• <tr id="yyy80"></tr>
  • <sup id="yyy80"></sup>
  • <tfoot id="yyy80"><noscript id="yyy80"></noscript></tfoot>
  • 99热精品在线国产_美女午夜性视频免费_国产精品国产高清国产av_av欧美777_自拍偷自拍亚洲精品老妇_亚洲熟女精品中文字幕_www日本黄色视频网_国产精品野战在线观看 ?

    Optimized Load Balancing Technique for Software Defined Network

    2022-08-24 12:56:34AashishKumarDarpanAnandSudanJhaGyanendraPrasadJoshiandWoongCho
    Computers Materials&Continua 2022年7期

    Aashish Kumar, Darpan Anand, Sudan Jha, Gyanendra Prasad Joshiand Woong Cho

    1Department of Computer Science and Engineering, Chandigarh University, Mohali, Punjab, 140413, India

    2School of Computational Sciences, CHRIST (Deemed to be University), Ghaziabad, 201003, India

    3Department of Computer Science and Engineering, Sejong University, Seoul, 05006, Korea

    4Department of Automotive ICT Convergence Engineering, Daegu Catholic University, Gyeongsan, 38430, Korea

    Abstract: Software-defined networking is one of the progressive and prominent innovations in Information and Communications Technology.It mitigates the issues that our conventional network was experiencing.However, traffic data generated by various applications is increasing day by day.In addition, as an organization’s digital transformation is accelerated, the amount of information to be processed inside the organization has increased explosively.It might be possible that a Software-Defined Network becomes a bottleneck and unavailable.Various models have been proposed in the literature to balance the load.However, most of the works consider only limited parameters and do not consider controller and transmission media loads.These loads also contribute to decreasing the performance of Software-Defined Networks.This work illustrates how a software-defined network can tackle the load at its software layer and give excellent results to distribute the load.We proposed a deep learning-dependent convolutional neural networkbased load balancing technique to handle a software-defined network load.The simulation results show that the proposed model requires fewer resources as compared to existing machine learning-based load balancing techniques.

    Keywords: SDN; software-defined networks; load balancing; performance enhancement

    1 Introduction

    In today’s era, traditional networks are unable to provide new services and fast innovation.Apart from this, the complexity of the network increases, which makes it harder to manage, and the data plane and control plane are coupled with each other as in router or switches.Software is deployed on dedicated hardware and, therefore, vendor-specific.Adding new Services requires costly new hardware.So, there is a need for a new approach to solving the issue discussed earlier.Therefore, a Softwaredefined Network (SDN) solves these problems, decoupling the control plane and information plane,making the network more agile, automated, open to all, and reducing business costs.However, today,the world is generating many data, it might be possible that the SDN network may get flooded with packets, and networks become congested and bottleneck.So, there is a need to balance the load in SDN [1].Let us discuss the architecture of SDN.In SDN, a logically centralized controller controls several routing devices, as shown in Fig.1 [1-3].An important function of SDN architecture, given in Fig.1, are as follows:

    ?Information Plane:Switching devices such as routers, switches, etc., are part of the information plane.It is responsible for processing and forwarding the packets as per the rules defined in the forwarding table.

    ?Controller plane:It is also known as“NETWORK BRAIN”and consists of controllers.It takes a decision where a packet is to be forwarded.The control plane’s functions are system configuration, management, and routing table information exchange.It can be deployed in a different geographical area.

    ?Management plane:It is used for management and configure as per user’s demand.

    Still, there was a problem with a standard protocol that helps communication between the information plane and controller plane.The OpenFlow protocol was created in 2008 to address the problem [4].The Main Building blocks of Openflow are as follows:

    ?Flow table:It is the important com ponent of OpenFlow switches/routers.When a packet arrives at a particular switch, the header field of the packet must match with flow table entries (shown in Fig.2, which includes various phases, layers, and its header description [5].

    ?Port:Openflow characterized ports are capable of forwarding the packets to the controller,flooding, etc.

    ?Messages:The OpenFlow controller and switch represent messages that exchange between the OpenFlow controller and switches.Messages can be symmetric messages, asynchronous messages, or initiated by the controller.For example, PACKET IN, ECHO, PORT STATUS message.

    Figure 2: Structure of flow table

    As illustrated in this manuscript, load-balancing is required through smart and intelligent algorithms that can tackle the load at its software layer and give excellent results to distribute the load.We propose a deep learning-dependent convolutional neural network (CNN)-based load balancing technique to handle the load in SDN.Further, we compare the test results with existing machine learning-based load balancing techniques.

    An issue with SDN is that its controller can be overloaded by traffic.Therefore, we will discuss load balancing in the controller using a load balancer in SDN in a later section.

    2 Load Balancing Fundamentals

    IIt is a procedure that is utilized to distribute the load among servers or any other computing device.Maximum throughput,minimum response time, maximum utilization of resources, avoiding overload, and avoiding crashing are the main objective of load balancing.Load balancing can be performed by hardware, software, or a combination of both.Dedicated software load balancers are manufacturerspecific and are costly.It can run on the hardware or virtual machine, as shown in Fig.3.In our proposed methodology, we deal with the issue of load balancing.

    Figure 3: Load balancer

    Intelligent Load balancing

    In today’s world, due to the huge increase in traffic, the demand of users, complexity of the application.Therefore, the business model needs to a robust, reliable, scalable, less complex, and adapt to new challenges with reduced cost.Hence there is a need for an intelligent load balancer.An intelligent load balancer distributes the packet or provides the optimized path considering various parameters.An intelligent load balancer can be achieved using Machine learning techniques.There are various machine learning techniques available, out of which we have taken a CNN to build our model.

    3 Related Work

    This section looks at the research that has been published in the literature.It includes pros and cons, as well as the parameter they used for load balancing.Tab.1 summarizes the algorithm used and the basis of the decision.

    Table 1: Related work on SDN load balancing

    Table 1: Continued

    Table 1: Continued

    The details of Tab.1 are described as follows: Li-Der Chou et al.presented an algorithm based on genetics, and a coefficient is utilized to assess the performance of the algorithm.They have selected the roulette-wheel selection, single-point cross-over, and single-point mutation and compared it with the round-robin algorithm [6].

    Zhang et al.presented a strategy called Hybrid Routing.It uses both destination and explicit routing.In this methodology, the intricacy of the explicit routing diminishes, and the performance of the methodology is enhanced [7].

    Zhou et al.proposed DALB (dynamic and adaptive algorithm) algorithm having distributed controller.Thistakes O(n)as overhead.It is tested on attributes such as throughput and average arrival time of the packet [8].

    Wang et al.presented a load balancing methodology.The proposed model has three standard parts: a load collector, load balancer, and switch migrator.The load collector (Cbench) diverts the traffic to the controller.A controller has the threshold of the 70% of the controller information transmission [9].At the point when the controller beats the threshold, the load balancer picks the controller (lightest weight).

    Hadar Sufiev proposed a multi-regulator load adjusting model referred to as dynamic routing.Two sorts of controllers are the Super regulator andRegular regulator in the model.The load balancing occurs at two levels, one at the Super controller and the second at the Regular Controller.The supercontroller continuously checks the gathering whether they have passed the threshold or not.On the off chance that when load passes the limit, at that point, the segment calculation runs [10].

    He et al.combined fog computing with software-defined networks.Internet of Vehicles is integrated with Fog Computing-Software defined technology.The proposed methodology increases the QoS and reduces latency [11].

    Hu et al.presented a load balancing technique called switch migration algorithm(SMDM).SMDM can be divided into three phases.At that point, it predicts migration costs and migration efficiency.It decides whether to carry out migration movement activities or not [12].

    Zhong et al.introduced a methodology called Load Balancing Server Response Time (LBBSRT).It first gathers every server’s reaction time at that point, picks the server, which has a steady reaction time.The proposed methodology is simple and efficiently balances the load among the controller and low response time [13].

    Rangisetti et al.presented a combined version of Long-Term Evolution (LTE) Radio Access Network (RAN) and SDN and came up with a QALB algorithm.The algorithm calculates the neighboring cell’s load, the user’s QoS profile Equipment, and estimated throughput to balance the load.For more than 80 percent of the cell.The proposed algorithm provides a better QoS data rate [14].

    Aly provided a load balancing methodology called controller adaptive load balancing (CALB).In this methodology, the super controller and slave controller are connected and, in turn, connected with n switches.It is used for both load balancing and fault tolerance [16].

    Wilson Prakash et al.proposed an approach called Dynamic Agent-based load balancing(DALB), which is based on Backpropagation artificial neural network (BPANN) [15].BPNN is integrated into the Controller agent.The BPNN takes the three inputs: CPU utilization, Memory utilization, and response time [18].

    We propose an intelligent load balancing model (CNN 1D) for a software-defined network that takes 16 parameters, as discussed in the latter part of the section, including backward and forward parameters.We have used a convolution neural network that is less complex and takes less time than existing researcher models and machine learning models trained using KMeans.Let us discuss each element of the proposed model in detail.

    4 Proposed Methodology

    Based on the previous work, we conclude that for load balancing, researchers take the load of the controller or take the load of transmission media.One thing common in both of them is that they have taken two or three-parameter for decision making.Nevertheless, not considered a load of transmission media and controllers load simultaneously.However, in our proposed methodology,we have considered both the loads and based on a CNN, we have used TensorFlow and Keras to implement the model.CNN is used because it is less complex and gives accurate results than other machine learning techniques.

    Our main goal is to suggest a strategy for the Software-Defined System that considers all of the various parameters.The architecture of the suggested model is based on the dataset taken from the kaggle.com of Universidad Del Cauca Popayan, Colombia [19].

    Clients send a request for a specific service to the server under our proposed method.It first passes the software load balancer, where the algorithm is implemented.Using CIC Flowmeter, the first algorithm will acquire flow statistics such as IP addresses, ports, inter-arrival periods, and so on.Based on these features, it determines the cluster value of the request and sends it to the relevant servers.In the following papes, we will go through each point in detail.

    4.1 Process Flow of the Controller for SDN Traffic

    In the proposed model, when a client sends a particular server service, it initially sends it to the controller.The load balancer checks the packet size of the packet and according to the flow statistics of both backward and forward direction of the flow captured using CIC Flowmeter.It will send to a router (server) based on the optimized results obtained through the CNN model and threshold of the router (server).Here, an important point to be considered is that the threshold of the router (server) is based on how many bytes per second it can process and the total number of the packet it can handle at any instance of time shown in Fig.4.Let us examine the dataset used in the model in the following segment.

    Figure 4: Proposed solution for software-defined network load balancing

    4.1.1 DataSet from Kaggle

    We used the dataset provided by Zhao et al.[19].Fundamentally, it was caught from the Universidad Del Cauca, Popayan, Colombia United States of America.These data are from the TCP/UDP traffic caught at various hours on various days during the morning and evening of 2017.There is an aggregate of 3577296 packets.Out of these, 100000 packets are used to train our machine.This dataset has 87 features.We have taken 16 parameters to train and test the model, as shown in Tab.2.

    Table 2: Parameters of the dataset

    Visualization of the dataset is summarized in Fig.5 [6].Fig.5 shows the correlation among the parameters and calculated using the Pearson correlation coefficient.In Fig.6, all the parameters are taken on the x-axis and y-axis, generating correlations for exploratory analysis.

    We took some parameters to visualize using the Seaborn module in Fig.5.All the above statistics are helpful to perform Exploratory data analysis.In the following section, we will examine Data preprocessing.

    Figure 5: Correlation among the parameters in the dataset

    4.1.2 Data Preprocessing

    Data preprocessing is the procedure to process the raw information with the goal that it tends to be utilized effectively with the three phases as Data Cleaning, Data Transformation, and Data Reduction

    We first check the missing and boisterous information in our dataset, just as information sort of every parameter.These are accomplished by the panda’s module, the Seaborn module of Python.Whatever we get the result, we performed hashing on this dataset.To standardize information, with the goal that it can lie somewhere in the range of 0 and 1.One of the famous and vital techniques to normalize the dataset is Min-Max scaling.

    Min-max scalar formula is as follows:

    where x is the ith value,max (x)andmin (x)are maximum and minimum points in the particular column of a dataset, respectively, and z is the normalized value that lies between 0 to 1.This can be obtained bysklearn.preprocessing.MinMaxScalerclass.

    4.2 Convolutional Neural Network (CNN or ConvNets)

    A CNN is a subset of deep learning which is inspired by biological processes where connectivity between neurons is similar to that of the animal cortex.Commonly used CNN is CNN 1D and CNN 2D.

    Figure 6: Dataset analysis

    COMMON LAYERS are: let us take an example of our model by running a CNN 1D on 16 features.

    Input Layer:It contains input data of different dimensions and is based onCNN1D,CNN2D, etc.

    Convolution Layer:The primary function detects the feature and yields the output by calculating the multiplication between filters and data size (in matrix form) as Fig.6.Then slightly move the filter based on the stride size and perform the same operation.This process continues till we are at the end of the image.The output volume of the convolutional layer can be measured as:

    wherenl= Output dimension oflthlayer,-1=Output dimension of (l-1)thlayer,s[l]= Stride used in the layer,f[l]= Size of the filter, p[l-1]= Padding added of the l-1 layer, As shown in Fig.7 there is images of sizenh×nwbe 6×6 and the size of filter(f)= 3, stride (s) = 1 and padding (p)=0.We overlap the filter to the information, perform component-wise multiplication, and add the outcome.Move the filter according to the stride and perform the same calculation until it reaches the end of the image.

    Figure 7: Convolutional layer

    Activation Function Layer:The activation function is applied to each element of the previous layer output.Examples are RELU.Sigmoid.Tanh, Leaky RELU, etc.The output remains unchanged.Here we have used RELU and Sigmoid function

    Pooling Layer:The main objective is to diminish the size of the previous convolutional layer output, which makes the estimation fast, diminishes memory.There are two types of operation: max pooling and average pooling.It has three parameters, i.e., size, stride, and type of operations.An example of the Average pooling layer and Max pooling layer is shown in Fig.8.Max pooling is a pooling activity that chooses the most significant component from the district of the element map secured by the filter.Average pooling registers the normal of the components present in the highlighted map area secured by the channel.Along these lines, while max-pooling gives the most conspicuous component in a specific fix of the element map, normal pooling gives the normal of highlights present in a fix.

    Fully-Connected Layer:It is a deep neural system layer that takes contribution from the past layer and registers the class scores, and yields the 1-D cluster of size equal to the number of classes as shown in Fig.9.All the layer that has been discussed with their output is shown in Fig.10.In the proposed model, we have used two Convolution layers: conv1d 1 and conv1d 2.The output is obtained using Eq.(2).HereMAX POOLING is used(as depicted max pooling1d 1).Then we flatten the result obtained usingMAXPOOLING LAYER (shown as flatten) 1.We have used four dense layers (shown as dense 1, dense 2, dense 3, dense 4) and get the output in dense 4 layers (all the layers shown in Fig.10.

    Figure 8: Pooling layer

    Figure 9: Performance of convolutional neural network model

    Figure 10: Elbow method in KMEANS

    5 Training Model

    The Convolution neural network is trained on the training data for 30 epochs to get an accuracy of 98.94 percent, and testing is performed on 20 percent of the data.Here we have used the loss function as Binary cross entropy [20], while the Accuracy and Confusion matrix is used as a metric to test the model’s accuracy.Now the forward and backward flow features (16 parameters) are used to obtain optimized results.

    5.1 Training Results of CNN Model

    We tested the model on 43000 packets, out of which 20 percent are used for testing purposes,and the remaining 80 percent are used for training.The output has been shown in Figs.10 and 11.In contrast, the performance has been illustrated in Fig.10.A total of 9 layers are used to train the proposed model.We used 90, 60 filters in the convolution layer with a kernel size of 2, and in the pooling scenario, a pool size of two is used.The results are summarized in Tab.3 and also shown in short form in Fig.11.

    Figure 11: (Continued)

    Figure 11: KMeans method in KMEANS (a) KMean results for n=1000 pkts (b) KMean results for n=5000 pkts (c) KMean results for n=10000 pkts

    5.2 Comparison

    We have our compared model with the Machine learning model trained using KMeans.It is tested on 5000, 10000, 20000, 50000 packets (shown in Figs.10 and 11 First, in a machine learning model trained using KMeans, the elbow method is used to find an optimal number of servers (as shown in Fig.10).The trained machine requires four servers to balance the load as concluded from the elbow method, and in Fig.11, color legends show that numbers of servers and particular requests will go to the particular server.

    The primary objective of any machine learning model is the efficiency parameters or performance parameters.The Tab.3 illustrates the accuracy of the system is 98.94% while the validation accuracy for the system is more than 99%, i.e., 99.22%.It illustrates the correctness of the system.In terms of the loss percentage which is very important parameter specially for the loadbalancing techniques for software defined network.This loss parameter is 3.61% and therefore the final parameter is validation loss which is 2.52%.The clustering details are illustrated Figs.10 and 11.The Fig.10 is illustrating the elbow method to validate the number of clusters while the Fig.11 is illustrating the clusters formed during the loadbalancing process.The system is tested for 1000, 5000 and 10000 packets respectively but the clutering is showing good in figure.

    Table 3: Result of proposed convolutional neural network model

    From this, we conclude that it requires four servers to balance the load obtained using KMeans,whereas our model can give optimized results on two servers only.

    6 Conclusion

    Load balancing is an essential and crucialmechanism for distributed computing-based technology in current network administration and application.A software-defined network is a technology to monitor and manage distributed computing requirements through various network management approaches, such as layered architecture, network virtualization, and orchestration.As the load is changed, the technique to handle this issue is also required to change.Therefore, we detailed the dynamic and intelligent method to handle load balancing for SDN in this work.The proposed techniques are based on deep learning, i.e., in particular, convolution neural networks.The proposed system is tested for an increasing load.The simulation results show that the proposed model performs better than the literature’s existing static load balancing methods.We will plan to embed the code for the proposed technique through the SDN load balancer at the control plane.

    Funding Statement:This work was supported by Ulsan Metropolitan City-ETRI joint cooperation Project [21AS1600], Development of intelligent technology for key industries and autonomous human-mobile-space autonomous collaboration intelligence technology].

    Conflicts of Interest:The authors declare that they have no conflicts of interest to report regarding the present study.

    精品国内亚洲2022精品成人| 欧美日韩亚洲国产一区二区在线观看| 国产精品日韩av在线免费观看| 91av网一区二区| 网址你懂的国产日韩在线| 欧美一级a爱片免费观看看| 午夜爱爱视频在线播放| 欧美性猛交黑人性爽| 亚洲美女黄片视频| 一个人看的www免费观看视频| 九色国产91popny在线| 亚洲最大成人手机在线| 国产一区二区三区在线臀色熟女| 日本色播在线视频| 九九在线视频观看精品| 午夜视频国产福利| 亚洲欧美精品综合久久99| 午夜福利视频1000在线观看| 淫秽高清视频在线观看| 在线观看一区二区三区| 国产极品精品免费视频能看的| 国产免费男女视频| 真实男女啪啪啪动态图| 久久草成人影院| a在线观看视频网站| a级毛片免费高清观看在线播放| 午夜a级毛片| 精品人妻偷拍中文字幕| 国产乱人伦免费视频| 99热精品在线国产| 亚洲av电影不卡..在线观看| 性色avwww在线观看| 免费大片18禁| 丰满人妻一区二区三区视频av| 午夜福利在线在线| 亚洲中文字幕日韩| 中国美白少妇内射xxxbb| 国产av一区在线观看免费| aaaaa片日本免费| 中亚洲国语对白在线视频| 亚洲av日韩精品久久久久久密| 色噜噜av男人的天堂激情| 1000部很黄的大片| 看免费成人av毛片| 可以在线观看毛片的网站| 精品日产1卡2卡| 欧美色视频一区免费| 男女之事视频高清在线观看| 国产av不卡久久| www日本黄色视频网| 午夜福利在线观看免费完整高清在 | 久久久久久久久久久丰满 | 亚洲久久久久久中文字幕| 能在线免费观看的黄片| 最近中文字幕高清免费大全6 | 99在线人妻在线中文字幕| 婷婷色综合大香蕉| 亚洲中文字幕一区二区三区有码在线看| 熟妇人妻久久中文字幕3abv| 精品久久久久久久久久久久久| 一进一出抽搐动态| 欧美在线一区亚洲| 国产av不卡久久| 久久这里只有精品中国| 97碰自拍视频| 亚洲成人免费电影在线观看| 国产伦一二天堂av在线观看| 精品久久久久久久久久免费视频| 精品日产1卡2卡| 国产美女午夜福利| 最后的刺客免费高清国语| 18禁在线播放成人免费| 国产成年人精品一区二区| av国产免费在线观看| 成人av一区二区三区在线看| 亚洲中文字幕日韩| 久久久久国产精品人妻aⅴ院| 日日夜夜操网爽| 日本三级黄在线观看| 少妇人妻精品综合一区二区 | 日韩亚洲欧美综合| 69人妻影院| 18禁在线播放成人免费| 少妇的逼好多水| 一进一出好大好爽视频| 精华霜和精华液先用哪个| 亚洲一区二区三区色噜噜| 三级男女做爰猛烈吃奶摸视频| 欧美+日韩+精品| 精品久久久久久久人妻蜜臀av| 日韩欧美国产在线观看| 性欧美人与动物交配| 午夜激情欧美在线| 日韩亚洲欧美综合| 18禁裸乳无遮挡免费网站照片| 欧美一区二区亚洲| 国产亚洲91精品色在线| 免费av毛片视频| 黄片wwwwww| 国产一区二区三区av在线 | 久久久精品欧美日韩精品| 国产亚洲91精品色在线| 久久精品国产亚洲av香蕉五月| 亚洲精品一区av在线观看| 精品一区二区三区人妻视频| 黄色日韩在线| 国产伦在线观看视频一区| 波多野结衣巨乳人妻| 中文字幕久久专区| av黄色大香蕉| 内地一区二区视频在线| 亚洲成a人片在线一区二区| 美女cb高潮喷水在线观看| 精品久久久久久久末码| av在线天堂中文字幕| 高清日韩中文字幕在线| 18禁黄网站禁片免费观看直播| 久久热精品热| 999久久久精品免费观看国产| 免费无遮挡裸体视频| 伦精品一区二区三区| 日韩欧美在线乱码| 午夜精品一区二区三区免费看| 中出人妻视频一区二区| 国内精品久久久久久久电影| 美女cb高潮喷水在线观看| 狠狠狠狠99中文字幕| 中文字幕熟女人妻在线| 欧美一区二区国产精品久久精品| АⅤ资源中文在线天堂| 亚洲一级一片aⅴ在线观看| 草草在线视频免费看| 黄片wwwwww| 自拍偷自拍亚洲精品老妇| 亚洲精品亚洲一区二区| 免费在线观看影片大全网站| av.在线天堂| 亚洲成人免费电影在线观看| 在线a可以看的网站| 婷婷亚洲欧美| 不卡视频在线观看欧美| 色5月婷婷丁香| 国产伦在线观看视频一区| 亚洲性久久影院| 观看免费一级毛片| 色av中文字幕| 亚洲人成网站在线播| 国产精品伦人一区二区| 成年女人毛片免费观看观看9| 国产亚洲欧美98| 麻豆精品久久久久久蜜桃| 51国产日韩欧美| 99在线视频只有这里精品首页| 亚洲四区av| 春色校园在线视频观看| 成人无遮挡网站| 一个人观看的视频www高清免费观看| 亚洲av一区综合| 听说在线观看完整版免费高清| 欧美精品国产亚洲| 亚洲一级一片aⅴ在线观看| 高清日韩中文字幕在线| 99久久中文字幕三级久久日本| 黄片wwwwww| 男女啪啪激烈高潮av片| 人人妻,人人澡人人爽秒播| 国产高清视频在线播放一区| 永久网站在线| 国产一级毛片七仙女欲春2| 99视频精品全部免费 在线| 久久香蕉精品热| 国产精品亚洲一级av第二区| 国产老妇女一区| .国产精品久久| 国产精品98久久久久久宅男小说| 两性午夜刺激爽爽歪歪视频在线观看| 又爽又黄无遮挡网站| 亚洲va在线va天堂va国产| 欧美性猛交黑人性爽| a级毛片a级免费在线| 露出奶头的视频| 黄色视频,在线免费观看| 我要看日韩黄色一级片| 深爱激情五月婷婷| 观看美女的网站| 欧美黑人巨大hd| 中亚洲国语对白在线视频| 亚洲aⅴ乱码一区二区在线播放| 亚洲欧美日韩卡通动漫| 欧美激情在线99| 99国产精品一区二区蜜桃av| 少妇猛男粗大的猛烈进出视频 | 日日撸夜夜添| 精品久久久久久久久av| 天美传媒精品一区二区| 女人十人毛片免费观看3o分钟| 亚洲欧美清纯卡通| 午夜福利视频1000在线观看| 久久午夜亚洲精品久久| 别揉我奶头~嗯~啊~动态视频| 别揉我奶头~嗯~啊~动态视频| 无人区码免费观看不卡| 一卡2卡三卡四卡精品乱码亚洲| 九九爱精品视频在线观看| 午夜久久久久精精品| 日韩大尺度精品在线看网址| 赤兔流量卡办理| 熟妇人妻久久中文字幕3abv| 亚洲色图av天堂| 成年女人看的毛片在线观看| 日韩精品青青久久久久久| 人人妻人人澡欧美一区二区| 一a级毛片在线观看| 日韩欧美一区二区三区在线观看| 亚洲精品色激情综合| 国产私拍福利视频在线观看| 美女高潮喷水抽搐中文字幕| 在线免费观看的www视频| 久久久久久久午夜电影| 男女之事视频高清在线观看| 日本 欧美在线| 在线免费十八禁| 色视频www国产| 特大巨黑吊av在线直播| 又黄又爽又刺激的免费视频.| 成人一区二区视频在线观看| 国产精品99久久久久久久久| 婷婷精品国产亚洲av| 三级国产精品欧美在线观看| 日日啪夜夜撸| 热99在线观看视频| 亚洲国产欧洲综合997久久,| 听说在线观看完整版免费高清| 夜夜爽天天搞| 国产人妻一区二区三区在| 男人狂女人下面高潮的视频| 九九爱精品视频在线观看| 精品国内亚洲2022精品成人| 两性午夜刺激爽爽歪歪视频在线观看| 自拍偷自拍亚洲精品老妇| 三级男女做爰猛烈吃奶摸视频| 亚洲成人中文字幕在线播放| 极品教师在线免费播放| 非洲黑人性xxxx精品又粗又长| 天堂av国产一区二区熟女人妻| 久久久久久九九精品二区国产| 麻豆国产97在线/欧美| 狂野欧美白嫩少妇大欣赏| 色综合婷婷激情| 日韩欧美国产一区二区入口| 欧美日韩中文字幕国产精品一区二区三区| 成人性生交大片免费视频hd| 精品久久久久久久人妻蜜臀av| 日韩欧美国产一区二区入口| 狂野欧美激情性xxxx在线观看| 久9热在线精品视频| 在线观看舔阴道视频| 亚洲av美国av| 午夜久久久久精精品| 无遮挡黄片免费观看| 久久午夜亚洲精品久久| 一级黄色大片毛片| 夜夜爽天天搞| 国产精品国产高清国产av| 国产一区二区三区在线臀色熟女| 国产伦精品一区二区三区四那| 香蕉av资源在线| 中出人妻视频一区二区| 国产午夜精品久久久久久一区二区三区 | 国国产精品蜜臀av免费| 成人美女网站在线观看视频| 免费av毛片视频| 成年女人毛片免费观看观看9| 在线观看av片永久免费下载| 1024手机看黄色片| 999久久久精品免费观看国产| 色播亚洲综合网| 久久久久免费精品人妻一区二区| 久久久久久大精品| 99久久精品一区二区三区| 美女黄网站色视频| 中文在线观看免费www的网站| 精品久久久久久久久亚洲 | 男女下面进入的视频免费午夜| 国产日本99.免费观看| 国产一级毛片七仙女欲春2| 男女视频在线观看网站免费| 色综合色国产| 99国产精品一区二区蜜桃av| 成人高潮视频无遮挡免费网站| a级毛片a级免费在线| 欧美一级a爱片免费观看看| 黄色女人牲交| 精品久久久久久成人av| 亚洲精品日韩av片在线观看| 成人鲁丝片一二三区免费| 亚洲 国产 在线| 日本一二三区视频观看| 成人特级黄色片久久久久久久| 丝袜美腿在线中文| 婷婷亚洲欧美| 91久久精品国产一区二区成人| 成人永久免费在线观看视频| 免费大片18禁| 欧美激情国产日韩精品一区| 观看美女的网站| netflix在线观看网站| 亚洲熟妇中文字幕五十中出| 国产色爽女视频免费观看| 国产主播在线观看一区二区| 免费黄网站久久成人精品| 少妇人妻一区二区三区视频| 中文字幕精品亚洲无线码一区| 国内精品美女久久久久久| 最后的刺客免费高清国语| 成年版毛片免费区| 国产精品久久久久久av不卡| 一个人看的www免费观看视频| 日本爱情动作片www.在线观看 | 国产一区二区三区av在线 | 国产精品一区www在线观看 | 中文字幕免费在线视频6| 精品久久久久久,| 免费在线观看成人毛片| 午夜免费成人在线视频| 51国产日韩欧美| 老熟妇仑乱视频hdxx| 日本一二三区视频观看| 亚洲第一区二区三区不卡| 日韩欧美精品免费久久| 成人特级黄色片久久久久久久| 一区二区三区免费毛片| 亚洲专区国产一区二区| 18禁裸乳无遮挡免费网站照片| 亚洲精品一卡2卡三卡4卡5卡| 岛国在线免费视频观看| 亚洲色图av天堂| 成人午夜高清在线视频| 国产精品国产高清国产av| 嫩草影院新地址| 欧美日韩国产亚洲二区| 亚洲av第一区精品v没综合| 中国美白少妇内射xxxbb| 淫秽高清视频在线观看| 中文资源天堂在线| 一个人免费在线观看电影| 国产伦在线观看视频一区| 国产精品精品国产色婷婷| 18+在线观看网站| 国产一区二区亚洲精品在线观看| 人人妻人人看人人澡| 色综合站精品国产| 啦啦啦韩国在线观看视频| 亚洲熟妇中文字幕五十中出| 日日夜夜操网爽| 国产成人福利小说| 中文字幕av在线有码专区| 色哟哟·www| 人妻丰满熟妇av一区二区三区| 日韩欧美三级三区| 啦啦啦观看免费观看视频高清| 国产精品一区二区三区四区久久| 窝窝影院91人妻| 亚洲精品粉嫩美女一区| 精品久久久久久久人妻蜜臀av| 国产一区二区激情短视频| 国产精品野战在线观看| 国产一区二区三区在线臀色熟女| 欧美日本视频| 日韩欧美免费精品| 国产毛片a区久久久久| 欧美zozozo另类| 白带黄色成豆腐渣| 性色avwww在线观看| 亚洲人成网站在线播放欧美日韩| 亚洲四区av| 亚洲中文字幕一区二区三区有码在线看| 给我免费播放毛片高清在线观看| 一进一出抽搐gif免费好疼| 麻豆成人午夜福利视频| 日韩欧美国产一区二区入口| 国产亚洲欧美98| 国产精品无大码| 老司机午夜福利在线观看视频| 18禁黄网站禁片免费观看直播| 老司机福利观看| 国内揄拍国产精品人妻在线| 亚洲第一电影网av| 99热只有精品国产| 国产精品野战在线观看| 99九九线精品视频在线观看视频| 成人高潮视频无遮挡免费网站| 亚洲aⅴ乱码一区二区在线播放| 国产白丝娇喘喷水9色精品| 嫩草影院入口| 永久网站在线| 在线免费观看不下载黄p国产 | 亚洲美女搞黄在线观看 | 给我免费播放毛片高清在线观看| 日韩欧美 国产精品| bbb黄色大片| 国产精品久久久久久亚洲av鲁大| 国产精品伦人一区二区| 久久精品国产亚洲av香蕉五月| 精品人妻一区二区三区麻豆 | 乱人视频在线观看| 欧美日韩国产亚洲二区| 免费搜索国产男女视频| 在线观看免费视频日本深夜| 国产淫片久久久久久久久| 悠悠久久av| 少妇人妻精品综合一区二区 | 亚洲国产精品合色在线| 久久久色成人| 日本黄大片高清| 好男人在线观看高清免费视频| 悠悠久久av| 少妇人妻精品综合一区二区 | 伊人久久精品亚洲午夜| 夜夜爽天天搞| 久久精品91蜜桃| 成人av在线播放网站| 成人国产一区最新在线观看| 老司机午夜福利在线观看视频| 一区福利在线观看| 国产在线精品亚洲第一网站| 少妇的逼好多水| 18+在线观看网站| 日本免费a在线| 黄色配什么色好看| 亚洲欧美清纯卡通| 又粗又爽又猛毛片免费看| 久久精品国产亚洲网站| 亚洲欧美日韩无卡精品| 国产久久久一区二区三区| 免费电影在线观看免费观看| 亚洲不卡免费看| 国产高清激情床上av| 国产又黄又爽又无遮挡在线| 亚洲精品粉嫩美女一区| 久久久久久久久久成人| 啪啪无遮挡十八禁网站| 久久精品91蜜桃| 国产精品美女特级片免费视频播放器| 久久人人精品亚洲av| 亚洲久久久久久中文字幕| 日韩欧美三级三区| 国产精品人妻久久久久久| 一夜夜www| 又紧又爽又黄一区二区| 乱系列少妇在线播放| 无遮挡黄片免费观看| 亚洲不卡免费看| 国产高清视频在线观看网站| 免费av观看视频| 校园人妻丝袜中文字幕| 在线天堂最新版资源| 国产精品美女特级片免费视频播放器| 无遮挡黄片免费观看| 日本在线视频免费播放| 亚洲精品粉嫩美女一区| 两性午夜刺激爽爽歪歪视频在线观看| 日本 欧美在线| 三级国产精品欧美在线观看| 美女高潮的动态| 久久婷婷人人爽人人干人人爱| 天堂网av新在线| 亚洲性夜色夜夜综合| 免费无遮挡裸体视频| 久久午夜亚洲精品久久| 亚洲乱码一区二区免费版| 又爽又黄无遮挡网站| 深夜a级毛片| 亚洲欧美精品综合久久99| 亚洲国产精品久久男人天堂| 亚洲国产高清在线一区二区三| 亚洲最大成人av| 成人精品一区二区免费| 变态另类丝袜制服| 国产高清三级在线| 精华霜和精华液先用哪个| 日韩欧美在线二视频| 国产高清激情床上av| 精品人妻一区二区三区麻豆 | 亚洲色图av天堂| 69av精品久久久久久| 又黄又爽又免费观看的视频| 偷拍熟女少妇极品色| 中国美白少妇内射xxxbb| 国产一区二区三区在线臀色熟女| 亚洲中文日韩欧美视频| 亚洲av日韩精品久久久久久密| 国产探花在线观看一区二区| 欧美精品国产亚洲| 黄色欧美视频在线观看| 波多野结衣高清作品| 亚洲欧美日韩高清在线视频| 国产高清视频在线观看网站| 亚洲在线自拍视频| 日本精品一区二区三区蜜桃| 亚洲人成网站高清观看| 亚洲精品色激情综合| 亚洲在线自拍视频| 精品午夜福利在线看| 国内精品久久久久精免费| 欧美精品国产亚洲| 免费在线观看日本一区| 亚洲熟妇熟女久久| 成人欧美大片| 乱人视频在线观看| 色综合婷婷激情| 少妇裸体淫交视频免费看高清| 亚洲精华国产精华液的使用体验 | 国产乱人伦免费视频| 女人十人毛片免费观看3o分钟| 亚洲欧美日韩东京热| 丰满人妻一区二区三区视频av| 日韩欧美国产在线观看| 国产精品国产三级国产av玫瑰| 精品久久久久久久久久免费视频| 亚州av有码| 午夜精品一区二区三区免费看| 亚洲av成人av| 大又大粗又爽又黄少妇毛片口| 精品久久久久久久久av| 少妇的逼水好多| 噜噜噜噜噜久久久久久91| 久久久久久久久中文| 天堂动漫精品| 国产欧美日韩精品一区二区| 欧美性猛交黑人性爽| 尾随美女入室| 91在线精品国自产拍蜜月| 欧美不卡视频在线免费观看| 精品久久久久久久末码| 99视频精品全部免费 在线| 毛片女人毛片| 国产一级毛片七仙女欲春2| 麻豆成人午夜福利视频| 97碰自拍视频| 久久久国产成人精品二区| 热99在线观看视频| 久久久精品大字幕| 欧美成人性av电影在线观看| 亚洲国产高清在线一区二区三| 亚洲一区高清亚洲精品| 波多野结衣高清作品| 欧美精品啪啪一区二区三区| 乱系列少妇在线播放| 国产精品av视频在线免费观看| 亚洲中文日韩欧美视频| 成人无遮挡网站| 欧美日本亚洲视频在线播放| 男女那种视频在线观看| 嫩草影院新地址| 国产高潮美女av| 国产久久久一区二区三区| 国产精品爽爽va在线观看网站| 欧美区成人在线视频| 亚洲一区高清亚洲精品| 久久人人爽人人爽人人片va| 久久久久国产精品人妻aⅴ院| 久99久视频精品免费| 国产免费av片在线观看野外av| 国产亚洲精品av在线| 免费观看的影片在线观看| 美女免费视频网站| 日韩一区二区视频免费看| 精品久久久久久久末码| 最好的美女福利视频网| 黄色欧美视频在线观看| 一个人观看的视频www高清免费观看| 欧美+亚洲+日韩+国产| 一本精品99久久精品77| 少妇裸体淫交视频免费看高清| 我的老师免费观看完整版| 亚洲男人的天堂狠狠| 狂野欧美激情性xxxx在线观看| 国产精品久久久久久久电影| 91麻豆av在线| 丰满乱子伦码专区| 亚洲中文字幕日韩| 精品日产1卡2卡| 亚洲无线在线观看| 国产精品久久久久久久电影| 99久国产av精品| 有码 亚洲区| 国产老妇女一区| 成年女人毛片免费观看观看9| 国产精品三级大全| 国产精品亚洲一级av第二区| 久久草成人影院| 琪琪午夜伦伦电影理论片6080| 国产亚洲精品久久久久久毛片| 中出人妻视频一区二区| 欧美日韩中文字幕国产精品一区二区三区| 九九在线视频观看精品| 亚洲欧美精品综合久久99| 亚洲无线在线观看| 国产熟女欧美一区二区| 狂野欧美白嫩少妇大欣赏| 一级a爱片免费观看的视频| 天堂影院成人在线观看| 天堂√8在线中文| 中国美白少妇内射xxxbb| 毛片一级片免费看久久久久 | 久久人人爽人人爽人人片va| 波多野结衣巨乳人妻| 丝袜美腿在线中文| 亚洲第一区二区三区不卡| 麻豆av噜噜一区二区三区| 国产高清激情床上av| 亚洲在线观看片| 久久久久久国产a免费观看| 日韩国内少妇激情av|