A VNF-based multidimensional resource demand forecasting method based on the fusion of synchronous and asynchronous spatiotemporal characteristics.

Abstract

This invention proposes a VNF multidimensional resource demand prediction method based on the fusion of synchronous and asynchronous spatiotemporal characteristics. The implementation steps are as follows: obtaining auxiliary sample sets, training sample sets, and test sample sets; constructing a VNF multidimensional resource demand prediction network model based on the fusion of synchronous and asynchronous spatiotemporal characteristics and iteratively training it; and obtaining the VNF multidimensional resource demand prediction results. In the process of iteratively training the prediction network model and obtaining the prediction results, the synchronous and asynchronous spatiotemporal feature fusion module replaces the normalized adjacency matrix of the graph with a synchronous and asynchronous adjacency matrix. This allows for the simultaneous fusion of synchronous spatial features, asynchronous spatial features, and temporal features of VNFs, avoiding the shortcomings of existing technologies that ignore asynchronous spatial features between different VNFs, and effectively improving the prediction accuracy of VNF multidimensional resource demand.

Description

Technical Field

[0001] This invention belongs to the field of network and communication technology, and relates to a VNF multidimensional resource demand prediction method based on the fusion of synchronous and asynchronous spatiotemporal characteristics, which can be applied to predict VNF ​​multidimensional resource demand. Background Technology

[0002] Network Functions Virtualization (NFV) is a network architecture and concept that uses virtualization technology to separate traditional network functions (such as routing, switching, firewalls, load balancing, etc.) from dedicated hardware devices and deploy them on general-purpose hardware devices. Virtualized Network Functions (VNFs) are concrete implementations within the NFV framework, representing specific instances of network functions in a virtualized environment. Each VNF implements a specific network function, such as a firewall, router, or switch, and no longer depends on dedicated hardware but runs on a virtualization platform. Multiple VNFs are linked together in a specific order to form a Service Functions Chain (SFC). Through the SFC, data flows through a series of VNFs in a predetermined order, thereby enabling complex network services. The CPU, memory, and bandwidth requirements of running VNFs are highly dynamic. On the one hand, without prediction, resource over-allocation may occur, leading to resource waste; on the other hand, insufficient resources may cause VNFs to malfunction, affecting the quality of network services. By predicting the CPU, memory, and bandwidth resource requirements (multi-dimensional resource requirements) of VNFs, resource scheduling can be performed in advance based on future resource needs, achieving elastic scaling of resources.

[0003] In July 2024, Tang Lun et al. proposed a prediction model (DT-TCN) based on Graph Convolutional Neural Network (GCN) and Temporal Convolutional Network (TCN) in their patent application CN119030877A entitled "A Method for Active Migration of Virtual Network Functions Based on Resource Prediction." This model extracts spatiotemporal features based on the SFC topology to improve the accuracy of VNF resource demand prediction. GCN extracts spatial features based on the graph adjacency matrix, and TCN is used to extract global information from the time series. The drawback of this method is that when using GCN to extract spatial features, it can only extract the synchronous spatial correlations between different VNFs, ignoring the asynchronous effects between different VNFs. Summary of the Invention

[0004] The purpose of this invention is to overcome the shortcomings of the existing technology and propose a VNF multidimensional resource demand prediction method based on the fusion of synchronous and asynchronous spatiotemporal characteristics, which aims to improve the accuracy of VNF multidimensional resource demand prediction.

[0005] To achieve the above objectives, the technical solution adopted by the present invention includes the following steps:

[0006] (1) Obtain the auxiliary sample set, training sample set, and test sample set:

[0007] Normalize the multidimensional resource demand data for each of the T time steps in each of the N VNFs, and then partition the multidimensional resource demand data for each of the S time steps in each VNF. The first U data blocks of length M and their labels are then used to form the training sample set, and the remaining V data blocks are used to form the test sample set. The unpartitioned normalized N×T data blocks are then used as the test sample set. P An auxiliary sample set is composed of multiple dimensions of resource demand data, where N≥1, and T=S+T P T≥10000, S=U+V+M, M≥5;

[0008] (2) Construct a VNF multidimensional resource demand prediction network model based on the fusion of synchronous and asynchronous spatiotemporal characteristics:

[0009] A VNF multidimensional resource demand prediction network model O is constructed, which includes a cascaded enhanced feature extraction module, a time delay correlation extraction module, a synchronous and asynchronous adjacency matrix construction module, and a synchronous and asynchronous spatiotemporal feature fusion module.

[0010] (3) Iteratively train the prediction network model:

[0011] The prediction network model O is iteratively trained using auxiliary and training sample sets to obtain the trained prediction network model O. * ;

[0012] (4) Obtain the VNF multidimensional resource demand forecast results:

[0013] The test sample set is used as the trained prediction network model O. * The input is propagated forward to obtain the predicted value for each test sample.

[0014] Compared with the prior art, the present invention has the following advantages:

[0015] In the process of iteratively training the prediction network model and obtaining the prediction results, the synchronous and asynchronous spatiotemporal feature fusion module of this invention replaces the normalized adjacency matrix of the graph with a synchronous and asynchronous adjacency matrix. It can simultaneously fuse the synchronous spatial features, asynchronous spatial features and temporal features of VNF, avoiding the defect of existing technologies that ignore the asynchronous spatial features between different VNFs, and effectively improving the prediction accuracy of VNF multidimensional resource requirements. Attached Figure Description

[0016] Figure 1 This is a flowchart illustrating the implementation of the present invention;

[0017] Figure 2 This is a comparison chart of simulation results showing the prediction accuracy of the present invention and existing technologies. Detailed Implementation

[0018] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments.

[0019] Reference Figure 1 The present invention includes the following steps:

[0020] Step 1: Obtain the auxiliary sample set, training sample set, and test sample set:

[0021] Step 1a) For each of the N VNFs, the multidimensional resource demand data X = {X1, X2, ..., Xn} consists of T consecutive time steps. n ,…,X N Perform Min-Max normalization, where N = 6, T = 15000, and the time interval for each time step is 30 seconds. It is VNF n A total of T time steps of multidimensional resource demand data. It is VNF n Multidimensional resource demand data at time step t Representing VNF in sequence n The CPU resource requirements, memory requirements, and bandwidth requirements at time step t are... VNF n For the r-th time step, r∈{cpu,mem,bdw} The normalization formula is:

[0022]

[0023] in, yes The normalized result, x r,min x r,maxThese are the minimum and maximum values ​​among N×T r-type resource demand data, respectively. Data normalization scales the data proportionally to make it the same data interval and range, so as to reduce the impact of scale, features, distribution differences, etc. on the model, which can effectively improve the convergence speed and fitting effect of the model.

[0024] Step 1b) Using a sliding window of length M+1 time steps, the multidimensional resource demand data of each normalized VNF (Variable Functional Array) is divided into S time steps with a step size of one time step. The first M data points from the M+1 time steps of multidimensional resource demand data in each of the N VNFs after each division are used as data blocks, with the last data point serving as the label. This process is repeated SM times to obtain SM data blocks of size N×M and their labels. The first U data blocks of length M after division and their labels form the training sample set, and the remaining V data blocks form the test sample set. The undivided normalized N×T data blocks are then used as the test sample set. P A multidimensional resource demand data set is used to extract enhanced features, where S = 13000 and T... p =2000, M=9; Each sample obtained from the above partitioning is the historical resource demand data of 6 VNFs, each containing 9 time steps. The corresponding label is the multidimensional resource demand data of the next time step of the 6 VNFs. That is, the multidimensional resource demand data of the first 9 time steps of each VNF is used to predict the multidimensional resource demand data of the VNF in the next time step.

[0025] Step 2: Construct a VNF multidimensional resource demand prediction network model O, which includes a cascaded enhanced feature extraction module, a time-delay correlation extraction module, a synchronous / asynchronous adjacency matrix construction module, and a synchronous / asynchronous spatiotemporal feature fusion module.

[0026] Step 2a) Constructing an Enhanced Feature Extraction Module: Since the multidimensional resource requirements of each VNF—namely, CPU resource requirements, memory resource requirements, and bandwidth resource requirements—change over time with different trends, it is necessary to extract a time series for each VNF that reflects the overall trend of these three dimensions of features, i.e., an enhanced feature, to analyze the latency correlation between different VNFs; VNF n The enhanced features can be represented as in The calculation formula is:

[0027]

[0028] It is VNF n At the t p With t p -1 time step of the Euclidean distance of multidimensional resource demand data, representing VNF nThe overall change in 3D resource demand data at this time step relative to the 3D resource demand data at the previous time step; The sign is determined by a voting mechanism, that is: in the three-dimensional resource demand data, if more than one resource demand data point has increased relative to the previous time step, then... It is positive; otherwise, it is negative.

[0029] Step 2b) Constructing the latency correlation extraction module: The latency correlation extraction module includes a cascaded latency correlation calculation submodule and a maximum correlation latency extraction submodule;

[0030] The time-delay correlation calculation submodule calculates the enhanced feature set EF = {EF1, EF2, ... EF} obtained by the enhanced feature extraction module. n ,…,EF N}, calculate each pair of adjacent VNFs separately i and VNF j Enhanced features EF i With EF j,Δt Delay correlation The calculation formula is:

[0031]

[0032] Where DR(·) represents the operation of calculating the Spearman correlation coefficient, and Δt is the VNF. j Enhanced features EF j Relative to VNF i Enhanced features EF i The delayed time step; Spearman correlation coefficient is a statistic that measures the strength of the monotonic relationship between two time series and can also be used for nonlinear analysis; the value of Spearman correlation coefficient is between -1 and 1, and the closer it is to 1, the stronger the correlation between the two series.

[0033] The maximum correlation delay extraction submodule calculates each pair of adjacent VNFs based on the delay correlation calculation submodule. i and VNF j Calculate the latency correlation between them, and the latency when their latency correlation is at its maximum:

[0034]

[0035] Where Δt0 is The maximum value of Δt is 1≤i,j≤N,i≠j, and max(@) represents the operation of taking the maximum value;

[0036] Finally, the maximum correlation delay extraction submodule obtains the maximum correlation delay set D = {D} i,j |1≤i,j≤N}.

[0037] Step 2c) Constructing the synchronous and asynchronous adjacency matrix construction module: The synchronous and asynchronous adjacency matrix construction module needs to construct the maximum correlated delay set D = {D} extracted by the delay correlation extraction module. i,j We construct the synchronous and asynchronous adjacency matrix M using |1≤i,j≤N}. SA M SA Includes M×M sub-blocks of dimension N×N, and sub-block A in row r and column c. r,c The construction method is as follows:

[0038] When r = c:

[0039] A r,c =A adj

[0040] When r≠c:

[0041]

[0042] Among them, A r,c It is a block matrix M SA The sub-block in row r and column c, A adj It is the adjacency matrix of the graph representing the service function chain. It is matrix A r,c The element in row a and column b, D a,b ∈D;

[0043] M SA It will be used in the synchronous and asynchronous spatiotemporal feature fusion module to extract synchronous and asynchronous spatiotemporal features.

[0044] Step 2d) Constructing the synchronous and asynchronous spatiotemporal feature fusion module: including a cascaded synchronous and asynchronous spatial feature fusion graph convolution submodule and a long short-term memory (LSTM) submodule. The LSTM submodule includes N parallel LSTM networks.

[0045] The synchronous and asynchronous spatial feature fusion graph convolutional submodule consists of two stacked graph convolutional networks (GCNs). The activation function of each GCN layer is ReLU, and the adjacency matrix of each GCN layer is multiplied by the synchronous and asynchronous adjacency matrix M. SA The alternative forward propagation formula for the synchronous and asynchronous spatial feature fusion graph convolutional submodule is:

[0046] H = ReLU(M) SA ReLU(M SA XW1)W2)

[0047] Where ReLU(·) represents the activation function, X = [X t-M+1 ,X t-M+2 ,…,X t ] T It is a training sample, where It is the multidimensional resource requirement data of N VNFs at the t-th time step. and These are learnable weight parameters.

[0048] Use M SA The purpose of replacing the adjacency matrix is ​​that when performing graph convolution using the adjacency matrix, each VNF can only fuse the features of its neighboring VNFs at that time step, i.e., synchronous spatial features, and does not take into account the asynchronous features of neighboring VNFs. By using a synchronous-asynchronous adjacency matrix instead of the adjacency matrix, during the graph convolution process, the VNFs... i Features extracted at time step t for:

[0049]

[0050] in, It is VNF i Features at time step t It is at time step t that all VNF i The sum of features of adjacent VNFs, It is related to VNF i Adjacent VNFs k At t+D i,k The characteristics of the time step, where D i,k It is VNF i With VNF k The maximum correlation delay between them, W∈{W1,W2}; then yes Hidden features that combine synchronous and asynchronous spatial features.

[0051] The Long Short-Term Memory (LSTM) submodule consists of N parallel LSTM networks. Each LSTM network comprises two stacked LSTM layers and one fully connected layer. The number of neurons in each LSTM layer is set to N, and the number of neurons in the fully connected layer is set to 1. The N LSTM networks each receive N hidden features from VNFs, where VNFs... n The hidden features are It contains a total of M time steps.

[0052] Step 3: Iteratively train the prediction network model:

[0053] Step 3a) Initialize the number of iterations to k, the maximum number of iterations to K, K = 100, and the prediction network model O for the k-th iteration. k The weight is W k And let k = 1;

[0054] Step 3b) The enhanced feature extraction module extracts the enhanced features EF for each VNF based on the auxiliary sample set.n Then, the N enhanced features of VNF constitute the enhanced feature set EF = {EF1, EF2, ..., EF}. n ,…,EF N}, The delay correlation extraction module extracts the delay correlation of EF; the synchronous / asynchronous adjacency matrix construction module constructs a synchronous / asynchronous adjacency matrix M of dimension MN×MN using the extracted maximum correlation delay set D. SA , where EF n It is VNF n Enhanced features, It is the tth p Enhanced features at each time step;

[0055] Step 3c) The synchronous and asynchronous spatiotemporal feature fusion module uses M SA Using the adjacency matrix as the basis for graph convolution of the training samples, we obtain the hidden features H = [H1, H2, ..., H]. n ,…,H N ] T The N parallel LSTM networks in the LSTM submodule extract H1, H2, ..., H respectively. n ,…,H N The temporal characteristics are used to obtain the predicted value Y for each training sample. pred ,in It is VNF n Hidden features;

[0056] Step 3d) employs the mean squared error loss function and uses the true label Y of each training sample. true and predicted value Y pred Calculate the loss value (Loss) of the prediction network model, and then apply the Loss to W. k The prediction network model O is updated to obtain the prediction model for this iteration. k The formula for calculating the loss value (Loss) and its application to W. k The update formulas for the update are as follows:

[0057]

[0058] m k =β1m k-1 +(1-β1)g k

[0059]

[0060]

[0061] in They represent VNFs respectively n The predicted and actual values ​​of multidimensional resource demand data. α = 0.01 represents the learning rate, g k The loss function with respect to parameter W k Find the gradient after partial derivative, m k and v k These are the first-order and second-order moment estimates of the gradient in momentum form, respectively. and These are the first-order moment estimates and second-order moment estimates after bias correction, where β1 = 0.9 and β2 = 0.999 are the attenuation rates, and ε = 10. -8 It is a constant.

[0062] Step 3e) Determine if k = K. If so, obtain the trained prediction network model O. * Otherwise, let k = k + 1, O k =O, and proceed to step 3c).

[0063] Step 4, Obtain the VNF multidimensional resource demand prediction results: Use the test sample set as the trained prediction network model O * The input is propagated forward to obtain the predicted value for each test sample.

[0064] The technical effects of the present invention will be further explained below with reference to simulation experiments:

[0065] 1. Simulation experimental conditions:

[0066] The simulation experiment software platform of this invention is as follows: the experimental computer operating system is Windows 10 (64-bit), the memory is 32GB, the processor configuration is: Intel(R) Core(TM) i7-10870H CPU@2.20GHz 2.21GHz, it is implemented using Python language, version is Python 3.8, the experimental running environment is JetBrains PyCharm Community Edition 2020.2, and the deep learning framework is PyTorch 2.1.0.

[0067] 2. Simulation content and result analysis:

[0068] The prediction accuracy of this invention is compared with that of an existing resource-predictive-based digital twin-assisted virtual network function active migration method through simulation. The results are as follows: Figure 2 As shown:

[0069] Reference Figure 2 (a), (b), and (c) represent three indicators of prediction accuracy: mean squared error (MSE), mean absolute error (MAE), and coefficient of determination (R²-SCORE). The horizontal axis represents the type of resource demand, and the vertical axis represents the values ​​of MSE, MAE, and R²-SCORE, respectively.

[0070] The formulas for calculating MSE, MAE, and R2-SCORE are as follows:

[0071]

[0072]

[0073] Where, x t This represents the true value of the sampling point at sampling time t. This represents the predicted value of the sampling point at sampling time t. This represents the mean of all sampling points included in the training sample.

[0074] from Figure 2 As can be seen, the MSE and MAE values ​​of the present invention are closer to 0, and the R2-SCORE is closer to 1, indicating that the prediction accuracy of the present invention is significantly higher than that of the prior art.

Claims

1. A VNF multidimensional resource demand forecasting method based on the fusion of synchronous and asynchronous spatiotemporal characteristics, characterized in that, Includes the following steps: (1) Obtain the auxiliary sample set, training sample set, and test sample set: Normalize the multidimensional resource demand data for each of the T time steps in each of the N VNFs, and then partition the multidimensional resource demand data for each of the S time steps in each VNF. The first U data blocks of length M and their labels are then used to form the training sample set, and the remaining V data blocks are used to form the test sample set. The unpartitioned normalized N×T data blocks are then used as the test sample set. P An auxiliary sample set is composed of multiple dimensions of resource demand data, where N≥1, and T=S+T P T≥10000, S=U+V+M, M≥5; (2) Construct a VNF multidimensional resource demand prediction network model based on the fusion of synchronous and asynchronous spatiotemporal characteristics: Construct a VNF multidimensional resource demand prediction network model O, comprising a cascaded enhanced feature extraction module, a time-delay correlation extraction module, a synchronous / asynchronous adjacency matrix construction module, and a synchronous / asynchronous spatiotemporal feature fusion module. The synchronous / asynchronous adjacency matrix construction module is used to construct the maximum correlated time delay set D = {D...} extracted by the time-delay correlation extraction module. i,j Construct an asynchronous adjacency matrix M for |1≤i,j≤N}. SA M SA Includes M×M sub-blocks of dimension N×N, and sub-block A in row r and column c. r,c The construction method is as follows: When r = c: A r,c =A adj When r≠c: Among them, A r,c It is a block matrix M SA The sub-block in row r and column c, A adj It is the adjacency matrix of the graph representing the service function chain. It is matrix A r,c The element in row a and column b, D a,b ∈D; (3) Iteratively train the prediction network model: The prediction network model O is iteratively trained using auxiliary and training sample sets to obtain the trained prediction network model O. * ; (4) Obtain the VNF multidimensional resource demand forecast results: The test sample set is used as the trained prediction network model O. * The input is propagated forward to obtain the predicted value for each test sample.

2. The method according to claim 1, characterized in that, The normalization of multidimensional resource demand data for each of the N VNFs, comprising T time steps, described in step (1), wherein the VNFs... n Multidimensional resource demand data at time step t Including VNF n CPU resource requirements data Memory requirement data and bandwidth demand data right The formula for normalization is: in, yes The normalization result, It is VNF n Resource demand data of type r at time step t, X r,min X r,max These are the minimum and maximum values ​​among N×T multidimensional resource demand data of type r.

3. The method according to claim 1, characterized in that, The method for partitioning the multidimensional resource demand data of each normalized VNF across S time steps described in step (1) is as follows: Using a sliding window of length M+1 time steps, the multidimensional resource demand data of each normalized VNF is divided into S time steps with one time step as the step size. The first M data from the M+1 time steps of multidimensional resource demand data in each of the N VNFs after each division are used as data blocks, and the last one is used as the label of the data block. A total of SM divisions are performed to obtain SM data blocks of size N×M and their labels.

4. The method according to claim 1, characterized in that, The VNF multidimensional resource demand prediction network model based on the fusion of synchronous and asynchronous spatiotemporal characteristics described in step (2) includes: The delay correlation extraction module includes a cascaded delay correlation calculation submodule and a maximum correlation delay extraction submodule; The synchronous and asynchronous spatiotemporal feature fusion module includes a cascaded synchronous and asynchronous spatial feature fusion graph convolution submodule and a long short-term memory (LSTM) submodule. The LSTM submodule consists of N parallel LSTM networks.

5. The method according to claim 1, characterized in that, The iterative training of the prediction network model O described in step (3) is implemented as follows: (5a) The initial number of iterations is k, the maximum number of iterations is K, K≥50, and the prediction network model O in the k-th iteration is... k The weight is W k And let k = 1; (5b) The enhanced feature extraction module extracts the enhanced features EF for each VNF based on the auxiliary sample set. n Then, the N augmented features of VNF constitute the augmented feature set EF = {EF1, EF2, ..., EF}. n ,…,EF N }, The delay correlation extraction module extracts the delay correlation of EF; the synchronous / asynchronous adjacency matrix construction module constructs a synchronous / asynchronous adjacency matrix M of dimension MN×MN using the extracted maximum correlation delay set D. SA , where EF n It is VNF n Enhanced features, It is the tth p Enhanced features at each time step; (5c) Synchronous and asynchronous spatiotemporal feature fusion module with M SA Using the adjacency matrix as the basis for graph convolution of the training samples, we obtain the hidden features H = [H1, H2, ..., H2]. n ,…,H N ] T The N parallel LSTM networks in the LSTM submodule extract H1, H2, ..., H respectively. n ,…,H N The temporal characteristics are used to obtain the predicted value Y for each training sample. pred ,in It is VNF n Hidden features; (5d) The mean squared error loss function is used, and the true label Y of each training sample is used. true and predicted value Y pred Calculate the loss value (Loss) of the prediction network model, and then apply the Loss to W. k The prediction network model O is updated to obtain the prediction model for this iteration. k ; (5e) Determine if k = K. If so, obtain the trained prediction network model O. * Otherwise, let k = k + 1, O k =O, and proceed to step (5c).

6. The method according to claim 5, characterized in that, The t-th step described in step (5b) p Enhanced features at each time step The calculation formula is as follows: Here, |·| represents the absolute value operation.

7. The method according to claim 5, characterized in that, The delay correlation extraction module described in step (5b) extracts the delay correlation of EF. The implementation steps are as follows: The latency correlation calculation submodule calculates every two adjacent VNFs in the enhanced feature set EF. i and VNF j Enhanced features EF i With EF j,△t Delay correlation The maximum correlation delay extraction submodule is based on Calculate latency correlation Maximum delay D i,j The maximum correlation delay set D = {D} is obtained. i,j |1≤i,j≤N}, where D i,j The calculation formulas are as follows: Where DR(·) represents the operation of calculating the Spearman correlation coefficient, and Δt is the VNF. j Enhanced features EF j Relative to VNF i Enhanced features EF i The delayed time step, Δt0 is The maximum value of Δt is 1≤i,j≤N,i≠j, and max(·) represents the operation of taking the maximum value.

8. The method according to claim 5, characterized in that, The hidden feature H mentioned in step (5c) is calculated using the following formula: H=ReLU(M SA ReLU(M SA XW1)W2) Where ReLU(·) represents the activation function, X represents the training sample, and W1 and W2 are learnable weight parameters.

9. The method according to claim 5, characterized in that, The formula for calculating the loss value Loss of the prediction network model in step (5d), and the application of W... k The update formulas for the update are as follows: m k =β1m k-1 +(1-β1)g k in They represent VNFs respectively n The predicted and actual values ​​of multidimensional resource demand data. α represents the learning rate, g k The loss function with respect to parameter W k Find the gradient after partial derivative, m k and v k These are the first-order and second-order moment estimates of the gradient in momentum form, respectively. and For the first-order moment estimate and the second-order moment estimate after bias correction, β1 and β2 are the decay rates, and ε is a constant.

Images