• <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 ?

    An optimizing algorithm of static task scheduling problem based on hybrid genetic algorithm①

    2016-12-05 01:31:18LiuYuSongJianWenJiayan
    High Technology Letters 2016年2期

    Liu Yu (柳 玉), Song Jian, Wen Jiayan

    (The Scientific Research Department, Naval Marine Academy, Guangzhou 510430, P.R.China)

    ?

    An optimizing algorithm of static task scheduling problem based on hybrid genetic algorithm①

    Liu Yu (柳 玉)②, Song Jian, Wen Jiayan

    (The Scientific Research Department, Naval Marine Academy, Guangzhou 510430, P.R.China)

    To reduce resources consumption of parallel computation system, a static task scheduling optimization method based on hybrid genetic algorithm is proposed and validated, which can shorten the scheduling length of parallel tasks with precedence constraints. Firstly, the global optimal model and constraints are created to demonstrate the static task scheduling problem in heterogeneous distributed computing systems(HeDCSs). Secondly, the genetic population is coded with matrix and used to search the total available time span of the processors, and then the simulated annealing algorithm is introduced to improve the convergence speed and overcome the problem of easily falling into local minimum point, which exists in the traditional genetic algorithm. Finally, compared to other existed scheduling algorithms such as dynamic level scheduling(DLS), heterogeneous earliest finish time(HEFT), and longest dynamic critical path(LDCP), the proposed approach does not merely decrease tasks schedule length, but also achieves the maximal resource utilization of parallel computation system by extensive experiments.

    genetic algorithm, simulated annealing algorithm, parallel computation, directed acyclic graph

    0 Introduction

    A parallel application can be abstracted as tasks sets, in which the elements are organized as a partial order and is in serial or parallel working mode. The task scheduling system allocates tasks to specified computers or processors and implement in parallel, based on some rules and constraints. In general, task scheduling algorithms are classified into two classes: static and dynamic, the former outputs scheduling schems in a compiling state, while the latter is made at run time. Static scheduling algorithms are widely used for their simplicity, high execution efficiency and low cost[1]. The task scheduling is to find a kind of tasks assignment scheme which can make application finished in the shortest time, and has been approved as an NP-complete problem. The results of scheduling decisions directly affect the total performance of distributed computing environment, even possibly offset the gains from distributed platform in worse case. Therefore, the design and implementation of an excellent task scheduling algorithm need to be focused and researched by computer scientists.

    There are several static tasks scheduling algorithms: graphs[2], queuing theory[3], math programming[4]and heuristics, etc., where the last is the most popular. According to the characteristics of task retained, there are two types of heuristic algorithms: task duplication and non-task duplication, and it has been proved that the former is better than the latter[5], typical examples including TDS[6]and OSA[7]algorithms. For example, The TDS assign the nodes with their pre-order nodes at the same processor to decrease parallel executing time. As an improved version of TDS, the OSA achieves similar results by removing some unnecessary constraints. In summary, the mathematical models for previous static tasks scheduling algorithms belong to the integer-programming applications essentially. Although the heuristic and branching-bounding algorithms were introduced later to improve the algorithms speed and efficiency in some extent, their computing time remains unbearable, especially when facing many more tasks and processors, and their results are easy to produce early maturity and fall into local extreme points. Therefore it is difficult to take full advantage of high performance computing environment.

    Genetic algorithm is a kind of highly parallel, randomized, adaptive search algorithm by learning nature selection and evolution mechanism from nature, which has excellent robustness and is especially suitable for complicated, non-linear problems puzzled by the traditional searching algorithm[8]. In this paper, by adding constraints such as heterogeneity, idle time interval and so on, an optimization model of static task scheduling is proposed in the heterogeneous distributed computing systems. In this model, by us real matrix coding seeds, combined with the simulated annealing in the process of evolution, the genetic population can achieve rapid convergence. As a result, the mixed integer nonlinear static task scheduling problems with multiple variables and constraints can be successfully solved and achieve to their global optimal solutions. Finally, results and performance of the proposed algorithm are tested by specific computational experiments and compared with other widely used scheduling algorithms.

    The remainder of this paper is organized as follows: in Section 1, the research problem and some necessary assumptions are defined. Section 2 introduces the optimized mathematical model of task scheduling problem. Section 3 considerably depicts hybrid genetic algorithm. Section 4 conducts two sets of experiments to acquire the performance of the algorithm. Finally a conclusion and an overview of future work are given in Section 5.

    1 Problem definition

    Definition 1 The heterogeneous distributed computing system (HeDCSs) is a computing environment with heterogeneous computing nodes or heterogeneous interconnection network in which computational nodes are located in different geographical districts[9].

    Hypothesis 1 Not considering the data transfer type, style, interface and other compiler details, a parallel application, can be represented by a directed acyclic graph (DAG). DAG is defined by the tuple (T,E), whereT={t0,t1,…,tN-1} is a set ofNtasks andE={eij=(ti,tj)|ti,tj∈T,i≠j} is a set of edges. Each edgeeijrepresents a precedence constraint and a communication message between taskstiandtj. That means ifeij∈E, then the execution oftjcannot be started beforetifinishes its execution. A task with no parents is called an entry task, and a task with no children is called an exit task. Associated with each edgeeij, there is a valuewijthat represents the amount of data to be transmitted from tasktito tasktj.

    Hypothesis 2 Representing heterogeneity only through the difference of computing resources of processors which are fully connected. LetP={p0,p1,…,pM-1} denote a set of processors in the HeDCSs environment,Mis the number of processors, computing ability of tasks is described by two-dimensional matrixC=[cimorc(ti,pm)]N×M, wherecimdenotes the tasktiexecution time needed in the processorpmto reflect heterogeneity of computation.

    Hypothesis 3 The calculating cost is monotonic, if ?ti∈T,?m, k∈{0,1,…,M-1} and m≠k,cim≥cik, then ?tj∈T, cjm≥cjk.

    Hypothesis 4 If the network data transmission rate is fixed, and the communication overhead of tasks executed in the same processors is zero, the communication cost between the two tasks can be measured by their interactive data amount.

    To sum up, static task scheduling problem in heterogeneous distributed computing environment may achieve the following optimization goal:

    ① The time to complete all the parallel tasks is shortest;

    ② Communication overhead between parallel tasks is most economical;

    ③ Average waiting time of tasks is the shortest;

    ④ Balance processors allocation as far as possible

    ⑤ Improve the operation efficiency of the computing system as far as possible.

    2 Problem modeling

    If static tasks scheduling problemZcan be seen as two procedures: processors allocation and computing task executing time in the processor, it establishes its mathematical model through analyzing time relationship between the first executed task and the last one:

    Z=min(F(tiN-1, PA(tiN-1))-S(ti0, PA(ti0)))

    ti0, ti1,…, tiN-1∈T (1)

    The constraint conditions:

    (1) if ?i, j∈{0,1,…,N-1}, ?eij∈E, then F(ti, PA(ti))≥S(tj, PA(tj)).

    (2) if ?i, j∈{0,1,…,N-1}, ?m∈{0,1,…,M-1} PA(ti)=PA(tj)=m, theneij=0.

    (3) F(tiβ, PA(tiβ))=S(tiβ, PA(tiβ))+c(tiβ, PA(tiβ)), β∈{0,1,…,N-1}.

    (4) if ?Pm∈P,?i, j∈{0,1,…,N-1}, PA(ti)=PA(tj)=Pm, then [S(ti,tj), F(ti,tj)]∩[S(ti,tj), F(ti,tj)]=?.

    (5) S(ti0, PA(ti0))=0.

    (6)PA(·) is a surjective function that is mapped from definition domain {0,1,…,N-1} to value domain {0,1,…,M-1}.

    In the above conditions, PA(tik), S(tik, PA(tik)) and F(tik, PA(tik)) respectively denote the processor assigned starting execution time and completion execution time of task tik. The solution of Eq.(1), ti0, ti1,…, tiN-1, shows a kind of parallel task scheduling sequence. The first item in the constraints list shows that the starting time of any child tasks must be less than their parents’ completed time. The second shows the communication overhead between two tasks assigned in the same processor is zero. The third shows that completion time of a task equals to the sum of its starting time and computation time needed. The first three conditions in fact explain that the final solution must meet the definition of HeDCSs model. The fourth illustrates that when a task begins to run, it cannot be interrupted until it is finished, namely each parallel task is viewed as an atomic one in the HeDCSs. The fifth shows that tasks begin from zero point and in the last condition, the functionPAdenotes which processor is allocated to the task.

    Definition 2 CPU idle time refers to execution time intervals of tasktiand its successortjin processorpm, denoted asIdle(ti,tj,pm) andIdle(ti,tj,pm)≥0.

    IfIdle(ti,tj,pm)=0, it means that tasktjis immediately started when tasktiis finished, otherwise the computing procedure oftiandtjis discontinuous, namely there are time fragments of CPU that are wasted.

    Definition 3S(ti,pm) represents the start time of tasktiexecuted in processorpmfor a specified scheduling scheme.

    When a task has parent nodes, it begins running until all parents are completed.

    Definition 4F(ti,pm) represents the completion time oftiexecuted in processorpmfor a specified scheduling schema, andF(ti,pm)=S(ti,pm)+c(ti,pm).

    Therefore, the optimal decision for the shortest scheduling length of static tasks is equivalent to computing the shortest scheduling length of parallel tasks relationship diagram, which is a mixed integer, nonlinear programming problem with multi-variables and many constraints. Genetic algorithm has highly parallel, random and self-adaptive search performance and can solve this kind of problem more effectively.

    3 Model resolution with hybrid genetic algorithm

    Such traditional linear optimization as heuristic algorithm, intelligent simulated annealing algorithm, taboo search algorithm and neural network try to seek a local approximate solution or satisfactory solution. It is a fresh way to tackle the problem of task scheduling in HeDCSs by the fusion of genetic algorithm with global searching characteristics and traditional ones with fast convergence speed.

    Firstly, randomly generated initial population in a solution of the problem space, then the genetic annealing operation of the initial population equation can be defined as

    GASA=(S(o),A,L,R,PTP,ρ,φ,TP,ξ)

    whereS(o) denotes the initial population andAis its population size,Lis the individual code length,Rrepresents selection operator,PTPis used to describe state transition of random (probability) matrix,ρis a set of genetic operator,φrepresents fitness function, TP={TPt, t=0,1,2,…, f} describes the temperature sequence,ξis the last condition of this algorithm.

    When the algorithm runs, it first calculates the individual fitness function value in the initial population, in which probabilityPTPdecides whether an individual should be remained in gene pool or participated in the following hybridization evolution choice. The resulting population of genetic evolution is processed as the same rules implementing mutation operation. According to the Metropolis criterion, individuals staying in the gene pool or participating in the hybrid mutation are chosen to constantly lower their temperature. The process is repeated for each newly generated offspring, until a termination condition of this algorithm is satisfied.

    3.1 Individuals with binary matrix code

    Taking the scheduling sequence of tasks as an individual in the genetic algorithm, thek-th individual Gkcoded by binary matrix can be expressed as

    (2)

    wherePidenotes the allocated processor of specified taskti,pi,mis the element in thei-th row andm-th column that showstiis placed on processorpmto run.

    According to the 4-th constraint of Eq.(1), CPU is only allowed to perform a task in a period until finished, so elementpi,mcan be deduced to meet the following:

    (3)

    3.2 Generation of initial population

    The initialization of genetic population is done by row sequences of encoding matrix of individuals and tasks are ordered afterwards by graph levels first and branches quantity later.

    (1) To compute the task level. Any nodes without independence relation constitute a level in DAG, which can be calculated by

    (4)

    where symbolHl(ti) denotes the level of taskti,Tentryis a set of entry points, function max is to return the maximal value of definition domain andPred(ti) is also a set of parent nodes oftiin DAG.

    (2) Given that there areLlevels in DAG, and thei-th level containsδinodes, all tasks can be arranged by their level grade: tγ(11), tγ(12),…, tγ(1δ1), tγ(21), tγ(22),…, tγ(2δ2),…, tγ(L1), tγ(L2),…, tγ(LδL).

    (3) Tasks in each level are organized with full permutation, and then do Descartes product. The result is 2δ1-1×2δ2-1×…×2δL-1valid combinations.

    (4) In view of feasibility and fairness in tasks assignment of processors, each task can be run on arbitrary processor. So, the above combinations can produce the total 2δ1-1×2δ2-1×…×2δL-1×MLinitial individuals.

    3.3 Individuals adjustment measurement

    Taking individual Gkas an example, specific adjustment measurements are given in the following:

    Rule 1 If all tasks meet the demand of the earliest completion time, they should be unchanged.

    Rule 2 If a task violates executing dependency constraints, the corresponding value in thePθ-th row should be adjusted as

    ① ?ta, tβ∈T,Idle(tα,tβ,pm)≥c(ti,pm);

    ② ?γ,δ∈T,tδ∈Succ(tγ), Idle(tγ,tδ,pm)≥c(ti,pm), Idle(tγ,tδ,pm)≥Idle(tα,tβ,pm), whereSucc(ti) is a set of successor nodes of taskti.

    Generally, the start time of the task executed in the processor is affected by the computation cost, communication cost and the length of processor idle time pieces. By properly exerting above three rules to compute the finish time of tasks in each processor, the value of element in thePθ-th row corresponding to a processor that can make the task finish with earliest speed is set to 1.

    3.4 The selection of fitness function

    When appling the above individual adjustment measurement, only the fourth constraint is possibly unsatisfied because the parallel tasks in execution cannot be interrupted. Meanwhile in order to accelerate the convergence, the following fitness function of an individual is established

    (5)

    whereGkdenotes an individual in the population,Siis penalty degree while tasktiviolates the constraint that executing process cannot be interrupted,Zis set to the computational cost ofti,μis the penalty factor,mis the number of the violation of that constraint, andAis the positive constant to represent penalty coefficient whentiviolates that constraint on one time.

    3.5 The definition of genetic operator

    1) The selection operator

    P(Gk) is the selected probability of individualGkin the genetic process, and defined as

    (6)

    2) The crossover operator

    Step 1 Generate random crossover factorω2in [0, 1], and an integerjin [0,N-1] which denotes the position of crossover.

    Step 2 Begin to crossover and produce two individualsD1,D2. The result is

    Step 3 ModifyD1andD2in accordance with rules in Section 3.3, and computeF(D1) andF(D2).

    Step 4 Employ a local tournament method to generate offspringO1,O2, where they satisfy the following conditions:

    O1∈{C1,D1}, F(O1)=max(F(C1), F(D1)),

    O2∈{C2,D2}, F(O2)=max(F(C2), F(D2)).

    3)The mutation operator

    Step 1 Generate a random factorβin [0, 1], and an integerγin [0, N-1] which denotes the position of mutation.

    Step 2 Generate a random binary mutation factorω3from the set {0, 1}, whereω3=1 represents mutation occurrence, otherwise no variation.

    Step 3 Implement mutation operation and generate offspringQexpressed as

    Q=[P0C1P1C1…ω3×β×PγC1Pγ+1C1…PN-1C1]T

    Step 4 ModifyQin accordance with rules in Section 3.3, and computeF(Q).

    Step 5 Employ a local tournament method to generate offspringS, where they satisfy that E1∈{C1, Q}, F(S)=max(F(C1), F(Q)).

    4 Performance evaluation

    4.1 Performance metrics

    1)Normalized schedule length (NSL)

    For the convenience of analysis and fairness of comparison, a standardized processing approach is introduced and given as

    (7)

    whereTCPis a collection of nodes on the critical path in DAG.

    2)Speedup[10]

    The speedup of a parallel application is defined as

    (8)

    whereMis the number of processors contained in HeDCSs. Meanwhile, 1≤S≤M.

    4.2 Testing environment

    Hardware environment of parallel computation is composed of DELL PowerEdge series blade servers. Software communication interface is chosen to MPICH 1.5. Parallel applications are simulated by 2000 random graphs where generation parameters are seen in Table 1.

    Table 1 Generation parameters of random graphs

    Topcuoglu[11]has proved that among popular static algorithms in HeDCSs, HEFT and DLS are better than CPOP. Mohammad[12]further compared the performance difference between LDCP, HEFT and DLS. Through the same performance metrics used in the above work, NSL and Speedup, detailed comparison experiments between our proposed algorithm and the other three algorithms are also done.

    4.3 Performance results on random graphs

    1) Comparison of NSL for four algorithms

    Generating 2000 DAGs with parameters from Table 1, the static task scheduling schemes are calculated respectively by the proposed algorithm, LDCP, HEFT and DLS. Furthermore, their NSL is computed by Eq.(7), the last statistical result is shown in Table 2.

    Table 2 Performance results: a standardized schedule length (NSL)

    Note:the data in a cell indicates comparison results on NSL using the left algorithm and the top one, including the number and proportion.

    It is shown in Table 2 that for the given tested data, the number of applications whose NSL value used by the proposed algorithm is 1593, 1761, 1845, respectively, shorter than those of the LDCP, HEFT, and DLS and 79.65%, 88.05%, 92.25% of proportion respectively relative to the gross data. The proposed algorithm can shorten the execution time of a parallel application, improve the utilization performance of processors, and has general character. In addition, experiments also achieve the same result as Ref.[9], and the performance on NSL of other three algorithms is ordered by LDCP, HEFT, and DLS from high to low. Hence, correctness of the proposed algorithm is verified.

    2) The influence of the generated parameters of random graphs on scheduling algorithm

    In order to further validate proposed algorithm influences on different types of parallel applications, two parameters of CCR and number of task nodes are changed to randomly generate 2000 DAG again. The difference of average NSL and speedup of four algorithms is shown in Fig.1.

    Fig.1 Effects of generation parameters of random graphs on scheduling algorithms

    It is shown that the average NSL value of the proposed algorithm is shorter than those of DLS, HEFT, and LDCP by (27.03%, 19.04%, 10.52%), (11.98%, 3.68%, 2.08%), (10.89%, 9.09%, 2.87%), (10.08%, 7.30%, 3.50%) and (9.67%, 5.08%, 2.94%), respectively. The first value of each parenthesized pair is the improvement achieved by the proposed algorithm over the DLS, while the second value is the improvement over the HEFT algorithm and the last value is relative to the LDCP algorithm. This convention for representing results will be adhered throughout this paper, unless an exception is explicitly noted. The average speedup of the proposed algorithm is higher than those of DLS, HEFT, and LDCP by (15.92%, 14.73%, 9.84%), (11.91%, 7.27%, 6.90%), (16.92%, 9.35%, 8.19%), (11.89%, 4.60%, 1.11%) and (30.08%, 11.11%, 6.67%), when the CCR is equal to 0.1, 0.5, 1.0, 2.0, 5.0, respectively. In these experiments, the proposed algorithm outperforms the others on NSL, and their NSL trend is directly proportional to CCR, while their speedup is in the reverse situation. This result fully satisfies the physical meaning of CCR and indicates that the proposed algorithm is effective. On the other hand, as the number of parallel tasks increases, the NSL and speedup value of the four algorithms mentioned above become greater, but the proposed algorithm in this paper is relatively more advantageous. Experiments on random graphs conformably show that our algorithm is effective and general.

    5 Conclusions

    It is a challenging work in high performance computing domain for parallel task scheduling under heterogeneous distributed computing environment, which directly affects the existing resource utilization of heterogeneous computing system. In this paper, by constructing a mathematical model of static task scheduling, the tasks scheduling problem in HeDCSs is further researched, and is solved properly based on hybrid genetic algorithm. We discuss and achieve such technical details of the algorithm as individual coding, generation of initial population, individual adjustment, selection of fitness function, and definition of genetic operator, etc., and verify the generality and the effectiveness of the algorithm by abundant experiments. The further research effort is to focus on the parallelization of genetic algorithm self and influences of all generation parameters of random graphs on scheduling algorithm, therefore make further improvement on the performance and pertinence of task scheduling algorithm in HeDCSs.

    [1] Nelissen G, Su H. An optimal boundary fair scheduling.Real-TimeSystems, 2014,50(4): 456-508

    [2] Liu M, Chu C B, Xu Y F, et al. An optimal online algorithm for single machine scheduling to minimize total general completion time.JournalofCombinatorialOptimization, 2012,23(2): 189-195

    [3] Lombardi M, Milano M. Optimal methods for resource allocation and scheduling: a cross-disciplinary survey.Constraints, 2012,17(1): 51-85

    [4] Regnier P, Lima G, Massa E, et al. Multiprocessor scheduling by reduction to uniprocessor: an original optimal approach.Real-TimeSystems, 2012,2013(4): 436-474

    [5] Epstein L, Hanan Z H. Online scheduling with rejection and reordering: exact algorithms for unit size jobs.JournalofCombinatorialOptimization, 2014,28(4): 875-892

    [6] Darbha S, Agrawal P D. Optimal scheduleing algorithm for distributed-memory machines.IEEETransactionsonParallelandDistributedSystems, 1998,9(1): 87-95

    [7] Park C I, Choe T Y. An optimal scheduling algorithm based on task duplication.IEEETransactionsonParallelandDistributedSystems, 2002,51(4): 444-448

    [8] Falzon G, Li M. Enhancing genetic algorithms for dependent job scheduling in grid computing environments.JournalofSupercomputing, 2012,62(1): 290-314

    [9] Arabnejad H, Barbosa J G. List scheduling algorithm for heterogeneous systems by an optimistic cost table.IEEETransactionsonParallelandDistributedSystems, 2014,25(3): 682-694

    [10] Paredes R U, Cazorla D, Sanchez J L, et al. A comparative study of different metric structures in thinking on GPU implementations.LectureNotesinEngineeringandComputerScience, 2012: 312-317

    [11] Topcuoglu H, Hariri S, Wu M Y. Performance-effective and low-complexity task scheduling for heterogeneous computing.IEEETransactionsonParallelandDistributedSystems, 2002,13(3): 260-274

    [12] Mohammad I D, Nawwaf K. A high performance algorithm for static task scheduling in heterogeneous distributed computing systems.JournalofParallelandDistributedComputing, 2008,19(8): 399-409

    Liu Yu, born in 1982. He received his Ph.D degree from Naval Engineering University in 2011. He also received his M.S. degree from Naval Aeronautical Engineering Institute in 2008. His research focuses on military modeling and simulation, operational system and high performance computing technology.

    10.3772/j.issn.1006-6748.2016.02.008

    ①Supported by the National Natural Science Foundation of China (No. 61401496).

    ②To whom correspondence should be addressed. E-mail: game_liuyu@163.comReceived on Dec. 20, 2014

    亚洲国产精品成人综合色| 男的添女的下面高潮视频| 如何舔出高潮| 午夜福利在线在线| 看片在线看免费视频| 精品欧美国产一区二区三| 国产黄a三级三级三级人| 我的老师免费观看完整版| 亚洲精品456在线播放app| 美女xxoo啪啪120秒动态图| 国产精品无大码| 黄色日韩在线| 免费人成在线观看视频色| 亚洲欧美一区二区三区国产| 亚洲18禁久久av| 91午夜精品亚洲一区二区三区| 秋霞伦理黄片| 国产麻豆成人av免费视频| 精品午夜福利在线看| 日本黄大片高清| 欧美极品一区二区三区四区| 亚洲av熟女| 亚洲18禁久久av| 人体艺术视频欧美日本| 丝袜喷水一区| 我要看日韩黄色一级片| 黄片无遮挡物在线观看| 免费av不卡在线播放| 国产探花在线观看一区二区| 中文字幕精品亚洲无线码一区| 一区二区三区四区激情视频| 国内揄拍国产精品人妻在线| 久久久久国产网址| 尤物成人国产欧美一区二区三区| 最近手机中文字幕大全| 日韩成人av中文字幕在线观看| 久久久久久伊人网av| 身体一侧抽搐| 亚洲人成网站在线观看播放| 九九爱精品视频在线观看| 亚洲国产色片| 欧美性感艳星| 亚洲自偷自拍三级| 波多野结衣高清无吗| 国产白丝娇喘喷水9色精品| 国产精品电影一区二区三区| 欧美高清性xxxxhd video| 国产亚洲精品久久久com| 国产一区二区在线av高清观看| 亚洲精品日韩av片在线观看| 国产精品久久久久久久电影| 国产精品不卡视频一区二区| 嘟嘟电影网在线观看| 国产真实乱freesex| 欧美激情在线99| 亚洲av男天堂| 波野结衣二区三区在线| 夜夜爽夜夜爽视频| 婷婷色av中文字幕| 午夜福利视频1000在线观看| 视频中文字幕在线观看| 永久网站在线| 少妇的逼水好多| 中文在线观看免费www的网站| 99热6这里只有精品| 男女视频在线观看网站免费| 亚洲国产欧美在线一区| 草草在线视频免费看| 男的添女的下面高潮视频| 久久鲁丝午夜福利片| 国产精品嫩草影院av在线观看| 精品久久国产蜜桃| 亚洲国产精品sss在线观看| 国产精品一区二区三区四区免费观看| 精品人妻一区二区三区麻豆| 熟女人妻精品中文字幕| 在线观看美女被高潮喷水网站| 国产一区二区在线观看日韩| 黄色日韩在线| av专区在线播放| 少妇人妻精品综合一区二区| 深爱激情五月婷婷| 亚洲欧美成人综合另类久久久 | 亚洲av不卡在线观看| 一本久久精品| 国国产精品蜜臀av免费| 波野结衣二区三区在线| 美女被艹到高潮喷水动态| 国产一级毛片在线| 一边摸一边抽搐一进一小说| 久久人人爽人人爽人人片va| 日韩中字成人| 两个人的视频大全免费| 岛国毛片在线播放| 大又大粗又爽又黄少妇毛片口| 欧美高清性xxxxhd video| 久久人人爽人人片av| 99久久无色码亚洲精品果冻| АⅤ资源中文在线天堂| 欧美日韩在线观看h| 午夜福利在线观看免费完整高清在| 欧美日韩国产亚洲二区| 尤物成人国产欧美一区二区三区| 青春草国产在线视频| 精品久久国产蜜桃| www.色视频.com| 3wmmmm亚洲av在线观看| 男的添女的下面高潮视频| 国产av码专区亚洲av| 尾随美女入室| 大又大粗又爽又黄少妇毛片口| 99久久成人亚洲精品观看| 麻豆精品久久久久久蜜桃| 好男人在线观看高清免费视频| 久久精品久久精品一区二区三区| АⅤ资源中文在线天堂| 少妇丰满av| 亚洲欧美精品专区久久| 国产精品久久久久久久电影| 久久精品国产亚洲av涩爱| 别揉我奶头 嗯啊视频| 嘟嘟电影网在线观看| 丝袜美腿在线中文| 一级二级三级毛片免费看| av免费在线看不卡| 日韩欧美精品v在线| 寂寞人妻少妇视频99o| 特级一级黄色大片| 国产色婷婷99| 成人二区视频| 精品人妻熟女av久视频| 亚洲精品影视一区二区三区av| 高清av免费在线| 看非洲黑人一级黄片| 波野结衣二区三区在线| 日韩欧美在线乱码| 天堂影院成人在线观看| 久久人人爽人人片av| 国产亚洲av嫩草精品影院| 观看美女的网站| 最新中文字幕久久久久| 村上凉子中文字幕在线| 日本黄色视频三级网站网址| 青春草视频在线免费观看| 嘟嘟电影网在线观看| 午夜精品在线福利| 18禁裸乳无遮挡免费网站照片| 日韩强制内射视频| 久久久精品欧美日韩精品| 国产视频首页在线观看| 亚洲伊人久久精品综合 | 大又大粗又爽又黄少妇毛片口| av国产久精品久网站免费入址| 亚洲国产欧美人成| 麻豆国产97在线/欧美| 亚洲精品乱久久久久久| 蜜臀久久99精品久久宅男| 成人亚洲欧美一区二区av| 国产欧美日韩精品一区二区| 18禁裸乳无遮挡免费网站照片| 亚洲成人中文字幕在线播放| 日韩欧美在线乱码| 成人综合一区亚洲| 级片在线观看| 亚洲不卡免费看| 亚洲av中文字字幕乱码综合| 两性午夜刺激爽爽歪歪视频在线观看| 老司机影院毛片| 日韩欧美 国产精品| 免费播放大片免费观看视频在线观看 | 亚洲精品国产成人久久av| 丰满人妻一区二区三区视频av| 久久精品91蜜桃| av天堂中文字幕网| 免费一级毛片在线播放高清视频| 精品久久久噜噜| 国产成人a区在线观看| 久久精品国产亚洲av天美| 亚洲人成网站在线播| 99热6这里只有精品| 亚洲欧美日韩卡通动漫| 观看美女的网站| 乱人视频在线观看| 亚洲内射少妇av| 欧美bdsm另类| 村上凉子中文字幕在线| 天天躁夜夜躁狠狠久久av| 我的老师免费观看完整版| 国产精品一区二区在线观看99 | 三级国产精品欧美在线观看| 亚洲精品乱码久久久久久按摩| 日本免费在线观看一区| 内射极品少妇av片p| 91狼人影院| 午夜视频国产福利| kizo精华| 国产亚洲精品av在线| 国产精品不卡视频一区二区| 国产精品99久久久久久久久| 成人高潮视频无遮挡免费网站| 日日摸夜夜添夜夜添av毛片| 老师上课跳d突然被开到最大视频| 精品99又大又爽又粗少妇毛片| 一级av片app| 干丝袜人妻中文字幕| 亚洲成人久久爱视频| 亚洲欧美精品专区久久| 久久99蜜桃精品久久| 欧美日本亚洲视频在线播放| 又粗又爽又猛毛片免费看| 久久久亚洲精品成人影院| 国产精品一区二区性色av| 亚洲第一区二区三区不卡| 亚洲经典国产精华液单| 久久这里有精品视频免费| 尤物成人国产欧美一区二区三区| 亚洲不卡免费看| 成年av动漫网址| 国产乱来视频区| 午夜激情欧美在线| 麻豆成人午夜福利视频| av女优亚洲男人天堂| 亚洲人与动物交配视频| 欧美丝袜亚洲另类| 日韩大片免费观看网站 | 国产成人aa在线观看| 亚洲精品影视一区二区三区av| av女优亚洲男人天堂| 插阴视频在线观看视频| 人妻少妇偷人精品九色| 国产又黄又爽又无遮挡在线| 伦理电影大哥的女人| av天堂中文字幕网| 精华霜和精华液先用哪个| 精品国产三级普通话版| av免费在线看不卡| 日本爱情动作片www.在线观看| 久久6这里有精品| 国产精品一及| 五月伊人婷婷丁香| 免费观看a级毛片全部| 久久精品夜夜夜夜夜久久蜜豆| 黄片wwwwww| 色综合亚洲欧美另类图片| 亚洲综合色惰| 日本与韩国留学比较| 午夜福利网站1000一区二区三区| 日本免费在线观看一区| 亚洲av中文av极速乱| 久久欧美精品欧美久久欧美| 欧美三级亚洲精品| 人妻夜夜爽99麻豆av| 久久久久久九九精品二区国产| 男的添女的下面高潮视频| 一个人看的www免费观看视频| 欧美不卡视频在线免费观看| 国产综合懂色| 午夜爱爱视频在线播放| 麻豆成人午夜福利视频| 日韩 亚洲 欧美在线| 久99久视频精品免费| 久久午夜福利片| 亚洲国产色片| 亚洲成人久久爱视频| 丝袜美腿在线中文| 日本爱情动作片www.在线观看| 国产黄片视频在线免费观看| 99久久精品国产国产毛片| 国产老妇伦熟女老妇高清| 国产真实乱freesex| 99在线视频只有这里精品首页| 久久午夜福利片| 一个人观看的视频www高清免费观看| 亚洲内射少妇av| av在线天堂中文字幕| 久久精品久久精品一区二区三区| 男人和女人高潮做爰伦理| 日韩国内少妇激情av| videossex国产| 国产精品一区www在线观看| 免费看av在线观看网站| 国产在线男女| 永久免费av网站大全| 久久99蜜桃精品久久| 成人漫画全彩无遮挡| 国产麻豆成人av免费视频| 秋霞伦理黄片| 国产三级中文精品| 国产亚洲最大av| 一级黄片播放器| 亚洲成av人片在线播放无| 国产三级在线视频| 国产美女午夜福利| 日本一本二区三区精品| 亚洲经典国产精华液单| 国产精品一及| 中文字幕亚洲精品专区| 久久国产乱子免费精品| 成年免费大片在线观看| 国产精品国产三级专区第一集| 亚洲四区av| 国产毛片a区久久久久| 乱人视频在线观看| 免费av观看视频| 久久精品影院6| 精品人妻一区二区三区麻豆| 亚洲怡红院男人天堂| 超碰97精品在线观看| 91午夜精品亚洲一区二区三区| 国产亚洲av嫩草精品影院| 亚洲成色77777| 免费观看性生交大片5| 中文字幕av在线有码专区| 国产成人精品婷婷| 成人综合一区亚洲| 亚洲精品色激情综合| 蜜桃久久精品国产亚洲av| 国产真实伦视频高清在线观看| 能在线免费观看的黄片| ponron亚洲| 亚洲四区av| 美女脱内裤让男人舔精品视频| 欧美性感艳星| 国产精品久久电影中文字幕| 国语自产精品视频在线第100页| 99热网站在线观看| 99在线视频只有这里精品首页| 人人妻人人看人人澡| 可以在线观看毛片的网站| 欧美丝袜亚洲另类| 搞女人的毛片| 亚洲一级一片aⅴ在线观看| 禁无遮挡网站| 国产激情偷乱视频一区二区| 国产一区二区在线观看日韩| 亚洲婷婷狠狠爱综合网| 国产伦精品一区二区三区四那| 午夜福利高清视频| 综合色丁香网| 可以在线观看毛片的网站| 男女下面进入的视频免费午夜| 国产精品久久视频播放| 国产伦精品一区二区三区四那| 男女下面进入的视频免费午夜| 亚洲欧美中文字幕日韩二区| 看免费成人av毛片| 精品久久国产蜜桃| 免费黄网站久久成人精品| 身体一侧抽搐| 岛国在线免费视频观看| 午夜精品一区二区三区免费看| 日本色播在线视频| 欧美成人一区二区免费高清观看| 精品久久久噜噜| 91av网一区二区| 午夜老司机福利剧场| 老女人水多毛片| 纵有疾风起免费观看全集完整版 | 国产色婷婷99| 国产熟女欧美一区二区| 国产 一区精品| 丰满人妻一区二区三区视频av| 亚洲欧美日韩高清专用| 欧美xxxx性猛交bbbb| 日韩在线高清观看一区二区三区| 九九在线视频观看精品| 亚洲伊人久久精品综合 | 国产成人aa在线观看| 91精品一卡2卡3卡4卡| 内射极品少妇av片p| 国产 一区精品| 免费大片18禁| 久热久热在线精品观看| 久久久久久久久久成人| 91aial.com中文字幕在线观看| 18禁在线播放成人免费| 青春草国产在线视频| 久久精品91蜜桃| 综合色丁香网| 国产伦精品一区二区三区四那| 国产精品久久久久久精品电影| 日韩一区二区三区影片| 最近中文字幕2019免费版| 日本猛色少妇xxxxx猛交久久| 人妻系列 视频| av福利片在线观看| 日韩强制内射视频| 高清毛片免费看| 欧美一区二区精品小视频在线| 狂野欧美激情性xxxx在线观看| 蜜桃亚洲精品一区二区三区| 永久网站在线| 亚洲av电影在线观看一区二区三区 | 男女视频在线观看网站免费| 精品久久久噜噜| 精品久久久久久久久亚洲| 欧美xxxx性猛交bbbb| 国产精品乱码一区二三区的特点| 亚洲内射少妇av| 久热久热在线精品观看| 1000部很黄的大片| 观看美女的网站| 秋霞伦理黄片| 男的添女的下面高潮视频| 成人漫画全彩无遮挡| 欧美成人精品欧美一级黄| ponron亚洲| 国语对白做爰xxxⅹ性视频网站| 亚洲无线观看免费| 亚洲伊人久久精品综合 | 赤兔流量卡办理| 国产成人91sexporn| 免费在线观看成人毛片| 少妇高潮的动态图| 中国国产av一级| 老师上课跳d突然被开到最大视频| 亚洲精品乱码久久久久久按摩| 久久精品久久精品一区二区三区| 国产精品蜜桃在线观看| 日本免费一区二区三区高清不卡| 欧美一区二区国产精品久久精品| 啦啦啦啦在线视频资源| 国产精品野战在线观看| 91久久精品电影网| 99九九线精品视频在线观看视频| 成人二区视频| 两性午夜刺激爽爽歪歪视频在线观看| 久久99热这里只有精品18| av女优亚洲男人天堂| 成人高潮视频无遮挡免费网站| 亚洲av.av天堂| 99热这里只有是精品在线观看| 波野结衣二区三区在线| 国产男人的电影天堂91| 亚洲欧美精品综合久久99| 一区二区三区高清视频在线| 99国产精品一区二区蜜桃av| 欧美变态另类bdsm刘玥| 最新中文字幕久久久久| 国产中年淑女户外野战色| 夫妻性生交免费视频一级片| 国产高清国产精品国产三级 | 天天躁夜夜躁狠狠久久av| 免费在线观看成人毛片| 日本免费一区二区三区高清不卡| 亚洲av熟女| 天堂av国产一区二区熟女人妻| 中文欧美无线码| 成人漫画全彩无遮挡| 亚洲国产精品国产精品| 国产成人精品婷婷| 色5月婷婷丁香| 成人午夜高清在线视频| 欧美日韩一区二区视频在线观看视频在线 | 小说图片视频综合网站| 成人av在线播放网站| 国产av在哪里看| 欧美成人免费av一区二区三区| 美女国产视频在线观看| 一级毛片aaaaaa免费看小| 中国国产av一级| 国产精品一区二区三区四区免费观看| 精品免费久久久久久久清纯| 国产精品电影一区二区三区| 国产一区二区在线观看日韩| 国产视频首页在线观看| 久久久久网色| 国产午夜精品一二区理论片| 精品一区二区免费观看| 亚洲精品456在线播放app| 欧美日韩综合久久久久久| 国产亚洲5aaaaa淫片| 成人特级av手机在线观看| 免费看日本二区| 久久99精品国语久久久| 高清毛片免费看| 亚州av有码| 天堂av国产一区二区熟女人妻| 亚洲国产欧美人成| 欧美日韩国产亚洲二区| 国模一区二区三区四区视频| 天天躁夜夜躁狠狠久久av| 久久99热6这里只有精品| 亚洲欧洲国产日韩| 精品久久久噜噜| 嫩草影院精品99| 少妇熟女aⅴ在线视频| 一边摸一边抽搐一进一小说| 大话2 男鬼变身卡| 精品无人区乱码1区二区| av免费观看日本| 级片在线观看| 狂野欧美激情性xxxx在线观看| 国产精品乱码一区二三区的特点| av国产久精品久网站免费入址| 国产一级毛片七仙女欲春2| 国产乱来视频区| 久久久久精品久久久久真实原创| 亚洲三级黄色毛片| 91在线精品国自产拍蜜月| 夜夜看夜夜爽夜夜摸| 韩国av在线不卡| 亚洲美女视频黄频| 99久久精品一区二区三区| 狠狠狠狠99中文字幕| 国产熟女欧美一区二区| a级毛色黄片| www.色视频.com| 国产在线男女| 天天躁日日操中文字幕| 中文字幕av成人在线电影| 亚洲av电影不卡..在线观看| 少妇的逼好多水| 日本熟妇午夜| 欧美色视频一区免费| 91av网一区二区| 国产精品麻豆人妻色哟哟久久 | 1024手机看黄色片| 日韩欧美精品v在线| 国产 一区 欧美 日韩| 中国国产av一级| 小说图片视频综合网站| 国产乱来视频区| 免费av毛片视频| 久久久欧美国产精品| 亚洲婷婷狠狠爱综合网| 午夜日本视频在线| 高清在线视频一区二区三区 | 日本一本二区三区精品| 国产精品女同一区二区软件| 精品国产一区二区三区久久久樱花 | 天堂网av新在线| 午夜亚洲福利在线播放| 国内精品美女久久久久久| 亚洲精品456在线播放app| 亚洲自偷自拍三级| 日本熟妇午夜| 亚洲国产色片| 免费播放大片免费观看视频在线观看 | 大话2 男鬼变身卡| 久热久热在线精品观看| 天堂√8在线中文| 久久鲁丝午夜福利片| 99在线视频只有这里精品首页| 精品久久久久久久久亚洲| 99热精品在线国产| 91精品伊人久久大香线蕉| 久久久亚洲精品成人影院| 别揉我奶头 嗯啊视频| 久久久久久大精品| 精品国产一区二区三区久久久樱花 | 国产在线男女| 好男人视频免费观看在线| 日日啪夜夜撸| 高清在线视频一区二区三区 | 欧美潮喷喷水| 性插视频无遮挡在线免费观看| 亚洲精品自拍成人| 又爽又黄a免费视频| 欧美性感艳星| 寂寞人妻少妇视频99o| 特大巨黑吊av在线直播| 国产精品三级大全| 看十八女毛片水多多多| 秋霞在线观看毛片| 在线免费观看的www视频| 熟妇人妻久久中文字幕3abv| 99在线视频只有这里精品首页| 男女下面进入的视频免费午夜| 亚洲av中文字字幕乱码综合| 丰满少妇做爰视频| 91aial.com中文字幕在线观看| 国产日韩欧美在线精品| 女人十人毛片免费观看3o分钟| 国产乱人视频| 乱码一卡2卡4卡精品| 午夜亚洲福利在线播放| 边亲边吃奶的免费视频| 别揉我奶头 嗯啊视频| 青春草亚洲视频在线观看| 99视频精品全部免费 在线| 又爽又黄无遮挡网站| 国产精品一区二区三区四区久久| 国产精品国产高清国产av| 午夜激情欧美在线| 亚洲综合色惰| 一个人看视频在线观看www免费| 2021少妇久久久久久久久久久| www日本黄色视频网| 小蜜桃在线观看免费完整版高清| 性色avwww在线观看| 97人妻精品一区二区三区麻豆| 欧美97在线视频| 欧美激情在线99| 男的添女的下面高潮视频| 在线播放国产精品三级| 观看美女的网站| 久久热精品热| 最近最新中文字幕大全电影3| 日日干狠狠操夜夜爽| 日韩精品青青久久久久久| 又黄又爽又刺激的免费视频.| 亚洲欧美日韩卡通动漫| 亚洲精品日韩在线中文字幕| 一级爰片在线观看| 亚洲av熟女| 神马国产精品三级电影在线观看| 国产综合懂色| 欧美激情在线99| 汤姆久久久久久久影院中文字幕 | 高清视频免费观看一区二区 | 啦啦啦韩国在线观看视频| 欧美xxxx黑人xx丫x性爽| 亚洲国产精品sss在线观看| 亚洲av中文字字幕乱码综合|