CN101695050A - Dynamic load balancing method based on self-adapting prediction of network flow - Google Patents

Abstract

The invention provides a dynamic load balancing method based on self-adapting prediction of network flow. The scheme includes: observing historical data regularity of network load flowing into a load balancing switch or a switching software in a certain time cycle; obtaining a parameter value of self-adapting algorithm in a prediction program to form a computing formula; then substituting load observation value at present moment into the formula to predict load value at the next moment; and distributing flow for a rear-end server at real time according to prediction value, thereby enabling the network load to be regulated in advance, avoiding lag effect, constantly keeping the network load in a comparatively balancing state, greatly strengthening self-adapting self-regulating capability of network to load, and being adaptable to networks with a certain time cycle regularity at occasions such as regular-period network backup and the like.

Description

A kind of dynamic load balancing method of adaptive-flow prediction Network Based

Technical field

The present invention relates to a kind of Computer Applied Technology, is exactly a kind of network load balancing method of server cluster specifically.

Technical background

Current, in each server part, its data traffic and calculating strength are big, make single equipment can't bear at all, and how to finish the rational traffic carrying capacity distribution of realization between a plurality of network equipments of said function, make it to be unlikely to occur one equipment is busy excessively and other equipment is not given full play to the situation of disposal ability, just become a problem, therefore load-balancing mechanism also arises at the historic moment.

Load balancing is based upon on the existing network infrastructure, and it provides a kind of cheap effective method expansion servers bandwidth and has increased throughput, strengthens the network data-handling capacity, improves network more flexible and availability.It mainly finishes following task: solve the network congestion problem, service provides nearby, realizes the geographical position independence; For the user provides better visit quality; Improve speed of response of server; Improve server and other efficiency of resource; Avoid the network key position single point failure or the like to occur.

More present said load balancing refers to equilibrium (sharing in other words) measure is carried out in the load of access server.

In present load balancing is used, substantially all be to lag behind to regulate, only be the measured value of present load is regulated, in fact load has changed when making adjusting, the easy like this response lag that causes, regulating effect is not obvious, sometimes even play the effect of adverse effect.

Summary of the invention

The dynamic load balancing method that the purpose of this invention is to provide a kind of server cluster of volume forecasting Network Based can make offered load be regulated in advance, has avoided hysteresis effect, makes offered load remain on the state of comparison equilibrium constantly.

The objective of the invention is to realize in the following manner: the parameter value that obtains adaptive algorithm by the historical data in certain hour cycle of offered load, form the formula of predictor, load measured value and above-mentioned parameter value substitution formula with current time, obtain predicted value to next offered load constantly, thereby make load balancing software adjust the weights of the dispatching algorithm of rear end in advance, realize load balancing according to this predicted value.

Concrete steps are as follows:

At first by based on the network flow monitoring software of the operating system of load equalizer of snmp protocol own or the network flow monitoring record of load balancing software, comprise the historical data of the network traffics of inflow load-equalizing switch or switch software in the certain hour cycle, the computing formula of adaptive algorithm in the substitution predictor, obtain the parameter of adaptive algorithm, when predicting, the load measured value substitution formula of current time is predicted next load numerical value constantly, and the load balancing module of the balanced software of load measured value input load of replacement current time comes back-end server is carried out assignment of traffic;

Adaptive algorithm is three layers of BP neural network algorithm, and the predictor formula of next of current time network traffic load constantly is:

${\mathrm{INPUT}}_n=\underset{i=0}{\overset{n-1}{\mathrm{\Σ}}}{x}_i{\mathrm{INPUT}}_i$

INPUT wherein _i(i=0,1,2,3 ... n-1) be current time network traffics, apart from the last time interval of current time, preceding two time intervals, first three time interval ... the network traffic load in preceding n-1 the time interval, x _iThen be the weights that obtain by adaptive algorithm, the calculating by top weighting autoregression formula obtains the network traffics INPUT to next time interval _nPredicted value, x _iComputational process with reference to the neural net introduction, the size in the selected and time interval of n value was decided according to the time cycle that this offered load changes;

Specific algorithm is described below:

X=(x ₀, x ₁... x _N-1) ^T, x ∈ R ⁿ, be input vector, hidden layer n ₁Individual neuron, x ' ∈ R ⁿ, x '=(x ' ₀, x ' ₁..., x ' _N-1) ^T, output neuron y ∈ R ^m, y=(y ₀, y ₁..., y _M-1) ^T, the weights between input layer and the hidden layer are w _Ik, threshold value is θ _i, the weights between hidden layer and output layer are w _Kj, threshold value is θ _l, the output of hidden neuron is satisfied so:

The neuronic output of output layer is satisfied

Network is estimated about the error of p sample:

The process of three layers of BP neural network algorithm is made up of the forward-propagating and the error back propagation of pattern, in the forward-propagating process, input information is successively handled and is passed to output layer through hidden neuron, output layer can not obtain desired output, then change back-propagation process over to, error between actual value and the network output is returned by original connection path, by revising the interneuronal weight of each layer error is reduced, and then change the forward-propagating process over to, repeated calculation like this is till error is less than set point.

Different effect of the present invention is: possess stronger self-learning capability, its parameter can change with the continuous variation of historical data, and the formula of predictor also changes thereupon, thereby guarantees that predicted value under any circumstance all has higher accuracy.Predict that by the predictor of adaptive algorithm next flows into the offered load of load-equalizing switch or switch software constantly, have the occasion of time cycle property variation characteristics in load variations, can carry out volume forecasting according to the offered load variation tendency in this time cycle more exactly and regulate in advance.

Description of drawings

The schematic diagram of the dynamic load balancing method of the server cluster of Fig. 1 volume forecasting Network Based;

Three layers of BP neural network algorithm of Fig. 2 schematic diagram.

Embodiment:

Elaborate with reference to Figure of description:

Before will being connected on the load balancing strategy based on the volume forecasting module of adaptive algorithm, use next network load prediction value is constantly passed to the load balancing strategy, rather than current offered load measured value.

Illustrate:

Can input and output be arranged regarding a flight data recorder as based on the predicting network flow module of adaptive algorithm (neural net).Such as certain day 18 load flow as output, certain day 17: 30 same day, 17: 40,17: 50 load flow was as input, can obtain being calculated by input one group of parameter of the neural net of output so.When today, 18 load flow was predicted, 17: 30 today that has observed, 17: 40,17: 50 load flow input neural network just can be predicted 18 load flow by the parameter value that obtains previously.Certainly the parameter value that only obtains according to one day historical data enough accurately need not carry out a large amount of tests and obtain more parameter value, and simple average or weighted average obtain more accurate and effective parameter again.

The learning process of other adaptive learning algorithms is similar.Use which kind of adaptive learning algorithm unimportant, it is a flight data recorder, and it is just passable to access output.

The predictor formula of next of current time network traffic load constantly is

${\mathrm{INPUT}}_n=\underset{i=0}{\overset{n-1}{\mathrm{\Σ}}}{x}_i{\mathrm{INPUT}}_i$

INPUT wherein _i(i=0,1,2,3 ... n-1) be current time network traffics, apart from the last time interval of current time, preceding two time intervals, first three time interval ... the network traffic load in preceding n-1 the time interval, x _iThen be the weights that obtain by adaptive algorithm, the calculating by top weighting autoregression formula obtains the network traffics INPUT to next time interval _nPredicted value.x _iComputational process with reference to the neural net introduction.The size in the selected and time interval of n value was decided according to the time cycle that this offered load changes.

With three layers of BP neural net is example, implementation procedure as shown in Figure 2:

The adaptive algorithm of three layers of BP network as shown in Figure 2, x=(x ₀, x ₁... x _N-1) ^T, n=6 represents (x on certain day 17: 30 same day respectively here ₀), 17: 35 (x ₁) ... 17: 55 (x _N-1) the offered load flow, hidden layer n ₁Individual neuron, x ' ∈ R ⁿ, x '=(x ' ₀, x ' ₁..., x ' _N-1) ^T, output neuron y ∈ R ^m, y=(y ₀, y ₁..., y _M-1) ^T, m=1 represents 18 offered load flow here, and the weights between input layer and the hidden layer are w _Ik, threshold value is θ _j, (generally getting 0), the weights between hidden layer and output layer are w _Kj, threshold value is θ _l, the output of hidden neuron is satisfied so:

$x_k^{\′}=f(\underset{i}{\overset{n-1}{\mathrm{\Σ}}}{w}_{\mathrm{ik}}{x}_i-{\mathrm{\θ}}_j)$

The neuronic output of output layer is satisfied

$y_l=f(\underset{k=0}{\overset{m-1}{\mathrm{\Σ}}}{w}_{\mathrm{kj}}{x}_k^{\′}-{\mathrm{\θ}}_l)$

Network is estimated about the error of p sample:

$E_p=\frac{1}{2}\underset{j=1}{\overset{m}{\mathrm{\Σ}}}{(y_{\mathrm{pj}}-o_{\mathrm{pj}})}^2$

Here 17: 30 that observed in the past, 17: 35 ... 17: 55 load flow substitution x=(x ₀, x ₁... x _N-1) ^T, 18 load flow substitution y=(y ₀, y ₁..., y _M-1) ^T, separate respectively

With

Obtain parameter value w _IkAnd w _KjThereby,, when the hidden layer number gets 1, can obtain formula

${\mathrm{INPUT}}_n=\underset{i=0}{\overset{n-1}{\mathrm{\Σ}}}{x}_i{\mathrm{INPUT}}_i$

Parameter value x in the formula _i=w _Ik* w _KjSo just can be big 17: 30 of the present, 17: 35 ... 17: 55 subnetwork load measured value substitution formula, thus 18 the predicting network flow value of today obtained.

As shown in Figure 1, the step of removing in the frame of broken lines is exactly general load balancing, the measured value of current network flow is directly given the dispatching algorithm of rear end and is carried out load balancing, our dynamic load leveling based on prediction then is to carry out step in the frame of broken lines earlier to obtain predicted value to next network traffics constantly, and the dispatching algorithm of giving the rear end this predicted value is carried out load balancing again.

Here " a certain amount of historical data " is the offered load discharge record in the certain hour section, and by study and the identification to these records, neural net can obtain one group of parameter, and this parameter can be used to predict.Obtain after the parameter, the load measured value of current time as input, is simulated with neural network algorithm, can obtain predicted value next load flow constantly.With the input of predicted value as load equalizer, just be equivalent in advance network traffics be adjusted, have in load under the situation of certain hour periodic law, the result has higher order of accuarcy.

Claims (1)

1. the dynamic load balancing method of an adaptive-flow Network Based prediction, it is characterized in that, obtain the parameter value of adaptive algorithm by the historical data in certain hour cycle of offered load, form the formula of predictor, with the load measured value and the above-mentioned parameter value substitution formula of current time, obtain predicted value, thereby make load balancing software adjust the weights of the dispatching algorithm of rear end in advance according to this predicted value next offered load constantly, realize load balancing, concrete steps are as follows:

At first by based on the network flow monitoring software of the operating system of load equalizer of snmp protocol own or the network flow monitoring record of load balancing software, comprise the historical data of the network traffics of inflow load-equalizing switch or switch software in the certain hour cycle, the computing formula of adaptive algorithm in the substitution predictor, obtain the parameter of adaptive algorithm, when predicting, the load measured value substitution formula of current time is predicted next load numerical value constantly, and the load balancing module of the balanced software of load measured value input load of replacement current time comes back-end server is carried out assignment of traffic;

Adaptive algorithm is three layers of BP neural network algorithm, and the predictor formula of next of current time network traffic load constantly is:

${\mathrm{INPUT}}_n=\underset{i=0}{\overset{n-1}{\mathrm{\Σ}}}{x}_i\mathrm{INPU}{T}_i$

INPUT wherein _i(i=0,1,2,3 ... n-1) be current time network traffics, apart from the last time interval of current time, preceding two time intervals, first three time interval ... the network traffic load in preceding n-1 the time interval, x _iThen be the weights that obtain by adaptive algorithm, the calculating by top weighting autoregression formula obtains the network traffics INPUT to next time interval _nPredicted value, x _iComputational process with reference to the neural net introduction, the size in the selected and time interval of n value was decided according to the time cycle that this offered load changes;

Specific algorithm is described below:

X=(x ₀, x ₁... x _N-1) ^T, x ∈ R ⁿ, be input vector, hidden layer n ₁Individual neuron, x ' ∈ R ⁿ, x '=(x ' ₀, x ' ₁..., x ' _N-1) ^T, output neuron y ∈ R ^m, y=(y ₀, y ₁..., y _M-1) ^T, the weights between input layer and the hidden layer are w _Ik, threshold value is θ _j, the weights between hidden layer and output layer are w _Kj, threshold value is θ _l, the output of hidden neuron is satisfied so:

The neuronic output of output layer is satisfied

Network is estimated about the error of p sample:

The process of three layers of BP neural network algorithm is made up of the forward-propagating and the error back propagation of pattern, in the forward-propagating process, input information is successively handled and is passed to output layer through hidden neuron, output layer can not obtain desired output, then change back-propagation process over to, error between actual value and the network output is returned by original connection path, by revising the interneuronal weight of each layer error is reduced, and then change the forward-propagating process over to, repeated calculation like this is till error is less than set point.

Images