• <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 Distribution Management System for Relational Databases in Cloud Environments

    2013-07-14 01:21:10SzeYaoLiChunMingChangYuanYuTsaiSethChenJonathanTsaiandWenLungTsai

    Sze-Yao Li, Chun-Ming Chang, Yuan-Yu Tsai, Seth Chen, Jonathan Tsai, and Wen-Lung Tsai

    1. Introduction

    People rely more and more on information systems in their daily life with the development and integration of information and communication technologies. The incurred transactions and generated data have thus been increased dramatically. Cloud computing is one of the technologies developed sharing resources among tenants to improve system performance and to save operational cost. In addition to newly developed applications, legacy systems are also considered to transplant to cloud environments.

    On a cloud platform, programs can be replicated to multiple nodes easily to distribute the processing load. This is achieved via the mechanism of virtual machine. To avoid becoming the next hot spot of processing, the database should also be distributed properly.

    In order to handle huge amount of request for data processing, several key-value store systems, such as Google’s Bigtable[1], Yahoo’s PNUT[2], Amazon’s Dynamo[3], and other similar systems[4], have been developed. Data in these systems are replicated or partitioned across different servers so that the data store can be scaled in and out easily as needed. However, this kind of system tends to achieve system availability at the sacrifice of data consistency. This is known as the CAP (consistency,availability, and partition tolerance) theory[5]or PACELC[6].

    Although systems with key-value store perform well in large-scale queries and simple updates, they can hardly provide ACID (atomic, consistent, isolated, and durable)guarantee for important database transactions[7]. For transaction processing systems with relational databases to run in the cloud environment, an effective tool is needed to help managing the partition and/or replication of data and providing ACID guarantee at the same time. With the aid of this tool, it may take less effort in migrating legacy systems to the cloud platforms. Moreover, professionals in relational database can be shifted and utilized with less training.

    When distributing data to different nodes, the decisions of how to spread data on those nodes affect the performance and complexity of succeeding query and update tasks significantly. Partitioning data to more nodes can distribute the workload at the cost of increased complexity in data allocation, result integration, and transaction coordination. On the other hand, although replicating data to different nodes makes the system affordable to large amount of similar queries,synchronization among data nodes for data modification becomes a problem to ensure consistency. Since the transaction behavior of various applications differs from each other, it is important to partition or replicate data according to the characteristics of an application system such that better performance can be achieved.

    This paper proposes a transaction-based database partition system for relational databases, which is named as database distribution manager (DBDM). The proposed DBDM has two basic modules which provide database partition and data migration functionalities, respectively.The database partition module advises strategies pointing out whether a table in a database should be partitioned or replicated, and transferred to which nodes. The algorithm in this module provides solutions according to behaviors of the application system, including content and frequency of each transaction, tables used in each structured query language (SQL) statement, and relations among tables in the database. In principle, decisions made by this module let tables use in a single SQL statement together so that join operations can be accomplished within one node. This will avoid massive data movement among nodes.

    The data migration module, in turn, actually moves data according to the difference between current data distribution and the proposed one. The migration can be scheduled in a progressive manner by moving a limited amount of data at a time. In this way, the number of interferences to the operation of the system can be reduced.The amounts of data to be moved in one step and the time interval between two consecutive steps are set as parameters to the migration module.

    The proposed system was implemented to verify its feasibility. Preliminary experiments show that the database is effectively distributed and tables are grouped according to transactions of the application system. Further experiments will be conducted to verify the effectiveness on load distribution and performance enhancement.

    In the following, related works are discussed in the next section. The proposed system is introduced in Section 3.Section 4 illustrates the implementations and some design issues are discussed. Section 5 concludes this paper.

    2. Related Works

    In recent researches in database partition, Rao et al.devised a tool named DB2 design advisor to automatically select optimal solution for data partitioning[8],[9]. This tool used workload as input and suggested a method of distribution in four facets, namely, indexing, materialized view, database partition, and multi-table clustering. To take into account of the dependency among them, the authors designed a hybrid method to reduce execution time without degrading system quality. These methods took advantage of parallelism and were suitable for applications involving analysis of massive data such as data warehousing and mining. However, it is hard to do parallel transactions due to the locking and synchronization problems incurred by the ACID requirements.

    Agrawal et al. proposed a technique for automatic partitioning database horizontally and vertically[10]. The technique was also implemented and tested on Microsoft SQL Server. Although system performance and manageability factors were considered, their method was only focused on a single node. Extending their algorithm for partition with multiple nodes needs more efforts.

    NewSQL was proposed by several researchers for on-line transaction processing (OLTP)[11]. The purpose of NewSQL is to distribute the workload while still guaranteeing the ACID requirements. The efforts are principally focused on searching for an optimal execution plan for data inquiry and manipulation. As mentioned in the previous section, the efficiency of statement execution depends on how the data are distributed. Thus, a good partition strategy is the basis of efficient query execution.

    Some researches discussed data lookup problems in query processing[12],[13]. Improvements were made on the synchronization protocol such that distributed transactions can be conducted more efficiently. However, these improvements also rely on proper distribution of data to obtain maximal performance.

    There were also researches working on server allocation after partitioning the database. Apers used a model to evaluate cost functions under different server configurations. Thus, how to distribute data among the servers can be decided with this model[14]. Genetic algorithm has also been adopted for the server allocation problem[15]. The researchers found that their algorithms outperformed greedy heuristics under several different partition sizes. Menon formulated the server allocation as an integer programming problem[16]. Storage and processing capabilities were considered as constraints to the objectives. In addition to the factors considered in these researches, current distribution of the database should also be taken into account since the difference between current status and the proposed distribution determines the amount of data movement, which will affect the system performance during the redistribution.

    In summary, many factors should be considered in order to properly partition a database. These factors include the cost of ensuring ACID requirement, data migration for partitioning, storage and capacity of servers, server allocation, and data lookup and transmission to complete a query. In the database partition system proposed by this paper, transaction behavior and database relations are used to help judge the distribution strategy of data, such that the majority of factors can be taken into consideration. The architecture of the proposed system is introduced in the next section.

    3. System Architecture

    The architecture of the proposed DBDM is shown in Fig. 1 along with its operation environment. In this environment, DBDM is operated outside the distributed database management system (DBMS). It retrieves database schema and statistics of the application system,evaluates partition strategies, and migrates data among data nodes according to the partition decision. Distributed database (DB) administrator can monitor the current partition status from the interface. The administrator can also direct the behavior of partition and migration process by adjusting threshold values and/or parameters. Some of the system information, such as available servers and their capacities, are also set by the administrator.

    Fig. 1. System architecture of the database distribution manager.

    Since this research focuses on database partition and migration, PostgreSQL DBMS and pgpool, both developed by PostgreSQL Global Development Group, are used as the distribution DBMS to simplify the setup of experiment environment. The proposed DBDM can be used with other distributed DBMS as well by modifying the interfaces between DBDM and the target distributed DBMS.

    The DBDM comprises five modules. The “Partition Algorithm” module performs the evaluation of current database status and the proposition of partition strategy.The “Migration Manager” module then actually moves records in the database in response to the proposed partition strategy. Two types of information are needed for these two modules to complete their missions. The static information of the database, such as table schema and relationships among tables, are provided by the “Schema Extractor”module. The “Statistics Collector” module in turn gathers the runtime information, for example, the number of records and distribution of key values. Finally, the“Management Interface” module is in charge of the communication with the administrator.

    The two main functionalities of DBDM, which are the“Partition Algorithm” and the “Migration Manager”, are introduced in more detail next.

    3.1 Partition Algorithm

    The objective of the “Partition Algorithm” module is maximizing system performance under the constraint of guaranteeing the ACID requirement. System performance is in turn affected by transaction processing time and the workload. Thus, the procedure and characteristics of handling a transaction are analyzed first so that the algorithm can be tailored for the requirements of transaction processing.

    The steps to process a distributed transaction are 1)parsing, 2) execution planning and optimizing, 3)distributed command dispatching, 4) result collecting and summarizing, and 5) distributed transaction committing.The operation analysis is focused on those steps which are related to the distributed process, namely, data allocation,data transmission, and distributed transaction commitment.

    Consider the data allocation process first. Two methods are generally used in partitioning and locating the records in a table. One method is for storing and retrieving data according to a hash function of the key. The advantage of this method is its constant time for locating data. However,it is hard to find a suitable hash function for partitioning data in irregular sizes. The other method is for dividing the table into consecutive pieces according to the key.Although it suffers from increased data locating time as the number of partitions increases, this method can partition data in arbitrary lengths. Since the distribution of key values in a table depends on the characteristics of the application and is not predictable, the tables are divided into consecutive segments in this research.

    Next, the data transmission among data nodes is discussed. The most time-consuming part in a query is the join operation, which includes cross product operations.Although the final result may be few, the amount of intermediate data is large due to the cross products.Massive data transmission will be inevitable if tables involved in a join operation reside in different nodes.Conversely, the transmission of the query results only is necessary if those tables are stored in the same node.Therefore, the “Partition Algorithm” module should strive to put together tables within the same query statement.

    Finally, distributed transaction commitment needs synchronization among data nodes which participate in the transaction to guarantee the ACID requirements. Due to the data transmission problem, keeping tables of the same transaction together can effectively reduce the communications among nodes.

    Two strategies of database partitioning can be concluded from the above discussions. First, the tables involved in the same query statement should be kept together so that massive data transmission among nodes can be avoided. Next, the table is divided consecutively if partitioning is inevitable.

    However, grouping tables together makes data centralized instead of distributed. The group itself should be divided if it becomes the transaction hot spot. To allow the partition of the group while keeping related tables together, replication and horizontal partitioning are used. In a group to be partitioned, the tables containing elementary information are replicated since they act as lookup tables and normally contain less data. Moreover, the tables with master-detail relationships are partitioned in pairs. For example, the table storing items of sales orders is partitioned according to how the table storing headers of sales orders is partitioned. In this way, data in a group is distributed over multiple nodes while those related records needed in join operations still keep intact.

    After analyzing the behavior of transaction processing and its performance concerns, the “Partition Algorithm” is designed and its modular structure is depicted in Fig. 2. The algorithm is divided into three main modules with three supporting modules, respectively. The main modules,namely “Grouping Tables”, “Assign Servers to Groups”,and “Partition Horizontally”, perform the steps of database partition, whereas the supporting modules provide the needed information derived from outside and/or previous results.

    Among the three main modules, the “Grouping Tables”module first inspects every SQL statement and puts tables together if they are involved in a join operation. Next, the“Assign Servers to Groups” module allocates servers for each group according to the transaction frequencies of the groups and the capabilities of the servers such that each server will get a portion of transaction load proportional to its capability. Finally, a group is partitioned horizontally into consecutive pieces by the last module if it is dispatched to more than one node.

    3.2 Migration Manager

    After the way of database partition is determined, the“Migration Manager” module will actually move those data to the proposed locations. This module includes three phases as shown in Fig. 3. The “Analyzing Differences”module takes current distribution of data and partition decisions generated by the “Partition Algorithm” module as its inputs. It compares each table in these inputs to determine the ranges of records which should be moved from one node to the other. Then, the “Calculate Amounts to Move” module evaluates the total amounts of data to be moved from one node to the other for every pair of nodes according to the analyzed differences. The migration action will be divided into several smaller movements if the system performance is concerned. As the last step, the“Move Data & Set Parameters” module composes physical instructions in response to the movement information passed from the previous step. The migration instructions are sent to each node involved in the migration process.The distributed DBMS is also informed regarding the changes on data distribution.

    Fig. 2. Modular structure of the “Partition Algorithm”.

    It is noteworthy that the system performance is affected by the selection, deletion, transmission, and insertion process during the migration. The effect on the performance varies with different operation environments.Thus, the same migration request which can be done in one shot may need to be executed progressively in a different environment. Moreover, the requirements in response time or system throughput which is tight to some applications may have wider tolerance in other systems.

    To deal with this variety, the system leaves the decisions to the DB administrator by providing two parameters, namely, AmountInOneMove and IntervalBetweenTwoMoves. The former specifies the upper limit of data in one move. The later tells how much time the system should wait before the next movement can proceed.

    The administrator can set AmountInOneMove to a small value and set IntervalBetweenTwoMoves to a large value so that the migration is hidden from the users. The drawback of this setting is that it takes a longer time to achieve the desired partition. This method is suitable for frequent redistribution since relatively few data are changed,which implies few data to migrate. In the other situation where the system might be running intermittently, the administrator can set AmountInOneMove to a huge number and run the partition and migration process in the periodical maintenance so that the migration can be done in the whole piece.

    4. Implementations

    This section introduces the results of physically implementing the proposed system. Some design issues are also discussed as a guideline for further research.

    4.1 Implementation Results

    The proposed system was implemented on Linux using PostgreSQL and pgpool as the distributed DBMS. Part of the transactions and data of a hospital’s inventory system were included as test data. Preliminary tests showed that the system partitioned the database and did the migration as expected. More data and queries are currently being prepared to further test the robustness and performance.

    The original configuration of the system contains three nodes with the same computation power, and a forth node with 80% power of the previous nodes is added in the new configuration. A snapshot of the DBDM exhibiting the state of the database is shown in Fig. 4. The “current status”column describes the current distribution of the database,which may be either the previously calculated result or the status after several migration steps, depending on the migration strategy. The “objective distribution” column shows the newly suggested distribution obtained from the“Partition Algorithm”. The percentages in both columns represent the portion of the transaction load. The effect of adding a new node can be seen by comparing the “current status” and the “objective distribution” columns.

    It can be observed that the distribution of table ITEM_EXCH has been changed due to different node configurations. The percentage of ITEM_EXCH allocated to sever are01 is different from the other since there are other groups of tables co-located on that node. The percentage of ITEM_EXCH on server are02 to server are04 in the “objective distribution” shows that they are indeed allocated by their respective capabilities.

    4.2 Discussions

    One of the strategies for the proposed “Partition Algorithm” is grouping together the tables related by join operations. However, almost all tables are related in a relational data model, which means that the grouping process may result in one group with all tables in most cases. Taking a closer examination on the relational data model, it is found that the tables which are most frequently joined are the ones storing basic information such as employee, department, product, and customer. These tables are usually served as referencing lists and contain relatively few records. Removing this kind of tables from the relational data model makes the remaining tables separate into smaller groups, which are really related in the application.

    Fig. 4. Partial results of the implemented DRDM system.

    According to the above observation, the system replicates the most frequently joined tables to the related groups while grouping the rest tables. In the current system,the frequency of joins involved is used to judge whether the table should be replicated or not. A parameter is introduced to indicate and adjust this frequency. Identifying those tables automatically is a possible direction in the future.

    A second designing consideration is how to map transaction distributions into data distribution. To simplify the implementation, the proposed system assumes that each transaction accesses those records in a table evenly. This is obviously not the case in physical situations. For example,trains in peak hours are queried more often than others and undelivered sales orders are accessed more frequently than historical ones. To cope with this, various access patterns should be observed and recorded so that the “Partition Algorithm” can distribute the data in accordance with the transaction load.

    Finally, the modular design is important to the flexibility of the proposed system. In the global view of the system architecture (see Fig. 1), the “Schema Extractor”and the “Statistics Collector” modules are separated from the partition and migration modules purposely, so that they can be replaced easily to adopt other DBMS as needed. The command generation process in the “Migration Manager”module (see Fig. 3) is also isolated for the same reason.

    The transaction frequencies used in the “Partition Algorithm” module (see Fig. 2) is isolated from the steps of partition calculation. Furthermore, the unit of those figures is transformed to percentages. Through this arrangement,the transaction frequencies used for the partitions can be changed to other measures without affecting the algorithm.Conversely, the proposed architecture can be used as a platform to test different partition strategies by simply replacing the algorithm body.

    5. Conclusions

    In cloud environments, a transaction processing application needs to keep ACID guaranteeing and handle huge amount of data efficiently. Providing a partition strategy to the distributed database is important for fulfilling these requirements. This paper proposes a database distribution management system, DBDM, which distributes the data according to the transaction behaviors of the application system. It splits tables into groups so that join operations can be conducted in a group. Moreover,large groups are horizontally partitioned in pairs by observing the master-detail relationships between pairs of tables. These strategies avoid massive transmission of records in join operations among data nodes.

    The proposed DBDM system was also implemented to verify its feasibility. Preliminary experiments showed that the DBDM performed the database partition and migration so that the transactions were distributed over the nodes.Extensive experiments will be conducted to evaluate the effectiveness on the performance improvements.

    The DBDM system was modularly designed so that it can be easily adapted to different DBMS. Modules which deal with the “Partition Algorithm” can also be replaced to evaluate the performance of different algorithms.

    [1] F. Chang, J. Dean, S. Ghemawat, et al., “Bigtable: a distributed storage system for structured data,” in Proc. of the 7th Symposium on Operating Systems Design and Implementation, Seattle, 2006, pp. 205-218.

    [2] B. F. Cooper, R. Ramakrishnan, U. Srivastava, et al.,“PNUTS: Yahoo!’s hosted data serving platform,” in Proc.the 34th Int. Conf. on Very Large Data Bases, Auckland,2008, pp. 1277-1288.

    [3] G. DeCandia, D, Hastorun, M, Jampani, et al., “Dynamo:Amazon’s highly available key-value store,” in Proc. of the 21st ACM SIGOPS Symposium on Operating System Principles, Stevenson, 2007, pp. 205-220.

    [4] D. Agrawal, A. E. Abbadi, S. Antony, and S. Das, “Data management challenges in cloud computing infrastructures,”Lecture Notes in Computer Science, vol. 5999, pp. 1-10,2010.

    [5] S. Gilbert and N. Lynch, “Brewer’s conjecture and the feasibility of consistent, available, partition-tolerant web services,” ACM SIGACT News, vol. 33, no. 2, pp. 51-59,2002.

    [6] D. J. Abadi, “Consistency tradeoffs in modern distributed database system design: CAP is only part of the story,”Computer, vol. 45, no. 2, pp. 37-42, 2012.

    [7] M. Stonebraker, N. Hachem, and P. Helland, “The end of an architectural era: (It's time for a complete rewrite),” in Proc.of the 33rd Int. Conf. on Very Large Data Bases, Vienna,2007, pp. 1150-1160.

    [8] J. Rao, C. Zhang, G. Lohman, et al., “Automating physical database design in a parallel database,” in Proc. of Int. Conf.on Management of Data and Symposium on Principles Database and Systems, Madison, 2002, pp. 558-569.

    [9] D. C. Zilio, J. Rao, S. Lightstone, et al., “DB2 design advisor: integrated automatic physical database design,” in Proc. of the 30th Int. Conf. on Very Large Data Bases,Toronto, 2004, pp. 1087-1097.

    [10] S. Agrawal, V. Narasayya, and B. Yang, “Integrating vertical and horizontal partitioning into automated physical database design,” in Proc. of the ACM SIGMOD Int. Conf. on Management of Data, Paris, 2004, pp. 359-370.

    [11] A. Pavlo, C. Curino, and S. Zdonik, “Skew-aware automatic database partitioning in shared-nothing, parallel OLTP systems,” in Proc. of the ACM SIGMOD Int. Conf. on Management of Data, Scottsdale, 2012, pp. 61-72.

    [12] A. Thomson, T. Diamond, S.-C. Weng, et al., “Fast distributed transactions for partitioned database systems,” in Proc. of the ACM SIGMOD Int. Conf. on Management of Data, Scottsdale, 2012, pp. 1-12.

    [13] A. L. Tatarowicz, C. Curino, E. P. C. Jones, et al., “Lookup tables: Fine-grained partitioning for distributed databases,”in Proc. of the 28th IEEE Int. Conf. on Data Engineering,Washington, 2012, pp. 102-113.

    [14] P. M. G. Apers, “Data allocation in distributed database systems,” ACM Trans. on Database System, vol. 13, no. 3,pp. 263-304, 1998.

    [15] A. L. Corcoran and J. Hale, “A genetic algorithm for fragment allocation in a distributed database systems,” in Proc. of 1994 ACM symposium on Applied computing,Phoenix, 1994, pp. 247-250.

    [16] S. Menon, “Allocating fragments in distributed databases,”IEEE Trans. on Parallel and Distributed Systems, vol. 16,no. 7, pp. 577-585, 2005.

    高清午夜精品一区二区三区| 三级国产精品欧美在线观看| 高清日韩中文字幕在线| 91精品国产国语对白视频| 国产乱来视频区| 亚洲三级黄色毛片| 国产白丝娇喘喷水9色精品| 亚洲av日韩在线播放| 亚洲av男天堂| 在线观看免费高清a一片| 亚洲人与动物交配视频| 99久久精品一区二区三区| 五月开心婷婷网| 日韩av在线免费看完整版不卡| 黑人高潮一二区| 亚洲国产最新在线播放| 欧美高清性xxxxhd video| 国产毛片在线视频| 久久99蜜桃精品久久| 欧美另类一区| 精品一区二区三区视频在线| 国产日韩欧美在线精品| 国产爱豆传媒在线观看| 伦理电影免费视频| 麻豆精品久久久久久蜜桃| 成人午夜精彩视频在线观看| 亚洲一区二区三区欧美精品| 亚洲国产精品一区三区| 一级av片app| 人人妻人人爽人人添夜夜欢视频 | 日韩av在线免费看完整版不卡| 精品午夜福利在线看| 性色avwww在线观看| 欧美成人午夜免费资源| 熟女电影av网| 日日撸夜夜添| 亚洲国产最新在线播放| 成人综合一区亚洲| 日韩一本色道免费dvd| 日韩av免费高清视频| 国产老妇伦熟女老妇高清| 亚洲人成网站高清观看| 国产永久视频网站| 青青草视频在线视频观看| 青春草国产在线视频| 91精品一卡2卡3卡4卡| 亚洲激情五月婷婷啪啪| 久久综合国产亚洲精品| 亚洲人成网站在线播| 久久鲁丝午夜福利片| 成人黄色视频免费在线看| 视频中文字幕在线观看| 亚洲美女搞黄在线观看| 熟女电影av网| 久久久久视频综合| 亚洲欧美中文字幕日韩二区| 秋霞伦理黄片| 国产永久视频网站| 视频中文字幕在线观看| 亚洲av在线观看美女高潮| 三级经典国产精品| 一本—道久久a久久精品蜜桃钙片| 国模一区二区三区四区视频| 国产成人精品一,二区| 久久人妻熟女aⅴ| 免费不卡的大黄色大毛片视频在线观看| videossex国产| 亚洲精品一区蜜桃| av在线蜜桃| 又粗又硬又长又爽又黄的视频| 日本欧美国产在线视频| 小蜜桃在线观看免费完整版高清| 国产真实伦视频高清在线观看| 久久久久久久精品精品| 精品亚洲乱码少妇综合久久| 黑丝袜美女国产一区| 国产成人91sexporn| 精品久久久精品久久久| 精品久久久噜噜| 永久网站在线| 美女主播在线视频| 有码 亚洲区| 我要看黄色一级片免费的| 黄色视频在线播放观看不卡| 伦精品一区二区三区| 国产一区二区三区av在线| 天堂8中文在线网| 男人狂女人下面高潮的视频| 女人久久www免费人成看片| 免费播放大片免费观看视频在线观看| 我要看黄色一级片免费的| 日韩视频在线欧美| 亚洲精品,欧美精品| a级毛片免费高清观看在线播放| 久久久久国产精品人妻一区二区| 99re6热这里在线精品视频| 少妇被粗大猛烈的视频| 高清欧美精品videossex| 国产成人a区在线观看| 色哟哟·www| 丝袜脚勾引网站| 国产av一区二区精品久久 | 熟女人妻精品中文字幕| 亚洲,欧美,日韩| 精华霜和精华液先用哪个| 91精品国产国语对白视频| 又爽又黄a免费视频| 一级毛片黄色毛片免费观看视频| 啦啦啦在线观看免费高清www| 丝袜脚勾引网站| 欧美高清成人免费视频www| 国产精品无大码| 久久韩国三级中文字幕| 老师上课跳d突然被开到最大视频| 国产高清不卡午夜福利| 激情五月婷婷亚洲| 精品99又大又爽又粗少妇毛片| 欧美最新免费一区二区三区| 国产欧美另类精品又又久久亚洲欧美| 久久久久网色| 亚洲精品一二三| 亚洲欧美日韩无卡精品| 中文字幕精品免费在线观看视频 | 日本av手机在线免费观看| 欧美极品一区二区三区四区| 两个人的视频大全免费| 久久综合国产亚洲精品| 国产色爽女视频免费观看| 人妻夜夜爽99麻豆av| 激情 狠狠 欧美| 三级国产精品片| 日韩,欧美,国产一区二区三区| 亚洲av成人精品一二三区| 亚洲综合精品二区| 中文资源天堂在线| 日韩不卡一区二区三区视频在线| 乱系列少妇在线播放| 亚洲性久久影院| 国内揄拍国产精品人妻在线| 精品久久久久久久末码| 久久久精品免费免费高清| 亚洲美女黄色视频免费看| 一二三四中文在线观看免费高清| 尤物成人国产欧美一区二区三区| 日产精品乱码卡一卡2卡三| 亚洲色图综合在线观看| 狠狠精品人妻久久久久久综合| 大陆偷拍与自拍| 国产国拍精品亚洲av在线观看| 国产乱来视频区| 亚洲色图av天堂| 亚洲精品乱久久久久久| 国产乱来视频区| 国产精品.久久久| 丰满人妻一区二区三区视频av| 久久久久精品性色| 大香蕉久久网| 国产精品人妻久久久影院| 乱系列少妇在线播放| 日韩国内少妇激情av| 色婷婷av一区二区三区视频| 99热全是精品| 精品酒店卫生间| 十分钟在线观看高清视频www | 国产欧美日韩一区二区三区在线 | 一个人看的www免费观看视频| 亚洲av福利一区| 99热全是精品| 亚洲av二区三区四区| 国产亚洲91精品色在线| 内地一区二区视频在线| 亚洲精品日韩在线中文字幕| 亚洲精品久久午夜乱码| 成人一区二区视频在线观看| 日韩制服骚丝袜av| 韩国av在线不卡| 国产在线免费精品| 欧美另类一区| 欧美高清性xxxxhd video| 国产免费一级a男人的天堂| 欧美一区二区亚洲| 黄色欧美视频在线观看| 亚洲av成人精品一二三区| 国产精品久久久久久精品电影小说 | 国产精品国产三级专区第一集| 亚洲aⅴ乱码一区二区在线播放| 国产免费又黄又爽又色| 久久精品国产自在天天线| 大香蕉97超碰在线| 国产成人一区二区在线| 国产男人的电影天堂91| 蜜桃久久精品国产亚洲av| 国产女主播在线喷水免费视频网站| 深爱激情五月婷婷| 丝袜喷水一区| 国产真实伦视频高清在线观看| 日韩免费高清中文字幕av| 国产精品人妻久久久久久| 99久久精品热视频| 韩国高清视频一区二区三区| 亚洲av日韩在线播放| 久久久久久久精品精品| 小蜜桃在线观看免费完整版高清| 亚洲精品乱久久久久久| 在线亚洲精品国产二区图片欧美 | 亚洲成人av在线免费| 夫妻午夜视频| 国产爽快片一区二区三区| 纵有疾风起免费观看全集完整版| 亚洲高清免费不卡视频| 内射极品少妇av片p| av又黄又爽大尺度在线免费看| 国产精品国产三级国产专区5o| 国产av码专区亚洲av| 一级二级三级毛片免费看| 久久久a久久爽久久v久久| 汤姆久久久久久久影院中文字幕| av不卡在线播放| 伦精品一区二区三区| 国产一区亚洲一区在线观看| 如何舔出高潮| 久久热精品热| 欧美xxxx黑人xx丫x性爽| 日韩在线高清观看一区二区三区| 久久人人爽人人爽人人片va| 99久久人妻综合| 日韩av在线免费看完整版不卡| 美女内射精品一级片tv| 亚洲在久久综合| 一边亲一边摸免费视频| 老女人水多毛片| 少妇人妻一区二区三区视频| 最近2019中文字幕mv第一页| av.在线天堂| 校园人妻丝袜中文字幕| 国产午夜精品一二区理论片| 777米奇影视久久| 嘟嘟电影网在线观看| 亚洲精品aⅴ在线观看| 卡戴珊不雅视频在线播放| 国产av一区二区精品久久 | 香蕉精品网在线| 人妻夜夜爽99麻豆av| 久久国产亚洲av麻豆专区| 国产伦精品一区二区三区四那| 精品一区二区免费观看| 亚洲国产成人一精品久久久| 国产精品无大码| 亚洲国产精品成人久久小说| 精品国产露脸久久av麻豆| 少妇裸体淫交视频免费看高清| 精品一区二区三卡| 一本—道久久a久久精品蜜桃钙片| 国产精品久久久久久精品古装| 我的老师免费观看完整版| 亚洲aⅴ乱码一区二区在线播放| 另类亚洲欧美激情| 国产一区亚洲一区在线观看| av线在线观看网站| 美女主播在线视频| 婷婷色av中文字幕| 国产亚洲5aaaaa淫片| 国产精品熟女久久久久浪| 性高湖久久久久久久久免费观看| 啦啦啦中文免费视频观看日本| 小蜜桃在线观看免费完整版高清| 免费久久久久久久精品成人欧美视频 | 亚洲中文av在线| 免费av中文字幕在线| 97在线人人人人妻| av国产免费在线观看| 欧美精品亚洲一区二区| 日本黄大片高清| 91精品伊人久久大香线蕉| 18+在线观看网站| 亚洲自偷自拍三级| 亚洲美女搞黄在线观看| 国产 一区精品| 精品人妻熟女av久视频| 寂寞人妻少妇视频99o| 亚洲婷婷狠狠爱综合网| 国产黄片美女视频| 日本爱情动作片www.在线观看| 国产精品国产av在线观看| xxx大片免费视频| 老司机影院成人| 国产男女超爽视频在线观看| 啦啦啦视频在线资源免费观看| 在线播放无遮挡| 亚州av有码| av线在线观看网站| 国产精品秋霞免费鲁丝片| 亚洲精品中文字幕在线视频 | 成人国产麻豆网| 国产乱来视频区| 中文天堂在线官网| 丝袜脚勾引网站| 亚洲电影在线观看av| 一二三四中文在线观看免费高清| 99re6热这里在线精品视频| 久久久国产一区二区| 中文在线观看免费www的网站| 国产一区二区在线观看日韩| 成人二区视频| 日本与韩国留学比较| 最近的中文字幕免费完整| 久久人妻熟女aⅴ| 亚洲婷婷狠狠爱综合网| .国产精品久久| 99九九线精品视频在线观看视频| 人人妻人人看人人澡| 一边亲一边摸免费视频| 一级毛片aaaaaa免费看小| 一级毛片 在线播放| 免费黄网站久久成人精品| 黄片无遮挡物在线观看| 色视频www国产| 亚洲精品视频女| 一本一本综合久久| 人妻夜夜爽99麻豆av| 国产在视频线精品| 中国美白少妇内射xxxbb| 日本黄色片子视频| 国产又色又爽无遮挡免| 免费观看a级毛片全部| 亚洲综合精品二区| 国产在线男女| 美女主播在线视频| 最近手机中文字幕大全| 久久久欧美国产精品| 少妇被粗大猛烈的视频| 亚洲av免费高清在线观看| 久久久欧美国产精品| 国内揄拍国产精品人妻在线| 日韩亚洲欧美综合| a级一级毛片免费在线观看| 亚洲天堂av无毛| 亚洲欧美清纯卡通| 蜜桃在线观看..| 久久精品国产自在天天线| 国产精品久久久久成人av| 熟女av电影| 亚洲怡红院男人天堂| 亚洲欧美中文字幕日韩二区| 黄色配什么色好看| 免费av中文字幕在线| 97在线人人人人妻| 美女主播在线视频| 99热全是精品| 一个人看的www免费观看视频| 丰满乱子伦码专区| 亚洲图色成人| 久久久午夜欧美精品| 午夜福利视频精品| 成人美女网站在线观看视频| 日本与韩国留学比较| 国产成人a∨麻豆精品| 国产色婷婷99| 免费观看a级毛片全部| 婷婷色综合www| 久久久久久久久久成人| 亚洲怡红院男人天堂| 简卡轻食公司| 看免费成人av毛片| 久久久国产一区二区| 如何舔出高潮| 午夜福利网站1000一区二区三区| 亚洲精品久久午夜乱码| 亚洲av成人精品一区久久| 亚洲真实伦在线观看| videossex国产| 国产成人午夜福利电影在线观看| 亚洲av男天堂| 妹子高潮喷水视频| 欧美日韩在线观看h| 亚洲成人一二三区av| 国产精品福利在线免费观看| 国产精品久久久久久精品古装| 黄片无遮挡物在线观看| 亚洲国产成人一精品久久久| 日韩制服骚丝袜av| 又黄又爽又刺激的免费视频.| 成人国产av品久久久| 在线观看免费高清a一片| 亚洲色图av天堂| 好男人视频免费观看在线| 欧美日韩综合久久久久久| 精品久久久精品久久久| 日本黄大片高清| 男人爽女人下面视频在线观看| 午夜福利网站1000一区二区三区| 免费黄色在线免费观看| 中文欧美无线码| 香蕉精品网在线| 日韩伦理黄色片| 亚洲精品乱码久久久v下载方式| 精品久久久久久电影网| 国产日韩欧美在线精品| 久久久精品94久久精品| 人妻一区二区av| 久久青草综合色| 国产一区二区三区综合在线观看 | 男的添女的下面高潮视频| 狂野欧美激情性bbbbbb| 欧美精品一区二区免费开放| 中文字幕制服av| 99久久精品热视频| 精品久久久噜噜| 精品久久国产蜜桃| 麻豆成人av视频| 女人十人毛片免费观看3o分钟| 国产成人精品福利久久| 国语对白做爰xxxⅹ性视频网站| 肉色欧美久久久久久久蜜桃| 国产成人aa在线观看| 免费高清在线观看视频在线观看| 久久久久久久大尺度免费视频| 亚洲美女视频黄频| 国产日韩欧美在线精品| 国产视频首页在线观看| 久久精品国产亚洲av涩爱| 亚洲欧美日韩卡通动漫| 日本免费在线观看一区| 中文字幕精品免费在线观看视频 | 如何舔出高潮| 亚洲伊人久久精品综合| 夜夜爽夜夜爽视频| 久久影院123| 人妻制服诱惑在线中文字幕| 国语对白做爰xxxⅹ性视频网站| 国产成人精品久久久久久| 欧美精品亚洲一区二区| 成人黄色视频免费在线看| 亚洲精品乱码久久久v下载方式| 成人国产av品久久久| 免费黄网站久久成人精品| 国产成人免费观看mmmm| 久久婷婷青草| 天堂8中文在线网| 看十八女毛片水多多多| 国产精品久久久久成人av| 国内少妇人妻偷人精品xxx网站| 免费在线观看成人毛片| 26uuu在线亚洲综合色| 日日啪夜夜爽| 日韩av免费高清视频| 久久久久久久久久人人人人人人| 夜夜骑夜夜射夜夜干| 身体一侧抽搐| 天堂俺去俺来也www色官网| 日韩中字成人| 少妇精品久久久久久久| a级毛片免费高清观看在线播放| 国产精品99久久久久久久久| 美女内射精品一级片tv| 日韩欧美 国产精品| 欧美成人精品欧美一级黄| 精品国产三级普通话版| 国产色爽女视频免费观看| 黑人猛操日本美女一级片| 欧美人与善性xxx| 18禁在线无遮挡免费观看视频| 国产成人精品久久久久久| 亚洲精品第二区| 精品久久久久久久久亚洲| 老师上课跳d突然被开到最大视频| 精品久久久久久久久av| 国产一区二区三区综合在线观看 | 久久久午夜欧美精品| 国产欧美日韩一区二区三区在线 | 搡女人真爽免费视频火全软件| 国产亚洲av片在线观看秒播厂| 有码 亚洲区| 在线精品无人区一区二区三 | 亚洲激情五月婷婷啪啪| 久久久久久久国产电影| 欧美精品国产亚洲| 国产成人一区二区在线| 一级毛片我不卡| 91久久精品电影网| 如何舔出高潮| 在线观看三级黄色| 欧美激情国产日韩精品一区| 亚洲精品中文字幕在线视频 | 亚洲精品国产色婷婷电影| 国产欧美亚洲国产| 男人和女人高潮做爰伦理| 亚洲精品久久久久久婷婷小说| 18禁动态无遮挡网站| 亚州av有码| 中文资源天堂在线| 能在线免费看毛片的网站| 日韩av不卡免费在线播放| 极品教师在线视频| 777米奇影视久久| 久久亚洲国产成人精品v| 99久久综合免费| 国产爱豆传媒在线观看| 亚洲精品国产av蜜桃| 国产一区亚洲一区在线观看| 人人妻人人爽人人添夜夜欢视频 | 一级黄片播放器| 日韩不卡一区二区三区视频在线| 国产欧美日韩精品一区二区| 亚洲aⅴ乱码一区二区在线播放| 身体一侧抽搐| 如何舔出高潮| 午夜激情久久久久久久| 国产一区二区三区综合在线观看 | 亚洲欧美日韩无卡精品| 国产精品麻豆人妻色哟哟久久| 搡老乐熟女国产| 成人免费观看视频高清| 熟女人妻精品中文字幕| 亚洲欧美成人综合另类久久久| 一级毛片aaaaaa免费看小| 国产高清有码在线观看视频| 两个人的视频大全免费| 亚洲国产成人一精品久久久| 少妇的逼好多水| 九草在线视频观看| 下体分泌物呈黄色| 久久精品久久久久久噜噜老黄| 精品视频人人做人人爽| 亚洲aⅴ乱码一区二区在线播放| 永久免费av网站大全| 亚洲精品中文字幕在线视频 | 久久人人爽av亚洲精品天堂 | 国产精品一二三区在线看| 99热网站在线观看| 日韩成人伦理影院| 日本一二三区视频观看| 深夜a级毛片| 欧美一级a爱片免费观看看| 天堂8中文在线网| 亚洲国产高清在线一区二区三| 国产视频内射| 中国三级夫妇交换| 99re6热这里在线精品视频| 2021少妇久久久久久久久久久| 日本av免费视频播放| 亚洲av成人精品一区久久| 日韩成人av中文字幕在线观看| 成年美女黄网站色视频大全免费 | 亚洲精品第二区| 99久久精品热视频| 熟女av电影| 春色校园在线视频观看| 久久国产乱子免费精品| 精品一区二区三卡| 一二三四中文在线观看免费高清| 日韩一本色道免费dvd| 国产在线一区二区三区精| 搡老乐熟女国产| 熟女电影av网| 一级av片app| 高清av免费在线| 看十八女毛片水多多多| 99热6这里只有精品| 一级毛片 在线播放| 欧美+日韩+精品| 中文乱码字字幕精品一区二区三区| 欧美日韩精品成人综合77777| 成人一区二区视频在线观看| 六月丁香七月| 亚洲精品,欧美精品| 在线观看一区二区三区激情| 国产高清有码在线观看视频| 在线观看av片永久免费下载| av线在线观看网站| 丰满迷人的少妇在线观看| 永久网站在线| 亚洲人成网站在线播| 成人特级av手机在线观看| 人体艺术视频欧美日本| 男人狂女人下面高潮的视频| 日本猛色少妇xxxxx猛交久久| 久久女婷五月综合色啪小说| 久久99热6这里只有精品| xxx大片免费视频| 色综合色国产| 久久久久性生活片| 精品午夜福利在线看| 亚洲中文av在线| 麻豆乱淫一区二区| 黑人高潮一二区| 精品酒店卫生间| 久久精品国产亚洲网站| 国产精品人妻久久久影院| 色综合色国产| a级毛色黄片| 久久人人爽人人爽人人片va| 久久久久久久久久久免费av| 国产精品国产三级国产专区5o| 国产精品人妻久久久影院| 亚洲中文av在线| a级毛色黄片| 啦啦啦在线观看免费高清www| 成人亚洲精品一区在线观看 | 日韩av不卡免费在线播放| 在线观看人妻少妇| 国精品久久久久久国模美| 这个男人来自地球电影免费观看 | 中文字幕人妻熟人妻熟丝袜美| 欧美一区二区亚洲| 欧美精品一区二区大全| 国产深夜福利视频在线观看| 天堂8中文在线网| 成人毛片a级毛片在线播放| 久久久久久久久久久丰满| 免费大片18禁| 久久久久久九九精品二区国产| 在线观看人妻少妇| 91久久精品国产一区二区成人| 只有这里有精品99|