• <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 ?

    A Novel Workload-Aware and Optimized Write Cycles in NVRAM

    2022-08-24 03:27:32ShriTharanyaaSharmilaandSaravanaKumar
    Computers Materials&Continua 2022年5期

    J.P.Shri Tharanyaa,D.Sharmila and R.Saravana Kumar

    1Department of ECE,Bannari Amman Institute of Technology,Tamil Nadu,India

    2Department of CSE,Jai Shriram Engineering College,Tamil Nadu,India

    3Department of ECE,Bannari Amman Institute of Technology,Tamil Nadu,India

    Abstract: With the emergence of the Internet of things (IoT), embedded systems have now changed its dimensionality and it is applied in various domains such as healthcare, home automation and mainly Industry 4.0.These Embedded IoT devices are mostly battery-driven.It has been analyzed that usage of Dynamic Random-Access Memory (DRAM) centered core memory is considered the most significant source of high energy utility in Embedded IoT devices.For achieving the low power consumption in these devices, Non-volatile memory (NVM) devices such as Parameter Random Access Memory (PRAM) and Spin-Transfer Torque Magnetic Random-Access Memory (STT-RAM) are becoming popular among main memory alternatives in embedded IoT devices because of their features such as high thickness, byte addressability, high scalability and low power intake.Additionally,Non-volatile Random-Access Memory(NVRAM)is widely adopted to save the data in the embedded IoT devices.NVM,flash memories have a limited lifetime,so it is mandatory to adopt intelligent optimization in managing the NVRAM-based embedded devices using an intelligent controller while considering the endurance issue.To address this challenge, the paper proposes a powerful,lightweight machine learning-based workload-adaptive write schemes of the NVRAM, which can increase the lifetime and reduce the energy consumption of the processors.The proposed system consists of three phases like Workload Characterization, Intelligent Compression and Memory Allocators.These phases are used for distributing the write-cycles to NVRAM,following the energy-time consumption and number of data bytes.The extensive experimentations are carried out using the IoMT (Internet of Medical things)benchmark in which the different endurance factors such as application delay,energy and write-time factors were evaluated and compared with the different existing algorithms.

    Keywords: Internet of things;DRAM;PRAM;STT-RAM;machine learning;internet of medical things;endurance

    1 Introduction

    Internets of Things (IoT) technologies are growing exponentially, finding their application in various domains such as healthcare, automation and security.These devices are equipped with batteries, sensors, microcontrollers and transceivers.Since these devices are battery-driven, making the IoT node sensitive to battery lifetime is a tricky mechanism.Recent reports suggest that using DRAM-based memory systems is the main contributor to embedded devices’overall energy utility to solve this significant issue,Non-Volatile Random-Access Memory(NVRAM)finds its place in terms of DRAM memory systems.Non-volatile memory devices,including Phase Change Memory(PCM)as well as Resilient Random-Access Memory (ReRAM), are playing a vital role in building future memory systems due to their salient features like high scalability and low power consumption[1–5].However,NVRAMs suffer from shortcomings like less endurance and high energy read/write cycles[6]when compared with the existing DRAM systems.Hence these constraints impose serious issues in using the NVRAM as the complete replacement for DRAM.

    Several studies were investigated in improving the endurance of NVRAM in terms of reducing and effectively managing the memory write cycles[7,8].Flip-N-Write(FNW)[9]becomes the method that combines old and new data to lessen the number of bits that help reducing redundancy between similar caches.Moreover,frequency compression techniques[10,11]was used to compress the data bits for an efficient write mechanism.Specifically, these methods retain a template table in the program memory and include several typical word patterns, including a highly responsible template.If all the word patterns in the table are matched, each word in the in-memory cache is represented by a compact format,which decreases the number of text bits.The term patterns in the template chart are predetermined and cannot be updated,defined as static patterns.

    To achieve the above-mentioned challenge,methodologies such as Dynamic Frequency Compression techniques[12],F(xiàn)lash Translational Layer(FTL)mapping techniques[13–18]were proposed to reduce the high energy write levels,thus increasing the endurance levels of NVRAM.However,these methods find it difficult to handle many applications when ported in Embedded IoT devices.

    To improve the endurance of the memory systems for large applications, intelligent analysis of larger applications is required, which can impart flexible and adaptive write cycles.Addressing the observations from the literary works,we put forward the new adaptive NVRAM write schemes called WHEAL(Workload Hybrid Energy Adaptive Learning)mechanism.The proposed technique works on dynamic compression techniques based on the workload characterizations with an effective NAND flash memory management.Moreover,WHEAL also expanded the complex pattern granularity from 32 bits to 64 bytes.In specific,our interventions are summed up as follows.

    a.Intelligent Workload Characterization:In contrast to the previous schemes,which require more statistical and extraction methods for obtaining the data characteristics, we proposed a new powerful machine-learning algorithm to classify the data types based on the workload characterization.The proposed technique increases the versatility of data compression mechanisms in a cost-effective way.

    b.Workload-Energy-Aware NVRAM Compression Schemes:The proposed write schemes introduce the new methodology Dynamic Workload Compression(DWC),which writes data to fit into dynamic compressive schemes.The DWC expands the data pattern for compacting an allzero cache while leveraging the value localization.In addition,this methodology preserves the energy with the reduced mechanism.

    c.Endurance Aware Allocation Methods:The proposed schemes introduce the compression and assign the adaptive write cycles for the intelligent allocation of the compressed workload bits in the caches following its characteristics to increase the memory’s endurance.

    The rest of the paper is organized as follows:Section 2 presents the related works of various literature.Section-2 also converses about the research gaps found in the existing systems.The preliminary overviews of NVRAM-Based Cache Memories and the motivation behind our research are detailed in Section-3.The complete working mechanism of the proposed framework is presented in Section-4.The implementation results with comparative analysis,performance metrics are exhibited in Section 5.Lastly the paper is concluded in Section-6.

    2 Related Works

    Power and performance optimization on embedded architectures is still a critical challenge.Many existing approaches are focused on designing an efficient software kernel for targeted architecture.On the other hand, hardware optimization in terms of design space area and memory optimization is still an open challenge problem to be focused on nowadays.In this literature survey, we initially discussed the workload characterization on embedded architectures followed by non-volatile memory optimization.

    Workload Characterization can be utilized to foresee future asset pre-requisites that enhance the scope organization,task planning and resource utilization effectively.The remaining burden portrayal is commonly performed by utilizing two unique methodologies i.e.,execution-based and model-based methods.In the model-based framework,the workloads are characterized using the performance tool[19]in terms of memory occupation,registers and other peripheral usages of execution of applications on the hardware platforms.Wang et al.[20] introduced multi level cell based phase change memory architecture.The significant value of this framework is relies upon energy variations in programming.To further enhance the performance,this framework incorporates data comparison write and results shows that this framework outperformed other existing methodologies in terms of energy consumption utilized for write cycles.

    Shishira SR et al.studied performance improvement achieved through workload characterization[21].The author initially defined the workload characterization with its various classes such as CPU, GPU, cloud workloads, physical environment, resource utilization and how the execution characteristics differ among each workload.The various tasks have distinct traits and the choice of the proper forum for a specific workflow processing is still an NP-hard task.The major limitation is a review paper, in which hypothetically detailed workload characterization and its classification are detailed.Maria Calzarossa et al.[22] developed and used a series of static tracks to establish task models based on the network usage and processing times.

    Alexander et al.developed a workload characterization model that focuses on multimedia application workloads.Multiprocessor system-on-chip is targeted for processing multimedia workloads in real-time.The workloads are categorized and differentiated based on the variability characterization curves and workload transformations [23].The limitation of this work, the proposed framework is more suitable for multimedia workloads on the MPSoC platform.Memory, I/O workloads are not considered in this work.Zhang et al.[24]utilize regulation methods to describe the vector task of roles with dependent implementation sequence.This classification framework is orthogonal to our design in the context that we are not just attempting to model timeline relationships inside tasks triggered by dependent implementation.Baruah et al.studied a real time task model in the perspective of feasibility examination.This framework incorporates 2 models called generalized multi framework model and sporadic task model.It is a powerful framework and easily handle both static and dynamic tasks in real time[25].

    Writam Banerjee [26] considered real-time challenges on non-volatile memory gadgets with different structures.The later advance faced the “memory wall”, i.e., the speed crevice within rationale and memory.Overcoming these issues, primary framework execution bottleneck and essential confinements related to contracting gadget measure and expanded processing complexity,developing NVM(eNVM)with energizing models have been introduced in this article.As per the recent survey,eNVM devices called FeFET, PCM, STTRAM and RRAM are the most assured memory devices that can be used for high-efficient performance with low cost and power.Likewise, Jishen Zhao et al.reviewed the low-cost architecture of non-volatile memory, byte-addressed non-volatile memory(NVM)architectures,including spin-transfer torque memory(STTMRAM), phase-change memory(PCM) and resistive memory (ReRAM), as just a substitute for conventional memory systems used throughout the memory models[27].

    Yuncheng Guo et al.developed an NVM writing scheme called Dynamic Frequent Pattern Compression(DFPC)to dramatically decrease the number of write units and increase their lifespan.The DFPC model comprises various stages such as sampling and dynamic patterns to improve the compression of data.Then, an enhanced DFCP algorithm is developed to optimize the latency and energy efficiency of NVM devices[12].

    3 Methodology

    3.1 NVRAM Background

    NVM,Characteristics and Drawbacks unlike conventional charge-based memory such as DRAM and SRAM, evolving NVMs store information using series resistance memories with increased concentration and scalability.Therefore,NVMs can be commonly used in the primary memory[27].

    Because all NVMs store data by modifying physical characteristics,the write operation acquires more time and resources than that of the read operation,contributing to the imbalance of reading and writing operations.Also,the write operation uses the NVM cell,particularly at high frequency,which results in reduced NVM durability.NVM-based structures also have to lessen the number of write-bits in the writing process.As one of the successful NVM innovations,PCM technology uses tolerance of chalcogenide glass for data storage.

    The element used for phase shift is germanium, antimony and tellurium alloys, including Ge2Sb2Te5.The substance has two possible states, the state of crystalline (SET) and the state of amorphous(RESET).The sensitivity of the substance to various states is fundamentally different.In particular, the tolerance rate of the amorphous form is much greater than those of the crystal-state material.PCM stores binary information in states by using material tolerance differences.

    In need of executing a RESET(SET)procedure to write“0”(“1”)to the PCM cell,the PCM cell is warmed just above the melting point that melts the chalcogenide substance,followed by immediate chilling to adjust the structure.The ReRAM system utilizes a permeable dielectric in the Metal-Insulator Metal structure.This can be adjusted among low-resistance state(SET)or high-resistance state(RESET)while using the necessary voltages.

    3.2 Motivation

    (i) Achieving high-endurance with low power is still an Np-hard issue.Workload characterization scheduling improves resource utilization and performance through static model-based optimization,which misleads specific architecture and specific applications.Existing approaches are not suitable for mixed-critical workloads and are not designed for common architectures.

    (ii) Memory(NVRAM)optimization is another emerging problem in recent research.However,few researchers have addressed the various non-volatile memory optimization structures which are not implemented on experimental boards.

    (iii) There are many inherent problems with eNVM technologies, like cell-level and system-level durability, variability, device performance, extremely smooth framework architecture, etc.Nevertheless,the eNVM paves the way to overwrite primary memories,play SCM’s function,explore new technologies,explore brain-inspired computational systems and model hardware safety systems.

    (iv) The low-power eNVM could also be beneficial for sensors,including Smart devices.However,there are also several obstacles in developing large eNVM systems, such as the production process,components and tailored operation for various products.

    4 Proposed Work

    4.1 Proposed WHEAL-NVRAM Architecture

    The proposed WHEAL-NVRAM-based architecture has been shown in Fig.1.The new workload awareness-compression and new wear-levelling approaches have been integrated to wear out the low energy cycles.In addition, the new lightweight learning algorithms are incorporated in memory to categorize the nature of the workloads.Based on the nature of workloads,novel compression and new adaptive-levelling and allocation technique were adopted to reduce the write cycle in the NVRAM for enhancing its endurance.

    4.2 Workload Characterization

    This section discusses the workload parameters and intelligent workload characterization based on the lightweight machine learning algorithms.

    4.3 Workload Parameters

    Tab.1 illustrates the workload parameters which are extracted from the different input application programming threads.These workloads are extracted by using the PSutils software,which runs on the kernel’s architecture.

    These workloads are considered the micro architecture-independent workload parameters used as input to the proposed learning model for efficient workload characterization.The major advantage of using the micro architecture-independent workloads is to capture the actual inherent program behaviours and prove its usefulness in characterizing suites for emerging workloads.

    4.4 Machine Learning Based Workload Extraction Mechanism

    Several algorithms such as Clustering,principal component analysis,histogram analysis,correlation methods were used for typical workload characterization.However, an accurate and intelligent workload characterization technique is still mandatorily needed for a better compression and levelling technique.This purpose can be served by using the machine learning algorithms and the proposed system incorporates a single-layer feedforward learning model for effective workload characterization.

    Figure 1:Proposed framework for the WHEAL-architectures

    Table 1:Workload parameters used for the characterization

    The single-layer feedforward networks are based on the Extreme Learning Machines (ELM).Fig.2 shows the architectural diagram for the proposed ELM.

    This infrastructure uses a single hidden layer, high speed, precision, speed planning and highly speculative and standard feature estimation skills.Hence,these algorithms are portable to embedded memory storage systems.In the same kind of model,the‘K’neurons in the hidden layer are expected to deal via an activation function that is very distinguishable while the output layer component is linear.In ELM,hidden layers are need not be tuned.Loads of the hidden layer are randomly allocated(counting loads of the bits).It is not the situation that hidden nodes are meaningless,but they do not need to be tuned and hidden neuron parameters can be haphazardly generated even in advance.Those are,until having to take care of a learning data collection.For a single layer ELM,the performance of the method is specified by Eq.(1)

    where a is the input

    βbe the output weight vector

    H(a)is the hidden layer output

    Figure 2:Structure of Extreme Learning Machines

    To formulate the output vector S that is considered as the target,the eqn can further be derived as follows

    The general implementation of the ELM is based on the minimum non-linear least square approaches expressed in Eq.(5)

    where H*is the inverse of H known as Moore-Penrose generalized inverse

    The output can be derived as

    4.5 Disadvantages of Existing ELM

    Although Extreme Learning Machines seem to be effective both in preparation and practice,the key downside was its non-optimal calibration of input weights.ELM often uses several hidden units to change the weight values relative to many other traditional learning strategies,which can influence the accuracy of the identification.

    To address the above downside, a novel Whale algorithm is used to optimize input connection weight variables to reach the optimal accuracy of the task classification system.There seems to be a significant trend in Whale optimization in recent decades.This evolutionary algorithm model is a computational approximation of the action and activity of humpback whales in their quest for food and supplies.Whale Optimization Algorithm(WOA)has been influenced by the Bubble-net assault technique, whereby whales begin catching fish by forming spiral-shaped bubbles surrounding their fish to 12 meters depth from the sea and then dive back and caught their intended prey.Centred on the shortest distance of the whales,the discovery phase in this method is a randomized quest for food that can be numerically converted by modifying old strategies rather than picking the right ones by randomly assigning other alternatives.In addition to this curious action,WOA separates it from other evolutionary algorithms because it only expands two values.These attributes determine a rapid integration among extraction and exploration processes.Fig.3 shows Encircling attack prey searching methodology for humpback whales.

    Figure 3:Encircling attack prey searching methodology for humpback whales

    We will explain the computational formula for prey conquering,prey seeking,spiral bubble-net hunting and gathering in the later subsection.Surrounding Prey:By raising the number of loops from the beginning to the optimum number,humpback whales encircle the prey and refine their location in the path of the best solution.This action can be expressed mathematically as:

    If(p<0.5 and mod(U)<1)

    Then the position of the candidate X(t+1)is updated and

    where p=0.1(constant)X(t+1)is the best position in the current situation.U and D are calculated by the following Eqs.(9)and(10)

    where a decrease linearly from 2 to 0 and r is the randomly selected vector

    Prey Searching:In prey searching mechanism,X is replaced with the random variables Xrandomand mathematical equation are given by

    The encircling of prey and spiral updation of prey have been done during the exploration phase of the whale optimization algorithm.The mathematical expression for updation of new position during the spiral process is given by Eq.(14)

    Here,D is the distance among the new position and updated position in the new generation,b is the constant,which varies from the 0 to 1.

    4.6 Optimized ELM for Workload Characterization

    The main limitation of the ELM is that the non-optimum collection of hidden units can trigger the formation of increasingly prevalent that impair the prediction accuracy.The proposed ELM network implements a whale algorithm to maximize input weight and bias variables to address this issue.The benefit of the whale algorithm in ELM is that it improves the global minimum search path,which can be more effective than the current optimization techniques.Throughout this case,accuracy is used as a feature of fitness.If the classification accuracy is equivalent to the accuracy of the standard,then the output variables will be deemed correct else,they will be ignored and their iterations will begin.This optimized learning model saves energy consumption and area overhead when implemented as the API in the Embedded Systems kernel.Hence the Eq.(7)can be modified as

    Eq.(15)depicts the final output from the proposed learning models,which is used to categorize the different types of the workloads such as Very Heavy(VH),Heavy(H),Medium(M)and Normal(N).Before categorizing the workloads by the proposed learning models, workload thresholds are calculated using the mathematical expression.Then,these values are used to label the workloads,which makes an effective categorization using the proposed learning models.

    4.7 Threshold Decision

    After calculating the workload parameters using the PSutils tools in the Embedded systems,energy and instruction per clock cycles are calculated using the decision rule for categorizing the workloads.The energy of each workload is calculated by using the mathematical expression, which is given as follows

    where E(W)is the Energy of the workloads,V is the voltage of the CPU,I is the current consumption and IPC is Clock cycles for writing in the memory.

    The Eq.(16)is used for labelling the workloads that better categorize workloads by the optimized machines.Based on the energy calculated,proposed learning models categorizes the different workloads as very heavy (VH), heavy (H), medium (M) and normal (N), which act as the inputs to the compressor and allocators.

    4.8 Intelligent Workload Compression and Allocation

    After the extractions of the workloads, these workloads are compressed by the dynamic compression technique and new threshold-based workload levelling is adopted to allocate the application threads in the cache memories.The architecture which is used for implementing the compression and allocator technique is shown in Fig.4.

    Figure 4:Memory controller architecture for compression and allocator

    The proposed memory controller architecture consists of two compression engines, a workload pattern table and an allocator.All these are controlled by the timing circuits, which control the workload patterns to write in the buffers.The working mechanism of the WHEAL memory controller is given as follows.

    4.8.1 Workload Storage Pattern

    These storage systems store the workload categorized by the proposed learning models and the categorization bits.These categorization bits are used to differentiate the workloads and following that,compression will take place concurrently for the different loads.

    4.8.2 Compression Engine

    The approach splits the entire 64-byte cache line as 16 32-bit words to perform data compression for different workloads.Further,the categorization bits are used to check the type of workloads,which avoids the repeated type of workloads.Also,these compression techniques work based on sampling time which is provided by the timing control circuits.

    5 Experimentation

    The proposed framework is simulated in a GEM-5 simulator emulated on the Raspberry Pi 3 Model B+ hardware.To analyze the workload characteristics, IoMT benchmarks were considered and implemented on hardware Raspberry Pi Model B+.Then the proposed NV-RAM architecture is simulated on GEM-5 software.The complete specification used for the experimentation is presented in Tab.2.

    Table 2:Specifications used for the experimentation

    For analysis,IoMT benchmarks are taken as input workload datasets in which 70%of workloads are used for training and 30% are used for testing.The different analysis of the proposed learning model in categorizing the workloads are given as follows.

    Figs.5–8 show the performance analysis of the different machine learning models in categorizing the different workloads.Fig.5 shows the performance metrics of different learning models in categorizing the very heavy workloads in which the proposed algorithm has exhibited 98.5%accuracy,98%sensitivity and 97.5%specificity,which has an edge of performances of 2%than SVM(Support Vector machines),3.5%than KNN(K-nearest neighbourhood),4%than DT(Decision Tree),5%than ANN-1(Artificial Neural Networks-Backpropagation layer)and ANN-2(Artificial Neural Networks-Feed forward layer)and finally 6%than Na?ve Bayes algorithms.A similar fashion of performance is found in Figs.6–8 in categorizing the different types of workloads.The integration of the optimized hidden layers in ELM has proved its stability in categorizing the workloads with high efficiency.Moreover,to prove the performance of the proposed framework,time of categorizing is calculated,which is shown in the figures,

    Figure 5:Performance analysis for different learning models for categorizing the very heavy workloads using IoMT benchmarks

    Figure 6:Performance analysis for different learning models for categorizing the heavy workloads using IoMT benchmarks

    Figure 7:Performance analysis for different learning models for categorizing the medium workloads using IoMT benchmarks

    From the above analysis, it is found that the proposed model consumes less time with high performance,which proves that these models can be integrated into the hardware,which can consume lesser energy and lesser overhead.

    Figure 8:Performance analysis for different learning models for categorizing the normal workloads using IoMT benchmarks

    6 Conclusion

    The paper proposes the first-ever hybrid machine-learning-based approaches for increasing the endurance in NVRAM.The proposed WHEAL methodology incorporates different stages such as workload categorization,compression technique and a memory allocator.The first phase is workload categorization in which the new energy-saving optimized ELM technique is used for the workload categorization where the energy is taken as the major threshold.Also, the paper proposes the dual compression technique for compressing the bits and the memory allocator stores the loads by using adaptive write cycles in different caches.Extensive experimentations have been conducted and performance metrics such as accuracy of categorization,write frequency ratio and write latencies were analyzed and compared with other existing architectures such as Baseline architecture and A-CACHE controllers.As a result,the proposed model has increased the lifetime of NVRAMS by 50%greater than the A-CACHE controller and even>90%greater than baseline architectures.Eventhough this technique can increase NVRAM’s endurance,computational overheads need its improvisation for less complexity implementation.

    Funding Statement:The authors received no specific funding for this study.

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

    搡老乐熟女国产| 午夜免费男女啪啪视频观看| 国产精品 国内视频| 亚洲国产精品专区欧美| 大片免费播放器 马上看| 久热久热在线精品观看| 两个人免费观看高清视频| 久久国产精品大桥未久av| 国产精品一国产av| 国产日韩一区二区三区精品不卡 | 99热这里只有精品一区| 国产精品.久久久| 婷婷色麻豆天堂久久| 高清av免费在线| 美女内射精品一级片tv| 国产爽快片一区二区三区| 在线精品无人区一区二区三| 男人操女人黄网站| 日日爽夜夜爽网站| 国产爽快片一区二区三区| 91精品三级在线观看| 久久99精品国语久久久| 精品少妇久久久久久888优播| 国产视频内射| 欧美日本中文国产一区发布| 日本黄大片高清| 欧美日韩在线观看h| 亚洲av男天堂| 黑人高潮一二区| 成人国语在线视频| 777米奇影视久久| 亚洲av福利一区| 中国国产av一级| 18禁在线播放成人免费| 又粗又硬又长又爽又黄的视频| 一级a做视频免费观看| 狂野欧美激情性xxxx在线观看| 国产精品一区二区在线不卡| 亚洲,一卡二卡三卡| 99久久综合免费| 简卡轻食公司| 建设人人有责人人尽责人人享有的| 人妻系列 视频| 在线免费观看不下载黄p国产| 亚洲精品国产av蜜桃| 十八禁网站网址无遮挡| 成人国产麻豆网| 日日爽夜夜爽网站| 涩涩av久久男人的天堂| 国产免费一级a男人的天堂| 成人国产av品久久久| 久久国产精品大桥未久av| 蜜桃在线观看..| 青春草视频在线免费观看| 一级毛片aaaaaa免费看小| 高清在线视频一区二区三区| 美女中出高潮动态图| 新久久久久国产一级毛片| 免费少妇av软件| 秋霞伦理黄片| 久久鲁丝午夜福利片| 日本-黄色视频高清免费观看| 满18在线观看网站| av福利片在线| 能在线免费看毛片的网站| 精品人妻一区二区三区麻豆| 国产成人精品无人区| 2018国产大陆天天弄谢| 国产乱来视频区| a级片在线免费高清观看视频| 亚洲精品乱码久久久久久按摩| 国产亚洲午夜精品一区二区久久| 久久精品国产亚洲网站| 亚洲中文av在线| 久久久久精品久久久久真实原创| 亚洲国产最新在线播放| 丁香六月天网| 日韩av不卡免费在线播放| 一级毛片电影观看| 91精品国产九色| 免费大片黄手机在线观看| 少妇丰满av| 91精品伊人久久大香线蕉| 九九爱精品视频在线观看| 不卡视频在线观看欧美| 视频在线观看一区二区三区| 国产 一区精品| 免费大片黄手机在线观看| 只有这里有精品99| 国产亚洲欧美精品永久| 一区二区三区乱码不卡18| 新久久久久国产一级毛片| 美女中出高潮动态图| 亚洲第一区二区三区不卡| 亚洲国产av新网站| 国产av国产精品国产| 亚洲五月色婷婷综合| 亚洲欧美日韩另类电影网站| 免费少妇av软件| 欧美日韩一区二区视频在线观看视频在线| 69精品国产乱码久久久| av一本久久久久| 国产精品一区www在线观看| 精品亚洲成a人片在线观看| 九草在线视频观看| av.在线天堂| 欧美日韩亚洲高清精品| 少妇人妻 视频| 亚洲怡红院男人天堂| 97超视频在线观看视频| 亚洲综合色惰| 国产免费一级a男人的天堂| 在现免费观看毛片| 久久精品夜色国产| 亚洲高清免费不卡视频| 中文字幕制服av| 一本一本久久a久久精品综合妖精 国产伦在线观看视频一区 | 欧美精品亚洲一区二区| 亚洲国产毛片av蜜桃av| 91午夜精品亚洲一区二区三区| 少妇人妻 视频| 亚洲成人手机| 亚洲三级黄色毛片| 日韩一本色道免费dvd| 亚洲精品国产色婷婷电影| 高清欧美精品videossex| 亚洲精品乱码久久久v下载方式| 看免费成人av毛片| 日韩三级伦理在线观看| 狂野欧美白嫩少妇大欣赏| 日韩在线高清观看一区二区三区| 欧美日韩视频精品一区| 日韩欧美精品免费久久| 精品人妻在线不人妻| 永久网站在线| 男女国产视频网站| 香蕉精品网在线| 国产一区二区三区综合在线观看 | 啦啦啦中文免费视频观看日本| 亚洲av不卡在线观看| 欧美精品高潮呻吟av久久| 黄色视频在线播放观看不卡| 国产在线一区二区三区精| 精品国产露脸久久av麻豆| 中国三级夫妇交换| 三级国产精品片| 久久鲁丝午夜福利片| 亚洲美女视频黄频| 亚洲人与动物交配视频| 久久久久久久久久久免费av| 男女边吃奶边做爰视频| 91精品国产九色| 色网站视频免费| 亚洲精品色激情综合| 欧美精品亚洲一区二区| 亚洲av二区三区四区| 各种免费的搞黄视频| 亚洲国产精品成人久久小说| 国产成人freesex在线| 亚洲美女视频黄频| 午夜激情福利司机影院| 免费高清在线观看日韩| 国产高清国产精品国产三级| 久久精品国产a三级三级三级| 亚洲第一区二区三区不卡| 午夜久久久在线观看| 狂野欧美白嫩少妇大欣赏| 国产成人freesex在线| 久久久久久伊人网av| 国产片特级美女逼逼视频| 久久久欧美国产精品| 少妇被粗大猛烈的视频| 人妻少妇偷人精品九色| 成人免费观看视频高清| 一级毛片电影观看| 亚洲精品一区蜜桃| 超色免费av| 久久国产亚洲av麻豆专区| 乱码一卡2卡4卡精品| 母亲3免费完整高清在线观看 | 欧美精品一区二区免费开放| 夜夜骑夜夜射夜夜干| 秋霞伦理黄片| 精品一区二区免费观看| 少妇熟女欧美另类| 99国产综合亚洲精品| 精品久久久久久久久av| 亚洲精品乱码久久久v下载方式| 满18在线观看网站| 亚洲,一卡二卡三卡| 一个人免费看片子| 亚洲熟女精品中文字幕| 午夜激情av网站| 欧美激情国产日韩精品一区| 2018国产大陆天天弄谢| 亚洲精品国产av蜜桃| av.在线天堂| 狂野欧美激情性xxxx在线观看| 欧美成人午夜免费资源| 亚洲精品久久久久久婷婷小说| 黄色毛片三级朝国网站| 日韩精品免费视频一区二区三区 | 亚洲综合色惰| 久久精品久久久久久久性| 免费av不卡在线播放| 欧美日韩精品成人综合77777| 亚洲av电影在线观看一区二区三区| 久久久久久久大尺度免费视频| 国产av精品麻豆| 有码 亚洲区| 蜜桃国产av成人99| 人成视频在线观看免费观看| 国产成人免费无遮挡视频| 丝袜喷水一区| 哪个播放器可以免费观看大片| 人妻人人澡人人爽人人| 久久精品国产亚洲av涩爱| 国产成人免费观看mmmm| 亚洲成人一二三区av| 人体艺术视频欧美日本| 人人妻人人添人人爽欧美一区卜| 日韩精品免费视频一区二区三区 | 久久久久精品久久久久真实原创| 狂野欧美白嫩少妇大欣赏| av有码第一页| 99九九线精品视频在线观看视频| 精品一区二区三卡| 国产精品秋霞免费鲁丝片| 亚洲国产色片| 色网站视频免费| 国产亚洲最大av| 熟妇人妻不卡中文字幕| 亚洲精品,欧美精品| 能在线免费看毛片的网站| 久久这里有精品视频免费| 免费播放大片免费观看视频在线观看| 亚洲国产色片| 高清午夜精品一区二区三区| 高清不卡的av网站| 国产精品不卡视频一区二区| 国产成人精品婷婷| 精品亚洲乱码少妇综合久久| 成人黄色视频免费在线看| 一本一本综合久久| 久久久久久久久久人人人人人人| 欧美日韩视频高清一区二区三区二| 丁香六月天网| 日韩伦理黄色片| 九九在线视频观看精品| 人人妻人人添人人爽欧美一区卜| 日韩欧美一区视频在线观看| 美女国产视频在线观看| 国产精品蜜桃在线观看| 亚洲一区二区三区欧美精品| 热99国产精品久久久久久7| h视频一区二区三区| 久久精品久久久久久噜噜老黄| 人妻少妇偷人精品九色| 全区人妻精品视频| 久久韩国三级中文字幕| 久久久欧美国产精品| 99热网站在线观看| 国产欧美日韩一区二区三区在线 | 亚洲精品456在线播放app| 国产免费一级a男人的天堂| 久久国产亚洲av麻豆专区| 又黄又爽又刺激的免费视频.| 成人手机av| 大香蕉97超碰在线| 成人免费观看视频高清| 日本av手机在线免费观看| 国产免费又黄又爽又色| 日日撸夜夜添| 午夜精品国产一区二区电影| 制服诱惑二区| 亚洲精品第二区| 狠狠婷婷综合久久久久久88av| 综合色丁香网| 国产又色又爽无遮挡免| 男女边吃奶边做爰视频| 大又大粗又爽又黄少妇毛片口| 国产欧美日韩综合在线一区二区| 超碰97精品在线观看| 高清黄色对白视频在线免费看| 91久久精品国产一区二区成人| 97超碰精品成人国产| 97超视频在线观看视频| av黄色大香蕉| 少妇人妻 视频| 性色av一级| 永久免费av网站大全| 久久午夜福利片| 亚洲国产精品999| 欧美精品国产亚洲| 国产深夜福利视频在线观看| 五月伊人婷婷丁香| 婷婷色av中文字幕| av在线观看视频网站免费| 色婷婷久久久亚洲欧美| 中文字幕人妻丝袜制服| 18禁在线无遮挡免费观看视频| 91精品一卡2卡3卡4卡| 免费不卡的大黄色大毛片视频在线观看| 999精品在线视频| 国产永久视频网站| 亚洲四区av| 夫妻性生交免费视频一级片| 91精品国产国语对白视频| 久久久欧美国产精品| 中文精品一卡2卡3卡4更新| 99热这里只有精品一区| 日韩伦理黄色片| 欧美丝袜亚洲另类| 免费高清在线观看视频在线观看| 欧美精品亚洲一区二区| 交换朋友夫妻互换小说| 曰老女人黄片| 国产成人精品无人区| 精品卡一卡二卡四卡免费| 一本一本久久a久久精品综合妖精 国产伦在线观看视频一区 | 日本黄大片高清| 亚洲高清免费不卡视频| 人人妻人人澡人人爽人人夜夜| 国产精品女同一区二区软件| kizo精华| 一级二级三级毛片免费看| 美女中出高潮动态图| 亚洲av男天堂| 亚洲国产精品国产精品| 国产伦精品一区二区三区视频9| 日本黄色片子视频| 大香蕉久久成人网| 三级国产精品片| 日韩一本色道免费dvd| 女性被躁到高潮视频| 最近中文字幕2019免费版| 啦啦啦中文免费视频观看日本| 尾随美女入室| 99九九线精品视频在线观看视频| 午夜福利,免费看| 久久ye,这里只有精品| 男女无遮挡免费网站观看| av免费在线看不卡| videos熟女内射| 又大又黄又爽视频免费| 国产爽快片一区二区三区| 亚洲熟女精品中文字幕| 飞空精品影院首页| 久久精品国产自在天天线| 另类精品久久| 哪个播放器可以免费观看大片| 国产亚洲一区二区精品| 只有这里有精品99| 韩国高清视频一区二区三区| 免费人妻精品一区二区三区视频| 国产乱来视频区| 满18在线观看网站| 丝袜脚勾引网站| 国产免费福利视频在线观看| 久久久久网色| 纵有疾风起免费观看全集完整版| 曰老女人黄片| 男女啪啪激烈高潮av片| 久久精品国产亚洲网站| 国产午夜精品久久久久久一区二区三区| 久久精品夜色国产| 蜜臀久久99精品久久宅男| 久久久久国产精品人妻一区二区| 亚洲精品视频女| 久久久久国产精品人妻一区二区| 日产精品乱码卡一卡2卡三| 国产黄色视频一区二区在线观看| 亚洲精品自拍成人| 国产日韩欧美视频二区| 欧美日韩成人在线一区二区| 在线观看www视频免费| 超色免费av| 国产视频内射| 欧美日韩在线观看h| 麻豆精品久久久久久蜜桃| 国产精品偷伦视频观看了| 国产视频内射| 精品久久久久久久久亚洲| 丝袜在线中文字幕| 久久久久精品久久久久真实原创| av在线app专区| 成人国产av品久久久| 搡女人真爽免费视频火全软件| 18禁动态无遮挡网站| 婷婷色综合大香蕉| 亚洲国产精品国产精品| 久久久久视频综合| 日韩电影二区| 久久ye,这里只有精品| 日日摸夜夜添夜夜添av毛片| 校园人妻丝袜中文字幕| 人人妻人人澡人人爽人人夜夜| 精品国产国语对白av| 人体艺术视频欧美日本| a级毛片黄视频| 交换朋友夫妻互换小说| 少妇 在线观看| 大香蕉久久成人网| 亚洲精品国产av蜜桃| 搡女人真爽免费视频火全软件| 十八禁高潮呻吟视频| av免费在线看不卡| 精品视频人人做人人爽| 青春草亚洲视频在线观看| 亚洲人成网站在线观看播放| 18禁在线播放成人免费| 国产在线视频一区二区| 亚洲成人手机| 日本黄大片高清| 黄色视频在线播放观看不卡| 日本与韩国留学比较| 亚洲精品中文字幕在线视频| 久久久久精品久久久久真实原创| 丝袜喷水一区| 黑人猛操日本美女一级片| kizo精华| 黄色毛片三级朝国网站| 你懂的网址亚洲精品在线观看| 久久国内精品自在自线图片| 欧美精品一区二区大全| 国产日韩欧美视频二区| 26uuu在线亚洲综合色| 乱人伦中国视频| 最近最新中文字幕免费大全7| 在线观看一区二区三区激情| 国国产精品蜜臀av免费| 欧美+日韩+精品| 91精品国产国语对白视频| 人体艺术视频欧美日本| 黄色视频在线播放观看不卡| 一二三四中文在线观看免费高清| 国产有黄有色有爽视频| 简卡轻食公司| 亚洲精品日韩在线中文字幕| 欧美日韩国产mv在线观看视频| 久久久久国产网址| 国产极品粉嫩免费观看在线 | 一本色道久久久久久精品综合| 中文字幕亚洲精品专区| 国产极品天堂在线| 高清欧美精品videossex| 国产欧美亚洲国产| 夫妻午夜视频| 亚洲精品久久成人aⅴ小说 | 国产精品女同一区二区软件| 日本欧美视频一区| 日本色播在线视频| 国产老妇伦熟女老妇高清| 人体艺术视频欧美日本| 人人澡人人妻人| 中文字幕久久专区| 精品一区二区免费观看| 日韩亚洲欧美综合| 亚洲人成网站在线观看播放| 极品人妻少妇av视频| 国产精品一二三区在线看| 久久久久久久久久久免费av| 国产爽快片一区二区三区| 亚洲精品,欧美精品| 三上悠亚av全集在线观看| 国产成人91sexporn| 黄色欧美视频在线观看| 成人18禁高潮啪啪吃奶动态图 | 亚洲欧洲精品一区二区精品久久久 | 国产又色又爽无遮挡免| 啦啦啦视频在线资源免费观看| 美女脱内裤让男人舔精品视频| 国产精品一区www在线观看| 国产午夜精品久久久久久一区二区三区| 欧美成人午夜免费资源| 国产男女内射视频| 伊人久久精品亚洲午夜| 最后的刺客免费高清国语| 亚洲av欧美aⅴ国产| 麻豆精品久久久久久蜜桃| 人人妻人人澡人人看| 黄色欧美视频在线观看| 在线观看美女被高潮喷水网站| 老司机影院毛片| 黄色视频在线播放观看不卡| 观看av在线不卡| 又黄又爽又刺激的免费视频.| 亚洲精品国产色婷婷电影| 黄色配什么色好看| av不卡在线播放| 亚洲精品日韩在线中文字幕| 蜜桃久久精品国产亚洲av| 又粗又硬又长又爽又黄的视频| 日韩在线高清观看一区二区三区| 亚洲婷婷狠狠爱综合网| 日韩一本色道免费dvd| 久久久久久久大尺度免费视频| 18禁裸乳无遮挡动漫免费视频| 蜜桃在线观看..| 中文字幕av电影在线播放| 你懂的网址亚洲精品在线观看| 少妇 在线观看| 国产国语露脸激情在线看| 亚洲第一av免费看| 成人漫画全彩无遮挡| 久久久国产精品麻豆| 色婷婷久久久亚洲欧美| 亚洲熟女精品中文字幕| 国产成人精品一,二区| 黄片播放在线免费| 夜夜看夜夜爽夜夜摸| 三级国产精品片| 亚洲国产精品专区欧美| 国产精品人妻久久久久久| 嫩草影院入口| 国产永久视频网站| 九九爱精品视频在线观看| 自拍欧美九色日韩亚洲蝌蚪91| 亚洲精品一二三| 男男h啪啪无遮挡| 母亲3免费完整高清在线观看 | 毛片一级片免费看久久久久| 22中文网久久字幕| 亚洲美女搞黄在线观看| 欧美另类一区| .国产精品久久| 少妇人妻精品综合一区二区| 亚洲av成人精品一二三区| 日韩欧美一区视频在线观看| 午夜福利视频精品| 欧美精品一区二区大全| 熟妇人妻不卡中文字幕| 丰满乱子伦码专区| 色94色欧美一区二区| 51国产日韩欧美| 免费播放大片免费观看视频在线观看| 国产色爽女视频免费观看| 国产亚洲欧美精品永久| 两个人免费观看高清视频| 制服人妻中文乱码| 亚洲精品一二三| 综合色丁香网| 日韩,欧美,国产一区二区三区| 日本av手机在线免费观看| 欧美日本中文国产一区发布| 亚洲精品久久午夜乱码| 不卡视频在线观看欧美| 国产成人免费无遮挡视频| 亚洲人成77777在线视频| 国模一区二区三区四区视频| 免费观看av网站的网址| 国产黄色视频一区二区在线观看| 免费高清在线观看视频在线观看| 啦啦啦啦在线视频资源| av播播在线观看一区| 狠狠精品人妻久久久久久综合| 日本wwww免费看| 新久久久久国产一级毛片| 777米奇影视久久| 美女中出高潮动态图| 伦理电影免费视频| 亚洲欧美成人精品一区二区| 亚洲精品久久午夜乱码| videosex国产| 中文字幕精品免费在线观看视频 | 国产黄频视频在线观看| 晚上一个人看的免费电影| 色婷婷av一区二区三区视频| 久久国产精品男人的天堂亚洲 | 少妇 在线观看| 久久久久久久久大av| 欧美+日韩+精品| a级毛片黄视频| 少妇人妻 视频| a级片在线免费高清观看视频| 一本久久精品| 免费观看a级毛片全部| 青春草视频在线免费观看| 免费观看无遮挡的男女| 天天躁夜夜躁狠狠久久av| 亚洲av在线观看美女高潮| 亚洲欧洲精品一区二区精品久久久 | 欧美精品一区二区大全| 丰满饥渴人妻一区二区三| 日韩不卡一区二区三区视频在线| 一级爰片在线观看| 欧美人与善性xxx| 一区二区三区四区激情视频| 成人免费观看视频高清| 国语对白做爰xxxⅹ性视频网站| 国产女主播在线喷水免费视频网站| 99热6这里只有精品| av福利片在线| 伊人久久国产一区二区| 中国三级夫妇交换| 考比视频在线观看| 亚洲精品久久午夜乱码| 黑丝袜美女国产一区| 丝袜脚勾引网站| 欧美日韩成人在线一区二区| 欧美日韩亚洲高清精品| 亚洲伊人久久精品综合| 成年女人在线观看亚洲视频| 青青草视频在线视频观看| 亚洲中文av在线| 亚洲综合精品二区| 永久免费av网站大全| 久久久精品区二区三区| 搡女人真爽免费视频火全软件| 精品久久久噜噜| videos熟女内射| 最近中文字幕高清免费大全6| 男女啪啪激烈高潮av片| 久久久久久久亚洲中文字幕| 男男h啪啪无遮挡|