Review of Load Balancing Mechanisms in SDN-Based Data Centers

With the continuous expansion of the data center network scale, changing network requirements, and increasing pressure on network bandwidth, the traditional network architecture can no longer meet people’s needs. The development of software defined networks has brought new opportunities and challenges to future networks. The data and control separation characteristics of SDN improve the performance of the entire network. Researchers have integrated SDN architecture into data centers to improve network resource utilization and performance. This paper first introduces the basic concepts of SDN and data center networks. Then it discusses SDN-based load balancing mechanisms for data centers from different perspectives. Finally, it summarizes and looks forward to the study on SDN-based load balancing mechanisms and its development trend.

L2-LBMT:A Layered Load Balance Routing Protocol for Underwater Multimedia Data Transmission

Providing highly efficient underwater transmission of mass multimedia data is challenging due to the particularities of the underwater environment. Although there are many schemes proposed to optimize the underwater acoustic network communication protocols, from physical layer, data link layer, network layer to transport layer, the existing routing protocols for underwater wireless sensor network(UWSN) still cannot well deal with the problems in transmitting multimedia data because of the difficulties involved in high energy consumption, low transmission reliability or high transmission delay. It prevents us from applying underwater multimedia data to real-time monitoring of marine environment in practical application, especially in emergency search, rescue operation and military field. Therefore, the inefficient transmission of marine multimedia data has become a serious problem that needs to be solved urgently. In this paper, A Layered Load Balance Routing Protocol(L2-LBMT) is proposed for underwater multimedia data transmission. In L2-LBMT, we use layered and load-balance Ad Hoc Network to transmit data, and adopt segmented data reliable transfer(SDRT) protocol to improve the data transport reliability. And a 3-node variant of tornado(3-VT) code is also combined with the Ad Hoc Network to transmit little emergency data more quickly. The simulation results show that the proposed protocol can balance energy consumption of each node, effectively prolong the network lifetime and reduce transmission delay of marine multimedia data.

Research on Rolling Load Distribution Method based on Data Mining

A new method of establishing rolling load distribution model was developed by online intelligent information-processing technology for plate rolling. The model combines knowledge model and mathematical model with using knowledge discovery in database （KDD） and data mining （DM） as the start. The online maintenance and optimization of the load model are realized. The effectiveness of this new method was testified by offline simulation and online application.

Load Balance Strategy of Data Routing Algorithm Using Semantics for Deduplication Clusters

The backup requirement of data centres is tremendous as the size of data created by human is massive and is increasing exponentially.Single node deduplication cannot meet the increasing backup requirement of data centres.A feasible way is the deduplication cluster,which can meet it by adding storage nodes.The data routing strategy is the key of the deduplication cluster.DRSS（data routing strategy using semantics） improves the storage utilization of MCS（minimum chunk signature） data routing strategy a lot.However,for the large deduplication cluster,the load balance of DRSS is worse than MCS.To improve the load balance of DRSS,we propose a load balance strategy used for DRSS,namely DRSSLB.When a node is overloaded,DRSSLB iteratively migrates the current smallest container of the node to the smallest node in the deduplication cluster until this overloaded node becomes non-overloaded.A container is the minimum unit of data migration.Similar files sharing the same features or file names are stored in the same container.This ensures the similar data groups are still in the same node after rebalancing the nodes.We use the dataset from the real world to evaluate DRSSLB.Experimental results show that,for various numbers of nodes of the deduplication cluster,the data skews of DRSSLB are under predefined value while the storage utilizations of DRSSLB do not nearly increase compared with DRSS,with the low penalty（the data migration rate is only6.5% when the number of nodes is 64）.

基于SDN的数据中心网络流量负载均衡研究

目前数据中心网络(data center network,DCN)的负载均衡方法存在对大小流的调度缺乏全局实时检测等不足,部分大流会造成拥塞、负载不均衡和带宽碎片等问题.针对上述问题,提出了一种SDN网络流量负载均衡算法—DSA-D.首先,对流量进行分类,为大流计算所有源至目的主机可达路径的最短跳数路径集;然后,根据LLDP和ECHO测量链路时延以求得时延最优路径集;最后,采用概率拟合算法分配路径,实现数据中心网络流量负载均衡.在相同场景下的实验结果表明,与ECMP、Hedera和DIFF算法相比,DSA-D算法具有更好的吞吐量、链路带宽利用率和平均往返时延.

“数形结合”——测量及不确定度教学模式的改革与研究

理解测量的不确定度是大学物理实验课程一项重要的教学目标。在大学物理实验绪论课中,准备了两个版本的不确定性教学方式:即有无测量数据的图形化展示。课程结束后,对学生进行测试,研究数据的图形化展示是否能够提高学生对不确定度的理解。结果表明,有图形展示的班级学生比没有图形展示的班级学生得分更高。图形表示有助于数据的可视化,在进行测量数据分析时减少了学生的认知负荷,因此应该在教学中积极推广使用。

算力-热力灵活性协同的数据中心能量管理方法

随着数字经济产业快速发展,数据中心成为电能消耗的重要部分。多样化的数据处理需求、大量的计算用电及制冷用电,在带来高能耗问题的同时也为数据中心的运行调度提供了灵活性。因此,协同考虑计算灵活性和热灵活性,制定数据中心优化运行策略,可有效提高能源利用效率,提升数据中心运行经济性。该文提出一种考虑算力-热力灵活性协同的数据中心能量管理方法。首先,考虑数据负载灵活调度、冷却系统温度变化以及冷热通道热惯性等精细化要素,实现了信息技术设备和冷却设备的协同建模。其次,进行模型的线性化转换,降低了求解难度。在此基础上,实现了基于灵活性协同的数据中心能量管理,从而降低数据中心运行能耗及成本。最后通过算例验证了所提方法的有效性。

Load Shedding Strategy Based on Combined Feed-Forward Plus Feedback Control over Data Streams

In data stream management systems (DSMSs), how to maintain the quality of queries is a difficult problem because both the processing cost and data arrival rates are highly unpredictable. When the system is overloaded, quality degrades significantly and thus load shedding becomes necessary. Unlike processing overloading in the general way which is only by a feedback control (FB) loop to obtain a good and stable performance over data streams, a feedback plus feed-forward control (FFC) strategy is introduced in DSMSs, which have a good quality of service (QoS) in the aspects of miss ratio and processing delay. In this paper, a quality adaptation framework is proposed, in which the control-theory-based techniques are leveraged to adjust the application behavior with the considerations of the current system status. Compared to previous solutions, the FFC strategy achieves a good quality with a waste of fewer resources.

Load Balancing Fat-Tree on Long-Lived Flows:Avoiding Congestion in a Data Center Network

In a data center network （DCN）, load balancing is required when servers transfer data on the same path. This is necessary to avoid congestion. Load balancing is challenged by the dynamic transferral of demands and complex routing control. Because of the distributed nature of a traditional network, previous research on load balancing has mostly focused on improving the performance of the local network; thus, the load has not been optimally balanced across the entire network. In this paper, we propose a novel dynamic load-balancing algorithm for fat-tree. This algorithm avoids congestions to the great possible extent by searching for non-conflicting paths in a centralized way. We implement the algorithm in the popular software-defined networking architecture and evaluate the algorithm＇ s performance on the Mininet platform. The results show that our algorithm has higher bisection band- width than the traditional equal-cost multi-path load-balancing algorithm and thus more effectively avoids congestion.

A Strategy for Loading Oblique Photogrammetry Models and Multilayer Basemap Data

With the development of drone technology and oblique photogrammetry technology, the acquisition of oblique photogrammetry models and basemap becomes more and more convenient and quickly. The increase in the number of basemap leads to excessively redundant basemap tiles requests in 3D GIS when loading oblique photogrammetry models, which slows down the system. Aiming at improving the speed of running system, this paper proposes a dynamic strategy for loading basemap tiles. Different from existing 3D GIS which loading oblique photogrammetry models and basemap tiles inde-pendently, this strategy dynamically loads basemap tiles depending on different height of view and the range of loaded oblique photogrammetry models. We achieve dynamic loading of basemap tiles by predetermining whether the basemap tiles will be covered by the oblique photogrammetry models. The experimental results show that this strategy can greatly reduce the num-ber of redundant requests from the client to the server while ensuring the user’s visual requirements for the oblique photogrammetric model.

基于低秩张量补全的非侵入式负荷监测缺失数据修复方法

非侵入式负荷监测技术(non-intrusive load monitoring,NILM)作为实现智能电网用户侧细粒度感知的重要手段,有助于实现需求响应、提高“源-网-荷”互动效率和优化用能,助力实现“30·60目标”。高质量的量测信息是数据驱动型NILM的基础,但由于数据采集装置故障、通道拥塞以及延时等都会导致数据缺失,尤其是严重的连续性缺失,由此造成非侵入式负荷监测与分解的精度下降,影响用户画像、需求响应等高级应用。因此,针对该问题,提出了一种基于CP分解的正则化低秩张量补全的量测数据缺失修复方法。算法突破传统单维数据处理局限,对NILM多维量测数据构建了三阶观测张量,从而利用数据内部时序关联性和参量维度间电气关联性进行正则化低秩张量补全。并针对每次核范数计算过程中奇异值分解计算量过大问题,采用基于CP因子矩阵分解的核范数计算降低计算量,减少计算时长,并证明了变换的等效性。最后基于NILM公开数据集iAWE进行了实验,实验结果表明所提出的方法可以提高数据修复精度,在高缺失率和连续缺失情况下仍能有较好地补全效果,并且通过非侵入式负荷分解实验证明其可有效提高分解精度,对智能电网提升细粒度感知能力具有良好的实际意义。

基于双通道生成对抗网络的城市用电负荷缺失数据补全方法

用电负荷数据的完整性与有效性在负荷预测等应用中具有重要意义。传统的缺失数据补全方法缺乏对用电负荷和多种外部时空关联信息的挖掘,难以获得高精度的补全结果。文中提出了一种双通道生成对抗网络,对缺失的负荷数据进行补全。首先,根据负荷的周期性变化特征和时空关联性构建三阶负荷张量,并将影响负荷变化的多种外部因素构建为三阶辅助信息张量。然后,为满足两种张量的双输入需求,在生成对抗网络的输入层引入双通道机制,通过卷积与反卷积运算提取张量的特征;为提升网络对张量数据的训练效果和补全精度,将张量分解损失引入原始损失函数,并采用改进的混沌映射粒子群优化算法联合优化超参数和网络。最后,在真实负荷数据集上开展数据补全实验。结果表明,所提方法能够对随机缺失率不超过50%、连续缺失不超过3天的负荷数据进行准确补全。

基于生成对抗Transformer的电力负荷数据异常检测

电力负荷异常数据将给电力系统规划、负荷预测以及用能分析等带来较大的负面影响,因此亟须对负荷数据异常进行检测与识别。首先,针对电力负荷数据异常分类、原因及其特征开展分析。其次,改进传统Transformer编码器结构,采用多头注意力层代替掩码多头注意力层,同时移除前馈网络,以提高模型对负荷时序序列的全局注意力。基于生成对抗网络(generative adversarial networks,GAN)生成器与判别器的博弈结构,提出一种改进的GAN-Transformer模型,以更好地捕捉趋势性特征并加速模型收敛。然后,引入多阶段映射与训练方法,综合焦点分数打分机制,通过分阶段负荷序列重构帮助模型更好地提取负荷数据异常特征。最后,算例分析结果表明,GAN-Transformer模型在负荷数据异常检测精确率、召回率、F_(1)值以及训练时间方面均具有更优的性能,验证了所提方法的有效性和优越性。文中研究工作为基于深度学习进一步实现电力负荷数据异常分类与数据修复提供了有益参考。

面向负载均衡的动态均衡分区策略

针对MapReduce计算框架处理倾斜数据集时造成Reduce端出现负载不均衡现象,提出一种动态均衡分区策略。在mapper阶段提出基于分治法的数据切分原则处理任务传入的数据组;结合最佳适应算法思想设计动态分配原则逐步将切分后的数据块均衡分配到预分区链表中;根据分区索引分配到各Reduce节点上实现负载均衡。实验结果显示,动态均衡分区策略与两个基准模型相比任务执行时长平均降低了7.7%,表明动态均衡分区策略更好地解决了数据倾斜问题,降低了任务执行时间,验证了模型的有效性。

基于多模态数据的在线学习认知负荷评估

近年来,认知负荷过载成为影响在线学习效果的一个重要因素。为解决此问题,文章聚焦在线学习认知负荷评估,首先设计了基于多模态数据的在线学习认知负荷评估研究框架,包含多模态数据采集、多模态特征提取和评估模型构建三个部分。接着,文章通过实验,采集学习者在特定在线学习环境下的多模态数据,提取人脸表情特征、眼动追踪指标特征和脑电信号特征,构建在线学习认知负荷评估模型。之后,文章对评估模型的准确性进行验证,发现在五种机器学习算法模型中,KNN模型的综合性能最佳,故将此模型作为最终的在线学习认知评估模型;同时,文章对多模态数据进行了分析,发现多模态融合方法在认知负荷评估上具有优越性。文章构建的评估模型可赋能在线学习平台实现认知负荷的实时评估,进而实现基于认知负荷的个性化学习,增强学习动机,提升学习效果,促进在线教育质量提升。

基于双子空间PCA降维的脑力负荷分类

人类社会至今的飞速发展使得大量体力劳动被机械工程替代,工作者的任务重心也从体力劳动逐渐转变为脑力劳动,对操作者脑力负荷进行实时评估以增强工作效率在当下有着重大意义。目前人类对于脑力负荷评估共有3种方式,有研究表明,采用生物电信号进行脑力负荷分类效果较其余两种方法更客观。但脑电信号经过特征提取后维数极高,所需数据量和运算量巨大,需要对其进行降维。目前降维方面最广泛运用的两种算法为主成分分析(principal component analysis,PCA)和线性判别分析(linear discriminate analysis,LDA)。针对PCA的非监督性和LDA的特征冗余敏感性,提出一种二分类下基于双子空间主成分分析的降维算法,分别对不同类别的训练集数据进行主成分分析,并将所有训练集数据映射到生成的空间中,再次进行PCA-LDA降维,以此提高降维后数据的可分性。实验结果表明,双子空间PCA-LDA降维算法在二分类任务下测试集精度整体高于单子空间PCA-LDA算法,以此为脑力负荷分类领域和高维数据降维领域提供了新思路。

结构动载荷识别研究进展

作用于工程结构上的动载荷由于环境等限制难以通过直接测量的方式获取,基于动响应信息间接识别或重构动载荷已成为一种十分有效的途径。动载荷识别经过数十年的发展,已经形成了一系列行之有效的方法。本文总结了动载荷识别方法的研究历程及主要成果,系统性阐述了典型时、频域方法以及基于函数拟合思想、正则化策略、贝叶斯框架、数据驱动等动载荷识别方法,并讨论了各方法的优缺点以及适用范围。此外,还针对载荷识别过程中普遍存在的结构参数不确定问题以及输入条件不确定问题进行了总结。动载荷位置识别也是动载荷识别问题的重要组成部分,本文对现有位置识别方法进行了归纳分析。探讨了动载荷识别方法的工程应用,并分析了现有方法的局限性。结合当前实际工程应用中日益迫切的需求及动载荷识别领域面临的问题,展望了未来动载荷识别亟需攻克的技术难题以及可能的发展方向和重点领域。

基于实测数据的货车编队荷载作用特征分析

货车编队技术是货运交通智能化的重要研究方向之一。以货车编队的荷载作用为研究对象,在明确货车编队的横向、纵向控制模式的基础上,调研并总结了国内外货车编队的实测项目;基于AUTOPILOT项目的实测数据,分析了货车编队的横向和纵向荷载分布特征,并在此基础上分析了货车编队的轴载作用影响。结果表明:货车编队的轮迹横向分布标准差为7.14 cm,仅为传统货车的1/3,具有明显的“渠化”特征;货车编队在以100 km/h的速度行驶时,车头时距可低于0.8 s,约为相同速度下传统货车的27%;与传统货车相比,货车编队运行轴载作用次数具有“高频”特性,对现有路面结构的抗疲劳性能提出更高的要求。

面向大数据的可扩展正则采样并行排序算法

排序算法是计算机科学领域的一个基础算法,是大量应用的算法核心。在大数据时代,随着数据量的极速增长,并行排序算法受到广泛关注。现有的并行排序算法普遍存在通信开销过大、负载不均衡等问题,导致算法难以大规模扩展。针对以上问题,提出一种大规模可扩展的正则采样并行排序(scalable parallel sorting by regular sampling,ScaPSRS)算法,摒弃传统正则采样并行排序(parallel sorting by regular sampling,PSRS)算法中由一个进程负责采样的做法,转而让所有进程参与正则采样,选出p-1个分隔元素,将整个数据集划分成p个不相交的子集,然后实施并行排序,避免了单一进程的采样瓶颈。此外,ScaPSRS采用一种新的迭代更新策略选择p-1个分隔元素,保证划分的p个子集尽可能大小相同,从而确保p个进程对各自的子集进行本地排序时的负载均衡。在天河二号超级计算机上进行的大量实验表明,ScaPSRS算法能够成功地扩展到32000个内核,性能比PSRS算法和Hofmann等人提出的分区算法分别提升了3.7倍和11.7倍。

面向分布式数据库的算子并行优化策略

随着网络技术的不断发展,数据规模呈现爆发式增长,使得传统的单机数据库逐步被分布式数据库所取代。分布式数据库采用节点协同工作方式解决了大规模数据存储问题,但由于增加了节点间通信开销,查询效率却不如单机数据库。分布式架构下,存储节点的数据仅用作多备份的冗余,为系统故障时提供数据恢复,并未被利用起来改善查询效率。针对上述问题,提出了一种面向分布式数据库的算子并行优化策略,通过对关键物理算子进行拆分,将拆分后的子请求均匀分配到存储层多个节点,由多个节点并行处理,从而减少查询响应时间。上述策略已经在分布式数据库CBase上进行了应用,实验表明,提出的并行优化策略可显著缩短SQL请求查询时间,并提高系统资源利用率。