件需要解决多个难点,分别是计算加速部件资源如何调度以及异构节点间如何共享问题、多个任务同时调用计算加速部件而引起的资源争用问题和集群中对计算加速部件的状态监控与管理问题.为了解决这些问题,提出了动态节点捆绑策略、流水线式的计算加速部件任务调度等,实现了YARN对计算加速部件的支持,并通过实验验证了其有效性.

Yarn架构下基于GA的Web日志挖掘技术

提出了一种面向TB级别日志文件挖掘需求的日志挖掘技术。采用MapReduce编程模型设计并实现了一种基于子种群联姻策略的GA,并部署到Yarn架构上,使Yarn架构与GA有效结合。在算法Partition阶段,采用Round-Robin策略代替原有的哈希方法,使各个子种群基因更加趋向均匀分布,增加了子种群收敛效率和结果准确性,同时均衡各个节点运行负载,提高了系统性能。经测试,应用该技术的挖掘结果平均准确度达到93%以上,效率提升接近33%。

YARN平台上的并行主题标引算法

文档主题标引是当前个性化智能检索的重要前提,但面对大规模海量数据资源时,主题标引也成为性能瓶颈。当前在Map Reduce框架上设计实现的主题标引算法,通常存在启动任务耗时长,中间数据过多地进行磁盘IO等缺陷。为了解决此类问题,采用YARN(yet another resource negotiator)作为底层分布式资源管理平台,选择更加合适的计算框架来改善计算性能。针对文档主题标引算法计算步骤多、阶段性强的特点,选择有向无环图(directed acyclic graph,DAG)计算模型进行算法实现,避免不必要的作业拆分,从而减少中间结果的磁盘IO。另外,考虑到Map Reduce的排序策略耗时较多,而有些计算无需对结果排序,故可以改用基于Hash的数据归约策略来提高计算性能,但这又会带来随机读的问题。利用固态硬盘高速随机读的特性,设计相应的优化计算策略来解决随机读的问题。通过实验对比发现,以YARN为底层管理平台,在此基础上选择合适的计算框架并加以优化,可以有效改善分布式计算的性能。

基于YARN规范的智能电网大数据异常检测

文章描述了Hadoop早期版本在处理智能电网大数据上的不足,同时分析了YARN规范对Map-Reduce进行改进后对电网大数据处理的优点。文章详细讨论了YARNDPP平台中智能电网大数据处理的编码与实现及YARN-DPP的优势,并以IEEE 118节点的电网作为智能电网大数据异常检测的案例程序,对YARN-DPP平台的硬件环境与软件运行环境进行配置。实验结果表明,针对海量的智能电网状态安全大数据异常检测的程序,YARN-DPP平台具有较好的吞吐量与加速比,可以满足现代智能电网大数据异常检测的需求,在计算速度上比单机串行计算及Map-Reduce计算要快。

一种YARN和Spark框架的网格聚类方法

分布式计算为大数据的处理提供一种新的平台,能有效提升算法的执行速度。在DBSCAN算法基础上提出一种数据分网格算法,该算法将每个分区上的数据集划分成以Eps半径为边长的单元格数据块,将查找Eps邻域的范围缩小到数据对象的八个相邻单元格之内,从而提高查找Eps邻域的速度及聚类速度,具有较好的加速比和扩展率。同时还优化分区聚类合并方法。

一种基于YARN的高优先级作业调度实现方案

提交到YARN上的一个大数据作业会被切分为一个或者多个任务,任务是大数据作业申请资源和执行的基本单位[1]。在某些领域中存在需要对作业紧急度作有效区分使得高紧急度作业优先获得资源的需求,但是在现有的YARN资源调度策略中,对于提交到YARN上的高优先级作业缺乏资源优先分配和高质量的资源保障机制。本文在修改YARN原有资源调度方案的基础上,提出了一种基于YARN的高优先级作业调度实现方案。实验表明,提交到YARN上的高优先级作业执行效率提升了7%左右,证明设计方案行之有效。

一种Hadoop Yarn的资源调度方法研究

针对Hadoop Yarn资源调度问题,为提高集群作业执行效率,提出一种基于蚁群算法与粒子群算法的自适应Hadoop资源调度算法SRSAPH.SRSAPH中,通过Hadoop Yarn跳通信机制获取负载、内存、CPU速度等属性信息初始化信息素矩阵;同时,将粒子群算法的自我认知能力与社会认知能力引入到蚁群算法,提高算法的收敛速度;此外,根据蚁群算法全局最优解的波动趋势动态调整信息素挥发系数,提高解的精度.实验表明,采用SRSAPH进行资源调度,集群的作业执行时间缩短至少10%.

Hadoop YARN大数据计算框架及其资源调度机制研究

Hadoop 2.0提出一种新的资源管理系统YARN,它可以支持多种大数据计算框架(如Map Reduce、Storm、Spark等),目前已经成为腾讯、阿里等互联网公司部署大数据平台采用的主流计算框架。为此,文章在分析Hadoop YARN计算框架的基本结构和工作流程的基础上,重点研究Hadoop YARN中的资源调度机制,包括其资源调度器的模型与机制,及其目前已经实现的容量调度器(Capacity Scheduler)和公平调度器(Fair Scheduler)的设计思想、工作原理和脆弱性,并探讨其下一代欧米伽调度器(Omega Scheduler)的主要设计思想。

基于优先级权重的Hadoop YARN调度算法

为解决Hadoop现有调度器调度任务时不能根据任务的紧迫程度分配资源的问题,研究YARN中的资源调度机制,改进原调度器(Capacity Scheduler),提出一种基于优先级权重的Hadoop YARN(Yet Another Resource Negotiator)调度算法(Weight Scheduler Based on Priority)。为叶子队列设置队列优先级,结合队列资源利用率和队列优先级选择队列;将应用程序的初始权重设置为应用程序优先级的大小,通过等待时间判断是否更新权重,根据权重对队列中的应用程序进行排序,调度时优先为权重高的应用程序分配资源。实验结果表明,与原有调度算法相比,改进算法平均任务执行时间约减少141 s,平均等待时间减少34.5%,保证了用户执行任务的相对公平,提高了用户总体满意度。

一种Yarn框架下的异步双随机梯度下降算法

针对异步随机梯度下降算法在多核系统和主/从集群环境中的通信冲突问题,提出了异步双随机梯度下降算法.该算法主要通过离散各从节点在梯度求解时的计算量,分散各从节点对主节点的通信请求,有效减少模型训练过程中通信冲突次数,从而加快模型的训练速度;在Hadoop Yarn基础上提出了异步并行计算框架,进行了异步随机梯度下降算法以及异步双随机梯度下降算法的快速求解;在多核系统和Hadoop环境下,基于HIGGS数据集进行了异步随机梯度下降算法和异步双随机梯度下降算法的对比实验,结果表明,在保证模型准确率的前提下,异步双随机梯度下降算法比异步随机梯度下降算法具备更快的训练速度.

一种Hadoop YARN的资源调度机制

YARN是Hadoop中广泛应用的资源管理系统,支持MapReduce,Spark,Storm等多种计算框架,已成为大数据生态中的核心组件。然而,在Hadoop YARN现有的资源调度器中,采用基于资源预留的资源保障机制,会产生资源碎片,导致资源浪费。为提高集群的资源利用率和吞吐量,本文提出一种基于预约回填的资源分配机制。在该机制中,基于作业的优先级来决定是否对资源进行预约,并引入回填策略,在不影响预约作业执行的情况下,对资源进行回填使用。实验表明,使用基于预约回填的资源调度机制能够有效提高Hadoop YARN集群的资源利用率和吞吐量。

基于Spark+YARN的空气污染预报云平台设计

近年来我国空气污染导致的雾霾天气频繁发生,空气污染已经成为一个亟待解决的重要问题,其数值预报推广应用是一个难题,主要表现在数值预报系统的运行过程中计算量大,数据量大,在传统高性能计算集群中进行空气污染预报存在资源有限、复杂的并行化操作、批处理作业等待耗时等问题,尤其对于资源缺乏的研究团队,还存在基础设施昂贵而无法负担的问题。因此,如何利用有限的资源,为大气科学家提供一种基于云计算模式的高性能计算环境,提供可扩展、快速、廉价和动态可分配的计算和存储资源,是亟待解决的关键问题。研究一种基于Spark+YARN的空气污染数值预报的云平台,针对空气污染数值预报特点,为大气科学家提供一种空气污染数值预报大数据解决方案。

基于YARN的加权模糊C均值算法在土壤养分数据挖掘中的应用

基于数据挖掘的土壤肥力状况分析是当前针对土壤肥力状况研究最有效的途径之一,针对研究过程中三个有待改进的方面:土壤养分数据的几何状增长单机处理方式已无法胜任;由于Hadoop自身"中心化"严重,基于Hadoop的解决方案存在集群资源利用率低以及单点失效的风险;土壤养分的数据分析过程中没有考虑或没有较好的解决土壤养分数据空间变异性导致的结果偏差,有针对性的提出基于YARN的加权模糊C均值算法,实验表明该算法具有良好的扩展性。将该算法应用于吉林省长春市农安县的土壤肥力状况分析中,聚类效果明显,所得结果适合土壤肥力变化趋势的评价。

Analysis on the Characterization of the Twist Insertion Level and Its Effect of the Compact and Ring Spun Yarns

In this paper, the difference of the twist insertion levels of the compact and ring spun yarns of the same count is analyzed based on the relationship of the yarn twist factor to the twist angle and the bulk density. It is proposed that the relationship of the twist angle, twist level, and the yarn diameter should be synthetically considered when evaluating the twist insertion level of the compact yarn. The twist angle of the compact yarn is smaller than that of the ring yarn with the same twist insertion level. This results from the ordered fiber arrangement and compact structure of the compact yarn. Experiment was conducted to verify the conclusions. It is also discovered that the selection of the twist factor of the compact yarn, which can be usually lower than that of the ring yarn by 10%-15%, can be determined based on the yarn tensile strength.

Full‑Fiber Auxetic‑Interlaced Yarn Sensor for Sign‑Language Translation Glove Assisted by Artificial Neural Network

Yarn sensors have shown promising application prospects in wearable electronics owing to their shape adaptability, good flexibility, and weavability. However, it is still a critical challenge to develop simultaneously structure stable, fast response, body conformal, mechanical robust yarn sensor using full microfibers in an industrial-scalable manner. Herein, a full-fiber auxetic-interlaced yarn sensor(AIYS) with negative Poisson’s ratio is designed and fabricated using a continuous, mass-producible, structure-programmable, and low-cost spinning technology. Based on the unique microfiber interlaced architecture, AIYS simultaneously achieves a Poisson’s ratio of-1.5, a robust mechanical property(0.6 c N/dtex), and a fast train-resistance responsiveness(0.025 s), which enhances conformality with the human body and quickly transduce human joint bending and/or stretching into electrical signals. Moreover, AIYS shows good flexibility, washability, weavability, and high repeatability. Furtherly, with the AIYS array, an ultrafast full-letter sign-language translation glove is developed using artificial neural network. The sign-language translation glove achieves an accuracy of 99.8% for all letters of the English alphabet within a short time of 0.25 s. Furthermore, owing to excellent full letter-recognition ability, real-time translation of daily dialogues and complex sentences is also demonstrated. The smart glove exhibits a remarkable potential in eliminating the communication barriers between signers and non-signers.

Study on the Bending Properties of Woven Fabrics and Its Yarns

A theoretical analysis on the two extreme cases of yarn bending rigidity as well as the relationship between them is proposed. The significant influence of fiber-to-fiber interaction on yarn bending rigidity is revealed.Experimental observation on the relationship between fabric/yarn bending property and fiber fineness, yarn structure, fabric structure is given as we ll. Three problems are discussed and the following conclusion is thus obtained. There exists, similar t0 a rigid engineering rod, a linear relationship between yarn bending rigidity and the square of yarn fineness (expressed as yarn count in tex), but the increasing speed is much lower. Yarn bending rigidity in the fabric also to a great extent depends on the fabric density and yarn twist factor, and so on. Fabric's bending hysteresis primarily depends on its bending rigidity.

主资源公平调度算法在YARN中应用研究

Apache Hadoop YARN是一种新的通用的Hadoop资源管理器,是在Mapreduce1的基础上发展起来的,该引入为集群在利用率、资源统一管理和数据共享等方面带来了巨大好处。为此,文章在分析了Hadoop YARN计算框架的基本结构和工作流程,重点研究了Hadoop YARN中的多资源分配机制,尤其是主资源公平调度算法(Dominant Resource Fairness,DRF)在YARN中的应用。

Multi-objective Optimal Trajectory and Placement of a Robot for Cotton Yarns Handling

This study deals with a robot manipulator for yarn bobbin handling in the cotton yarns lattice distortion modification system.The aim is to achieve an operation of yarn bobbin handling with minimal execution time,energy consumption and jerk in motion together.The placement of the robot,in relation to the yarn bobbin stations,is also optimized in conjunction of trajectory optimization.Three possible techniques for building the handling traj'ectory were considered:the quaternion spherical linear interpolation in Cartesian space,the quintic polynomial spline and quintic B-spline in joint space.The genetic algorithm(GA) was used to optimize the trajectories of the robot,with a penalty function to handle nonlinear constraints associated in the robot motion.Two simulations of the optimal trajectory in joint space and the placement of robot were carried out and the results obtained were presented and discussed.It is concluded that the quintic polynomial spline constructs a better trajectory in joint space and the proper placement of robot makes better performance.

Radial Yarn Movement in Yarn Duct of Interlacer

In order to find the relations between the production performance of interlacer and its cross-sectional shapes of yarn duct,radial yarn movement in various yarn ducts of interlacer and its effect on performance of interlaced yarn were studied in this paper.Seven interlacers,which were different in the cross-sectional shape of yarn duct,were made.Yarn motion in yarn duct of interlace was observed with optical fibers and photo-sensors.Performance of the interlaced yarn was evaluated by the number and the strength of tangles.Experimental results disclose the relations between the yarn motions and the properties of the interlaced yarn.The interlacer with round yarn duct is capable of producing an interlaced yarn with bigger number of tangles,since it can make yarn transverse frequently the axis of air jet issuing from air jet nozzle.The interlacer with cornered shape of yarn duct makes yarn subjected to the action of compressed air for a longer time,which results in larger strength of tangles.