CN103036974B - Cloud computing resource scheduling method based on hidden Markov model and system - Google Patents

Abstract

The present invention relates to cloud computing correlative technology field, particularly relate to a kind of cloud computing resource scheduling method based on hidden Markov model and system, described method includes: model training step, and hidden Markov model is trained the model parameter obtaining being correlated with；Portfolio detection steps, carries out detection to the portfolio of cloud computing resources and obtains portfolio observation；Traffic prediction step, portfolio observation detection obtained inputs described hidden Markov model and carries out state computation, obtains predicted state, and described hidden Markov model uses the model parameter that model training step obtains；Scheduling of resource step, is scheduling cloud computing resources according to described predicted state.Owing to portfolio being predicted by hidden Markov model (Hidden Markov Model, HMM), thus carry out cloud computing resources scheduling according to the result of prediction, it is possible to carry out forward scheduling, can more effectively utilize resource.

Description

Cloud computing resource scheduling method based on hidden Markov model and system

Technical field

The present invention relates to cloud computing correlative technology field, particularly relate to a kind of cloud computing based on hidden Markov model Resource regulating method and system.

Background technology

Scheduling of resource is the key problem of cloud computing: the introducing of the technology such as cloud computing virtualization so that dispersion originally, free time Resource form unified, standardized virtual resource so that resource is allocated flexibly and is possibly realized；Resource optimization allotment can make The operation system being deployed on cloud computing resource pool can use resource according to the size of portfolio, can meet operation system Need, be avoided that the wasting of resources simultaneously.

Current resource regulating method is often conceived to virtual machine Optimization deployment on physical server, mainly solves It is the mapping the most reasonably determining virtual machine to node, is often viewed as a bin packing, i.e. find optimum by void Plan machine is assigned to the scheme of node, so that the use resource sum of virtual machine is less than what node can be provided by each node The upper limit, and the number of nodes used is optimum.The scholar such as Buyya propose based on economic model resource regulating method, the method carries Having gone out market-oriented cloud computing architecture and market-oriented Resource Distribution and Schedule method, this architecture passes through SLA Resource allocator realizes the negotiation between resource user and resource provider, it is achieved resources configuration optimization.Li Qiangs etc. propose The virtual machine of band application service level goal constraint places multiple-objection optimization genetic resources dispatching method etc..Menaud and Van etc. People is proposed for the management of virtual resource in cloud computing and proposes dynamic dispatching method, with overall scheduling time as target, by examining Consider reconfiguration time and virtual machine (vm) migration time, provide a kind of resource customary ways Entropy.Wei Guiyi etc. propose based on rich Playing chess opinion resource Resources allocation dispatching method and solve cloud computing resources assignment problem, this resource regulating method is for first with integer The resource requirement of single operation system is optimized by planing method, then uses evolution resource regulating method to solve multiple operation systems Complex optimization problem.Xu Xianghua et al. proposes a kind of cloud computing resources allocation strategy based on market mechanism, and designs one Individual demand and the equilibrium problem of supply processing market based on genetic price adjustment resource regulating method.Lin Weiwei et al. Propose a kind of cloud computing resource scheduling method based on dynamic reconfiguration virtual resources.The method is collected with cloud application monitor Cloud application load information be foundation, be then based on running the load capacity of virtual resource of cloud application and current negative of cloud application It is loaded into Mobile state decision-making, is cloud application dynamic reconfiguration virtual resources according to the result of decision-making.By reconfiguring void for cloud application The method intending resource realizes the dynamic adjustment of resource, it is not necessary to dynamically redistribute physical resource and stopping cloud application performs. Beloglazov etc. propose a kind of resource regulating method based on minimum power consumption, by the energy of the server before and after computation migration Consumption is optimized scheduling.

Above-mentioned resource regulating method major part belongs to static resource regulating method, i.e. before portfolio determines substantially Put, calculate the resources of virtual machine needed for portfolio, carry out resources of virtual machine and physical resource by optimizing resource regulating method Binding.In actual production environment, portfolio is dynamically change, and these resource regulating methods are often based upon portfolio Under big scene, deploying virtual machine is optimized.Therefore, in the case of most of portfolios are the most little, this type of resource is adjusted Degree method can cause virtual resource to leave unused, and can cause certain wasting of resources.The method that Lin Weiwei et al. and Li Qiang et al. propose Belonging to dynamic resource regulating method, the former, by detecting the utilization rate of current resource node, exceedes set threshold value Carry out resource increase and decrease；The latter has carried out particular study to WEB flow, obtains the process threshold value of WEB flow, and to virtual machine Monitoring performance, trigger resource mobilization when exceeding threshold value.Both resource regulating methods broadly fall into the money of passive-type Source dispatching method, i.e. can only be scheduling by virtual machine real resource consumption, can lead portfolio is busy when Cause scheduling and lag behind business demand so that business demand cannot timely respond to.

The said method overwhelming majority is all based on current resources situation and is scheduling definitiveness resource, and in cloud computing environment, industry Business measurer has certain uncertainty, detection portfolio to be scheduling frequently can lead to resource short-term wretched insufficiency or waste again.

Summary of the invention

Based on this, it is necessary to carry out the meter of scheduling of resource in cloud computing resources according to prediction portfolio for prior art Calculation problem, it is provided that a kind of cloud computing resource scheduling method based on hidden Markov model.

A kind of cloud computing resource scheduling method based on hidden Markov model, uses hidden Markov model to portfolio Situation of change is described and predicts, described method includes:

Model training step, is trained the model parameter obtaining being correlated with to hidden Markov model；

Portfolio detection steps, carries out detection to the portfolio of cloud computing resources and obtains portfolio observation；

Traffic prediction step, portfolio observation detection obtained inputs described hidden Markov model and carries out state Calculating, obtain predicted state, described hidden Markov model uses the model parameter that model training step obtains；

Scheduling of resource step, is scheduling cloud computing resources according to described predicted state.

Wherein in an embodiment, the portfolio of described cloud computing resources is network traffics parameter.

Wherein in an embodiment, described Traffic prediction step includes:

As a time slot, portfolio being carried out record every the t time, every N number of time slot constitutes an observation window, often Individual observation window carries out data process and obtains portfolio observation, and wherein, t is more than 0, and N is the positive integer more than or equal to 1；

Carry out data at each observation window to process and obtain portfolio observation and specifically include:

By i-th observation window T_iJth time slot be designated as t_i,j, corresponding traffic log value is x_{I, j}, wherein 1≤j ≤ N, calculates T_iThe meansigma methods of interior all record valuesWherein, 1≤j≤N；

Calculate i-th observation window T_iRecord value meansigma methods X_iWith the i-th-1 observation window T_i-1Record value meansigma methods X_i-1Difference, as i-th observation window T_iPortfolio observation the first parameter o_i,1=X_i-X_i-1；

Calculate i-th observation window T_iContrast C ON of interior all record values_i=∑_m,n|m-n|p_i,mn, as i-th Observation window T_iPortfolio observation the second parameter estimator value o_i,2, wherein, min (x_i,j)≤m,n≤max(x_i,j), 1≤j≤ N, p_i,mnIt is i-th observation window T_iThe situation that jth record value is m and+1 record value of jth is n at i-th observation window T_iThe frequency of middle appearance, p_i,mn=# (x_i,j=m, x_i,j+1=n)/(T-1), and wherein, min (x_i,j)≤m,n≤max(x_i,j)；

To o_i,2And o_i,2Carry out being converted to the portfolio observation of correspondence.

Wherein in an embodiment:

Described model parameter includes: state set, observation value set, state transition probability matrix, observation produce Probability matrix and initial probability distribution matrix；

Described state set S={s₁,s₂,s₃, wherein s₁Represent that portfolio increases state, s₂Represent that portfolio maintains an equal level shape State, s₃Represent that portfolio reduces state；

Described observation value set V={v₁,v₂,…v_K},K≥1；

Described state transition probability matrix A={a_mn, wherein a_mnRepresent from state s_mJump to state s_nProbability；

Described observation produces probability matrix B={b_m(k) }, wherein b_mK () represents in state s_mLower generation observation v_k's Probability；

Described initial probability distribution matrix π={ π_m, wherein π_mExpression original state is s_mProbability；

Described Traffic prediction step specifically includes:

Calculate the e observation window T_ePredicted state q_e:

Wherein, α_e(i)α_eIt is the e observation window T_eI-th forward variable, and:Wherein, o_eIt is the e observation window T_eTraffic prediction value.

Wherein in an embodiment, described scheduling of resource step specifically includes:

If predicted state is portfolio reduces state, then reclaim cloud computing resources；

If predicted state is portfolio increases state, then increase cloud computing resources；

Maintain an equal level state if predicted state is portfolio, then maintain cloud computing resources.

Wherein, described cloud computing resources can be the resources of virtual machine in cloud computing system.

A kind of cloud computing resources dispatching patcher based on hidden Markov model, including:

Model training module, for being trained the model parameter obtaining being correlated with to hidden Markov model；

Portfolio detecting module, obtains portfolio observation, concurrently for the portfolio of cloud computing resources is carried out detection Deliver to Traffic prediction module；

Traffic prediction module, for inputting described hidden horse by the portfolio observation that the detection of portfolio detecting module obtains Er Kefu model carries out state computation, obtains predicted state and is sent to scheduling of resource module, and described hidden Markov model is adopted The model parameter obtained with model training module；

Scheduling of resource module, for adjusting cloud computing resources according to the predicted state obtained from portfolio prediction module Degree.

Wherein in an embodiment:

Described portfolio detecting module includes:

Record sub module, for carrying out record as a time slot to portfolio every the t time；

Portfolio observation calculating sub module, constitutes an observation window for every N number of time slot, enters at each observation window Row data process and obtain portfolio observation, specifically for:

By i-th observation window T_iJth time slot be designated as t_i,j, corresponding traffic log value is x_{I, j}, wherein 1≤j ≤ N, calculates T_iThe meansigma methods of interior all record valuesWherein, 1≤j≤N；

Calculate i-th observation window T_iRecord value meansigma methods X_iWith the i-th-1 observation window T_i-1Record value meansigma methods X_i-1Difference, as i-th observation window T_iPortfolio observation the first parameter o_i,1=X_i-X_i-1；

Calculate i-th observation window T_iContrast C ON of interior all record values_i=Σ_m,n|m-n|p_i,mn, as i-th Observation window T_iPortfolio observation the second parameter estimator value o_i,2, wherein, min (x_i,j)≤m,n≤max(x_i,j), 1≤j≤ N, p_i,mnIt is i-th observation window T_iThe situation that jth record value is m and+1 record value of jth is n at i-th observation window T_iThe frequency of middle appearance, p_i,mn=# (x_i,j=m, x_i,j+1=n)/(N-1), and wherein, min (x_i,j)≤m,n≤max(x_i,j)；

To o_i,2And o_i,2Carry out being converted to the portfolio observation of correspondence and being sent to Traffic prediction module.

Wherein in an embodiment:

Described model parameter includes: state set, observation value set, state transition probability matrix, observation produce Probability matrix and initial probability distribution matrix；

Described state set S={s₁,s₂,s₃, wherein s₁Represent that portfolio increases state, s₂Represent that portfolio maintains an equal level shape State, s₃Represent that portfolio reduces state；

Described observation value set V={v₁,v₂,…v_K},K≥1；

Described state transition probability matrix A={a_mn, wherein a_mnRepresent from state s_mJump to state s_nProbability；

Described observation produces probability matrix B={b_m(k) }, wherein b_mK () represents in state s_mLower generation observation v_k's Probability；

Described initial probability distribution matrix π={ π_m, wherein π_mExpression original state is s_mProbability；

Described Traffic prediction module specifically for:

Calculate the e observation window T_ePredicted state q_e:

Wherein, α_e(i)α_eIt is the e observation window T_eI-th forward variable, and:Wherein, o_eIt is the e observation window T_eTraffic prediction value.

Wherein in an embodiment, described scheduling of resource module includes:

Portfolio reduces scheduling sublayer module, is that portfolio reduces state for predicted state, reclaims cloud computing resources；

Portfolio increases scheduling sublayer module, is that portfolio increases state for predicted state, increases cloud computing resources；

Portfolio maintains an equal level scheduling sublayer module, is that portfolio maintains an equal level state for predicted state, maintains cloud computing resources.

The commonly used monitor mode to existing resource of current technology is scheduling, and can only accomplish resource deficiency In the case of carry out intervention schedule.Above-mentioned a kind of cloud computing resource scheduling method based on hidden Markov model, due to by hidden Portfolio is predicted by Markov model (Hidden Markov Model, HMM), thus carries out according to the result of prediction Cloud computing resources is dispatched, it is possible to carries out forward scheduling, can more effectively utilize resource.

Accompanying drawing explanation

Fig. 1 is the flow chart of the cloud computing resource scheduling method based on hidden Markov model of the present invention；

Fig. 2 is the block diagram of the cloud computing resources dispatching patcher based on hidden Markov model of the present invention；

Fig. 3 is the portfolio detection time segmentation schematic diagram of the present invention；

Fig. 4 is that the contrast of the present invention calculates schematic diagram；

Fig. 5 is the flow chart of the resource regulating method of the present invention；

Fig. 6 is the first quantization observation and the code conversion figure of the second quantization observation of the present invention.

Detailed description of the invention

The present invention will be further described in detail with specific embodiment below in conjunction with the accompanying drawings.

The flow chart of the cloud computing resource scheduling method based on hidden Markov model of the present invention is shown such as Fig. 1.

A kind of cloud computing resource scheduling method based on hidden Markov model, described method includes:

Step S101, is trained the model parameter obtaining being correlated with to hidden Markov model；

Step S102, carries out detection to the portfolio of cloud computing resources and obtains portfolio observation；

Step S103, portfolio observation detection obtained inputs described hidden Markov model and carries out state computation, Obtaining predicted state, described hidden Markov model uses the model parameter that model training step obtains；

Step S104, is scheduling cloud computing resources according to described predicted state.

In the HMM of the present embodiment, state set S has 3 states, wherein s₁Represent that portfolio increases state, s₂Represent Portfolio maintains an equal level state, s₃Represent that portfolio reduces state.

For the training of HMM, carry out to algorithm before and after employing, specific as follows:

Before and after HMM to algorithm computation complexity be O (M²T), wherein T is the length of training and forecasting sequence.

Definition forward variable and computational methods equation below 1 thereof:

Wherein, λ represents and works as front mould Type；

Backward variable-definition and be calculated as follows formula 2:

${\mathrm{\β}}_t\left(i\right)\stackrel{\mathrm{\Δ}}{=}P\left(o_{t+1}^T\right|q_t=s_i,\mathrm{\λ})=\left\{\begin{array}{cc}1,& t=T,1\≤i\≤M\\ \underset{j=1}{\overset{M}{\Σ}}{a}_{ij}{\mathrm{\β}}_{t+1}\left(j\right)b_i\left(o_{t+1}\right),& t\≠T,1\≤i\≤M\end{array}\right.;$

Definition intermediate variable γ_t(i) be the t state be s_iProbability, be calculated as follows formula 3:

${\mathrm{\γ}}_t\left(i\right)\stackrel{\mathrm{\Δ}}{=}P(q_t=s_i|o_1^T,\mathrm{\λ})=\frac{{\mathrm{\α}}_t\left(i\right){\mathrm{\β}}_t\left(i\right)}{{\Σ}_{i=1}^M{\mathrm{\α}}_t\left(i\right){\mathrm{\β}}_t\left(i\right)};$

Definition intermediate variable ξ_t(i is j) from state s in t_iJump to s_jProbability, be calculated as follows formula 4:

${\mathrm{\ξ}}_t(i,j)\stackrel{\mathrm{\Δ}}{=}P(q_t=s_i,q_{t+1}=s_j|o_1^T,\mathrm{\λ})=\frac{{\mathrm{\α}}_t\left(i\right)a_{ij}{b}_j\left(o_{t+1}\right){\mathrm{\β}}_{t+1}\left(j\right)}{{\Σ}_{i=1}^M{\Σ}_{j=1}^M{\mathrm{\α}}_t\left(i\right)a_{ij}{b}_j\left(o_{t+1}\right){\mathrm{\β}}_{t+1}\left(j\right)}.$

In evaluation problem, by the probability solution equation below 5 of the given observation sequence of this model generation:

$P\left(O\right|\mathrm{\λ})={\Σ}_{i=1}^M{\mathrm{\α}}_T\left(i\right);$

In decoding problem, most probable produces the status switch Q=q of given observation sequence₁q₂…q_TIn each state permissible Obtained by equation below 6:

$q_t=\underset{1\≤i\≤M}{\mathrm{argmax}}\[{\mathrm{\γ}}_t\left(i\right)\],1\≤t\≤T.$

The model parameter mated most with observation sequence, can be estimated by Baum-Welch algorithm, first by sight Order-checking rowIntermediate variable γ is calculated with "current" model parameter Ω=(A, B, π)_t(i) and ξ_t(i, j), then by formula (3.7- 9) model parameter updated is obtainedCalculate "current" modelProduce the probability of given observation sequenceAs Fruit does not has significant change, i.e. compared with model beforeFor threshold value), then it is assumed that receive Hold back model parameter Ω=(A, B, π) and be the parameter mating most this observation sequence；Otherwise, orderRepeat said process, directly To convergence.Model parameter calculation during convergence is as follows:

${\widehat{\mathrm{\π}}}_m=\frac{{\mathrm{\γ}}_1\left(m\right)}{{\Σ}_m{\mathrm{\γ}}_1\left(m\right)},{\hat{a}}_{ij}=\frac{\underset{t=1}{\overset{T-1}{\Σ}}{\mathrm{\ξ}}_t(i,j)}{\underset{t=1}{\overset{T-1}{\Σ}}{\mathrm{\γ}}_t\left(i\right)},{\hat{b}}_i\left(k\right)=\frac{\underset{s.t.o_t=v_k}{\overset{T}{\underset{t=1,}{\Σ}}}{\mathrm{\γ}}_t\left(i\right)}{\underset{t=1}{\overset{T-1}{\Σ}}{\mathrm{\γ}}_t\left(i\right)}.$

In step s 102, the portfolio to cloud computing resources detects that to obtain the mode of portfolio observation as follows: The description of portfolio can pass through cloud by network traffics parameter (include access bandwidth, request number, connect the parameters such as number) Calculate the load-balancing device on resource pool uniform outlet, firewall box is acquired.

As it is shown on figure 3, at system run duration, portfolio is carried out record every the t time as a time slot, the most N number of Time slot constitutes an observation window, carries out data process at each observation window and obtains portfolio observation, and wherein, t is more than 0, N For the positive integer more than or equal to 1；

Carry out data at each observation window to process and obtain portfolio observation and specifically include:

1) by i-th observation window T_iJth time slot be designated as t_i,j, corresponding traffic log value is x_{I, j}, wherein 1≤ J≤T, works as T_iAfter the record value of all time slots all records, calculate T_iThe meansigma methods of interior all record valuesWherein, 1≤j≤N；

2) i-th observation window T is calculated_iRecord value meansigma methods X_iWith the i-th-1 observation window T_i-1Record value average Value X_i-1Difference (X₀It is set to 0), as i-th observation window T_iPortfolio observation the first parameter o_i,1=X_i-X_i-1；

3) i-th observation window T is calculated_iContrast C ON of interior all record values_i=∑_m,n|m-n|p_i,mn, as i-th Individual observation window T_iPortfolio observation the second parameter estimator value o_i,2=CON_T=∑_m,n|m-n|p_i,mn, wherein, min (x_i,j) ≤m,n≤max(x_i,j), 1≤j≤N, p_i,mnIt is i-th observation window T_iJth record value be m and+1 record value of jth is The situation of n is at i-th observation window T_iThe frequency of middle appearance, p_i,mn=# (x_i,j=m, x_i,j+1=n)/(N-1), and wherein, min (x_i,j)≤m,n≤max(x_i,j)；

Fig. 4 lists one and calculates p_i,mnMethod, once there is the observation window T of 10 time slots_iIn, obtain such as figure The traffic log value on the left side, then:

Insert in the table on the right side of figure.

4), after being disposed, observation O of this observation window is obtained_i=(o_i,1,o_i,2), record Oi and record value meansigma methods X_i, due to observation O_iIt is a two-dimentional variable, needs to quantify in record value to the transformation process of observation, the most right o_i,2And o_i,2Carry out being converted to the portfolio observation of correspondence.To o_i,2And o_i,2Carry out the mode changed, the common skill in this area Art personnel can use various ways to change.In the present embodiment, following conversion regime is used: for the first parameter estimator Value o_i,1, predetermined minimum is minO₁(actual appearance is less than minO₁Observation think and be equivalent to minO₁), maximum is maxO₁(actual appearance is more than maxO₁Observation think and be equivalent to maxO₁) to interval [minO₁, maxO₁] it is equally divided into length For L₁M₁Decile, for falling at [minO₁, minO₁+L₁] the first parameter estimator value o_i,1, it is believed that its corresponding first quantization observation Value is v_1,1, for falling at [minO₁+L₁* (k-1), minO₁+L₁* k] the first parameter estimator value o_i,1, it is believed that its correspondence first Quantization observation is v_1,k, wherein 2≤k≤M₁.In like manner, the second parameter estimator value o can be preset_i,2Minima be minO₂, Big value is maxO₂, to interval [minO₂, maxO₂] it is equally divided into a length of L₂M₂Decile, in the same way to the second parameter Observation o_i,2Quantify.Carrying out code conversion as shown in Figure 6 after quantization, observation is as shown in the numerical value in Fig. 6 medium square. By coding can by and be the variable portfolio observation that is converted to one-dimensional variable.Such as: in Fig. 6, abscissa is that the first parameter is seen Measured value o_i,1Quantization scale, vertical coordinate is the second parameter estimator value o_i,2Quantization scale, if then the first parameter estimator value o_i,1 At (minO₁+1*L₁, minO₁+2*L₁In), and the second parameter estimator value o_i,2At (minO₂+1*L₁, minO₂+2*L₁In), the most right The portfolio observation answered is M₁+2。

Step S102 of the present embodiment, it is also possible to use additive method to realize, the most do not carry out time slot dividing, or do not adopt Being observed value by modes such as watch windows to gather, those skilled in the art, after reading technical scheme, have After limit time test, also can obtain similar method.

And the present embodiment has carried out time slot dividing, and use observation window carry out data calculating mode, be due to often Individual time slot is smaller, if each time slot once calculates, then data volume is the hugest, therefore, it is divided into observation Window, calculates in each observation window, uses average and contrast to carry out traffic characteristic description, can reflect that flow becomes Changing general trend, also can embody the severe degree of changes in flow rate in certain time, the calculating of average and contrast all can be by one Secondary forward direction iterative manner realizes, and i.e. ensure that the accuracy of data, it also avoid googol and calculates according to amount.

In step s 103, Traffic prediction uses the HMM trained and the observation of portfolio detecting module transmission Value carries out state computation.According to formula 6, the e observation window T_e, predictive value O_eLast time point corresponding, corresponding prediction State computation is as follows:

$q_e=\underset{1\≤i\≤M}{\arg \max }\[{\mathrm{\γ}}_t\left(i\right)\]=\arg \max \[\frac{{\mathrm{\α}}_e\left(i\right){\mathrm{\β}}_e\left(i\right)}{\underset{1\≤i\≤M}{{\Σ}_{i=1}^M}{\mathrm{\α}}_e\left(i\right){\mathrm{\β}}_e\left(i\right)}\]\arg nax\[\frac{{\mathrm{\α}}_e\left(i\right)}{\underset{1\≤i\≤M}{{\Σ}_{i=1}^M}{\mathrm{\α}}_e\left(i\right)}\].$

Therefore it may only be necessary to calculate forward variable.T in system start-up₀Moment starts to calculate forward variable, the most often Individual observation window terminates all to update by iterative manner forward variable.The forward variable of e observation window is calculated as follows:

Wherein, o_eIt is the e observation window T_ePortfolio Predictive value.

Therefore, by comparing the forward variable value of each state, current operation amount state in which can be obtained, correspondence It is portfolio increase, fair, the state of minimizing.

In step S104, resource regulating method based on prediction is as it is shown in figure 5, include:

Step S510, comparison prediction state and current state, process accordingly according to comparative result:

Reduce state (representing that portfolio is on a declining curve) when being in portfolio, perform step S520；

Maintain an equal level state (represent portfolio keep constant) when system is in portfolio, do not make any process, perform step S530；

Reduce state (representing that portfolio is in rising trend) when system is in portfolio, perform step S540；

Step S520, it may be judged whether have multiple stage virtual machine, if it has, then perform step S521, otherwise performs step S530；

Step S521, closes 1 virtual machine, recovery section resource, terminates scheduling；

Step S530, maintains the statusquo, and terminates scheduling；

Step S540, it is judged that existing resource is enough, if enough, performs step S541, otherwise performs step S542；

Step S541, starts virtual machine, terminates scheduling；

Step S542, maintains the statusquo, and alerts, and terminates scheduling.

Owing to system uses HMM to carry out data prediction, therefore can do sth. in advance the budget trend to resource, thus by increase or Reduce virtual machine cloud computing resources is scheduling.

The square frame signal of the cloud computing resources dispatching patcher of based on hidden Markov model of the present invention as shown in Figure 2 Figure.

System uses the HMM of a three condition to be predicted portfolio, three state the most corresponding portfolio increases, industry Business amount maintains an equal level and portfolio reduces.In the training stage, it is trained obtaining a stable forecast model by historical data, The traffic forecast stage, the data genaration observation gathered by existing network, judge the shape at operation system according to forecast model State, determines the need for increasing and decreasing virtual machine according to predicted state.

A kind of cloud computing resources dispatching patcher 200 based on hidden Markov model, including:

Model training module 210, for being trained the model parameter obtaining being correlated with, this module to hidden Markov model History according to actual volume gathers data, is trained model, obtains stable model so that model can preferably Joining portfolio, the system free time when, model is updated by working model training module；

Portfolio detecting module 220, obtains portfolio observation for the portfolio of cloud computing resources is carried out detection, and Being sent to Traffic prediction module, this module is responsible for collecting the portfolio between external user 25 and resource node 26, to business Amount data carry out pretreatment and form observation；

Traffic prediction module 230, for inputting institute by the portfolio observation that portfolio detecting module 220 detection obtains State hidden Markov model and carry out state computation, obtain predicted state and be sent to scheduling of resource module, described hidden Markov Model uses the model parameter that model training module 210 obtains, and this module is responsible for being analyzed observation, according to forecast model The portfolio of next stage is predicted；

Scheduling of resource module 240, for entering cloud computing resources according to the predicted state obtained from portfolio prediction module Row scheduling, this module according to Traffic prediction module 230 predict the outcome and resource is suitable for present situation and is scheduling, specifically include new Increase virtual machine, reduce virtual machine, and virtual machine (vm) migration etc..

Portfolio detecting module 220 includes:

Record sub module 221, for carrying out record as a time slot to portfolio every the t time；

Portfolio observation calculating sub module 222, constitutes an observation window for every N number of time slot, at each observation window Mouth carries out data process and obtains portfolio observation, specifically for:

By i-th observation window T_iJth time slot be designated as t_i,j, corresponding traffic log value is x_{I, j}, wherein 1≤j ≤ N, calculates T_iThe meansigma methods of interior all record valuesWherein, 1≤j≤N；

Calculate i-th observation window T_iRecord value meansigma methods X_iWith the i-th-1 observation window T_i-1Record value meansigma methods X_i-1Difference, as i-th observation window T_iPortfolio observation the first parameter o_i,1=X_i-X_i-1；

Calculate i-th observation window T_iContrast o of interior all record values_i,2=CON_T=∑_m,n|m-n|p_i,mn, as I-th observation window T_iPortfolio observation the second parameter estimator value o_i,2, wherein, min (x_i,j)≤m,n≤max(x_i,j), 1 ≤ j≤T, p_i,mnIt is i-th observation window T_iThe situation that jth record value is m and+1 record value of jth is n i-th see Survey window T_iThe frequency of middle appearance, p_i,mn=# (x_i,j=m, x_i,j+1=n)/(T-1), and wherein, min (x_i,j)≤m,n≤max (x_i,j)；

To o_i,2And o_i,2Carry out being converted to the portfolio observation of correspondence and being sent to Traffic prediction module 230.

Described Traffic prediction module 230 specifically for:

Calculate the e observation window T_ePredicted state q_e:

Wherein, α_e(i)α_eIt is the e observation window T_eI-th forward variable, and:Wherein, o_eIt is the e observation window T_eTraffic prediction value.

Scheduling of resource module 240 includes:

Portfolio reduces scheduling sublayer module 241, is that portfolio reduces state for predicted state, reclaims cloud computing resources；

Portfolio increases scheduling sublayer module 242, is that portfolio increases state for predicted state, increases cloud computing resources；

Portfolio maintains an equal level scheduling sublayer module 243, is that portfolio maintains an equal level state for predicted state, maintains cloud computing resources.

The present invention utilizes more ripe HMM model, delineates portfolio, and utilizes the trend prediction method of HMM Portfolio it is predicted and dispatches, it is possible to carrying out forward scheduling, can more effectively utilize resource.

Embodiment described above only have expressed the several embodiments of the present invention, and it describes more concrete and detailed, but also Therefore the restriction to the scope of the claims of the present invention can not be interpreted as.It should be pointed out that, for those of ordinary skill in the art For, without departing from the inventive concept of the premise, it is also possible to make some deformation and improvement, these broadly fall into the guarantor of the present invention Protect scope.Therefore, the protection domain of patent of the present invention should be as the criterion with claims.

Claims (8)

1. a cloud computing resource scheduling method based on hidden Markov model, it is characterised in that use hidden Markov mould Portfolio situation of change is described and predicts by type, and described method includes:

When the system free time, hidden Markov model is trained the model parameter obtaining being correlated with；

The portfolio of cloud computing resources is carried out detection and obtains portfolio observation；

Portfolio observation detection obtained inputs described hidden Markov model and carries out state computation, obtains predicted state, Described hidden Markov model uses the model parameter obtained in model training step；

According to described predicted state, cloud computing resources is scheduling.

Cloud computing resource scheduling method based on hidden Markov model the most according to claim 1, it is characterised in that institute The portfolio stating cloud computing resources is network traffics parameter.

Cloud computing resource scheduling method based on hidden Markov model the most according to claim 1, it is characterised in that institute State Traffic prediction step to include:

As a time slot, portfolio being carried out record every the t time, every N number of time slot constitutes an observation window, in each sight Survey window carries out data process and obtains portfolio observation, and wherein, t is more than 0, and N is the positive integer more than or equal to 1.

Cloud computing resource scheduling method based on hidden Markov model the most according to claim 3, it is characterised in that:

Described model parameter includes: state set, observation value set, state transition probability matrix, observation produce probability Matrix and initial probability distribution matrix；

Described state set S={s₁, s₂, s₃, wherein s₁Represent that portfolio increases state, s₂Represent that portfolio maintains an equal level state, s₃ Represent that portfolio reduces state；

Described observation value set V={v₁, v₂..., v_k..., v_K, K >=1, v_kFor kth observation；

Described state transition probability matrix A={a_mn, wherein a_mnRepresent from state s_mJump to state s_nProbability；

Described observation produces probability matrix B={b_m(k) }, wherein b_mK () represents in state s_mLower generation observation v_kProbability；

Described initial probability distribution matrix π={ π_m, wherein π_mExpression original state is s_mProbability；

Described Traffic prediction step specifically includes:

Calculate the e observation window T_ePredicted state q_e:

Wherein, α_eI () is the e observation window T_eI-th forward variable, and:Wherein, o_eIt is the e observation window T_eTraffic prediction value.

Cloud computing resource scheduling method based on hidden Markov model the most according to claim 4, it is characterised in that institute State scheduling of resource step to specifically include:

If predicted state is portfolio reduces state, then reclaim cloud computing resources；

If predicted state is portfolio increases state, then increase cloud computing resources；

Maintain an equal level state if predicted state is portfolio, then maintain cloud computing resources.

6. a cloud computing resources dispatching patcher based on hidden Markov model, it is characterised in that including:

Model training module, for being trained the model parameter obtaining being correlated with when the system free time to hidden Markov model；

Portfolio detecting module, obtains portfolio observation for the portfolio of cloud computing resources is carried out detection, and is sent to Traffic prediction module；

Traffic prediction module, for inputting described hidden Ma Erke by the portfolio observation that the detection of portfolio detecting module obtains Husband's model carries out state computation, obtains predicted state and is sent to scheduling of resource module, and described hidden Markov model uses mould The model parameter that type training module obtains；

Scheduling of resource module, for being scheduling cloud computing resources according to the predicted state obtained from portfolio prediction module.

Cloud computing resources dispatching patcher based on hidden Markov model the most according to claim 6, it is characterised in that:

Described model parameter includes: state set, observation value set, state transition probability matrix, observation produce probability Matrix and initial probability distribution matrix；

Described state set S={s₁, s₂, s₃, wherein s₁Represent that portfolio increases state, s₂Represent that portfolio maintains an equal level state, s₃ Represent that portfolio reduces state；

Described observation value set V={v₁, v₂... v_K, K >=1, v_kFor kth observation；

Described state transition probability matrix A={a_mn, wherein a_mnRepresent from state s_mJump to state s_nProbability；

Described observation produces probability matrix B={b_m(k) }, wherein b_mK () represents in state s_mLower generation observation v_kProbability；

Described initial probability distribution matrix π={ π_m, wherein π_mExpression original state is s_mProbability；

Described Traffic prediction module specifically for:

Calculate the e observation window T_ePredicted state q_e:

Wherein, α_eI () is the e observation window T_eI-th forward variable, and:Wherein, o_eIt is the e observation window T_eTraffic prediction value.

Cloud computing resources dispatching patcher based on hidden Markov model the most according to claim 7, it is characterised in that institute State scheduling of resource module to include:

Portfolio reduces scheduling sublayer module, is that portfolio reduces state for predicted state, reclaims cloud computing resources；

Portfolio increases scheduling sublayer module, is that portfolio increases state for predicted state, increases cloud computing resources；

Portfolio maintains an equal level scheduling sublayer module, is that portfolio maintains an equal level state for predicted state, maintains cloud computing resources.