CN102137023A - Multicast congestion control method based on available bandwidth prediction - Google Patents

Abstract

The invention discloses a multicast congestion control method based on available bandwidth prediction, which aims at solving the technical problem that the existing multicast congestion control method has hysteretic congestion control. The method comprises the following steps: a layered multicast mechanism is adopted; and a layered index number strategy is selected, i.e., the accumulated sum of the minimum rates from the low layer to the high layer meets the law that index number is progressively increased. By regulating the transmission interval of data packets, a source end detects the available bandwidths from the source end to a receiving end under the condition that extra data packets are not introduced, a least square support vector machine is used for predicting the available bandwidth of next time, flow is regulated in advance according to the prediction result, and the multicast congestion can be effectively controlled.

Description

Multicast Congestion Control method based on the available bandwidth prediction

Technical field

The present invention relates to a kind of Multicast Congestion Control method, particularly a kind of Multicast Congestion Control method based on the available bandwidth prediction.

Background technology

Document " Kwon G I; Byers J W.Smooth multirate multicast congestion control.IEEEINFOCOM March 2003; San Francisco; USA; IEEE Communications Society; 2003:1022-1032 " discloses a kind of Multicast Congestion Control method, i.e. SMCC (Smooth Multicast Congestion Control) control method.This method is incorporated into single-rate mechanism TFMCC (TCP Friendly Multicast Congestion Control) in the layered multicast, and TFMCC is based on TCP throughput equation (1) calculation expectation speed:

$R_{\mathrm{req}}(p,\mathrm{RTT},s)=\frac{s}{\mathrm{RTT}(\sqrt{\frac{2p}{3}}+12\sqrt{\frac{3p}{8}})p(1+32p^2)}---\left(1\right)$

In the formula, s is the size of bag, and p is a packet loss, and RTT is a round-trip delay.Calculating to expected rate is finished at receiving terminal, so each receiving terminal needs estimated parameter p and RTT.SMCC causes congested easily, and main cause has following 2 points:

1, the assumed condition of TCP throughput equation is difficult satisfies, and through type (1) calculates is not the actual reception ability of each receiving terminal, and promptly the source end is to the available bandwidth of receiving terminal.

2,, have only part receiving terminal representative (hereinafter to be referred as CLR) regularly to send feedback in this method to the source end for fear of the feedback implosion.The receiving terminal of non-CLR is not owing to sending feedback, so difficult accurately estimation RTT.So the expected rate receiving ability actual with it that each receiving terminal obtains exists than large deviation.Emulation shows that this speed usually surpasses available bandwidth, even surpasses link bandwidth, thereby causes congested generation.

3, there is the phenomenon of " congested regulation and control hysteresis " in the document disclosed method, is the receiving ability in each receiving terminal past because method all estimates, and goes to regulate the problem that hysteresis can appear regulating and control in flow according to the receiving ability in past; And multicast is a periodic adjustment speed with " wheel " generally, and the cycle of " wheel " is generally a plurality of RTT, so that the hysteresis quality of having aggravated congested regulation and control.

Summary of the invention

In order to overcome existing Multicast Congestion Control method, the invention provides a kind of Multicast Congestion Control method based on the available bandwidth prediction to the problem that congested regulation and control lag behind.This method adopts layered multicast mechanism, selects index layering strategy for use, promptly satisfies the exponential increasing rule by low layer to high-rise minimum-rate accumulation sum.The source end is by adjusting the transmission interval of packet, under the condition of not introducing the excessive data bag, record available bandwidth from the source end to each receiving terminal, and use least square method supporting vector machine to predict next available bandwidth constantly, in advance flow is regulated according to predicting the outcome, thereby realization is to effective control of Multicast Congestion.

The technical solution adopted for the present invention to solve the technical problems is: a kind of Multicast Congestion Control method based on the available bandwidth prediction is characterized in comprising the steps:

Step 1: the source end starts i wheel, and the sign that begins as epicycle of packet of mark, i=1 wherein, and 2,3 ...

Step 2: the transmission that the source end is adjusted each layer data bag according to each layer minimum speed limit at interval makes the speed of data flow be exponential decrease, and each layer data flow rate coverage is 1.5 times of this layer maximum rate 0.5 times to this layer minimum-rate.

Step 3: the source end is with constant rate of speed R _iContinue to send data.

Step 4: after each receiving terminal was received every opening flag of taking turns, the time of the packet that record is received was according to time delay information calculations available bandwidth a separately _i

Step 5: the available bandwidth of prediction next round specifically comprises following operation:

(1) measures the average One Way Delay o that this is taken turns _iWith average packet loss ratio p _i

(2) available bandwidth, average One Way Delay and average packet loss ratio sequence are carried out phase space reconfiguration, that is:

1. make A _i=[a _i, a _I-1, a _I-2..., a _I-(m-1)], in the formula, A _iExpression available bandwidth vector, m embeds dimension;

2. make h ₁=a ₁, h _i=0.5a _i+ 0.5h _I-1, i=2 ..., n, h _iThe exponent-weighted average value of expression available bandwidth has reflected the size of population of available bandwidth in the past; By h _iConstitute weighted average vector H _i=[h _i, h _I-1, h _I-2], represent average One Way Delay vector;

3. make P _i=[p _i, p _I-1, P _I-2], P _iRepresent average One Way Delay vector;

4. make OWD=[o _i, o _I-1, o _I-2], OWD _iExpression average packet loss ratio vector;

5. constitute the input of sample by reproducing sequence, i.e. x _i=[A _i, H _i, P _i, OWD]; Constitute the output of sample, i.e. y by next available bandwidth value constantly _i=a _I+1Thereby, obtain sample (x _i, y _i);

(3) to the input x of sample _i=[A _i, H _i, P _i, OWD _i] carry out core pivot element analysis, realize dimensionality reduction noise reduction to data, obtain carrying out the sample input x behind the core pivot element analysis _{KPCA, i}Sample input x _{KPCA, i}Output y with sample _i=a _I+1Constitute new sample (x _{KPCA, i}, y _i);

(4) with new samples (x _{KPCA, i}, y _i) train the least square method supporting vector machine model as training set;

(5) with the available bandwidth of the model prediction next round that trains;

(6) upgrade training set and least square method supporting vector machine model;

Step 6: according to different updating network state expected rate R _e

(1) stationary phase: A _c〉=R _c,

(2) congested omen: A _c〉=R _c,

(3) network jitter: A _c＜R _c,

(4) network congestion: A _c＜R _c,

In the formula, A _cThe current actual available bandwidth that expression is measured, A _pThe next round available bandwidth of expression prediction, R _cThe cumulative speed of representing current adding multicast layer,

Be the epicycle expected rate,

Be the next round expected rate, T represents cycle of taking turns, and Δ R is the rate increase factor, and its value is s/RTT.

Step 7: if Less than the minimum-rate that requires when anterior layer, then carry out immediately and move back layer operation.

Step 8: use the exponential weighting random timer to carry out the feedback inhibition operation.

Step 9: the source end sends to add a layer synchronous points behind timer expiry receiving that first feedback back starts timer.

Step 10: the expected rate after certain receiving terminal upgrades

When reaching the adding speed of more high-rise requirement, then add layer operation.

Step 11: after synchronous points, the source end continues with speed R _iSend data,, get back to step 1, start the i+1 wheel when epicycle finishes.

The invention has the beneficial effects as follows: owing to adopt layered multicast mechanism, select index layering strategy for use, promptly satisfy the exponential increasing rule to high-rise minimum-rate accumulation sum by low layer.The source end is by adjusting the transmission interval of packet, under the condition of not introducing the excessive data bag, record available bandwidth from the source end to each receiving terminal, and use least square method supporting vector machine to predict next available bandwidth constantly, in advance flow is regulated according to predicting the outcome, thereby realized effective control Multicast Congestion.

Below in conjunction with drawings and Examples the present invention is elaborated.

Description of drawings

Fig. 1 is employed multicast layering of the inventive method and rate-allocation schematic diagram.

Fig. 2 is the employed artificial network topological diagram of the inventive method.

Fig. 3 is the control timing figure that the present invention is based on the Multicast Congestion Control method of available bandwidth prediction.

Fig. 4 distributes the relative time delay of measuring the available bandwidth packet in the inventive method.

Fig. 5 is each layer data bag string structure of multicast of measuring available bandwidth in the inventive method.

Fig. 6 is the model and the parameter update schematic diagram of available bandwidth prediction in the inventive method.

Fig. 7 is the realization block diagram of the available bandwidth prediction in the inventive method.

Fig. 8 is the result of the inventive method emulation experiment 1, the packet loss at node R 1 place.

Fig. 9 is the result of the inventive method emulation experiment 2, the present invention and SMCC response and stability contrast.

Embodiment

Following examples are with reference to Fig. 1～9.Present embodiment is a layered multicast mechanism, and the speed bottom is a basic unit, and the minimum-rate that adds basic unit is B ₀, be that receiving terminal adds the desired minimum-rate of multicast.Present embodiment adopts index layering strategy, promptly from the (L of basic unit ₀Layer) to n layer (L _nLayer) minimum-rate accumulation sum satisfies B _n=2 ⁿB ₀Form.Then certain receiving terminal will add L _nLayer, the minimum-rate of requirement is B _nEach receiving terminal adds multicast all must add basic unit, and each receiving terminal adds the different numbers of plies according to the speed that its expectation receives then, and the number of plies of adding is many more, and the throughput of acquisition is big more, and the service quality that obtains is just high more.

The multicast source end is S1, and receiving terminal is D1, D2; 4 groups of TCP flow points are not made up of TSi-TRi (i=1,2,3,4).Multicast data flow and 4 groups of shared bottleneck bandwidths of TCP stream by R1-R2.

Specific implementation process is:

1, control timing.

Present embodiment was called with time interval of double available bandwidth measurement takes turns.Each takes turns early stage (for the epicycle cycle 10%～20%) for " bandwidth measurement phase "; During this period, the source end is adjusted packet at interval to measure the available bandwidth from the source end to each receiving terminal." bandwidth measurement phase " back is " rate stabilization phase ", and during this period, it is constant that each layer data speed keeps.Its control timing concerns that as shown in Figure 3 wherein each timing node is explained as follows:

T0: the sign that end mark packet in source begins as epicycle starts new one and takes turns.In new one took turns, the source end was provided with scope according to each layer speed, adjusted packet at interval, thereby made the available bandwidth of the receiving terminal energy measurement different range that adds different layers.

T1: the available bandwidth measurement process finishes, and the source end sends an end mark.After each receiving terminal is received end mark, calculate the available bandwidth A of epicycle from the source end to this receiving terminal _c, represent the current actual available bandwidth of measurement, and further predict the available bandwidth A of next round _p, then adjust expected rate and be

And startup exponential weighting random timer T _RCarry out feedback inhibition.

T2: after the source end is received first feedback packet in the epicycle, with the receiving terminal that sends this feedback packet as representative, and with this representative information (ID number and

) multicast is given full group membership, and startup timer T _S, begin to collect feedback information, and the renewal of representing.

T3:T _ROvertime, each receiving terminal is judged its expected rate

Whether reach the more high-rise rate requirement of adding, wait for that then synchronous points application adding is more high-rise if reach; If do not reach then inoperation.

T4:T _SOvertime, end mark packet in source is as synchronous mark.

T5: i.e. application added more high-rise after each receiving terminal was received synchronous mark.

(1) feedback inhibition and timer T _RSetting.

Each receiving terminal is upgrading expected rate After, as the receiving terminal of last round of CLR directly with it

Feed back to the source end; And the receiving terminal of other non-CLR starts random timer T _RThis random timer T _RValue by the expected rate of each receiving terminal

Carry out exponential weighting, the timer that the more little receiving terminal of expected rate is assigned with is short more, thereby can guarantee the slowest recipient's the easier arrival of feedback source end.Timer T when non-CLR receiving terminal _RAfter overtime, if do not receive this layer CLR information or the expected rate of the current C LR that receives greater than the expected rate of oneself, then send its feedback, otherwise do not send feedback;

(2) timer T _SSetting.

Owing to have the receiving terminal lower than CLR expected rate in each is taken turns, the Gu Yuan end allows at T after receiving first feedback _S" status " of competition CLR, T are brought in other receptions in time _SThe value operated by rotary motion (RTT is 2*RTT) in the scope interval.

2, the measurement of available bandwidth.

In " bandwidth measurement phase ", the source end is adjusted packet makes its bag string satisfy rule that speed successively decreases to measure the available bandwidth from the source end to each receiving terminal at interval, and each receiving terminal receives that (the initial data bag of bidding note is P to the epicycle opening flag ₀) after, write down other packets P _iThe time of advent.Definition bag P _iτ in relative time delay _iFor: bag P _iOne Way Delay deduct the bag P ₀One Way Delay.As shown in Figure 4, in whole bag string, when i hour, streaming rate is bigger, generally greater than available bandwidth, then this moment detection packet relative time delay can increase rapidly; Along with i increases, streaming rate constantly reduces, and when being reduced to the available bandwidth size, corresponding bag will be tending towards 0 (flex point place among the figure) relative time delay, and this moment, the Mean Speed of bag string was available bandwidth.Therefore, the Mean Speed excursion of data flow is the measuring range of available bandwidth.

Because there is random delay in network, has jitter the relative time delay of detection packet.Therefore, the present invention adopts the relative time delay that front and back three bags weighted-average method is calculated each bag, calculates bag P _iRelative time delay

Formula be:

$\stackrel{\‾}{{\mathrm{\τ}}_i}=0.25{\mathrm{\τ}}_{i-1}+0.5{\mathrm{\τ}}_i+0.25{\mathrm{\τ}}_{i+1}$ (i＝1，2，3，…，n)(2)

Therefore, record the relative time delay of k bag when certain receiving terminal

Be tending towards 0, then its available bandwidth is:

$A_c=\frac{k\·s}{\mathrm{\Δt}}---\left(3\right)$

Wherein s is long for bag, and Δ t receives packet P for this receiving terminal ₀With P _iTime difference.

Because multicast has isomerism, the receiving ability of each receiving terminal has nothing in common with each other, so the distribution of available bandwidth is bigger.If all receiving terminals all adopt identical bag string to measure available bandwidth, then require the Mean Speed of bag string that very big excursion is arranged, and Measuring Time is longer, also may cause network congestion.Therefore present embodiment is by regulating the transmission interval of each layer data bag of multicast, thereby make the receiving terminal that adds certain one deck can obtain the available bandwidth measurement scope that adapts with this layer transmission rate, as shown in table 1, to realize the available bandwidth measurement of a large amount of heterogeneous receiving terminals.

The concrete structure of each layer data bag as shown in Figure 5, first of base layer data bag string is spaced apart g ₁, i g at interval then _i=β ^I-1G ₁, wherein β is the interval growth factor, generally between 1 to 1.2, then the base layer data bag sends and satisfies the exponential increasing rule at interval its span.Enhancement layer (L ₁Layer and L ₁The layer that layer is above) packet evenly inserts between the packet of its lower floor (empty frame is represented the packet of lower floor among the figure), then adds L _nThe receiving terminal of layer will receive L _nThe packet of layer and all lower floors, these packets are formed new bag string and are still satisfied the rule that speed is successively decreased.Packet P received in each receiving terminal record _iTime, and calculate P with formula (2) _iRelative time delay

Find

After the flex point that changes, can calculate current available bandwidth A with formula (3) _c

[table 1] available bandwidth measurement scope (only listing preceding four layers)

Level number	This layer speed range	The bandwidth measurement scope
			L 0	(B 0，B 1)	(0.5B 0，1.5B 1)
L 1	(B 1，B 2)	(0.5B 1，1.5B 2)
			L 2	(B 2，B 3)	(0.5B 2，1.5B 3)
L 3	(B 3，B 4)	(0.5B 3，1.5B 4)

3, the prediction of available bandwidth.

(1) data preliminary treatment.

Adopt available bandwidth vector A in the present embodiment _i=[a _i, a _I-1, a _I-2..., a _I-m], available bandwidth exponential weighting mean value vector H _i=[h _i, h _I-1, h _I-2], average packet loss ratio vector P _i=[p _i, p _I-1, p _I-2] and One Way Delay vector OWD _i=[o _i, o _I-1, o _I-2], constitute sample input x _i=[A _i, H _i, P _i, OWD _i].

Present embodiment adopts core pivot element analysis, and (Kernel Principle Component Analysis is KPCA) to sample input x _i=[A _i, H _i, P _i, OWD _i] carry out obtaining x behind the dimensionality reduction noise reduction _{KPCA, i}, x _{KPCA, i}With sample output y _i=a _I+1Constitute new samples (x _{KPCA, i}, y _i).

(2) predict based on the available bandwidth of least square method supporting vector machine.

A) optimization of model parameter

Obtain sample x _{KPCA, i}, y _i) after, set up sample set S={ (x _{KPCA, i}, y _i) | i=1,2 ..., n} need select for use a kind of prediction algorithm that sample set S is learnt, and then predicts next round available bandwidth A _pBecause least square method supporting vector machine (Least Squares Support Vector Machine, LSSVM) training time few, the precision of prediction height is so present embodiment selects for use LSSVM that available bandwidth is predicted.In the model of LSSVM, the kernel function that present embodiment is selected for use be RBF: K (x, y)=exp (|| x-y|| ²/ 2 δ ²).

In model training, need to determine two parameters, be respectively: penalty factor c in the LSSVM model and nuclear parameter δ.Therefore present embodiment is selected fast convergence rate, parameter is few and amount of calculation is little particle cluster algorithm for use (Particle Swarm Optimization PSO) is optimized two parameters in the LSSVM model.The prediction relative mean square error RMSE (Relative Mean Squares Error) of m predicted value of definition is:

Wherein

Be y _iPredicted value, RMSE is optimized parameter δ and the c of LSSVM as the adaptive value of PSO.

B) model online updating

LSSVM is output as

K (x wherein _i, x) be kernel function, α _i(i=1,2 ..., be the Lagrange coefficient of introducing when finding the solution LSSVM model optimization problem n), its LSSVM model solution can be equivalent to finds the solution α and b in the formula (5).

$\left[\begin{array}{cc}0& e^T\\ e& K+(1/c)I\end{array}\right]\×\left[\begin{array}{c}b\\ a\end{array}\right]=\left[\begin{array}{c}0\\ y\end{array}\right]---\left(5\right)$

In the formula (5), make kernel function

Then K is that element is K _IjN * n rank nuclear matrix, vectorial e=(1 ..., 1) ^T, I is a unit matrix, vectorial α=(α ₁, α ₂..., α _n) ^T, vectorial y=(y ₁, y ₂..., y _n) ^T

Because the sequence of network availability bandwidth is a nonlinear time series, after after a while, the model of initial training can produce bigger error to the prediction of new samples.Therefore, model should upgrade with the renewal of sample.If sample of every renewal, LSSVM finds the solution again, then needs to recomputate nuclear matrix K, and also demand is separated the contrary of [K+ (1/c) I] simultaneously, and computation complexity is o (n ³) more than, amount of calculation is bigger.The characteristics that present embodiment is found the solution when the model modification according to LSSVM adopt recursion Calculation Method online updating model, and adopt different recursion modes according to the different phase of prediction, improve the renewal speed and the precision of prediction of model.

Order matrix G=K+ (1/c) I, then

$G=\left[\begin{array}{cccc}{G}_{11}& G_{12}& ...& G_{1n}\\ G_{21}& G_{22}& ...& ...\\ ...& ...& ...& G_{(n-1)n}\\ G_{n1}& ...& G_{n(n-1)}& G_{\mathrm{nn}}\end{array}\right]=\left[\begin{array}{cccc}{\mathrm{<figure-callout\; id="11"\; label="K"\; filenames="HSA00000472325500033.png"\; state="\{\{state\}\}">K</figure-callout>}}_{11}+1/c& K_{12}& ...& K_{1n}\\ K_{21}& K_{22}+1/c& ...& ...\\ ...& ...& ...& K_{(n-1)n}\\ K_{n1}& ...& K_{n(n-1)}& K_{\mathrm{nn}}+1/c\end{array}\right]---\left(6\right)$

Then separate formula (5) can get LSSVM separate for:

$\left\{\begin{array}{c}a=G^{-1}(y-\mathrm{be})\\ b=e^T{G}^{-1}y/e^T{G}^{-1}e\end{array}\right.---\left(7\right)$

So calculate α, the key of b is to find the solution G ^-1In calculating, utilize the result of calculation of last time to carry out recursion calculating, then can reduce amount of calculation greatly.Present embodiment adopts recursion Calculation Method online updating model according to the characteristics of matrix G, and adopts different recursion modes in the different phase of prediction.

(b1) incremental learning

In the incipient stage of available bandwidth prediction, because the sample of gathering is less, model need increase the sample size of training sample set with the increase of sample.Then at every turn more during new model, the dimension of matrix G adds 1.The sample size of supposing current training sample set is l, and when upgrading l+1 sample, preceding l * l the element of matrix G do not change, and just increased delegation and row behind original matrix, and its form is as follows:

$G^{l+1}=\left[\begin{array}{ccc}{G}^l& & G_{1(l+1)}\\ & & ...\\ G_{(l+1)1}& ...& G_{(l+1)(l+1)}\end{array}\right]=\left[\begin{array}{cc}{G}^l& \mathrm{\&beta;}\\ {\mathrm{\&beta;}}^l& g\end{array}\right]---\left(8\right)$

Wherein vectorial β=(G _{(l+1) 1}, G _{(l+1) 2}..., G _{(l+1) l}) ^T, scalar g=G _{(l+1) (l+1)}, G ^lWith G ^L+1Represent that respectively sample size is the matrix G of l and the matrix G behind sample of increase, the contrary (G of the original matrix when then the key of recursion calculating is to train from last time ^l) ^-1Derive (G ^L+1) ^-1, as can be known according to the method for inverting of matrix in block form

${\left(G^{l+1}\right)}^{-1}={\left[\begin{array}{cc}{G}^l& \mathrm{\&beta;}\\ {\mathrm{\&beta;}}^T& g\end{array}\right]}^{-1}=\left[\begin{array}{cc}{\left(G^l\right)}^{-1}& 0\\ 0^{\mathrm{<figure-callout\; id="0"\; label="T"\; filenames="HSA00000472325500031.png,HSA00000472325500033.png"\; state="\{\{state\}\}">T</figure-callout>}}& 0\end{array}\right]+\frac{1}{\mathrm{\&theta;}}\left[\begin{array}{cc}{\mathrm{EF}}^T& E\\ F^T& 1\end{array}\right]---\left(9\right)$

Wherein, vectorial E=-(G ^l) ^-1β, vectorial F=(β ^T(G ^l) ^-1) ^T, scalar θ=g+F ^Tβ.Therefore, the process of incremental learning is in the present embodiment: after obtaining new samples, at first calculate β and g; Again by (G ^l) ^-1Calculate E, F and θ, through type (9) is obtained (G then ^L+1) ^-1Last through type (7) calculates α and b.Therefore, in the incipient stage of available bandwidth prediction, the training sample of LSSVM constantly increases, and adopting the method for above incremental learning can be computation complexity from o (l ³) reduce to o (l ²), thereby the renewal speed of raising model.

(b2) iterative learning

Along with constantly carrying out of available bandwidth prediction, sample constantly increases, but training dataset can not constantly enlarge with the increase of sample.Then increase when a certain amount of when number of training, needing to adopt size is the sliding window new samples more of n, and up-to-date sample of promptly every increase just abandons a sample the oldest, thereby the sample number of maintenance training dataset is constant.The G matrix that obtains before and after the model modification is expressed as G respectively ^OldAnd G ^NewConsistent with incremental learning, during the iterative learning model modification, hope can be from (G ^Old) ^-1Derive (G ^New) ^-1Thereby, reduce computation complexity.Contrast G ^OldAnd G ^NewAs can be known, but the difference form of two matrixes turn to:

$G^{\mathrm{old}}=\left[\begin{array}{ccc}{\mathrm{<figure-callout\; id="11"\; label="G"\; filenames="HSA00000472325500033.png"\; state="\{\{state\}\}">G</figure-callout>}}_{11}& ...& G_{1n}\\ ...& & \\ G_{n1}& & \mathrm{\&Phi;}\end{array}\right]=\left[\begin{array}{cc}{\mathrm{<figure-callout\; id="11"\; label="G"\; filenames="HSA00000472325500033.png"\; state="\{\{state\}\}">G</figure-callout>}}_{11}& {\mathrm{\&rho;}}^l\\ \mathrm{\&rho;}& \mathrm{\&Phi;}\end{array}\right]\&RightArrow;\left[\begin{array}{ccc}\mathrm{\&Phi;}& & G_{1n}\\ & & ...\\ G_{n1}& ...& G_{\mathrm{nn}}\end{array}\right]---\left(10\right)$

Wherein, Φ representing matrix G ^OldIn the 2nd row to the capable matrix of forming with the 2nd element that is listed as l row of l.By formula (10) as can be known, if can pass through (G ^Old) ^-1Obtain Φ ^-1, the then available method identical with incremental learning is from Φ ^-1Derive (G ^New) ^-1, (the G that once calculates before establishing ^Old) ^-1Form is as follows:

${\left(G^{\mathrm{old}}\right)}^{-1}={\left[\begin{array}{cc}{\mathrm{<figure-callout\; id="11"\; label="G"\; filenames="HSA00000472325500033.png"\; state="\{\{state\}\}">G</figure-callout>}}_{11}& {\mathrm{\&rho;}}^T\\ \mathrm{\&rho;}& \mathrm{\&Phi;}\end{array}\right]}^{-1}=\left[\begin{array}{cc}\mathrm{\&gamma;}& N\\ M& Q\end{array}\right]---\left(11\right)$

Wherein, γ is (G ^Old) ^-1In first the row first element, M is (G ^Old) ^-1In the column vector of the 2nd to l element composition of first row, N is (G ^Old) ^-1In the row vector formed of the 2nd to l element of first row, Q represents (G ^Old) ^-1In the 2nd row to the capable matrix of forming with the 2nd element that is listed as l row of l.With reference to the method for inverting of the matrix in block form of formula as can be known:

${\left(G^{\mathrm{old}}\right)}^{-1}=\left[\begin{array}{cc}\mathrm{\&gamma;}& N\\ M& Q\end{array}\right]=\left[\begin{array}{cc}\mathrm{\&gamma;}& N\\ M& {\mathrm{\&Phi;}}^{-1}+\mathrm{MN}/\mathrm{\&gamma;}\end{array}\right]---\left(12\right)$

Then as can be known, Φ by formula (12) ^-1=Q-MN/ γ.Therefore, the process of iterative learning is: through type (12) is by (G ^Old) ^-1Calculate Φ ^-1, adopt the method for incremental learning again, by Φ ^-1Calculate (G ^New) ^-1, through type (7) calculates α and b again.

After available bandwidth prediction beginning a period of time, model promptly from the incremental learning step transition to the iterative learning stage, thereby guarantee that model can online in real time upgrade, and guarantees precision of prediction.Another parameter that the LSSVM training need be determined the time is the sample size n sample range of training sample set, i.e. sliding window size during iterative learning.In theory, the n value is big more, and the amount of information that LSSVM obtains is big more, and it is accurate more to predict; Yet along with the increase of n, the amount of calculation of training is big more, and the spent training time is many more.Take all factors into consideration the amount of calculation of available bandwidth accuracy of predicting and model modification, the value of n is generally 10 ²To 10 ³Between.

(3) realization flow of available bandwidth prediction.

Since when adopting the parameter of PSO algorithm optimization LSSVM, repeatedly train sample, very consuming time.Therefore can not behind sample of every renewal, just optimize again the parameter of LSSVM, and it is little because the data characteristics in the short time changes, also there is no need behind sample of every renewal parameters optimization again, the model after the parameter optimization still adapts within a certain period of time.Therefore, as shown in Figure 6, adopt a step recursion to calculate in the present embodiment and upgrade the LSSVM model, and model parameter is selected for use the multistep method for updating.

The specific implementation process of available bandwidth prediction as shown in Figure 7 in the present embodiment.The data that need to gather comprise current available bandwidth value, One Way Delay and packet loss.Obtain to carry out preliminary treatment after the data that acquired original arrives, promptly each argument sequence is carried out phase space reconfiguration.Utilize core principle component analysis to reduce the sample data noise in the multidimensional sample that obtains, extract the principal component that comprises data message, reduce the sample dimension simultaneously, then with the input of the principal component extracted as LSSVM.In the starting stage of prediction, adopt incremental learning, after sample size reaches window value n, adopt iterative learning.After upgrading k sample, be target to minimize the prediction relative mean square error, adopt the parameter of PSO algorithm optimization LSSVM, to realize accurately and available bandwidth prediction fast.

4, Ceng adding with withdraw from.

Measured value A at the epicycle available bandwidth _cPredicted value A with the next round available bandwidth _pAfter, upgrade expected rate R _eReceiving terminal adds layer and moves back layer and all be based on its expected rate R in the present embodiment _eCarry out, it is different with the time point that moves back layer just to add layer.In every the wheel, be not when receiving terminal decision adds layer, just to add layer operation, for realize each receiving terminal add layer synchronously, the source end is taken turns the later stage at this and is inserted synchronous points, each receiving terminal needs could carry out after synchronous points by the time and adds layer operation; But in the time of need moving back layer, need not wait for, directly operate, in order to avoid cause network congestion.

Work as R _eAt interval (B _i, B _I+1) in the time, can add application and add L _iLayer.But work as R _eNear the intersection of adjacent two layers during slight jitter, may cause receiving terminal to add continually and withdraw from certain one deck.For fear of this situation, present embodiment adopts conservative method, and regulation has only when expecting that speed is at scope (dB _i, B _I+1) when interior, receiving terminal could apply for adding L _iLayer, wherein, d is a damping factor, and is relatively good through experiment showed, that d gets 1.2 effects.Be reduced to the minimum-rate B of i layer when the expected rate value of certain receiving terminal of i layer _iThe time, this receiving terminal withdraws from the i layer.

5, emulation experiment.

Effect by emulation tool check the inventive method.

(1) emulation experiment 1.

The objective of the invention is to reduce network congestion, packet loss is an important parameter weighing the network congestion situation.The packet loss at emulation contrast the present invention and node R 1 place of SMCC before bottleneck link, simulation result as shown in Figure 8.

As shown in Figure 8, the packet loss of SMCC fluctuates between 0.017～0.056, and average packet loss ratio is about 0.0289; And packet loss of the present invention remains near 0.003, and jitter amplitude is less relatively, and average packet loss ratio is about 0.0037.This is because the present invention can predict available bandwidth, when available bandwidth descends, can shift to an earlier date regulations speed, avoids network congestion to take place, and has therefore reduced packet loss; And SMCC adopts expected rate that the TCP throughput equation calculates often greater than available bandwidth, causes congestedly easily, so its packet loss is higher.

(2) emulation experiment 2.

A beneficial effect of the present invention is that algorithm has reaction speed faster when available bandwidth changes greatly, i.e. response preferably.Topology shown in Figure 2 is adopted in experiment, and the link capacity between R1 and the R2 is replaced by 30Mbps.When 10s, D1, D2 add multicast group, start TCP1, TCP3 and TCP4 simultaneously; TCP2 starts when 100s, withdraws from when 150s.Therefore, only compete bottleneck link to D1 between the 100s, only compete bottleneck link to D1 between the 150s with two streams of TCP1, TCP2 at 100s with TCP1 at 10s.Because TCP2 starts when 100s, caused that the traffic carrying capacity of this link changes, three streams are shared the 18Mbps bottleneck bandwidth, and fair bandwidth is 6Mbps.Select for use SMCC and the present invention to experimentize as multicast data flow respectively, simulation time is 250s.

Fig. 9 shows that the throughput of receiving terminal D1 under two kinds of multicast data flows changes.As seen from the figure, owing to the startup of TCP2, the network fair bandwidth becomes and becomes 6Mbps by 9Mbps when 100s, and the present invention and SMCC can both withdraw from the 2nd layer quickly, add the 1st layer.As shown in the figure, the reaction time of the present invention is 0.91s, and the reaction time of SMCC is 2.04s.This is because SMCC rises at packet loss, finds network congestion, just its expected rate is reduced, and the present invention reduced its expected rate in advance, so the present invention is faster than the response speed of SMCC by the variation of prediction available bandwidth.Equally, because the withdrawing from of TCP2, the network fair bandwidth becomes 9Mbps again when 150s, and the reaction time of the present invention is 1.67s, and the reaction time of SMCC is 3.82s; This is because the present invention predicts available bandwidth, its expected rate of increase that therefore can be more Zao than SMCC, thus improve bandwidth availability ratio.

Another beneficial effect of the present invention is that its transmission rate big shake can not take place when available bandwidth generation slight jitter, promptly has stationarity preferably.Employing mean square deviation of speed in a period of time is investigated the stationarity of speed, the stage that big variation do not take place the network bandwidth in the analysis chart 9 as can be known, in time period, the mean square deviation of SMCC throughput is 0.3228Mbps at 0s～100s and 200s～300s, and maximal jitter amplitude is 0.89Mbps; And the mean square deviation of throughput of the present invention is 0.1761Mbps, and maximal jitter amplitude is 0.46Mbps.Therefore, compare with SMCC, the stationarity of throughput of the present invention is better, this is because the present invention has adopted Congestion Avoidance when calculating its expected rate, guarantee that expected rate can not surpass its available bandwidth, therefore can reduce congested generation, throughput is remained near the available bandwidth.And the expected rate that SMCC calculates based on the TCP throughput equation, and owing to its parameters of formula exists than large deviation, therefore its expected rate that obtains usually surpasses available bandwidth, cause the rising of packet loss, and the rising of packet loss causes the decline of expected rate, so the shake of the throughput of SMCC is bigger than the present invention.

Claims (1)

1. the Multicast Congestion Control method based on the available bandwidth prediction is characterized in that comprising the steps:

Step 1: the source end starts i wheel, and the sign that begins as epicycle of packet of mark, i=1 wherein, and 2,3,

Step 2: the transmission that the source end is adjusted each layer data bag according to each layer minimum speed limit at interval makes the speed of data flow be exponential decrease, and each layer data flow rate coverage is 1.5 times of this layer maximum rate 0.5 times to this layer minimum-rate;

Step 3: the source end is with constant rate of speed R _iContinue to send data;

Step 4: after each receiving terminal was received every opening flag of taking turns, the time of the packet that record is received was according to time delay information calculations available bandwidth a separately _i

Step 5: the available bandwidth of prediction next round specifically comprises following operation:

(1) measures the average One Way Delay o that this is taken turns _iWith average packet loss ratio p _i

(2) available bandwidth, average One Way Delay and average packet loss ratio sequence are carried out phase space reconfiguration, that is:

1. make A _i=[a _i, a _I-1, a _I-2..., a _I-(m-1)], in the formula, A _iExpression available bandwidth vector, m embeds dimension;

2. make h ₁=a ₁, h _i=0.5a _i+ 0.5h _I-1, i=2 ..., n, h _iThe exponent-weighted average value of expression available bandwidth has reflected the size of population of available bandwidth in the past; By h _iConstitute weighted average vector H _i=[h _i, h _I-1, h _I-2], represent average One Way Delay vector;

3. make P _i=[p _i, p _I-1, p _I-2], P _iRepresent average One Way Delay vector;

4. make OWD _i=[o _i, o _I-1, o _I-2], OWD _iExpression average packet loss ratio vector;

5. constitute the input of sample by reproducing sequence, i.e. x _i=[A _i, H _i, P _i, OWD _i]; Constitute the output of sample, i.e. y by next available bandwidth value constantly _i=a _I+1Thereby, obtain sample (x _i, y _i);

(3) to the input x of sample _i=[A _i, H _i, P _i, OWD _i] carry out core pivot element analysis, realize dimensionality reduction noise reduction to data, obtain carrying out the sample input x behind the core pivot element analysis _{KPCA, i}Sample input x _{KPCA, i}Output y with sample _i=a _I+1Constitute new sample (x _{KPCA, i}, y _i);

(4) with new samples (x _{KPCA, i}, y _i) train the least square method supporting vector machine model as training set;

(5) with the available bandwidth of the model prediction next round that trains;

(6) upgrade training set and least square method supporting vector machine model;

Step 6: according to different updating network state expected rate R _e

(1) stationary phase: A _c〉=R _c,

(2) congested omen: A _c〉=R _c,

(3) network jitter: A _c＜R _c,

(4) network congestion: A _c＜R _c,

In the formula, A _cThe current actual available bandwidth that expression is measured, A _pThe next round available bandwidth of expression prediction, R _cThe cumulative speed of representing current adding multicast layer,

Be the epicycle expected rate,

Be the next round expected rate, T represents cycle of taking turns, and Δ R is the rate increase factor, and its value is s/RTT;

Step 7: if

Less than the minimum-rate that requires when anterior layer, then carry out immediately and move back layer operation;

Step 8: use the exponential weighting random timer to carry out the feedback inhibition operation;

Step 9: the source end sends to add a layer synchronous points behind timer expiry receiving that first feedback back starts timer;

Step 10: the expected rate after certain receiving terminal upgrades

When reaching the adding speed of more high-rise requirement, then add layer operation;

Step 11: after synchronous points, the source end continues with speed R _iSend data,, get back to step 1, start the i+1 wheel when epicycle finishes.

Images