Network slicing configuration method based on traffic awareness

Abstract

The present invention belongs to the field of network communication technology, and specifically relates to a network slice configuration method based on traffic perception, and the specific steps are: reconstructing network traffic based on a compressed sensing end-to-end network traffic reconstruction method to obtain reconstructed network traffic A; obtaining network traffic B in historical network data, performing LSTM prediction network training, substituting the reconstructed network traffic A into the LSTM prediction network, performing traffic load prediction on the reconstructed network traffic A, and obtaining the predicted load value of each network element; combining NFV technology to realize intelligent allocation of network slice resources: comparing the load threshold of each server in the current system and the predicted load values ​​of all network elements, and performing resource allocation for the predicted load value of each network element. Beneficial effects: wide adaptability, low cost, reasonable allocation and planning, reasonable use and allocation of system resources, and accelerated system network operation speed and processing efficiency.

Description

Technical Field

[0001] The present invention belongs to the field of network communication technology, and specifically relates to a network slicing configuration method based on traffic perception. Background Art

[0002] The diversified development of 5G services in the future will have many uncertain factors such as channels, traffic requirements, and business scenarios. Flexible resource allocation methods such as 5G core network cloudification and network slicing are the magic weapon for current operators to improve user service quality and increase network performance. Many scholars have conducted relevant research in this field. Falahatraftar et al. proposed a heterogeneous vehicle network slicing method based on conditional generative adversarial networks in 5G application scenarios; Cui Ting et al. proposed a 5G slicing transmission scheme for diversified information in the field of agricultural network monitoring transmission; Barakabitze et al. proposed a 5G slicing scheme based on SDN and NFV and provided an open source code programmer; Abidi et al. proposed a 5G slicing scheme based on machine learning and deep learning, and realized optimal weighted feature extraction based on deep learning and proposed an effective network slicing hybrid learning algorithm; Gao Liang et al. proposed a network resource management method based on SDN, and its simulation results showed that the bandwidth resource utilization rate was significantly improved. Most of the above studies realize the allocation of network resources from the dynamic changes of business requests, but do not combine the perception of current business traffic, that is, they only use simple methods to predict the future load state of the network to allocate resources. The above method may cause the problem of frequent migration of VNF (Virtual Network Feature), that is, once the system has network traffic overload, in order to ensure service reliability, the system starts to start data redirection, thereby realizing rapid data migration. Frequent migration of VNF will inevitably occupy a large amount of computing resources and storage resources, reduce system performance, and increase service latency for users. Therefore, based on the above requirements, it is necessary to propose a solution to predict the dynamic demand for resources of future services, so as to reduce the frequency of VNF migration caused by dynamic changes in service requests. Summary of the invention

[0003] The present invention aims to provide a network slicing configuration method based on traffic awareness to reduce the VNF migration frequency caused by dynamic changes in service requests, thereby avoiding the problem of increased end-to-end delay caused by network function virtualization scheduling to a certain extent.

[0004] In order to achieve the above object, the scheme of the present invention is:

[0005] A network slice configuration method based on traffic awareness is specifically performed in the following steps:

[0006] S1: The end-to-end network traffic reconstruction method based on compressed sensing reconstructs the network traffic at time t to obtain the reconstructed network traffic A;

[0007] S2: Obtain the network traffic B at the time T0+1 in the historical network data, substitute it into the LSTM network framework with unknown network parameters, perform LSTM network training, obtain LSTM network parameters, and output the LSTM prediction network;

[0008] S3: Substitute the reconstructed network traffic A obtained in step S1 into the LSTM prediction network obtained in step S2 to perform traffic load prediction on the reconstructed network traffic A, and then obtain the predicted load value of each network element;

[0009] S4: Combine NFV technology to realize intelligent allocation of network slice resources: compare the load threshold of each server in the current system with the predicted load value of all network elements, and allocate resources based on the predicted load values ​​of all network elements.

[0010] Through the above design, the unknown network traffic is reconstructed using the end-to-end network traffic reconstruction method based on compressed sensing, avoiding the large consumption of network resources caused by directly measuring network traffic. It has wide applicability. Compared with the expensive price of directly measuring network traffic, the estimation method adopted by the present invention is low-cost and conducive to vigorous promotion. LSTM and attention mechanism are combined to realize traffic prediction, and finally NFV technology is used to realize intelligent allocation of network slice resources. The present invention predicts the network traffic of the next time period through the network traffic at the current moment, and combines the predicted load value and the load threshold of the server in the system to perform reasonable allocation and planning, reasonably utilize and allocate system resources, and speed up the system network operation speed and processing efficiency.

[0011] Furthermore, in step S1, the specific content of reconstructing the network traffic A by the end-to-end network traffic reconstruction method based on compressed sensing is:

[0012] S11: Generate a matrix G with all zero entries = {g ij} L×N ; and construct a random 0-1 matrix R = {r ij} L×N ; Thus, the random measurement matrix Φ is obtained, and the OD flow signal X(t) of the acquired network traffic A is obtained; and then the measurement vector Z(t) is obtained;

[0013] Among them, g ij =1, ifp(i=k)>α; r ij The value of is 0 or 1;

[0014] L<<N, L=Klog(N / K), K represents the sparseness of network traffic signal; N is the number of flows of output node OD flow in network traffic A; K<<N;

[0015] k=1,2,…L;i=1,2,…L,j=1,2,…N;α is the probability threshold, 0≤α≤1;

[0016] The calculation formula of the random measurement matrix Φ is:

[0017] Φ={φ ij} L×N =R·*G={r ij ·*g ij} L×N ;

[0018] The OD flow signal X(t) of the network traffic A is:

[0019]

[0020] x i (t) represents the flow rate of the i-th OD flow at the t-th moment in the time period T;

[0021] The calculation formula of the measurement vector Z(t) is: Z(t)=ΦX(t);

[0022] S12: construct a sparse transformation matrix Ψ of the end-to-end network traffic at time t in time period T using K-SVD;

[0023] The steps to solve the sparse transformation matrix Ψ are:

[0024] The sparse representation of the OD flow X(t) of the network traffic A is:

[0025] X(t)=ΨS(t); that is, Z(t)=ΦX(t)=ΦΨS(t);

[0026] S(t) is the sparse signal of the OD flow X(t) of the end-to-end network traffic in the sparse transformation matrix Ψ transformation domain;

[0027] Let D∈R N×N ,X(t)∈R N , S(t)∈R n； ;

[0028] D represents the network traffic dictionary;

[0029] Then X = (X(1), X(2), ..., X(Z)) represents a set of Z network traffic signals;

[0030] S = (S(1), S(2), ..., S(Z)) represents a sparse representation vector set of the flow x;

[0031] Then the solution of the unknowns D and S is expressed as:

[0032]

[0033] Through the least squares ALS transformation, the solutions of the unknowns D and S are further improved through the multi-constrained optimization problem as follows:

[0034]

[0035] D=(d 1 , d 2 , …, d N ); k = 1, 2, ..., N;

[0036] S=(s1,s2,…,s N ) T ; k = 1, 2, ..., N;

[0037] E k is due to removing the atom D from all N samples K Errors caused by the composition of

[0038] Update D by singular value decomposition SVD K and k The solutions of the unknowns D and S are further improved through multi-constrained optimization problems, and the new dictionary is updated New sparse signals for network traffic dictionaries Through the new dictionary The sparse decomposition is obtained; then combined with iteration to obtain the final dictionary The final dictionary is the sparse transformation matrix Ψ of the sparse representation of network traffic, that is

[0039] The multi-constrained optimization problem can be further expressed as:

[0040]

[0041] Ω k is a matrix of size Z×|w k |, and (w k (i), the term on i) is equal to 1, otherwise it is 0;

[0042] Definitions kR =s k Ω k , X k R =XΩ k and E k R =E kΩ k

[0043] s kR It is from k The shrinkage result of removing zero input in ;

[0044] X k R From X k The shrinkage result of removing zero input in ;

[0045] E k R From E k The shrinkage result of removing zero input in ;

[0046] X k R Currently in d k The atomic sample set used in;

[0047] E k R is due to the removal of atoms d in the sample k The errors caused by

[0048] s k R The length of |w k |;

[0049] X k R and E k R The size of the matrix is ​​x×|w k |;

[0050] S13: Calculate the reconstructed sparse signal at time t

[0051]

[0052] S14: Reconstruct the sparse signal obtained by combining step S13 Get the initial reconstructed network traffic

[0053]

[0054] S15: Initial reconstruction of network traffic Optimize and get the OD flow signal at time t

[0055]

[0056] S16: Determine whether all moments in the time period T are reconstructed. If all are completed, output the reconstructed network traffic A of the time period T; otherwise, set t=t+1 and return to step S13.

[0057] Furthermore, the specific steps of step S2 are:

[0058] Assume that there are network traffic data at time T0+1, and use the network traffic data at time t0 to predict the network traffic data at time t0+1, t0 = T, T is the reconstruction time of compressed sensing;

[0059] S21: Network traffic from time i0 to time t0+i0-1 Put the data into the LSTM network framework to get the output set of each cell in the LSTM network Among them, i0 takes values ​​from 1 to T0-t0+1;

[0060] S22: Put the output set H of each cell of the LSTM network into the attention mechanism to obtain a reasonable attention coefficient α;

[0061] The attention mechanism calculation formula is:

[0062] M = tanh(Hwu);

[0063] α=softmax(M);

[0064] in, D is the length of the generated LSTM layer; w and u are learning parameters; and A=t0; tanh is the activation function; softmax is the normalization function;

[0065] S23: Calculate the output R of the attention layer. The calculation formula is:

[0066] R=(H T α) T ;

[0067] in,

[0068] S24: Put the output R of the attention layer into the fully connected layer of the LSTM network framework to obtain the predicted network traffic at time t2

[0069]

[0070] Among them, the value of t2 is t0+1 to T0+1;

[0071] Among them, W fcis the weight matrix; b fc is bias;

[0072] S25: Use SSE and variance to train LSTM network parameters J:

[0073] is the actual value of network traffic.

[0074] The beneficial effects of this solution are:

[0075] Through this design, the end-to-end network traffic reconstruction method based on compressed sensing can realize the estimation of unknown network traffic and be used to plan the subsequent network system. It avoids the large consumption of network resources caused by directly measuring network traffic. It has wide applicability. Compared with the expensive price of directly measuring network traffic, the estimation method adopted by the present invention has low cost and is conducive to vigorous promotion.

[0076] When estimating traffic, we effectively utilize the advantages of compressed sensing in signal reconstruction. Based on the compressed sensing framework, we can reconstruct the original signal with high accuracy under the premise of sample value. And when predicting network traffic, we use the attention mechanism to select the information that is relatively important to the current task from all inputs. One advantage of the attention mechanism is that they can determine which part of the input needs to be focused, and then assign greater weight to the important part. Therefore, the prediction is more accurate. BRIEF DESCRIPTION OF THE DRAWINGS

[0077] Figure 1 The network slice configuration flow chart of the present invention;

[0078] Figure 2 It is a flow chart of network traffic reconstruction of the present invention;

[0079] Figure 3 Calculate the flow chart for the LSTM prediction network;

[0080] Figure 4 Schematic diagram of the construction framework of the measurement vector obtained by partially measuring the OD flow;

[0081] Figure 5 This is a schematic diagram of the forget gate data input and output;

[0082] Figure 6 Schematic diagram of input gate data input and output;

[0083] Figure 7 This is a schematic diagram of the data input and output of the output gate;

[0084] Figure 8 This is a diagram of the end-to-end resource allocation process based on NFV technology. DETAILED DESCRIPTION

[0085] The following is further described in detail through specific implementation methods:

[0086] A network slice configuration method based on traffic awareness, combined with Figure 1 It can be seen that the specific steps are as follows:

[0087] S1: The end-to-end network traffic reconstruction method based on compressed sensing reconstructs the network traffic at time t to obtain the reconstructed network traffic A;

[0088] See also Figure 2 In step S1, the specific content of reconstructing network traffic A by the end-to-end network traffic reconstruction method based on compressed sensing is:

[0089] S11: Generate a matrix G with all zero entries = {g ij} L×N ; and construct a random 0-1 matrix R = {r ij} L×N ; Thus, the random measurement matrix Φ is obtained, and the OD flow signal X(t) of the acquired network traffic A is obtained; and then the measurement vector Z(t) is obtained;

[0090] Among them, g ij =1, ifp(i=k)>α; r ij The value of is 0 or 1;

[0091] L<<N, L=Klog(N / K), K represents the sparseness of network traffic signal; N is the number of flows of output node OD flow in network traffic A; K<<N;

[0092] k=1,2,…L;i=1,2,…L;j=1,2,…N;α is the probability threshold, 0≤α≤1;

[0093] The calculation formula of the random measurement matrix Φ is:

[0094] Φ={φ ij} L×N =R·*G={r ij ·*g ij} L×N ;

[0095] The OD flow signal X(t) of the network traffic A is:

[0096]

[0097] x i (t) represents the flow rate of the i-th OD flow at the t-th moment in the time period T;

[0098] ci (t) represents the capacity of the path passed by the i-th OD flow at time t;

[0099] A(t) represents the routing information matrix at time t;

[0100] Y(t)=(y1(t),y2(t),…,y M1 (t)) represents the traffic of all links in the network; y M1 is the traffic of link No. M1;

[0101] M1 is the total number of links;

[0102] We use compressed sensing theory and framework to solve the end-to-end network traffic reconstruction problem shown in Eq. According to compressed sensing theory, we project the network traffic X(t) onto a set of metric bases Φ = {φ1, φ2, …φ j …,φ N}Then the calculation formula of the measurement vector Z(t) is: Z(t)=ΦX(t);

[0103] Z(t)=(z1(t),z2(t),…,z L (t)) represents the measurement data of compressed sensing; L<<N;

[0104] The size of the random measurement matrix Φ is L×N;

[0105] φ j represents the jth measurement basis; where j = 1, 2, ..., N;

[0106] S12: construct a sparse transformation matrix Ψ of the end-to-end network traffic at time t in time period T using K-SVD;

[0107] The steps to solve the sparse transformation matrix Ψ are:

[0108] The sparse representation of the OD flow X(t) of the network traffic A is:

[0109] X(t)=ΨS(t); that is, Z(t)=ΦX(t)=ΦΨS(t);

[0110] S(t) is the sparse signal of the OD flow X(t) of the end-to-end network traffic in the sparse transformation matrix Ψ transformation domain;

[0111] Let D∈R N×N ,X(t)∈R N , S(t)∈R n； ;

[0112] D represents the network traffic dictionary;

[0113] Then X = (X(1), X(2), ..., X(Z)) represents a set of Z network traffic signals;

[0114] S = (S(1), S(2), ..., S(Z)) represents a sparse representation vector set of the flow x;

[0115] Then the solution of the unknowns D and S is expressed as:

[0116]

[0117] Through the least squares ALS transformation, the solutions of the unknowns D and S are further improved through the multi-constrained optimization problem as follows:

[0118]

[0119] D=(d 1 , d 2 , …, d N ); k = 1, 2, ..., N;

[0120] S=(s1,s2,…,s N ) T ; k = 1, 2, ..., N;

[0121] E k is due to removing the atom D from all N samples K Errors caused by the composition of

[0122] Update D by singular value decomposition SVD K and k The solutions of the unknowns D and S are further improved through multi-constrained optimization problems, and the new dictionary is updated New sparse signals for network traffic dictionaries Through the new dictionary The sparse decomposition is obtained; then combined with iteration to obtain the final dictionary The final dictionary is the sparse transformation matrix Ψ of the sparse representation of network traffic, that is

[0123] The multi-constrained optimization problem can be further expressed as:

[0124]

[0125] Ω k is a matrix of size Z×|w k |, and (w k (i), the term on i) is equal to 1, otherwise it is 0;

[0126] Definitions kR =sk Ω k , X k R =XΩ k and E k R =E k Ω k

[0127] s kR It is from k The shrinkage result of removing zero input in ;

[0128] X k R From X k The shrinkage result of removing zero input in ;

[0129] E k R From E k The shrinkage result of removing zero input in ;

[0130] X k R Currently in d k The atomic sample set used in;

[0131] E k R is due to the removal of atoms d in the sample k The errors caused by

[0132] s k R The length of |w k |;

[0133] X k R and E k R The size of the matrix is ​​x×|w k |;

[0134] S13: Calculate the reconstructed sparse signal at time t

[0135]

[0136] S14: Reconstruct the sparse signal obtained by combining step S13 Get the initial reconstructed network traffic

[0137] S15: Initial reconstruction of network traffic Optimize and get the OD flow signal at time t

[0138]

[0139] S16: Determine whether all moments in the time period T are reconstructed. If all are completed, output the reconstructed network traffic A of the time period T; otherwise, set t=t+1 and return to step S13.

[0140] S2: Obtain the network traffic B at the time T0+1 in the historical network data, substitute it into the LSTM network framework with unknown network parameters, perform LSTM network training, obtain LSTM network parameters, and output the LSTM prediction network;

[0141] See also Figure 3 , the specific steps of step S2 are:

[0142] Assuming that there are historical network traffic data at time T0+1, the network traffic data at time t0 is used to predict the network traffic data at time t0+1; t0=T, T is the reconstruction time of compressed sensing;

[0143] S21: Network traffic from time i0 to time t0+i0-1 Put the data into the LSTM network framework to get the output set of each cell in the LSTM network Among them, i0 takes values ​​from 1 to T0-t0+1;

[0144] S22: Put the output set H of each cell of the LSTM network into the attention mechanism to obtain a reasonable attention coefficient α;

[0145] The attention mechanism calculation formula is:

[0146] M=tanh(Hwu)

[0147] α=softmax(M);

[0148] in, D is the length of the generated LSTM layer; w and u are learning parameters; and A=t0; tanh is the activation function; softmax is the normalization function;

[0149] S23: Calculate the output R of the attention layer. The calculation formula is:

[0150] R=(H T α) T ;

[0151] in,

[0152] S24: Put the output R of the attention layer into the fully connected layer of the LSTM network framework to obtain the predicted network traffic at time t2

[0153]

[0154] Among them, the value of t2 is t0+1 to T0+1;

[0155] Among them, W fc is the weight matrix; b fc is bias;

[0156] S25: Use SSE and variance to train LSTM network parameters J:

[0157] is the actual value of network traffic.

[0158] S3: Substitute the reconstructed network traffic A obtained in step S1 into the LSTM prediction network obtained in step S2 to perform traffic load prediction on the reconstructed network traffic A, and then obtain the predicted load value of each network element;

[0159] S4: Combine NFV technology to realize intelligent allocation of network slice resources: compare the load threshold of each server in the current system with the predicted load value of all network elements, and allocate resources based on the predicted load value of all network elements. Figure 8 .

[0160] In order to explain the accuracy of the above network traffic processing reconstruction and prediction in more detail, the following deduction process is used to illustrate:

[0161] 1. End-to-end network traffic reconstruction method based on compressed sensing to reconstruct network traffic

[0162] The end-to-end network traffic matrix reflects the inherent behavior of network users and the characteristics of network equipment for data transmission and information exchange. For a backbone network with n nodes, the number of OD flows is N = n 2 For the backbone network, the end-to-end network traffic can be modeled as:

[0163]

[0164] Where i = 1, 2, ..., N, xi(t) represents the flow rate of the i-th OD flow at time t, c i (t) represents the capacity of the path that the i-th OD flow passes through at time t, and A(t) represents the routing information matrix at time t. M(t)) represents the traffic of all links in the network, M is the total number of links, y M is the flow of the mth link. According to (2)-(4), the network flow x can be found in formula (1): i (t). However, the number of links in the actual network is much smaller than the number of OD flows, that is, M < < n. In other words, equation (2) is not a square matrix, resulting in a highly ill-posed inverse problem. Therefore, it is difficult to obtain an exact solution for end-to-end network flows based on equations (2)-(4).

[0165] In this embodiment, compressed sensing theory and framework are used to solve the end-to-end network traffic reconstruction problem shown in the equation.

[0166] According to the theory of compressed sensing, the network traffic X(t) is projected onto a set of metric bases Φ = {φ1, φ2, …, φ M0}. Then the measurement data can be obtained:

[0167] Z(t)=ΦX(t) (5)

[0168] where Z(t)=(z1(t),z2(t),…,z L (t)) represents the compressed sensing measurement data, L<<N. Φ is the random measurement matrix, the size is L×N. φ i Represents the i-th measurement basis, where i = 1, 2, …, M0, Φ is mainly used to project the network traffic X(t) into the measurement data matrix Z(t).

[0169] According to formula (5), if the low-dimensional measurement data Z(t) can be accurately obtained, then the end-to-end network traffic X(t) can be reconstructed using the compressed sensing theory. However, end-to-end network traffic is not sparse in practice. Existing traffic measurement literature shows that end-to-end network traffic is mainly composed of a few large outer diameter flows (elephant flows), which account for 80% of the entire outer diameter flow. If only the elephant flow is considered, according to the signal sparsity theory, the end-to-end network traffic X(t) can be sparsely represented by the transformation matrix Ψ as:

[0170]

[0171] Ψ=(ψ1,ψ2,…,ψ N ) represents the sparse transformation matrix of the network traffic signal;

[0172] S(t) is the sparse projection of the end-to-end network traffic X(t) in the Ψ transform domain; there are K non-zero rows (where K<<N).

[0173] According to equations (5) and (6), we can get equation (7):

[0174] Z(t)=ΦX(t)=ΦΨS(t) (7)

[0175] If the measurement signal Z(t) is obtained, the sparse signal S(t) can be accurately reconstructed according to the compressed sensing theory. Therefore, the end-to-end network traffic X(t) can be accurately reconstructed according to equation (6), which prompted the inventors to consider using compressed sensing according to equations (1)-(4) to solve the end-to-end network traffic reconstruction problem.

[0176] The random measurement matrix Φ is constructed as follows:

[0177] The random measurement matrix Φ has the following reconstruction properties:

[0178] 1. In order to extract the OD flow to be measured using the random measurement matrix Φ, the random measurement matrix Φ should be 0 or 1 and obey a random distribution. The random measurement matrix Φ equal to 1 indicates the OD flow to be measured, and the random measurement matrix Φ equal to 0 indicates the OD flow not to be measured.

[0179] 2. The random measurement matrix Φ represents the linear combination relationship between the measurement signal Z(t) and the end-to-end network traffic X(t). The row vector in the random measurement matrix Φ can have multiple non-zero entries, and the entries of the measurement signal Z(t) can be represented as a linear combination of multiple OD flows.

[0180] 3. The random measurement matrix Φ indicates that a certain OD flow is only related to a certain item in the measurement signal Z(t). Each column (row) vector in Φ has only one non-zero entry, and an OD flow is collected into only one entry in the measurement signal Z(t).

[0181] 4. The random measurement matrix Φ represents the aggregation relationship between the measurement signal Z(t) and the network traffic X(t). The OD flow is aggregated into the measurement signal Z(t).

[0182] 5. The random measurement matrix Φ represents the conversion relationship from high-dimensional network traffic X(t) to low-dimensional measurement signal Z(t).

[0183] Based on the above properties, the following Algorithm 1 is proposed: Construct a random measurement matrix Φ of end-to-end network traffic:

[0184] (1) According to Property 1, we obtain a completely zero matrix G = {g ij} L×N , where L<<N, L=K log(N / K), and K represents the sparsity of network traffic signals.

[0185] (2) According to properties 1 and 2, from the completely zero matrix G = {g ij} L×N Randomly select d positions from each column vector of , and set them as:

[0186] gij =1, if p(i=k)>α (8)

[0187] Where k is a variable, k = 1, 2, ..., L,i is also a variable, i = 1, 2, ..., L,j = 1, 2, ... N; α is the probability threshold, 0 ≤ α ≤ 1. For each column vector of the completely zero matrix G, d random positions are selected and updated according to equation (8). Experiments show that the smaller the value of d, the higher the sparsity of the measurement matrix finally constructed, and the more accurate the end-to-end network traffic reconstruction.

[0188] (3) According to attribute 3, let the number of randomly selected positions d = 1

[0189] (4) Construct a random 0-1 matrix R = {r ij} L×N , the random measurement matrix is ​​as follows:

[0190] Φ={φ ij} L×N =R·*G={r ij ·*g ij} L×M (9)

[0191] If R is a randomly generated 0-1 matrix, the rank of the random measurement matrix Φ obtained by formula (9) is not greater than the rank of the completely zero matrix G. Therefore, the linear correlation of the column vectors of the random measurement matrix Φ is changed compared with the completely zero matrix G. If R is a randomly generated 0-1 matrix, the sparsity of the random measurement matrix Φ obtained by formula (9) is enhanced. It can be seen from formula (9) that we can improve the sparsity of the matrix G by randomly generating the 0-1 matrix R. Subsequently, the accuracy of end-to-end network traffic reconstruction based on compressed sensing will be improved.

[0192] (5) A subset of the OD flow to be measured is determined by the random measurement matrix Φ; Eq. (11) indicates that by performing a "sum" calculation on the column vectors of the random measurement matrix Φ, the non-zero position corresponding to the outer diameter flow to be measured can be correctly obtained.

[0193] M OD ={OD j |φ ij ∩1=1} (11)

[0194] (6) According to properties 4 and 5, we will obtain the following measurement vector:

[0195]

[0196] Combination Figure 4 , for the end-to-end traffic flow vector The dark grey boxes represent the partially measured OD flows determined by the measurement matrix, and the white boxes represent the OD flows that need to be reconstructed.

[0197] The construction of the sparse transformation matrix Ψ is:

[0198] From the theory of compressed sensing, we know that since the traffic matrix itself is not sparse, it is impossible to correctly and accurately reconstruct the OD flow by only constructing a random measurement matrix. Therefore, it is necessary to find a suitable sparse transformation matrix Ψ so that the network traffic X(t) can be sparsely represented by the sparse transformation matrix Ψ.

[0199] We use the K-SVD algorithm to construct our sparse matrix by constructing an overcomplete dictionary. The K-SVD algorithm has strong sparse representation capabilities and low computational complexity.

[0200] According to K-SVD, let D∈R N×N ,X(t)∈R N and S(t)∈R n Denote the network traffic dictionary, network traffic signal, and sparse representation of network traffic, respectively. Therefore, X = (X(1), X(2), ..., X(Z)) represents a set of Z network traffic signals. S = (S(1), S(2), ..., S(Z)) represents the sparse representation vector set of x. In order to construct the sparse representation matrix of end-to-end network traffic, according to K-SVD, the solution of D and S can be expressed as the following multi-constrained optimization problem:

[0201]

[0202] N is the upper limit of the number of non-zero entries in the sparse representation of end-to-end network traffic. Equation (13) contains two unknown variables D and s. Alternating least squares (ALS) is used to obtain the optimal solution of the optimization process (13)-(16) shown in the above equation.

[0203] First, given a network traffic dictionary D, we determine a sparse representation S of the network traffic using the ALS method.

[0204] Then, the ALS method can be further used to better determine the network traffic dictionary D based on the obtained network traffic sparse representation S. Update one column of the network traffic dictionary D at a time. If you want to update the kth column of the network traffic dictionary D (where k = 1, 2, ..., N, D = (d 1 , d 2 , …, d N )), let the kth column of the network traffic sparse matrix S correspond to s k (where k = 1, 2, ..., N, S = (s1, s2, ..., s N ) T ), then the above multi-constraint optimization problem can be expressed as:

[0205]

[0206] Equation (17) divides the multiplication DS into the sum of K rank 1 matrices. Given the K-1 entries, the kth column is the vector to be processed. The matrix E k represents the error due to removing the atom D from all N samples K The components of . Equations (18)-(20) represent the constraints. Then, singular value decomposition (SVD) is used to update D K and k According to the singular value decomposition (SVD), we can find the closest E k The rank 1 matrix of . Then the error of equation (17) can be effectively reduced. However, the s obtained in this way k Most of the entries in will be non-zero vectors. Therefore, we cannot force an update to d k to satisfy the sparsity constraint. That is, due to the influence of zero values, when updating the vector at non-zero position, s k changes and divergence occurs. To solve this problem, delete s k All 0s in D and keep non-zero values. When using SVD to update D K and k , divergence can be avoided.

[0207] Define w k ={i|1≤i≤N, s ki ≠0} as d k The set of indices {X(1), X(2), …, X(Z)} of all signal sets used, i.e., s ki ≠0 when the point index (where s ki represents the entry in the kth row and ith column of the matrix S). Define Ω k is a matrix of size Z×|w k |, and (w k (i), i) is equal to 1, otherwise it is 0. Define s kR =s k Ω k , X k R =XΩ k and E k R =E k Ω k From s k , X and E k The shrinking result is to remove zero inputs from X. k R Currently in d k The atomic sample set used in E kR is due to the removal of atoms d in the sample k The error caused by , and the effect of removal is not considered. k R The length of |w k |.X k R and E k R The size of the matrix is ​​x×|w k |. Equations (17)-(20) can be rewritten as:

[0208]

[0209] After updating, you can get new New Sparse Network Traffic Dictionary It can be expressed through a dictionary The final dictionary can be obtained by iterative method. It can be regarded as a sparse representation matrix ψ of network traffic, that is,

[0210] End-to-end network traffic reconstruction content:

[0211] Based on the above discussion, the measurement matrix Φ and the sparse matrix Ψ are successfully constructed. Therefore, we can reconstruct the sparse signal S(t) at time t according to the compressed sensing theory. The reconstruction process can be expressed as the following optimization problem:

[0212]

[0213] Then according to formula (5), we can get the reconstruction result of the end-to-end network traffic at time t, that is:

[0214]

[0215] Due to the inherent characteristics of the complex end-to-end network traffic, the result obtained by equation (26) may be inaccurate and contain certain errors. In order to further reduce the reconstruction error, according to the end-to-end network traffic model represented by equations (1)-(4), the reconstruction method represented by equation (26) is expressed as a constrained optimization process. Based on this, it can be obtained:

[0216]

[0217] 2. Use LSTM network and attention mechanism to realize traffic prediction.

[0218] LSTM networks process long sequences by introducing a memory cell and three additional gates: input gate, forget gate, and output gate. Specifically, the input gate determines the information that goes into the memory cell. The output gate produces the output at each time step. The forget gate updates the state equation of the memory cell.

[0219] The calculation process of LSTM unit:

[0220] Forget gate: The first step in an LSTM is to decide what information to discard from the cell state. This decision is made through a layer called a forget gate. This gate reads h t-1 and x t , output a value between 0 and 1 for each cell in state C t-1 The number in the table is 1, which means "totally retain" and 0, which means "totally discard".

[0221] Combination Figure 5 It can be seen that for the forget gate data input and output process, the following relationship exists:

[0222] f t =σ(W f ·[h t-1 ,x t ]+b f )

[0223] where h t-1 represents the output of the previous cell, x t Represents the input of the current cell. σ represents the sigmoid function.

[0224] Input gate: The next step is to decide how much new information to add to the cell state. This requires two steps: first, a sigmoid layer called the "input gate layer" determines which information needs to be updated; a tanh layer generates a vector That is, the alternative content for updating. In the next step, we combine these two parts to update the state of the cell.

[0225] See also Figure 6 , is the input gate data input and output process, and the following relationship exists:

[0226] i t =σ(W i ·[h t-1 , x t ]+b i );

[0227]

[0228] Now it is time to update the old cell state, Ct-1 Updated to C t The previous steps have determined what will be done, and now we need to actually do it. We compare the old state with f t Multiplying, discarding the information we are sure we need to discard. Then add These are the new candidate values, varying according to how much we decide to update each state.

[0229] Output gate: The output gate consists of two parts. The first part also takes the hidden state of the previous moment and the input variable of the current moment as input, and controls the output range through the sigmoid function. The second part controls the output range of the hidden state through the tanh activation function, and then multiplies it with the output result of the first part to update the hidden state. Figure 7 The input-output relationship of the output gate is:

[0230] o t =σ(W o [h t-1 , x t ]+b o )

[0231] h t =o t *tanh(C t )

[0232] Attention mechanism:

[0233] Here, two learnable parameters w and u are used to improve the representation ability of the attention layer. The reasonable attention coefficient a indicates the importance of each element in H. In this case, the attention layer can assign more reasonable weights to different elements in the input H.

[0234] M = tanh(Hwu);

[0235] α=softmax(M);

[0236] R=(H T α) T .

[0237] in D is the length of the resulting LSTM layer, and is a learning parameter, A is set equal to t0, t0 = T, T is the number of reconstruction times in compressed sensing. The function tanh is the activation function, as shown in the following equation. Here, two learning parameters are used to better map the correlation between different output vectors, thereby generating the attention parameter Equation (6) is the softmax function used to normalize M to calculate a. is the output of the attention layer.

[0238]

[0239]

[0240] 3. Using NFV technology to achieve intelligent allocation of network slice resources

[0241] Server load is related to many factors. Considering that the server load of the 5G core network is mainly related to NFV network elements, this paper takes the traffic perception of different types of NFV network elements as input, and the traffic prediction of different NFV network elements as server load to achieve traffic-based load prediction, and then combines the threshold set by the system to migrate some NFV network elements with large changes to other servers. Based on the network function virtualization technology, the complete decoupling of software and hardware is achieved, thereby realizing the intelligent migration out and migration in based on NFV network elements. See Figure 8 , which is the process of intelligent allocation of network slice resources based on NFV technology.

[0242] Based on NFV technology, network equipment resources can be integrated into a unified resource pool, and the resource pool can be dynamically scheduled according to the abstracted network services and functional requirements. Figure 8 In the figure, it is shown that when NFV network element 4 and network element 5 share a server and the load is too large, the system migrates some NFV network elements out of the original server and migrates them to a server that meets the traffic demand according to the traffic perception level of NFV network element 4 and network element 5, thereby realizing intelligent allocation of network resources.

[0243] The above is only an embodiment of the present invention, and the common knowledge such as the known specific structure and characteristics in the scheme is not described in detail here. It should be pointed out that for those skilled in the art, several modifications and improvements can be made without departing from the structure of the present invention, which should also be regarded as the protection scope of the present invention, and these will not affect the effect of the implementation of the present invention and the practicality of the present invention. The specific implementation methods and other records in the specification can be used to interpret the content of the claims.

Claims

1. A network slice configuration method based on traffic awareness, characterized in that: Follow these steps: S1: The end-to-end network traffic reconstruction method based on compressed sensing reconstructs the network traffic at time t to obtain the reconstructed network traffic A; S2: Obtain the network traffic B at the time T0+1 in the historical network data, substitute it into the LSTM network framework with unknown network parameters, perform LSTM network training, obtain LSTM network parameters, and output the LSTM prediction network; S3: Substitute the reconstructed network traffic A obtained in step S1 into the LSTM prediction network obtained in step S2 to perform traffic load prediction on the reconstructed network traffic A, and then obtain the predicted load value of each network element; S4: Combine NFV technology to realize intelligent allocation of network slice resources: compare the load threshold of each server in the current system with the predicted load value of all network elements, and allocate resources according to the predicted load values ​​of all network elements; In step S1, the specific content of reconstructing the network traffic at time t by the end-to-end network traffic reconstruction method based on compressed sensing is as follows: S11: Generate a matrix G with all zero entries = {g ij } L×N ; and construct a random 0-1 matrix R = {r ij } L×N ; Thus, the random measurement matrix Φ is obtained, and the OD flow signal X(t) of the acquired network traffic A is obtained; and then the measurement vector Z(t) is obtained; Among them, g ij =1, if p(i=k)>α; r ij The value of is 0 or 1, fp(i=k)>α means that if the probability of i=k is greater than α, the value g in the i-th row and j-th column of the all-zero matrix ij =1; r ij Represents the value of the i-th row and j-th column in a 0-1 matrix; L<<N, L=K log(N / K), K represents the sparseness of network traffic signal; N is the number of flows of output node OD flow in network traffic A; K<<N; k=1,2,…L;i=1,2,…L,j=1,2,…N;α is the probability threshold, 0≤α≤1; The calculation formula of the random measurement matrix Φ is: Φ={φ ij } L×N =R·*G={r ij ·*g ij } L×N ; The OD flow signal X(t) of the network traffic A is: x i (t) represents the flow rate of the i-th OD flow at the t-th moment in the time period T; The calculation formula of the measurement vector Z(t) is: Z(t)=ΦX(t); Φ ij represents the value of the i-th row and j-th column of the random measurement matrix Φ; Y(t)=(y1(t),y2(t),…,y M1 (t)) represents the traffic of all links in the network; S12: construct a sparse transformation matrix Ψ of the end-to-end network traffic at time t in time period T using K-SVD; The steps to solve the sparse transformation matrix Ψ are: The sparse representation of the OD flow X(t) of the network traffic A is: X(t)=ΨS(t); that is, Z(t)=ΦX(t)=ΦΨS(t); S(t) is the sparse signal of the OD flow X(t) of the end-to-end network traffic in the sparse transformation matrix Ψ transformation domain; Let D∈R N×N ,X(t)∈R N , S(t)∈R N ; D represents the network traffic dictionary; Then X = (X(1), X(2), ..., X(Z)) represents a set of Z network traffic signals; S = (S(1), S(2), ..., S(Z)) represents a sparse representation vector set of the flow x; Z is the number of network traffic signals; Then the solution of the unknowns D and S is expressed as: Through the least squares ALS transformation, the solutions of the unknowns D and S are further improved through the multi-constrained optimization problem as follows: D=(d 1 ,d 2 ,…,d N );k=1,2,…,N; S=(s1,s2,…,s N ) T ;k=1,2,…,N; E k is due to removing the atom D from all N samples K Errors caused by the composition of Update D by singular value decomposition SVD K and k The solutions of the unknowns D and S are further improved through multi-constrained optimization problems, and the new dictionary is updated New sparse signals for network traffic dictionaries Through the new dictionary The sparse decomposition is obtained; then the final dictionary ^D is obtained by combining iterations; the final dictionary ^D is the sparse representation sparse transformation matrix Ψ of the network traffic, that is, ψ=^D; The multi-constrained optimization problem can be further expressed as: Ω k is a matrix of size Z×|w k |, and (w k (i), the term on i) is equal to 1, otherwise it is 0; Definitions kR =s k Ω k , X k R =XΩ k and E k R =E k Ω k s kR It is from k The shrinkage result of removing zero input in ; X k R From X k The shrinkage result of removing zero input in ; E k R From E k The shrinkage result of removing zero input in ; X k R Currently in d k The atomic sample set used in; E k R is due to the removal of atoms d in the sample k The errors caused by s k R The length of |w k |; X k R and E k R The size of the matrix is ​​x×|w k |; S13: Calculate the reconstructed sparse signal ^S(t) at time t; S14: Combine the reconstructed sparse signal ^S(t) obtained in step S13 to obtain the initial reconstructed network traffic To measure the signal; S15: Initial reconstruction of network traffic Optimize and get the OD flow signal at time t S16: Determine whether all moments in the time period T are reconstructed. If all are completed, output the reconstructed network traffic A of the time period T; otherwise, set t=t+1 and return to step S13.

2. The method for configuring a network slice based on traffic awareness according to claim 1, characterized in that: The specific steps of step S2 are: Assume that there are a total of historical network traffic data at time T0+1, and use the network traffic data with a data length of t0 to predict the network traffic data at time t0+1, t0 = T, T is the reconstruction time of compressed sensing; S21: Network traffic from time i0 to time t0+i0-1 Put the data into the LSTM network framework to get the output set of each cell in the LSTM network Among them, i0 takes values ​​from 1 to T0-t0+1; S22: Put the output set H of each cell of the LSTM network into the attention mechanism to obtain a reasonable attention coefficient α; The attention mechanism calculation formula is: M = tanh(Hwu); α=softmax(M); in, D is the length of the generated LSTM layer; w and u are learning parameters; and A=t0; tanh is the activation function; softmax is the normalization function; S23: Calculate the output R of the attention layer. The calculation formula is: R=(H T a) T ; in, S24: Put the output R of the attention layer into the fully connected layer of the LSTM network framework to obtain the predicted network traffic at time t2 Among them, the value of t2 is from t0+1 to T0+1; Among them, W fc is the weight matrix; b fc is bias; S25: Use SSE and variance to train LSTM network parameters J: is the actual value of network traffic.

Images