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.

Energy-balanced clustering protocol for data gathering in wireless sensor networks with unbalanced traffic load

Energy-efficient data gathering in multi-hop wireless sensor networks was studied,considering that different node produces different amounts of data in realistic environments.A novel dominating set based clustering protocol (DSCP) was proposed to solve the data gathering problem in this scenario.In DSCP,a node evaluates the potential lifetime of the network (from its local point of view) assuming that it acts as the cluster head,and claims to be a tentative cluster head if it maximizes the potential lifetime.When evaluating the potential lifetime of the network,a node considers not only its remaining energy,but also other factors including its traffic load,the number of its neighbors,and the traffic loads of its neighbors.A tentative cluster head becomes a final cluster head with a probability inversely proportional to the number of tentative cluster heads that cover its neighbors.The protocol can terminate in O(n/lg n) steps,and its total message complexity is O(n2/lg n).Simulation results show that DSCP can effectively prolong the lifetime of the network in multi-hop networks with unbalanced traffic load.Compared with EECT,the network lifetime is prolonged by 56.6% in average.

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).

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 toavoid congestion.Load balancing is challenged by the dynamic transferral of demands and complex routing control.Because ofthe distributed nature of a traditional network,previous research on load balancing has mostly focused on improving the perfor-mance of the local network;thus,the load has not been optimally balanced across the entire network.In this paper,we propose anovel dynamic load-balancing algorithm for fat-tree.This algorithm avoids congestions to the great possible extent by searching fornon-conflicting paths in a centralized way.We implement the algorithm in the popular software-defined networking architectureand 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.

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.

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.

Mining Rules from Electrical Load Time Series Data Set

The mining of the rules from the electrical load time series data which are collected from the EMS (Energy Management System) is discussed. The data from the EMS are too huge and sophisticated to be understood and used by the power system engineer, while useful information is hidden in the electrical load data. The authors discuss the use of fuzzy linguistic summary as data mining method to induce the rules from the electrical load time series. The data preprocessing techniques are also discussed in the paper.

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。

Optimization of support vector machine power load forecasting model based on data mining and Lyapunov exponents

According to the chaotic and non-linear characters of power load data,the time series matrix is established with the theory of phase-space reconstruction,and then Lyapunov exponents with chaotic time series are computed to determine the time delay and the embedding dimension.Due to different features of the data,data mining algorithm is conducted to classify the data into different groups.Redundant information is eliminated by the advantage of data mining technology,and the historical loads that have highly similar features with the forecasting day are searched by the system.As a result,the training data can be decreased and the computing speed can also be improved when constructing support vector machine(SVM) model.Then,SVM algorithm is used to predict power load with parameters that get in pretreatment.In order to prove the effectiveness of the new model,the calculation with data mining SVM algorithm is compared with that of single SVM and back propagation network.It can be seen that the new DSVM algorithm effectively improves the forecast accuracy by 0.75%,1.10% and 1.73% compared with SVM for two random dimensions of 11-dimension,14-dimension and BP network,respectively.This indicates that the DSVM gains perfect improvement effect in the short-term power load forecasting.

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。

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。

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。

工程降水对钻孔应变的干扰特征及干扰机制研究

抽水干扰在钻孔应变观测中常见且影响显著,为了能真实捕捉到地震前兆信息,须对其干扰机制进行研究.随着城市化进程,基坑工程及其伴随的工程降水逐步增多,工程降水对钻孔应变的干扰比普通单井抽水干扰更为严重.工程降水对钻孔应变观测有何干扰特征、干扰机制如何、干扰数据能否修复等问题对地震异常的探测具有重要意义.本文基于莱阳体应变受到附近御龙湾项目工程降水干扰的观测案例,研究了工程降水的干扰特征,修复了干扰畸变,通过数值模拟确定了干扰机制.结果表明:(1)工程降水在干扰体应变趋势变化的基础上叠加了方向明确、量级稳定的固体潮畸变,可以通过一阶差分逆运算进行修复.修复固体潮畸变并扣除背景速率后,工程降水干扰量级达-15.65×10^(-9)/d;(2)通过建立基于集中载荷模型的网格算法,模拟出工程降水在体应变异常初期和后期分别产生6.61×10^(9) N和1.73×10^(10) N的集中力载荷,干扰量级明显大于常见的单井抽水干扰,压性干扰集中于体应变仪与抽水井连线的垂直方向;(3)通过固体潮畸变特征及力学机制分析确定了工程降水造成固体潮畸变的干扰机制为"抽水引起压性趋势下降—抽水暂停引起张性台阶上升—抽水引起压性趋势下降"周期行为,固体潮畸变时间间隔(频次)和钻孔水位之间呈显著同步特征,线性系数为0.737.

基于输入特征稀疏化的图神经网络训练加速

图神经网络(graph neural network,GNN)是处理图数据的重要方法.由于计算复杂、图数据容量大,在大规模图上训练图神经网络依赖于CPU-GPU协作和图采样训练方法,其中图结构和特征数据存储在CPU内存中,而采样得到的子图及其特征则传输至GPU进行训练.然而,这种方法面临着严重的图特征数据加载瓶颈,显著降低了端到端训练性能,且图特征占用过多内存,严重限制了可训练的图规模.为了解决这些问题,我们提出了基于输入特征稀疏化的数据加载方法,显著减少CPU内存占用和跨PCIe总线传输的数据量,大幅缩短数据加载时间,加速GNN的训练,使其可以充分利用GPU计算资源.针对图特征和GNN计算特性,我们提出了适用于图特征数据的稀疏化方法,在压缩比和模型准确度之间达到平衡.我们在3个常见GNN模型和3个不同规模的数据集上进行了实验评估,包括最大的公开数据集之一MAG240M.结果表明,此方法将特征尺寸减小了一个数量级以上,并实现1.6-6.7倍的端到端训练加速,而模型准确度的降低不超过1%.此外,在仅使用4个GPU的情况下,仅需40 min就可以在MAG240M上完成GraphSAGE模型的训练并达到目标准确度.

某商业建筑高效制冷机房运行数据分析

制冷机房作为系统工程,其系统运行参数的变化(如冷冻水供回水温差)会对机房内不同设备的能效产生不同的影响,机房整体能效最优是机房节能运行的重要目标。针对现有建筑运行数据存储较多但对其分析较少的现状,采用数据挖掘技术,对某商业建筑制冷机房实际运行能耗、能效进行统计计算,并将其与负荷率、供回水温度、供回水温差等变量的相关性展开分析,揭示机房能效与各变量间的内在联系。研究结果显示:制冷机房单位面积制冷能耗约为16~17 kW·h/(m^(2)·a),冷水机组能耗占80%以上,冷水机组的能效表现主导了机房的整体能效表现,机房系统能效比(COPs)高于设计值6.06 kW·h/kW·h,达到了预期目标。机房整体能效与室外气温及负荷率、冷冻(却)水供回水温差、冷却水进水出水温度呈负相关关系,机房整体能效与冷冻水供回水温度主要呈正相关关系。冷水机组运行COP与机组负荷率和冷却水回水温度呈现一定相关性,且基本符合机组定冷却水温和AHRI变冷却水温工况下的选型曲线。

基于FFLD分析的汽车在极限凹坑路面工况下车身结构失效CAE新方法研究

在车身前期设计阶段,为了考核车辆在极限坑洼路面的冲击下对车身结构的影响,根据通用全球相关方坑道路试验标准,本文针对极限冲击的方坑#3路面,以某车型在方坑#3路试中出现了车身结构失效的问题为研究对象,找出了传统钣金失效准则CAE方法的不足,不能复现出试验失效的问题,进而通过大量料片试验建立了断裂成形极限图(fracture forming limit diagram,FFLD),来作为钣金失效准则的CAE新方法。随后基于该CAE新方法,通过车辆动力学科计算的方坑#3虚拟路谱作为载荷输入,对车身结构进行有限元仿真分析,成功复现了试验失效。根据CAE新方法分析的结果,对车身结构进行改进,最终通过了方坑#3路试验证,且试验和CAE仿真结果对标具有良好的一致性。证明了基于该方法,可以在车身开发前期阶段,准确预测出钣金在复杂变形条件下的真实破坏情况,从而减少了后期试验中出现的车身结构失效的风险。