A Routing Optimization Method and System Based on Attention Mechanism and Temporal Graph Convolutional Networks

Abstract

This invention discloses a routing optimization method and system based on attention mechanism and temporal graph convolutional network, belonging to the field of communication network technology. The method classifies network traffic. When the network traffic is a mouse flow, a trained A3T-GCN model is used to predict the network state of each link, and an improved butterfly optimization routing algorithm is used to optimize the links. The latency and volatility of the links are calculated based on the network state, and through iterative optimization, the link with the minimum latency and latency volatility is obtained. The mouse flow is then routed along the link corresponding to this optimal solution. When the network traffic is an image flow, the A3T-GCN model is used to predict the average remaining bandwidth of each link in the shortest hop count link set within the next w periods, and the link with the largest average remaining bandwidth is used for routing. This method effectively improves the overall efficiency and throughput of the network by optimizing the allocation and utilization of network resources.

Description

Technical Field

[0001] This invention relates to the field of communication network technology, specifically to a routing optimization method and system based on attention mechanisms and temporal graph convolutional networks. Background Technology

[0002] With the rapid development of internet technology, data centers, as the core infrastructure for information processing and storage, directly impact the quality of various services such as cloud computing, big data analytics, and online transaction processing. One of the main challenges facing data center networks (DCNs) is handling the ever-increasing and highly dynamic data traffic, especially in ensuring the efficiency and stability of data transmission. Traffic is typically characterized by uneven distribution and high burstiness, placing high demands on network routing optimization strategies. Currently, routing optimization techniques mainly face two types of traffic challenges: small flows (mouse flows) and large flows (elephant flows). Small flows require rapid transmission and are extremely sensitive to latency; while large flows, due to their large data volume and long duration, have higher bandwidth requirements. Traditional routing optimization methods, such as the Equal Cost Multipath (ECMP) strategy, allocate traffic by hashing the five-tuple (source IP address, destination IP address, source port number, destination port number, and protocol type) in data packets. However, this method does not optimize for the specific characteristics of the traffic, easily leading to resource conflicts and inefficiency.

[0003] To address these issues, the concept of Software-Defined Networking (SDN) was proposed and has been gradually applied to the design and optimization of data center networks. SDN achieves flexible management and efficient configuration of network resources by separating the network control layer (control plane) and the data forwarding layer (data plane). Centralized management of the control layer allows the network to dynamically adjust policies to adapt to different application needs, such as traffic analysis, route optimization, and security management. However, despite the potential for improved network management, SDN still has some limitations in practical deployments. For example, centralized control in SDN can become a source of performance bottlenecks or security risks. Furthermore, SDN typically relies on real-time monitoring and data analysis of network status, which requires efficient and scalable technical support to ensure timely adjustment and optimization of network policies.

[0004] Against this backdrop, in-depth research into the characteristics of data center networks and the development of novel routing optimization algorithms have become particularly important. For example, some studies have improved network adaptability and resource utilization by introducing finer-grained control units (such as flowlet or packet-level control). While these methods can theoretically improve network performance, they still need to overcome problems in practical applications, including out-of-order packets, inaccurate traffic prediction, and the complexity of scheduling algorithms. Summary of the Invention

[0005] To address the problems existing in the prior art, this invention provides a routing optimization method and system based on attention mechanism and temporal graph convolutional network. It predicts the congestion status of the entire network links through neural network and optimizes traffic routing for different traffic types.

[0006] This invention is achieved through the following technical solution:

[0007] A routing optimization method based on attention mechanism and temporal graph convolutional network includes the following steps:

[0008] Obtain the network latency and bandwidth of each link formed between switches in the network topology at each time point;

[0009] The network latency and bandwidth at each time point are used as network state information to construct a training dataset, which is then used to train the A3T-GCN model. The trained A3T-GCN model obtains global temporal and spatial correlations through an attention mechanism to predict the network state information of links at future time points.

[0010] When the network traffic is a mouse flow, the trained A3T-GCN model is used to predict the network status of each link, and the improved butterfly optimization routing algorithm is used to optimize the links. The latency and volatility of each link are determined according to the network status. The genetic algorithm is used to exchange links to realize link updates. After iterative optimization, the global optimal solution is obtained, that is, the link with the minimum latency and latency volatility. The mouse flow is routed along the link corresponding to the optimal solution.

[0011] When network traffic is in the form of object flow, determine the shortest hop count link set, then use the A3T-GCN model to predict the average remaining bandwidth of each link in the shortest hop count link set over the next w periods, and use the link object flow with the largest average remaining bandwidth for routing.

[0012] Preferably, the acquisition of network latency for each link includes:

[0013] The one-way delay T1 from the controller to each switch is used to acquire data.

[0014] The one-way delay T2 for acquiring data from each switch to the controller;

[0015] The total latency of acquiring data from the controller to each switch and back to the controller;

[0016] The network delay of the link is determined based on the one-way delay T1, the one-way delay T2, and the total delay.

[0017] Preferably, the delay matrix X is constructed based on the link delay between adjacent switches, and the network delay of the link is determined based on the delay matrix;

[0018] The expression for the delay matrix is ​​as follows:

[0019]

[0020] Where n is the number of switch nodes, and X is the element of the delay matrix. i,j The link delay d between adjacent switches (i,j) represents the link delay d between them. (i,j) .

[0021] Preferably, the acquisition of network bandwidth for each link includes:

[0022] Determine the total number of bytes rx received by the controller in each cycle. bytes Total number of bytes sent (tx) bytes Based on the time interval of the previous data collection cycle, the network bandwidth used by the current link is determined, and the network bandwidth of the link is obtained by subtracting the current network bandwidth used from the total network bandwidth of the link.

[0023] Preferably, the A3T-GCN model includes a graph convolutional network (GCN) and a gated recurrent network (GRU). The method for predicting the network state information of the A3T-GCN model is as follows:

[0024] A graph convolutional network (GCN) is used to obtain hidden states with spatial features from the network state information and input them into a gated recurrent network (GRU). The GRU obtains hidden states with spatial and temporal features from the network state information and uses a scoring function to determine the weights of the hidden states with spatial and temporal features.

[0025] We use the context vector of global network state change information from the attention function, and the weights of the hidden states and the context vector to predict the network state information of the links at future time steps.

[0026] Preferably, a pre-trained DNN model is used to classify the traffic in the network and determine whether the network traffic is mouse traffic or elephant traffic;

[0027] The training method for the DNN model is as follows:

[0028] Obtain the traffic characteristics of network traffic and construct a training dataset based on the traffic characteristics;

[0029] The traffic characteristics include source and destination IP, source and destination port, flow duration, and flow rate;

[0030] A DNN model is constructed, and a dropout mechanism is introduced into the DNN model. The DNN model is trained using a training dataset, and BCEWithLogitsLoss is used as the loss function of the DNN model. The trained DNN model is then used to classify the traffic in the network into image traffic and mouse traffic.

[0031] Preferably, the method for determining the delay and volatility of the link is as follows:

[0032] The A3T-GCN model predicts the latency of each link in the next w moments, thereby determining the average latency of each link in the network topology in the next w moments. The latency of each link through which the data packet passes in the next w moments is accumulated to obtain the total latency of the current link in the next w moments. The latency variability on each link is determined based on the total latency.

[0033] Preferably, the iteration of the global optimal solution includes:

[0034] The distances from the link to the positive and negative ideal solutions are determined based on the total latency and volatility. The positive ideal solution is the best performance value that we hope to achieve in the decision-making process, while the negative ideal solution is the worst performance value that we hope to avoid in the decision-making process.

[0035] Distance to the ideal solution:

[0036]

[0037] in, and These are the paths with the least latency and the paths with the least latency variability, respectively.

[0038] Distance to the negative ideal solution:

[0039]

[0040] in, and The path with the greatest latency and the path with the greatest latency variability;

[0041] Based on the ratio of the distances from the link to the positive ideal solution and the negative ideal solution, calculate the comprehensive score of each link, and select the link with the highest comprehensive score as the optimal solution.

[0042] Preferably, the Dijkstra algorithm is used to determine the set of shortest hop links.

[0043] A routing optimization system based on attention mechanism and temporal graph convolutional network, comprising:

[0044] The acquisition module is used to obtain the network latency and bandwidth of each link formed between switches in the network topology at each time point;

[0045] The prediction module is used to construct a training dataset by using network latency and bandwidth at each time point as network state information, and to train the A3T-GCN model. The trained A3T-GCN model obtains global temporal dynamics and spatial correlation through an attention mechanism to predict the network state information of the link at future time points.

[0046] Route optimization module,

[0047] When the network traffic is a mouse flow, the trained A3T-GCN model is used to predict the network status of each link. An improved butterfly optimization routing algorithm is used to optimize the links. The latency and volatility of each link are determined according to the network status. A genetic algorithm is used to exchange links to update the links. After iterative optimization, the global optimal solution is obtained, which is the link with the minimum latency and latency volatility. The mouse flow is then routed along the link corresponding to the optimal solution.

[0048] When network traffic is in the form of object flow, it is used to determine the shortest hop count link set. Then, the A3T-GCN model is used to predict the average remaining bandwidth of each link in the shortest hop count link set over the next w periods. The link object flow with the largest average remaining bandwidth is used for routing.

[0049] Compared with the prior art, the present invention has the following beneficial technical effects:

[0050] This invention provides a routing optimization method based on an attention mechanism and a temporal graph convolutional neural network (A3T-GCN). Compared to traditional routing optimization schemes that rely on prediction over a single time period, this method can predict the network state over multiple future time periods, thus improving network state awareness. Secondly, by utilizing A3T-GCN to predict the network state over multiple future time periods, and combining it with the butterfly optimization algorithm and Dijkstra's algorithm, intelligent path selection is performed for traffic. For latency-sensitive traffic, routes with high path stability are preferred; for traffic with high bandwidth requirements, paths with sufficient bandwidth are selected. This strategy significantly reduces network congestion and increases data transmission throughput. This method effectively improves the overall network efficiency and throughput by optimizing the allocation and utilization of network resources.

[0051] Furthermore, this invention uses deep neural networks (DNNs) for traffic classification optimization. By deeply learning and analyzing the characteristics of different traffic types, the DNN model of this invention can directly and accurately identify various types of traffic, effectively meeting the specific scheduling requirements of latency-sensitive and bandwidth-intensive traffic, thereby improving network responsiveness and quality of service. Attached Figure Description

[0052] Figure 1 This is a framework diagram of the adaptive routing optimization system of the present invention;

[0053] Figure 2 This is a flowchart of the adaptive routing optimization method of the present invention;

[0054] Figure 3 This is a diagram of the network status acquisition module of the present invention;

[0055] Figure 4This is a diagram of the T-GCN model of the present invention. Detailed Implementation

[0056] The present invention will now be described in further detail with reference to the accompanying drawings. These descriptions are intended to explain the invention and not to limit it.

[0057] See Figure 2-4 A routing optimization method based on attention mechanism and temporal graph convolutional network includes the following steps:

[0058] Step 1: Introduce the dropout mechanism into the DNN model and use the DNN model to classify the traffic in the network into mouse flow and elephant flow.

[0059] S1.1 Utilizing traffic characteristics obtained from the SDN environment: six key characteristics including source and destination IP, source and destination port, flow duration, and flow rate; constructing a training dataset based on these traffic characteristics.

[0060] In an SDN (Software-Defined Networking) environment, the architectural characteristics of SDN, such as separation of control and forwarding, network virtualization, and open programmability, make it easier to capture and analyze the characteristics of network traffic.

[0061] S1.2 Construct a DNN model, introduce the dropout mechanism into the DNN model, train the DNN model using the training dataset, and use the trained DNN model to classify network traffic into image traffic and mouse traffic. Use the DNN model to identify image traffic earlier, implement reasonable routing strategies, and promote load balancing.

[0062] DNNs consist of an input layer, multiple hidden layers, and an output layer. As the number of network layers increases, they can model and learn more complex tasks with greater precision, thereby improving the model's generalization ability. Specifically, to enhance the model's generalization ability and prevent overfitting, a dropout mechanism is introduced between the hidden layers. This strategy randomly discards a portion of neurons during training, forcing the network to learn more robust feature representations, thus enhancing the model's predictive ability on new and unseen data.

[0063] In a DNN network, to calculate the output of the hidden layer h1, it is necessary to determine the output value of the layer preceding h1 and the weights connecting these two layers. The layer preceding h1 is the input layer of the model, and the feature vector of the input layer is denoted as h0. The formula for calculating h0 is as follows: h1 = f(b1 + w1h0).

[0064] Where f represents the activation function, and here we use ReLU as the planning function, and b1 represents the bias term of the hidden layer h1.

[0065] Therefore, by analogy, the DNN model uses the output of this layer as the input of the next layer to achieve forward propagation of the signal until the model's predicted value is output. In order to determine the difference between the model's predicted value and the true value, a loss function is defined accordingly, and BCEWithLogitsLoss is adopted as the loss function:

[0066]

[0067] in, This represents the model output, y represents the true value, which is 0 or 1, N is the number of samples in the batch, and σ represents the sigmoid function.

[0068] Step 2: Obtain the network latency of each link formed between each switch in the network topology at each time point.

[0069] S2.1 Obtain the one-way latency from the controller to each switch.

[0070] In this embodiment, two switches are used as an example for explanation.

[0071] The controller periodically sends an echo_request to switches A and B in each cycle. When each switch receives the echo_request, it immediately records the current time and fills the time into the echo_reply message, thus obtaining the one-way latency of each switch.

[0072] S2.2 Obtain the one-way delay from each switch to the controller.

[0073] When a switch receives an echo_request, it immediately sends an echo_reply message to the controller. Upon receiving the echo_reply, the controller subtracts the timestamp from the message from the current time to obtain the round-trip time t between switches A and B and the controller. A and t B This gives the one-way delay from each switch to the controller.

[0074] S2.3 Obtain the latency from controller to switch A to switch B to controller.

[0075] Measurements are performed using LLDP (Link Layer Discovery Protocol), during which the controller periodically sends a sequence of data including the source address and timestamp t. sendLLDP packets. When the controller sends an LLDP packet to switch A, switch A floods the LLDP packet to neighboring switches. When a neighboring switch receives the LLDP packet, since there is no corresponding flow table rule for forwarding, it sends a packet-in message to the controller, which contains the LLDP packet header. When the controller receives the packet-in message, it can obtain the source address and timestamp t. receive The total latency t from controller to switch A to switch B to controller can be calculated by subtracting the timestamp in the message from the current time. lldp =t receive -t send .

[0076] S2.4 Calculate the latency between switch A and switch B.

[0077] The delay between switches A and B can be calculated from the various delays calculated in the first three steps. (A,B) =t lldp -(t A +t B )÷2.

[0078] On the controller, a time delay matrix of size n×n can be established;

[0079]

[0080] Where n is the number of switch nodes. The elements X of the delay matrix... i,j The link delay d between adjacent switches (i,j) represents the link delay d between them. (i,j) Note that the latency from the switch to itself is 0, therefore the element X on the diagonal... i,i All are 0. This delay matrix X can be constructed by calculating the link delay between adjacent switches, where X... i,j This represents the link delay from switch i to switch j.

[0081] Step 3: Collect the network bandwidth of each link;

[0082] 3a) In a software-defined networking (SDN) architecture, the OpenFlow protocol provides network switches with the ability to maintain flow table entries, including specific counters for collecting key information about flow table entries or ports (such as the number of bytes received, duration of existence, etc.). These counters allow for tracking changes in the number of matching bytes per unit of time, thereby accurately calculating the real-time transmission bandwidth of the data flow.

[0083] 3b) The controller periodically sends PortStats requests to all switches. Upon receiving a request, each switch sends back an EventOFPPortStatsReply event containing the status of all ports to the controller. The controller compares the data collected this time with the data collected previously to calculate the total number of bytes received, rx. bytes Total number of bytes sent (tx) bytes Based on the time interval since the last data collection, the network bandwidth used by the current link is calculated:

[0084]

[0085] 3c) Calculate the available network bandwidth B for each link. available That is, the total network bandwidth B of the link. total Subtract the currently used network bandwidth B available =B total -B useage .

[0086] Step 4: Use the network latency of each link formed by the switch at each time point obtained in Step 2, and the network bandwidth of each link collected in Step 3 as training data to train the A3T-GCN model. The trained A3T-GCN model predicts the network status of each link in the future time period, that is, the latency and bandwidth of the link.

[0087] See Figure 4 The A3T-GCN model combines the features of GCN and GRU, and uses an attention mechanism to obtain global network state information for effectively processing and predicting network state information. The training method of the A3T-GCN model is as follows:

[0088] S4.1. Abstract the entire data center network topology into an unweighted graph data structure G = (V, E). Treat each link as a node, where V represents the set of nodes V = (v1, v2, ..., v...). N N is the number of link nodes. E represents the set of variables. Then, the adjacency matrix A is used to represent the connection relationship A∈R between links. N×N The adjacency matrix contains only elements 0 and 1. If two links are not adjacent, the element is 0; if they are adjacent, the element is 1.

[0089] S4.2. Treat the time delay on the links in the topology as a characteristic attribute of the nodes, denoted as X∈R. N×P .

[0090] Where P represents the number of node attribute features, i.e., the length of the historical time series. X t ∈R N×tLet represent the time delay on each link at time t. Therefore, given a fixed network topology G and feature matrix X, it is possible to predict the network state over the next w periods. This problem can be abstracted into an abstract function f, [X]. t+1 ,X t+2 ,…,X t+w ] = f(G; (X) t-q ,…,X t-1 ,X t In this process, the link latency at time t+w is predicted using time series data of length q.

[0091] S4.3. Use a 2-layer GCN model to obtain spatial dependencies, as follows:

[0092]

[0093] in, Let I represent a self-connected adjacency matrix, where I is the identity matrix. This indicates the preprocessing stage, where W0∈R P×H Let W1 represent the weight matrix input to the hidden layer, P be the length of the feature matrix, H be the number of hidden units, and W1 ∈ R. H This represents the weight matrix from the hidden layer to the output layer. ReLU() represents the activation function Rectified LinearUnit, and σ(·) represents the activation function sigmoid.

[0094] S4.4. A soft attention model is used to learn the importance of network state information in each time period. Based on this, a context vector that can express the global trend of network state changes is calculated for future network state prediction tasks, i.e., the latency and bandwidth of each link in future time periods.

[0095] Suppose we have a time series X of network state information. i (i = 1, 2, ..., n), where n is the length of the time series.

[0096] The first step is to obtain the hidden state at each time step, output the hidden state with spatial features through the GCN, and then input it into the GRU to obtain the hidden state h with spatial and temporal features. i This is represented as H = (h1, h2, ..., h n ).

[0097] The second step is to use a scoring function to calculate the weight of each hidden state.

[0098] The third step involves using an attention function to compute a context vector C that represents changes in the global network state. t .

[0099] The fourth step is to use the weights of the hidden states and the context vector to obtain the final output.

[0100] In this paper, a multilayer perception is used as the scoring function. The hidden state h at each time step is... i The input is used to calculate the weights for each hidden state. The model uses two hidden layers to obtain the corresponding outputs. Each α... i The normalization calculation is performed using the Softmax function, where w1 and w2 represent the weights of the first and second layers, and b1 and b2 represent the deviations between the first and second layers.

[0101] e i =w2(w1H+b1)+b2,

[0102] Finally, the calculation process for the context vector in the attention function is as follows:

[0103] In the S4.5 and A3T-GCN models, the spatial dependencies of the input data are obtained through the GCN model, and then input into the GRU model to obtain the temporal dependencies. Simultaneously, an attention mechanism is introduced to weight the influence of historical network state information, thereby capturing global changes in the network state.

[0104] The T-GCN model is constructed by combining GCN and GRU. First, historical network state information is input into the T-GCN to obtain n hidden states h = {h_n} containing both spatial and temporal information. t-n ,…,h t-1 ,h t}

[0105] Step 5: Classify the traffic in the network according to the DNN model;

[0106] When the network traffic is a mouse flow, the A3T-GCN model trained in step 4 is used to predict the network status of each link. A genetic algorithm is introduced into the butterfly optimization routing algorithm to improve it. The improved butterfly optimization routing algorithm is used to perform local and global optimization of the links. The fitness function of the link is determined according to the network status. According to the fitness function, which includes the delay and volatility of each link, the genetic algorithm is used to exchange links to realize link updates. After iterative optimization, the global optimal solution is obtained, that is, the link with the minimum delay and delay volatility. The mouse flow is routed along the link corresponding to the optimal solution.

[0107] An improved butterfly routing algorithm is employed to efficiently find network links with minimal latency and latency variability. The improved butterfly routing algorithm treats each available link in the network as a butterfly, uses the predicted network state of each link as its fitness value, and uses this fitness value as a "fragrance" to symbolize the quality of the path. Local and global searches are performed on all links.

[0108] The butterfly moves randomly around its current location, simulating a local search to find a local optimum; it then moves towards the area with the strongest fragrance, simulating a global search to find a global optimum.

[0109] The core element of the improved butterfly-optimized routing algorithm is the stimulus strength, which is determined by the link's fitness function. The fitness function of a link is determined based on the predicted network state, comprehensively considering the link's latency and volatility, and the Topsis algorithm is used to evaluate the link's latency and volatility.

[0110] The improved butterfly optimization routing algorithm is executed in three phases: initialization, iteration, and obtaining the optimal solution (termination).

[0111] In the initialization phase, multiple links are constructed between the source and destination nodes to form an initial population. Links are generated by randomly selecting the next-hop node at each node. During the iteration phase, the butterfly algorithm decides whether to perform a global or local search based on random numbers. To prevent getting trapped in local optima, a crossover operator from genetic algorithms is used, mimicking the mating and recombination process in natural selection. Path updates are achieved by exchanging links (chromosomes). This exchange mechanism effectively eliminates links with low fitness while preserving effective paths. After reaching a predetermined number of iterations, the algorithm terminates, obtaining the global optimum, and the mouse flow is routed along the link corresponding to this optimum. Throughout the iteration process, the butterfly optimization algorithm effectively balances global and local searches by mimicking information propagation among butterflies and the evolutionary mechanism of genetic algorithms. This balance allows the algorithm to adapt to the complexity and unpredictability of the network environment, thus demonstrating high efficiency in routing selection and optimization.

[0112] The method for optimizing links using an improved butterfly routing algorithm to obtain links with minimal latency and latency variability is as follows:

[0113] A1) In nature, butterflies search for food through both global and local searches. The butterfly algorithm simulates this process. Because of unpredictable factors in nature, such as rain and wind, the butterfly algorithm uses a constant `p` to determine whether a butterfly performs a global or local search. For each iteration of the butterfly population, each butterfly generates a random number. If the current random number is greater than `p`, the butterfly performs a global search; otherwise, it performs a local search. In the global search, the butterfly moves towards the butterfly `g` with the highest fragrance value. * To move, denoted by , represents the solution vector, indicating the solution corresponding to the i-th butterfly in the t-th iteration. * This represents the optimal solution in the current iteration. i Let r represent the scent emitted by the i-th butterfly, where r is a random number between 0 and 1. (Local search) and This is the solution corresponding to the j-th and k-th butterflies in the t-th iteration. The function of r is the same as in the global search;

[0114] A2) To avoid getting trapped in local optima, the crossover operator from the genetic algorithm is used to update the butterfly population. The crossover operator in the genetic algorithm is generated by mimicking the mating and recombination process in the natural evolutionary process of organisms. In this scheme, each butterfly, i.e., each link, corresponds to a chromosome in the genetic algorithm, and each switch node on this link represents a gene on the chromosome. Any switch node on a chromosome corresponds to a path in the network topology along with its preceding and following nodes. Therefore, to ensure route connectivity, a single-point crossover method is chosen. Additionally, it is required that the two selected chromosomes, besides the source and destination nodes, have at least one common node, but they do not need to be in the same location. In the global update part, the crossover swaps parts of the links between the two selected chromosomes (one the path with the highest score, representing the global optimum, and the other the current link). The fitness of the two new links is compared with the two original links, eliminating the two paths with lower fitness values, achieving greedy evolution to ensure that the paths are updated in a more optimal direction. Similarly, in the local update part, the crossover function randomly swaps three chromosomes (one from the current path and two from random paths), resulting in three new chromosomes and the original three chromosomes. The three chromosomes with lower fitness are eliminated, and the original three chromosomes are retained.

[0115] A3) To measure the distance to the optimal solution during the optimization search process, latency and latency volatility in the network state of the data center network are used as the fitness function. The A3T-GCN model is used to predict the latency of each link at the next w time points, thereby determining the average latency of each link in the network topology at the next w time points.

[0116]

[0117] Where x ij The latency on the link between switch i and switch j.

[0118] Therefore, the total latency on the path at future w moments can be obtained by simply summing the latency of the links traversed by the data packet at future w moments. Latency volatility is represented by the entropy of each link. The entropy of each link is determined based on the total latency. Entropy represents the degree of uncertainty or disorder of a system. The larger the entropy value, the greater the disorder, and thus the greater the data volatility. In a fat-tree symmetric topology, there are multiple paths from the source to the destination node that can transmit data packets. If a path with relatively stable latency over a longer period and a smaller total latency can be selected, it avoids always choosing the optimal path, which can lead to a large influx of traffic into the originally idle optimal path, causing congestion and frequent rerouting, resulting in drastic fluctuations in path congestion.

[0119] A4) Collect and calculate the future average latency x of the path delay and path delay variability x entropy The relevant data is as follows. It is well known that the lower the total latency of a path, the shorter the transmission time of data packets along the path. The lower the latency variability of a path, the less likely data packets will encounter out-of-order delivery or congestion issues during transmission. Therefore, the Topsis algorithm is considered as the fitness function.

[0120] Step 1: Perform reverse standardization on these two metrics to map their values ​​to the range [0,1], so that the smaller the value, the closer the path is to 1 after standardization.

[0121] For the total path delay:

[0122]

[0123] Regarding the volatility of the path:

[0124]

[0125] In this process, the path to the minimum value will have a normalized value of 1, while the path to the maximum value will have a normalized value of 0.

[0126] Step 2: Determine the ideal solution of the Topsis algorithm and negative ideal solution

[0127] Among them, the ideal solution It represents the optimal performance value desired in decision-making. Negative ideal solution. It represents the worst performance value that we hope to avoid in decision-making.

[0128] For path i, the distance to the positive ideal solution The calculation is as follows:

[0129]

[0130] in, and This represents the ideal case where the latency and latency variability metrics achieve optimal performance, specifically the path with the minimum latency and the path with the minimum latency variability, respectively. Similarly, the path i is computed to the negative ideal solution D. i - Distance:

[0131]

[0132] in, and This represents the ideal situation where latency and latency volatility metrics perform optimally, specifically the path with the highest latency and the path with the highest latency volatility, respectively.

[0133] Step 3: Calculate the overall score for each path based on the ratio of the distances to the positive ideal solution and the negative ideal solution.

[0134]

[0135] A higher-scoring path represents a better choice. Paths are ranked based on their overall scores, and the highest-ranked path is selected as the optimal choice. This ensures that a path with both low total latency and minimal path fluctuation is chosen.

[0136] When network traffic is image flow, the Dijkstra algorithm is first used to calculate the set of all possible shortest hop count links. Then, the A3T-GCN model is used to predict the average remaining bandwidth of each link in the shortest hop count link set over the next w periods. Among these links, the path with the largest average remaining bandwidth is selected for routing, that is, the image flow is routed to the link with the largest remaining bandwidth to ensure optimal traffic transmission.

[0137] B1) First, use Dijkstra's algorithm to calculate the set of paths with the shortest number of hops, P = (p1, p2, ..., p...). N ), where p i Representing the i-th path, we also calculate the average remaining bandwidth of the path set P over the next t+w periods, summed to B = (B1, B2, ..., B...). N ),in From the set of average remaining bandwidths B over the next t+w periods, select the largest average remaining bandwidth B. i path p i Perform routing.

[0138] See Figure 1 Another method of the present invention also provides a routing optimization system based on attention mechanism and temporal graph convolutional network, including a collection module, a routing module and an execution module.

[0139] The data acquisition module collects and analyzes network state information for understanding and abstraction, providing optimization-based control decision support to the routing module. This plane integrates key components such as a database, a pre-trained A3T-GCN model, and a DNN model. The database stores bandwidth, latency, network topology, and traffic state information for each link within the data plane, for use in model prediction. The A3T-GCN model focuses on predicting network state, providing a basis for network traffic routing. The DNN model is used for traffic (i.e., mouse-elephant flow) classification.

[0140] Routing Module: The control plane, located in the middle layer of the SDN architecture, is primarily responsible for collecting and analyzing network information and, based on the current network state, routing traffic to ensure that data packets efficiently reach their intended destinations. The controller collects link bandwidth, latency, and flow state information from the data plane through the network state acquisition module and transmits this critical data to the control plane, providing the necessary parameters for model training and prediction. Furthermore, in the network traffic classification module, upon receiving a packet_in message, the control plane classifies the traffic based on the packet information and then adopts corresponding routing strategies based on the classification results.

[0141] Execution Module: The data plane is the foundational layer in the SDN architecture. Its task is to receive data packets from hosts and correctly route them to ports using the flow tables within the switches. Furthermore, the data plane is responsible for further processing of data packets as needed. Switches are configured according to the flow tables issued by the control plane and collect network state information by forwarding probe packets.

[0142] The present invention has the following beneficial effects:

[0143] 1. This invention employs an attention-based temporal graph convolutional neural network (A3T-GCN), which, compared to traditional routing optimization schemes that rely on prediction over a single time period, can predict the network state over multiple future time periods, thereby improving the network state awareness capability. This technique reduces the root mean square error (RMSE) and mean absolute error (MAE) by 10.27% and 13.12% respectively compared to the traditional temporal graph convolutional neural network (T-GCN), significantly improving the performance and efficiency of routing optimization.

[0144] 2. Optimization of Refined Traffic Classification and Scheduling Strategies: Addressing the issue of insufficient traffic feature classification in traditional solutions, this invention utilizes Deep Neural Networks (DNNs) for traffic classification optimization. By deeply learning and analyzing the characteristics of different traffic types, the DNN model of this invention can directly and accurately identify various traffic types, effectively meeting the specific scheduling needs of latency-sensitive and bandwidth-intensive traffic, thereby improving network responsiveness and quality of service.

[0145] 3. Improved Path Selection and Network Resource Management: This invention proposes an innovative scheduling strategy that utilizes A3T-GCN to predict network conditions over multiple future time periods, combining the Butterfly Optimization algorithm and Dijkstra's algorithm for intelligent path selection of traffic. For latency-sensitive traffic, routes with high path stability are preferred; for traffic with high bandwidth requirements, paths with sufficient bandwidth are selected. This strategy significantly reduces network congestion and improves data transmission throughput.

[0146] 4. Improved Network Efficiency and Throughput: This invention effectively improves the overall efficiency and throughput of the network by optimizing the allocation and utilization of network resources. This represents a significant technological advancement for data center networks that need to handle high-density data streams.

[0147] The above content is only for illustrating the technical concept of the present invention and should not be construed as limiting the scope of protection of the present invention. Any modifications made to the technical solution based on the technical concept proposed in this invention shall fall within the scope of protection of the claims of this invention.

Claims

1. A routing optimization method based on attention mechanism and temporal graph convolutional network, characterized in that, Includes the following steps: Obtain the network latency and available network bandwidth of each link formed between switches in the network topology at each time point; The network latency and bandwidth at each time point are used as network state information to construct a training dataset. The A3T-GCN model of attention temporal graph convolutional network is trained. The trained A3T-GCN model obtains global temporal and spatial correlation through the attention mechanism and predicts the network state information of the link at future time points. A pre-trained DNN model is used to classify the network traffic and determine whether the network traffic is mouse flow or elephant flow; The training method for the DNN model is as follows: Obtain the traffic characteristics of network traffic and construct a training dataset based on the traffic characteristics; The traffic characteristics include source and destination IP, source and destination port, flow duration, and flow rate; A DNN model is constructed, and a dropout mechanism is introduced into the DNN model. The DNN model is trained using a training dataset, and BCEWithLogitsLoss is used as the loss function of the DNN model. The trained DNN model is then used to classify the traffic in the network, dividing the traffic into image flow and mouse flow. When the network traffic is a mouse flow, the trained A3T-GCN model is used to predict the network status of each link, and the improved butterfly optimization routing algorithm is used to optimize the links. The latency and volatility of each link are determined according to the network status. The genetic algorithm is used to exchange links to realize link updates. After iterative optimization, the global optimal solution is obtained, that is, the link with the minimum latency and latency volatility. The mouse flow is routed along the link corresponding to the optimal solution. When network traffic is in the form of object flow, determine the shortest hop count link set, then use the A3T-GCN model to predict the average remaining bandwidth of each link in the shortest hop count link set over the next w periods, and use the link object flow with the largest average remaining bandwidth for routing.

2. The routing optimization method based on attention mechanism and temporal graph convolutional network according to claim 1, characterized in that, The acquisition of network latency for each link includes: The one-way delay T1 from the controller to each switch is used to acquire data. The one-way delay T2 for acquiring data from each switch to the controller; The total latency of acquiring data from the controller to each switch and back to the controller; The network delay of the link is determined based on the one-way delay T1, the one-way delay T2, and the total delay.

3. The routing optimization method based on attention mechanism and temporal graph convolutional network according to claim 2, characterized in that, This delay matrix X is constructed based on the link delay between adjacent switches, and the network delay of the link is determined based on the delay matrix. The expression for the delay matrix is ​​as follows: Where n is the number of switch nodes, and X is the element of the delay matrix. i,j The link delay d between adjacent switches (i,j) represents the link delay d between them. (i,j) ; element X on the diagonal i,i A value of 0 indicates that the latency from the switch to itself is 0.

4. The routing optimization method based on attention mechanism and temporal graph convolutional network according to claim 1, characterized in that, The acquisition of available network bandwidth for each link includes: Determine the total number of bytes rx received by the controller in each cycle. bytes Total number of bytes sent (tx) bytes Based on the time interval of the previous data collection cycle, the network bandwidth used by the current link is determined; the available network bandwidth of the link is obtained by subtracting the current network bandwidth used by the total network bandwidth of the link.

5. The routing optimization method based on attention mechanism and temporal graph convolutional network according to claim 1, characterized in that, The A3T-GCN model includes a graph convolutional network (GCN) and a gated recursive network (GRU). The method for predicting the network state information of the A3T-GCN model is as follows: A graph convolutional network (GCN) is used to obtain hidden states with spatial features from the network state information and input them into a gated recurrent network (GRU). The GRU obtains hidden states with spatial and temporal features from the network state information and uses a scoring function to determine the weights of the hidden states with spatial and temporal features. An attention function is used to compute a context vector representing global network state changes. The weights of the hidden states and the context vector are then used to predict the network state information of each link at future time steps.

6. The routing optimization method based on attention mechanism and temporal graph convolutional network according to claim 1, characterized in that, The method for determining the delay and volatility of the link is as follows: The A3T-GCN model predicts the latency of each link in the next w moments, thereby determining the average latency of each link in the network topology in the next w moments. The latency of each link through which the data packet passes in the next w moments is accumulated to obtain the total latency of the current link in the next w moments. Based on the total latency, the latency and volatility of each link are determined.

7. The routing optimization method based on attention mechanism and temporal graph convolutional network according to claim 1, characterized in that, The globally optimal solution obtained through iterative optimization includes: The distances from a link to the positive ideal solution and the negative ideal solution are determined based on the total delay and volatility of the link. The positive ideal solution is the best performance value that we hope to achieve in the decision-making process, and the negative ideal solution is the worst performance value that we hope to avoid in the decision-making process. Distance to the ideal solution: in, and These are the links with the lowest latency and the links with the lowest latency variability, respectively. Distance to the negative ideal solution: in, and For the link with the maximum latency and the link with the maximum latency variability; Based on the ratio of the distances from the link to the positive ideal solution and the negative ideal solution, calculate the comprehensive score for each link, and select the link with the highest comprehensive score as the optimal solution.

8. The routing optimization method based on attention mechanism and temporal graph convolutional network according to claim 1, characterized in that, Dijkstra's algorithm is used to determine the set of shortest hop links.

9. A routing optimization system based on attention mechanism and temporal graph convolutional network, characterized in that, include: The acquisition module is used to obtain the network latency and available network bandwidth of each link formed between switches in the network topology at each time point; The prediction module is used to construct a training dataset by using network latency and bandwidth at each time point as network state information, and to train the attention temporal graph convolutional network A3T-GCN model. The trained A3T-GCN model obtains global temporal dynamics and spatial correlation through the attention mechanism to predict the network state information of the link at future time points. The classification module uses a pre-trained DNN model to classify network traffic and determine whether the network traffic is mouse traffic or elephant traffic. The training method for the DNN model is as follows: Obtain the traffic characteristics of network traffic and construct a training dataset based on the traffic characteristics; The traffic characteristics include source and destination IP, source and destination port, flow duration, and flow rate; A DNN model is constructed, and a dropout mechanism is introduced into the DNN model. The DNN model is trained using a training dataset, and BCEWithLogitsLoss is used as the loss function of the DNN model. The trained DNN model is then used to classify the traffic in the network, dividing the traffic into image flow and mouse flow. Route optimization module, When the network traffic is a mouse flow, the trained A3T-GCN model is used to predict the network status of each link. An improved butterfly optimization routing algorithm is used to optimize the links. The latency and volatility of each link are determined according to the network status. A genetic algorithm is used to exchange links to update the links. After iterative optimization, the global optimal solution is obtained, which is the link with the minimum latency and latency volatility. The mouse flow is then routed along the link corresponding to the optimal solution. When network traffic is in the form of object flow, it is used to determine the shortest hop count link set. Then, the A3T-GCN model is used to predict the average remaining bandwidth of each link in the shortest hop count link set over the next w periods. The link object flow with the largest average remaining bandwidth is used for routing.

Images