A cdn access acceleration method based on a graph neural network

Abstract

This invention discloses a CDN access acceleration method based on graph neural networks, comprising: Step 1: collecting CDN network node status data and link status data and constructing a dynamic attribute graph; Step 2: responding to access requests, extracting user region identifiers and target content identifiers, and generating content instance virtual nodes; Step 3: filtering candidate edge nodes and constructing a candidate insertion graph; Step 4: performing random walks based on link status, node status, and path cumulative status to generate path sequences; Step 5: inputting path features into an improved RWNN model and performing differential path encoding to generate path representation results; Step 6: calculating access latency based on path representation results and determining the target deployment node; Step 7: performing content writing and request redirection and recording access data. This invention effectively reduces CDN access latency.

Description

Technical Field

[0001] This invention relates to the field of graph neural networks and content delivery networks, and in particular to a method for accelerating CDN access based on graph neural networks. Background Technology

[0002] With the continuous growth of internet services and the widespread deployment of applications such as video, live streaming, and large file distribution, content delivery networks (CDNs) play a crucial role in reducing access latency and alleviating pressure on origin servers. Existing technologies for CDN access acceleration primarily rely on rule-based caching deployment methods or statistical feature-based node selection methods for content distribution optimization. However, these methods commonly suffer from the following problems in real-world network environments: Traditional CDN caching strategies often employ content deployment methods based on access frequency or simple popularity assessments, which struggle to depict the complex topological relationships and link state changes between nodes in the network. This results in limited cache hit rate and access latency optimization performance in dynamic network environments. Existing methods typically make decisions based on static or local features, lacking the ability to model the global structure of access paths and failing to accurately reflect the overall impact of different deployment locations on access paths. Some studies have attempted to introduce graph neural networks for node representation learning, but most only model single graph structures, neglecting the impact of network topology changes before and after content deployment on path distribution, making it difficult for models to depict the changing patterns of actual access paths. Furthermore, existing path modeling methods based on random walks typically encode only single path sequences, lacking a mechanism for comparing differences before and after path deployment, and failing to combine link state, node load, and cumulative path state for joint modeling, resulting in insufficient path representation capabilities and consequently affecting the accuracy of access latency prediction. In addition, current technologies do not fully integrate multi-dimensional factors such as path-level latency changes, replication costs, and node load during content deployment decision-making, making it difficult to achieve refined node selection and dynamic optimization.

[0003] Therefore, how to provide a CDN access acceleration method based on graph neural networks is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0004] One objective of this invention is to propose a CDN access acceleration method based on graph neural networks. This invention constructs a dynamic attribute graph for CDN and introduces virtual nodes for content instances. By combining candidate insertion graphs with a dual-graph structure comparison, it models the path changes before and after content deployment. By integrating a random walk mechanism that combines link states, node states, and cumulative path states, and by introducing an improved RWNN model with multi-state path memory and differential path encoding, it enhances path representation capabilities. Furthermore, by combining path latency estimation and multi-factor deployment evaluation, it achieves accurate selection of candidate nodes, thereby effectively reducing access latency and the number of origin pulls.

[0005] According to an embodiment of the present invention, a CDN access acceleration method based on a graph neural network includes the following steps: Step 1: Collect CDN network node status data and link status data, construct a CDN dynamic attribute graph, and configure corresponding attributes for nodes and links; Step 2: Respond to the access request, extract the user region identifier and target content identifier, generate a virtual node for the content instance, and configure the content attributes; Step 3: Filter candidate edge nodes in the CDN dynamic attribute graph, insert the content instance virtual nodes into the graph structure corresponding to each candidate edge node, establish deployment association edges, and generate a candidate insertion graph; Step 4: For each candidate insertion graph, perform a random walk starting from the user access node in the graph structure before and after insertion. Determine the node transfer relationship based on the link status parameters, node status parameters, and cumulative path status, and generate the path sequence before deployment and the path sequence after deployment. Step 5: Construct a path feature sequence based on the pre-deployment path sequence and the post-deployment path sequence, input the path feature sequence into the improved RWNN model, perform difference calculation on the pre-deployment path representation and the post-deployment path representation, and combine the path state information to generate the path representation result; Step 6: Aggregate the path representation results to obtain the access latency value corresponding to each candidate insertion graph, and determine the target deployment node by combining the baseline access latency value, content replication cost and node load value; Step 7: Perform content writing and cache space adjustment based on the target deployment node, and redirect access requests to the target deployment node, while recording access latency and number of times the request is routed back to the origin.

[0006] Optionally, step one specifically includes: The system collects operational data from user access nodes, edge cache nodes, regional center nodes, and origin nodes in the CDN network within a preset time window. The operational data includes node identifier, cache usage, total cache capacity, number of requests processed, request response time, and number of times requests are retrieved from the origin server. The transmission data of each communication link within the time window is collected. The transmission data includes source node identifier, destination node identifier, round-trip time, transmission data volume, link occupancy time, number of packet losses, and total number of packets sent. The running data and transmitted data are aligned based on the node identifier and the source node identifier and destination node identifier, and missing and duplicate records are deleted. For each node, the cache utilization rate is calculated by the ratio of cache usage to total cache capacity, the node load value is calculated by the ratio of the number of requests processed to the duration of the time window, the average response latency is calculated by the ratio of the cumulative value of request response time to the number of requests processed, and the origin pull ratio is calculated by the ratio of the number of origin pulls to the number of requests processed. For each communication link, the link bandwidth is calculated by the ratio of the amount of data transmitted to the link occupancy time, the packet loss rate is calculated by the ratio of the number of packet losses to the total number of packets sent, and the link congestion value is calculated by the ratio of the link occupancy time to the time window duration. A node set is constructed using node identifiers, and an edge set is constructed using source node identifiers and destination node identifiers. The node set and edge set are then combined to form a dynamic attribute graph for the CDN. Cache utilization, node load, average response latency, and origin-wait ratio are written into the node attribute vector, while round-trip latency, link bandwidth, packet loss rate, and link congestion are written into the edge attribute vector.

[0007] Optionally, step two specifically includes: Receive an access request message and parse the source IP address, target content address identifier, and request timestamp in the access request message; The corresponding user region identifier is determined by matching the source IP address with the IP address and region mapping table constructed based on IP address geospatial data. The target content address identifier is searched in the content index table to determine the target content identifier, and the corresponding content size, content category and historical access records are extracted. A virtual node identifier for a content instance is constructed based on the user region identifier and the target content identifier, and the virtual node identifier for the content instance is added as a node identifier to the node set of the CDN dynamic attribute graph. The access fluctuation cycle is calculated based on the time series distribution of historical access records, and a statistical time window is determined based on the access fluctuation cycle. Within the statistical time window, the number of accesses to the target content and the number of cross-regional accesses are counted. The content popularity value is calculated by the ratio of the number of accesses to the duration of the time window. When the number of accesses is greater than zero, the cross-regional access ratio is calculated by the ratio of the number of cross-regional accesses to the number of accesses. A content attribute vector is constructed based on the content size, content category, content popularity value, and cross-regional access ratio, and the content attribute vector is written into the node attributes of the corresponding content instance virtual node.

[0008] Optionally, step three specifically includes: Extract the edge cache node set from the node set of the CDN dynamic attribute graph, and read the node attribute vector of each edge cache node; Candidate scores for each edge cache node are calculated based on node attribute vectors. The candidate scores are normalized according to node load, cache utilization, and average response latency. The normalized results are then subtracted by 1 and weighted and summed. Candidate nodes are sorted from largest to smallest according to their scores to form a candidate node sequence. The first K edge cache nodes are selected as the candidate edge node set according to the candidate node sequence. K is calculated by rounding up the product of the total number of nodes in the set and the candidate ratio. For each candidate edge node, the CDN dynamic attribute graph is copied to form a corresponding graph structure, and the virtual nodes of the content instances are added to the node set of the corresponding graph structure. Establish deployment association edges between content instance virtual nodes and corresponding candidate edge nodes. The edge attributes of the deployment association edges include content size and node remaining caching capacity. Establish a back-to-origin connection edge between the virtual node of the content instance and the origin node corresponding to the target content. The edge attributes of the back-to-origin connection edge include the origin access latency and the link bandwidth. The graph structures after adding content instance virtual nodes, deploying associated edges, and back-to-origin associated edges are output as a set of candidate insertion graphs.

[0009] Optionally, step four specifically includes: In each candidate insertion graph, the user access node is determined as the starting node of the random walk, and the path sequence is initialized in the graph structure before the insertion of the virtual node of the content instance and in the graph structure after the insertion of the virtual node of the content instance. Read the set of neighboring nodes of the current node, obtain the link status parameters of the corresponding link and the node status parameters of the neighboring nodes, and record the cumulative path delay of the current path; The link latency, link bandwidth, and link congestion level of adjacent nodes are normalized, and the node cache status, node remaining cache capacity, and node load value are also normalized. Based on the normalized link state parameters, node state parameters, and cumulative path delay, the link delay and link congestion level are processed with reverse weights, the link bandwidth, node buffer status and node remaining buffer capacity are processed with forward weights, and the node load value is processed with reverse weights. The weighted summation is then performed to obtain the transfer weight of each adjacent node. The transition weights of each adjacent node are normalized to form a transition probability distribution. The next node is selected based on the transition probability distribution, the selected node is added to the path sequence, and the current link delay is accumulated to the cumulative path delay. Starting from the current node, the node transfer process is executed cyclically, including determining adjacent nodes, processing state parameters, calculating transfer weights, and selecting nodes, until the cumulative path delay reaches the delay threshold or the target edge node or source node is reached, generating a single path sequence. Perform N random walks in the graph structure before and after insertion respectively to obtain the path sequence before deployment and the path sequence after deployment.

[0010] Optionally, step five specifically includes: Input the path feature sequences corresponding to the pre-deployment path sequence and the post-deployment path sequence into the improved RWNN model; The improved RWNN model includes a pre-deployment path coding branch, a post-deployment path coding branch, a differential path coding module, a multi-state path memory module, and a path representation output module; At the input end, the features of each path step in the pre-deployment path sequence and the post-deployment path sequence are vectorized and encoded respectively. The path step features include node cache status, node remaining cache capacity, node load value, link latency, link bandwidth, link congestion degree and cumulative path latency, forming the pre-deployment path input sequence and the post-deployment path input sequence. Input the pre-deployment path input sequence into the pre-deployment path encoding branch, and input the post-deployment path input sequence into the post-deployment path encoding branch. Perform sequence encoding calculation step by step in the two path encoding branches according to the path step order to generate the pre-deployment path hidden state sequence and the post-deployment path hidden state sequence. In the multi-state path memory module, the following processing is performed for each path step: Extract the link latency, link bandwidth, and link congestion level corresponding to the current path step, and construct a local link state memory vector; Based on the type of the node corresponding to the current path step, the region migration state memory vector is constructed by encoding the category among user access node, regional center node, edge cache node and source station node. Calculate the number of graph hops from the node corresponding to the current path step to the virtual node of the content instance and the source node, and construct the content proximity state memory vector; The local link state memory vector, the region migration state memory vector, and the content proximity state memory vector are concatenated with the current path step hidden state to generate the path step joint encoding result. The following processing is performed in the differential path encoding module: Extract the terminal hidden states of the pre-deployment path hidden state sequence and the post-deployment path hidden state sequence; calculate the element-wise difference between the post-deployment path terminal hidden state and the pre-deployment path terminal hidden state to generate a difference vector; concatenate the pre-deployment path terminal hidden state, the post-deployment path terminal hidden state, and the difference vector in sequence; perform a linear transformation on the concatenation result to generate the differential path encoding result. The path step joint encoding result is used as the path step level input, and the differential path encoding result is used as the path level feature input to the path representation output module. The weight coefficients corresponding to each path step are calculated based on the path step joint encoding result in the path step dimension, and a weighted summation calculation is performed according to the weight coefficients to generate the path representation result of the corresponding candidate insertion graph.

[0011] Optionally, step six specifically includes: Read the path representation results corresponding to each candidate insertion graph, and normalize the path representation results according to the path sequence dimension; The path delay estimate is calculated based on the normalized path representation results. The path delay estimate is calculated by linearly combining the path representation results with the mapping coefficients calculated based on the distribution of the path representation results. The access delay value corresponding to the candidate insertion graph is obtained by weighted summation of multiple path delay estimates corresponding to the same candidate insertion graph. The weight coefficients are determined based on the vector norm of the path representation result. The baseline access latency value is calculated by weighted summation of the path latency estimates for the corresponding path sequence when no content instance virtual nodes are inserted. The content replication cost is calculated by using the ratio of the content size to the remaining cache capacity of the target node, and weighted by combining the current load value of the node. Deployment evaluation metrics are constructed based on access latency, baseline access latency, content replication cost, and node load. The deployment evaluation metrics are obtained by calculating the difference between the baseline access latency and the access latency and subtracting the weighted result of the content replication cost and node load. The deployment evaluation metrics corresponding to each candidate edge node are compared, and the candidate edge node with the largest deployment evaluation metric value is selected as the target deployment node.

[0012] Optionally, step seven specifically includes: Receive the target deployment node identifier, and read the current cache usage, total cache capacity, and list of cached content of the target deployment node; Determine the relationship between the remaining cache capacity of the target deployment node and the size of the target content. When the remaining cache capacity is less than the size of the target content, sort the list of cached content in ascending order of content popularity value, select the content with the lowest popularity value and delete it, and update the cache usage at the same time until the remaining cache capacity is not less than the size of the target content. Write the target content to the cache space of the target deployment node, and update the cache usage and the list of cached content; Update the routing table entries in the user access node to map the target content address to the target deployment node address, and forward the access request to the target deployment node according to the updated routing table; Read the content index table in the cache space and perform a matching operation based on the target content identifier; when the matching result exists, read the corresponding content data from the cache space and return it; when the matching result does not exist, establish a back-to-origin connection between the target deployment node and the origin node and obtain the content data, and write the obtained content data into the cache space. During the access request processing, the request initiation time and response return time are recorded. The access latency is calculated by performing a difference between the response return time and the request initiation time, and the number of back-to-origin requests is recorded. The access latency and number of times the data returns to the origin are associated with the corresponding target deployment node identifier and stored, and then written into the access record data table.

[0013] The beneficial effects of this invention are: This invention constructs a dynamic attribute graph for CDN that includes node and link states, and introduces virtual nodes for content instances to explicitly model content deployment behavior. Combining a candidate insertion graph generation mechanism and a dual-graph structure comparison strategy, it addresses the difficulty in depicting path changes and the dynamic impact of network states before and after content deployment in existing technologies. It proposes a path modeling method based on the differences in graph structures before and after insertion, significantly improving the expressive power of access path changes and the accuracy of deployment decisions. In the path modeling stage, a state-driven random walk mechanism is introduced, fusing link state parameters, node state parameters, and cumulative path states to construct transition relationships, achieving adaptive characterization of path distribution in dynamic network environments. In the path encoding stage, an improved RWNN model is designed to construct… The system establishes pre-deployment and post-deployment path coding branches and introduces a multi-state path memory module and a differential path coding module. Through joint modeling of local link states, regional migration states, and content proximity states, as well as differential calculation of path end states, it effectively enhances the model's ability to perceive changes in path structure. In the path representation output stage, it generates stable path representation results through weighted aggregation at the path step level, and combines access latency estimation and a multi-factor deployment evaluation mechanism to achieve fine-grained screening of candidate edge nodes. Finally, it completes content deployment and access acceleration through dynamic adjustment of cache space and request redirection strategies, and outputs key indicators such as access latency and origin pull count, realizing structural awareness, path differential modeling, and intelligent optimization decision-making in the CDN access process. Attached Figure Description

[0014] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1This is a schematic diagram of the overall process of a CDN access acceleration method based on graph neural networks proposed in this invention. Figure 2 This is a schematic diagram of the structure of the improved RWNN model in the CDN access acceleration method based on graph neural networks proposed in this invention. Detailed Implementation

[0015] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0016] refer to Figures 1-2 A CDN access acceleration method based on graph neural networks includes the following steps: Step 1: Collect CDN network node status data and link status data, construct a CDN dynamic attribute graph, and configure corresponding attributes for nodes and links; Step 2: Respond to the access request, extract the user region identifier and target content identifier, generate a virtual node for the content instance, and configure the content attributes; Step 3: Filter candidate edge nodes in the CDN dynamic attribute graph, insert the virtual nodes of content instances into the graph structure corresponding to each candidate edge node, establish deployment association edges, and generate a candidate insertion graph; Step 4: For each candidate insertion graph, perform a random walk starting from the user access node in the graph structure before and after insertion. Determine the node transfer relationship based on the link status parameters, node status parameters, and cumulative path status, and generate the path sequence before deployment and the path sequence after deployment. Step 5: Construct a path feature sequence based on the pre-deployment path sequence and the post-deployment path sequence. Input the path feature sequence into the improved RWNN model, perform difference calculation on the pre-deployment path representation and the post-deployment path representation, and combine the path state information to generate the path representation result. Step 6: Aggregate the path representation results to obtain the access latency value corresponding to each candidate insertion graph, and determine the target deployment node by combining the baseline access latency value, content replication cost and node load value; Step 7: Perform content writing and cache space adjustment based on the target deployment node, and redirect access requests to the target deployment node, while recording access latency and number of times the request goes back to the origin server.

[0017] In this embodiment, step one specifically includes: Collect operational data of user access nodes, edge cache nodes, regional center nodes and origin nodes in the CDN network within a preset time window. The operational data includes node identifier, cache usage, total cache capacity, number of requests processed, request response time and number of times to retrieve data from the origin server. Collect transmission data of each communication link within the time window. The transmission data includes source node identifier, destination node identifier, round-trip time, transmission data volume, link occupancy time, number of packet losses, and total number of packets sent. Alignment of running and transmitted data is performed based on node identifiers and source and destination node identifiers, and missing and duplicate records are deleted. For each node, the cache utilization rate is calculated by the ratio of cache usage to total cache capacity, the node load value is calculated by the ratio of the number of requests processed to the duration of the time window, the average response latency is calculated by the ratio of the cumulative value of request response time to the number of requests processed, and the origin pull ratio is calculated by the ratio of the number of origin pulls to the number of requests processed. For each communication link, the link bandwidth is calculated by the ratio of the amount of data transmitted to the link occupancy time, the packet loss rate is calculated by the ratio of the number of packet losses to the total number of packets sent, and the link congestion value is calculated by the ratio of the link occupancy time to the time window duration. A node set is constructed using node identifiers, and an edge set is constructed using source node identifiers and destination node identifiers. The node set and edge set are then combined to form a dynamic attribute graph for the CDN. Cache utilization, node load, average response latency, and origin-wait ratio are written into the node attribute vector, while round-trip latency, link bandwidth, packet loss rate, and link congestion are written into the edge attribute vector.

[0018] In this implementation, node status data and link status data are continuously collected within a preset time window at a fixed sampling period, and various types of data are synchronized and aligned according to timestamps, so that the node running status and the corresponding link transmission status form a data mapping relationship at the same moment. During the alignment process, abnormal fluctuation data is smoothed by using a sliding window averaging method, and continuously missing data is filled in according to the average of adjacent time points. During the writing of node attributes and link attributes, a one-to-one mapping relationship is established according to the node identifier and the link identifier, and all parameters calculated within the same time window are uniformly written into the attribute vector and updated in chronological order, so that the node attributes and edge attributes in the CDN dynamic attribute graph remain synchronized with the changes over time.

[0019] In this embodiment, step two specifically involves: Receive access request messages and parse the source IP address, target content address identifier, and request timestamp in the access request messages; The corresponding user region identifier is determined by matching the source IP address with the IP address and region mapping table constructed based on IP address geospatial data. The target content identifier is searched in the content index table based on the target content address identifier, and the corresponding content size, content category and historical access records are extracted. Based on the user region identifier and the target content identifier, construct a virtual node identifier for the content instance, and add the virtual node identifier for the content instance as a node identifier to the node set of the CDN dynamic attribute graph; The access fluctuation cycle is calculated based on the time series distribution of historical access records, and a statistical time window is determined based on the access fluctuation cycle. Within the statistical time window, the number of accesses to the target content and the number of cross-regional accesses are counted. The content popularity value is calculated by the ratio of the number of accesses to the duration of the time window. When the number of accesses is greater than zero, the cross-regional access ratio is calculated by the ratio of the number of cross-regional accesses to the number of accesses. A content attribute vector is constructed based on the content size, content category, content popularity value, and cross-regional access ratio, and then the content attribute vector is written into the node attributes of the corresponding content instance virtual node.

[0020] In this implementation, the access fluctuation cycle is detected by sorting the historical access sequence of the target content by timestamp and then performing periodic detection. The sampling unit is divided into 60-second intervals, and no less than 3600 sampling points are continuously collected. The local access frequency average is calculated using a sliding window with a length of 300. The cycle length is determined based on the mean change amplitude. The statistical time window is a value within the range of 1.0 to 1.5 times the cycle length. The access count is the cumulative value of the number of requests for the target content within the statistical time window. The cross-region access count is the cumulative value of the number of requests where the user's region identifier does not match the content's historical service region identifier. When the access count is less than 5, the cross-region access ratio is 0. Content with a popularity value greater than 0.1 is marked as high-popularity content, and content with a popularity value less than 0.02 is marked as low-popularity content. When constructing the content attribute vector, the content size, content category code, content popularity value, and cross-region access ratio are written in sequence.

[0021] In this embodiment, step three specifically includes: Extract the edge cache node set from the node set of the CDN dynamic attribute graph, and read the node attribute vector of each edge cache node; Candidate scores for each edge cache node are calculated based on node attribute vectors. The candidate scores are normalized according to node load, cache utilization, and average response latency. The normalized results are then subtracted by 1 and weighted and summed. Candidate nodes are sorted from largest to smallest according to their scores to form a candidate node sequence. The first K edge cache nodes are selected as the candidate edge node set according to the candidate node sequence. K is calculated by rounding up the product of the total number of nodes in the set and the candidate ratio. For each candidate edge node, the CDN dynamic attribute graph is copied to form a corresponding graph structure, and the virtual nodes of the content instances are added to the node set of the corresponding graph structure. Establish deployment association edges between content instance virtual nodes and corresponding candidate edge nodes. The edge attributes of the deployment association edges include content size and node remaining caching capacity. Establish a back-to-origin connection edge between the virtual node of the content instance and the origin node corresponding to the target content. The edge attributes of the back-to-origin connection edge include the origin access latency and the link bandwidth. The graph structures after adding content instance virtual nodes, deploying associated edges, and back-to-origin associated edges are output as a set of candidate insertion graphs.

[0022] In this implementation, the node load value, cache utilization rate, and average response latency in the node attribute vector are linearly normalized with a minimum value of 0 and a maximum value of 1, respectively, and the weight coefficients are 0.4, 0.3, and 0.3, respectively; the candidate ratio is a value between 0.2 and 0.3, preferably 0.25. When the total number of nodes in the set is N, the number of candidate edge nodes K is determined by multiplying N by 0.25 and rounding up; when the remaining cache capacity of a node is less than 1.2 times the content size, no deployment association edge is established; the origin server access latency is the average value measured within the most recent 10 time windows, and the link bandwidth is the average value of the most recent 5 measurement results; when copying the CDN dynamic attribute graph, all nodes and link structures are retained, and the node set and edge set are updated after inserting the virtual node of the content instance, thereby forming multiple sets of candidate insertion graphs.

[0023] In this embodiment, step four specifically includes: In each candidate insertion graph, the user access node is determined as the starting node of the random walk, and the path sequence is initialized in the graph structure before the insertion of the virtual node of the content instance and in the graph structure after the insertion of the virtual node of the content instance. Read the set of neighboring nodes of the current node, obtain the link status parameters of the corresponding link and the node status parameters of the neighboring nodes, and record the cumulative path delay of the current path; The link latency, link bandwidth, and link congestion level of adjacent nodes are normalized, and the node cache status, node remaining cache capacity, and node load value are also normalized. Based on the normalized link state parameters, node state parameters, and cumulative path delay, the link delay and link congestion level are processed with reverse weights, the link bandwidth, node buffer status and node remaining buffer capacity are processed with forward weights, and the node load value is processed with reverse weights. The weighted summation is then performed to obtain the transfer weight of each adjacent node. The transition weights of each adjacent node are normalized to form a transition probability distribution. The next node is selected based on the transition probability distribution, the selected node is added to the path sequence, and the current link delay is accumulated to the cumulative path delay. Starting from the current node, the node transfer process is executed cyclically, including determining adjacent nodes, processing state parameters, calculating transfer weights, and selecting nodes, until the cumulative path delay reaches the delay threshold or the target edge node or source node is reached, generating a single path sequence. Perform N random walks in the graph structure before and after insertion respectively to obtain the path sequence before deployment and the path sequence after deployment.

[0024] In this implementation, link delay, link bandwidth, and link congestion are linearly normalized with a minimum value of 0 and a maximum value of 1, respectively. Node cache status is represented by binary encoding, and node remaining cache capacity and node load are normalized according to the maximum and minimum values ​​of the current node set. The weight coefficients for link delay, link bandwidth, link congestion, node cache status, node remaining cache capacity, and node load in the transition weight are 0.25, 0.20, 0.15, 0.15, 0.15, and 0.10, respectively. The delay threshold is 3.0 times the average link delay of all candidate insertions in the current graph. The number of random walks N is an integer between 20 and 50, preferably 30. Path generation is terminated when the cumulative path delay exceeds the delay threshold or the path length reaches 15.

[0025] In this embodiment, step five specifically includes: Input the path feature sequences corresponding to the pre-deployment path sequence and the post-deployment path sequence into the improved RWNN model; The improved RWNN model includes a pre-deployment path coding branch, a post-deployment path coding branch, a differential path coding module, a multi-state path memory module, and a path representation output module; At the input end, the features of each path step in the pre-deployment path sequence and the post-deployment path sequence are vectorized and encoded respectively. The path step features include node cache status, node remaining cache capacity, node load value, link latency, link bandwidth, link congestion degree and cumulative path latency, forming the pre-deployment path input sequence and the post-deployment path input sequence. Input the pre-deployment path input sequence into the pre-deployment path encoding branch, and input the post-deployment path input sequence into the post-deployment path encoding branch. Perform sequence encoding calculation step by step in the two path encoding branches according to the path step order to generate the pre-deployment path hidden state sequence and the post-deployment path hidden state sequence. In the multi-state path memory module, the following processing is performed for each path step: Extract the link latency, link bandwidth, and link congestion level corresponding to the current path step, and construct a local link state memory vector; Based on the type of the node corresponding to the current path step, the region migration state memory vector is constructed by encoding the category among user access node, regional center node, edge cache node and source station node. Calculate the number of graph hops from the node corresponding to the current path step to the virtual node of the content instance and the source node, and construct the content proximity state memory vector; The local link state memory vector, the region migration state memory vector, and the content proximity state memory vector are concatenated with the current path step hidden state to generate the path step joint encoding result. The following processing is performed in the differential path encoding module: Extract the terminal hidden states of the pre-deployment path hidden state sequence and the post-deployment path hidden state sequence; calculate the element-wise difference between the post-deployment path terminal hidden state and the pre-deployment path terminal hidden state to generate a difference vector; concatenate the pre-deployment path terminal hidden state, the post-deployment path terminal hidden state, and the difference vector in sequence; perform a linear transformation on the concatenation result to generate the differential path encoding result. The path step joint encoding result is used as the path step level input, and the differential path encoding result is used as the path level feature input to the path representation output module. The weight coefficients corresponding to each path step are calculated based on the path step joint encoding result in the path step dimension, and a weighted summation calculation is performed according to the weight coefficients to generate the path representation result of the corresponding candidate insertion graph.

[0026] In this implementation, the path step feature vector is set to 32 dimensions. Node attributes and link attributes are mapped to 16 dimensions respectively and then concatenated to form the input features. Both the pre-deployment path encoding branch and the post-deployment path encoding branch adopt a cyclic unit structure with a hidden layer dimension of 64, and the upper limit of the sequence length is 15. The local link state memory vector is set to 8 dimensions, the region migration state memory vector is set to 4 dimensions, and the content proximity state memory vector is set to 4 dimensions. The three types of memory vectors are concatenated with the hidden state of the path step and then uniformly mapped to 64 dimensions. The vector concatenation dimension in differential path encoding is 192, which is compressed to 64 dimensions through linear transformation. The path step weight coefficient is normalized, and its value ranges from 0 to 1. The number of path steps is 10 to 15. When the path step weight coefficient is lower than 0.05, the corresponding path step does not participate in the weighted calculation, thus obtaining a stable path representation result. Both the improved RWNN model and the RWNN model are based on random walks to generate path sequences and perform stepwise encoding on the path sequences. By performing sequence modeling on the node features and link features in the path, the path representation information in the graph structure is extracted. Both use a sequence encoding structure to recursively calculate the path step features and use the changes in the state of each node and link in the path to reflect the propagation characteristics in the graph structure, thereby realizing the expression of path-level features. Based on this, the improved RWNN model sets up a pre-deployment path encoding branch and a post-deployment path encoding branch to encode the path sequences before and after the insertion of virtual nodes of content instances. During the encoding process, a multi-state path memory module is introduced to construct a multi-dimensional memory vector from the link state, node type migration state, and content proximity state, and to jointly encode it with the hidden state of the path step. At the same time, a differential path encoding module is set up to perform subtraction, concatenation, and linear mapping on the pre-deployment path end state and the post-deployment path end state to form a differential path representation that reflects the path change characteristics. Through the above structural improvements, the model retains the ability to represent random walk paths while introducing a path state comparison mechanism and a multi-dimensional state memory mechanism. This enables the path representation to reflect changes in path structure and network state before and after content deployment, enhances the ability to represent access latency trends, and improves the ability to distinguish candidate edge nodes, thereby improving the accuracy and stability of CDN access acceleration decisions.

[0027] In this embodiment, step six specifically includes: Read the path representation results corresponding to each candidate insertion graph, and normalize the path representation results according to the path sequence dimension; The path delay estimate is calculated based on the normalized path representation results. The path delay estimate is calculated by linearly combining the path representation results with the mapping coefficients calculated based on the distribution of the path representation results. The access delay value corresponding to the candidate insertion graph is obtained by weighted summation of multiple path delay estimates corresponding to the same candidate insertion graph. The weight coefficients are determined based on the vector norm of the path representation result. The baseline access latency value is calculated by weighted summation of the path latency estimates for the corresponding path sequence when no content instance virtual nodes are inserted. The content replication cost is calculated by using the ratio of the content size to the remaining cache capacity of the target node, and weighted by combining the current load value of the node. Deployment evaluation metrics are constructed based on access latency, baseline access latency, content replication cost, and node load. The deployment evaluation metrics are obtained by calculating the difference between the baseline access latency and the access latency and subtracting the weighted result of the content replication cost and node load. The deployment evaluation metrics corresponding to each candidate edge node are compared, and the candidate edge node with the largest deployment evaluation metric value is selected as the target deployment node.

[0028] In this implementation, the path representation results are normalized to a minimum value of 0 and a maximum value of 1. The path latency estimate is calculated by multiplying each dimension of the path representation result element-wise with the mapping coefficient and then summing the results. The mapping coefficient is obtained by least-squares fitting of the historical path representation results and the actual measured path latency. The path weight coefficient is obtained by normalizing the path representation result using the L2 norm, and its value ranges from 0 to 1. The baseline access latency value is calculated by applying the same weight to the path latency estimate without inserting virtual nodes of content instances. In the content replication cost, the content size is measured in MB, and the remaining cache capacity of the node is measured in MB. When the remaining cache capacity is less than 1.1 times the content size, a penalty coefficient of 1.5 is added. When the node load value is greater than 0.8, a penalty coefficient of 1.2 is added. The weight coefficients of each deployment evaluation index are 1.0, 0.8, and 0.5, respectively.

[0029] In this embodiment, step seven specifically includes: Receive the target deployment node identifier, and read the current cache usage, total cache capacity, and list of cached content of the target deployment node; Determine the relationship between the remaining cache capacity of the target deployment node and the size of the target content. When the remaining cache capacity is less than the size of the target content, sort the list of cached content in ascending order of content popularity value, select the content with the lowest popularity value and delete it, and update the cache usage at the same time until the remaining cache capacity is not less than the size of the target content. Write the target content to the cache space of the target deployment node, and update the cache usage and the list of cached content; Update the routing table entries in the user access node to map the target content address to the target deployment node address, and forward the access request to the target deployment node according to the updated routing table; Read the content index table in the cache space and perform a matching operation based on the target content identifier; when the matching result exists, read the corresponding content data from the cache space and return it; when the matching result does not exist, establish a back-to-origin connection between the target deployment node and the origin node and obtain the content data, and write the obtained content data into the cache space. During the access request processing, the request initiation time and response return time are recorded. The access latency is calculated by performing a difference between the response return time and the request initiation time, and the number of back-to-origin requests is recorded. The access latency and number of times the data returns to the origin are associated with the corresponding target deployment node identifier and stored, and then written into the access record data table.

[0030] In this implementation, the total cache capacity is measured in GB, and the cache usage and remaining cache capacity are statistically analyzed in bytes. A cache space adjustment is triggered when the remaining cache capacity is less than 1.1 times the target content size. The content popularity value is calculated as the ratio of the number of accesses to the length of the statistical time window; content with a popularity value less than 0.02 is prioritized for deletion. A single cache deletion operation releases at least 0.5 times the target content size. The routing table update latency does not exceed 10ms. Origin connections are established using HTTP long connections, with an origin timeout of 200ms. Access latency is calculated with millisecond precision, and the timestamp resolution is 1ms. Requests with more than 3 consecutive origin pulls are marked as abnormal. The access record data table is indexed and stored according to the target deployment node identifier, and a batch write operation is performed every 100 records.

[0031] Example 1: To verify the feasibility of this invention in practice, it was applied to the actual CDN network environment of a large-scale video content distribution platform. This platform deploys over 300 edge caching nodes nationwide, covering more than 5 million users, with daily request volume reaching hundreds of millions. With the continuous growth of high-definition video and live streaming services, the network faces problems such as complex access paths, significant fluctuations in node load, and concentrated access to popular content, resulting in high latency for users in some areas and a significant increase in origin server backhaul pressure. Traditional caching deployment methods based on access frequency or static rules cannot respond promptly to changes in network status, leading to problems such as unreasonable path selection and low cache resource utilization. Therefore, the graph neural network-based CDN access acceleration method of this invention is introduced into this scenario to achieve dynamic optimization of access paths and content deployment strategies.

[0032] In practical applications, the system first collects operational data and link transmission data from each node in the CDN network, including node cache usage, node load, link latency, bandwidth, and congestion status, and constructs a dynamic attribute graph model. Then, when a user initiates a content access request, the system resolves the user's IP address to their region, identifies the target content identifier, and generates a corresponding virtual node for the content instance. This virtual node is then inserted into the graph structure corresponding to the candidate edge node, forming multiple candidate insertion graphs. For each candidate insertion graph, a random walk is performed based on link status, node status, and cumulative path latency to generate path sequences before and after content deployment. The paths are encoded using an improved RWNN model, and path change features are extracted using a differential path encoding mechanism. Multi-state path memory is also incorporated to enhance path representation capabilities.

[0033] After obtaining the path representation results, the access latency value corresponding to each candidate insertion graph is calculated through path latency estimation and weighted aggregation. Simultaneously, considering the baseline access latency, content replication cost, and node load, deployment evaluation metrics are calculated, and the optimal target deployment node is selected from multiple candidate nodes. Subsequently, content writing and cache adjustment are performed on the target node, and request redirection is implemented by updating the routing table, ensuring that user access requests preferentially hit edge cache nodes, thereby reducing origin requests and lowering access latency.

[0034] To verify the practical effectiveness of this invention, a comparative experiment was conducted with the method of this invention, traditional content popularity-based caching strategies, and GNN methods based on simple graph features. Under the same network environment and request data conditions, the performance of different methods in terms of access latency, origin retrieval rate, cache hit rate, and node load balancing was compared. The experimental results are shown in Table 1.

[0035] Table 1 Comparison of CDN Access Acceleration Effects - Experimental Data

[0036] As shown in Table 1, compared to traditional methods, this invention reduces average access latency by approximately 30.8% and maximum access latency by 29.8%, significantly improving the user experience. The origin server hit rate is reduced to 18.7%, effectively alleviating pressure on the origin server. The cache hit rate is increased to 81.5%, indicating more reasonable content deployment. Simultaneously, the standard deviation of node load decreases, indicating a more balanced network load. Furthermore, access latency is significantly reduced in both hot and non-hot areas, demonstrating that this invention can adapt to different access scenarios and achieve global optimization.

[0037] In summary, through application verification in a real CDN network environment, this invention can effectively solve the problems of insufficient path modeling capabilities, inaccurate deployment decisions, and poor dynamic adaptability of network status in existing technologies. It enables fine-grained modeling of access paths and intelligent optimization of content deployment strategies, thereby significantly reducing access latency, reducing the number of origin pulls, and improving overall system performance.

[0038] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A CDN access acceleration method based on graph neural networks, characterized in that, Includes the following steps: Step 1: Collect CDN network node status data and link status data, construct a CDN dynamic attribute graph, and configure corresponding attributes for nodes and links; Step 2: Respond to the access request, extract the user region identifier and target content identifier, generate a virtual node for the content instance, and configure the content attributes; Step 3: Filter candidate edge nodes in the CDN dynamic attribute graph, insert the content instance virtual nodes into the graph structure corresponding to each candidate edge node, establish deployment association edges, and generate a candidate insertion graph; Step 4: For each candidate insertion graph, perform a random walk starting from the user access node in the graph structure before and after insertion. Determine the node transfer relationship based on the link status parameters, node status parameters, and cumulative path status, and generate the path sequence before deployment and the path sequence after deployment. Step 5: Construct a path feature sequence based on the pre-deployment path sequence and the post-deployment path sequence, input the path feature sequence into the improved RWNN model, perform difference calculation on the pre-deployment path representation and the post-deployment path representation, and combine the path state information to generate the path representation result; Step 6: Aggregate the path representation results to obtain the access latency value corresponding to each candidate insertion graph, and determine the target deployment node by combining the baseline access latency value, content replication cost and node load value; Step 7: Perform content writing and cache space adjustment based on the target deployment node, and redirect access requests to the target deployment node, while recording access latency and number of times the request is routed back to the origin.

2. The CDN access acceleration method based on graph neural networks according to claim 1, characterized in that, Step one specifically involves: The system collects operational data from user access nodes, edge cache nodes, regional center nodes, and origin nodes in the CDN network within a preset time window. The operational data includes node identifier, cache usage, total cache capacity, number of requests processed, request response time, and number of times requests are retrieved from the origin server. The transmission data of each communication link within the time window is collected. The transmission data includes source node identifier, destination node identifier, round-trip time, transmission data volume, link occupancy time, number of packet losses, and total number of packets sent. The running data and transmitted data are aligned based on the node identifier and the source node identifier and destination node identifier, and missing and duplicate records are deleted. For each node, the cache utilization rate is calculated by the ratio of cache usage to total cache capacity, the node load value is calculated by the ratio of the number of requests processed to the duration of the time window, the average response latency is calculated by the ratio of the cumulative value of request response time to the number of requests processed, and the origin pull ratio is calculated by the ratio of the number of origin pulls to the number of requests processed. For each communication link, the link bandwidth is calculated by the ratio of the amount of data transmitted to the link occupancy time, the packet loss rate is calculated by the ratio of the number of packet losses to the total number of packets sent, and the link congestion value is calculated by the ratio of the link occupancy time to the time window duration. A node set is constructed using node identifiers, and an edge set is constructed using source node identifiers and destination node identifiers. The node set and edge set are then combined to form a dynamic attribute graph for the CDN. Cache utilization, node load, average response latency, and origin-wait ratio are written into the node attribute vector, while round-trip latency, link bandwidth, packet loss rate, and link congestion are written into the edge attribute vector.

3. The CDN access acceleration method based on graph neural networks according to claim 1, characterized in that, Step two specifically involves: Receive an access request message and parse the source IP address, target content address identifier, and request timestamp in the access request message; The corresponding user region identifier is determined by matching the source IP address with the IP address and region mapping table constructed based on IP address geospatial data. The target content address identifier is searched in the content index table to determine the target content identifier, and the corresponding content size, content category and historical access records are extracted. A virtual node identifier for a content instance is constructed based on the user region identifier and the target content identifier, and the virtual node identifier for the content instance is added as a node identifier to the node set of the CDN dynamic attribute graph. The access fluctuation cycle is calculated based on the time series distribution of historical access records, and a statistical time window is determined based on the access fluctuation cycle. Within the statistical time window, the number of accesses to the target content and the number of cross-regional accesses are counted. The content popularity value is calculated by the ratio of the number of accesses to the duration of the time window. When the number of accesses is greater than zero, the cross-regional access ratio is calculated by the ratio of the number of cross-regional accesses to the number of accesses. A content attribute vector is constructed based on the content size, content category, content popularity value, and cross-regional access ratio, and the content attribute vector is written into the node attributes of the corresponding content instance virtual node.

4. The CDN access acceleration method based on graph neural networks according to claim 1, characterized in that, Step three specifically involves: Extract the edge cache node set from the node set of the CDN dynamic attribute graph, and read the node attribute vector of each edge cache node; Candidate scores for each edge cache node are calculated based on node attribute vectors. The candidate scores are normalized according to node load, cache utilization, and average response latency. The normalized results are then subtracted by 1 and weighted and summed. Candidate nodes are sorted from largest to smallest according to their scores to form a candidate node sequence. The first K edge cache nodes are selected as the candidate edge node set according to the candidate node sequence. K is calculated by rounding up the product of the total number of nodes in the set and the candidate ratio. For each candidate edge node, the CDN dynamic attribute graph is copied to form a corresponding graph structure, and the virtual nodes of the content instances are added to the node set of the corresponding graph structure. Establish deployment association edges between content instance virtual nodes and corresponding candidate edge nodes. The edge attributes of the deployment association edges include content size and node remaining caching capacity. Establish a back-to-origin connection edge between the virtual node of the content instance and the origin node corresponding to the target content. The edge attributes of the back-to-origin connection edge include the origin access latency and the link bandwidth. The graph structures after adding content instance virtual nodes, deploying associated edges, and back-to-origin associated edges are output as a set of candidate insertion graphs.

5. The CDN access acceleration method based on graph neural networks according to claim 1, characterized in that, Step four specifically involves: In each candidate insertion graph, the user access node is determined as the starting node of the random walk, and the path sequence is initialized in the graph structure before the insertion of the virtual node of the content instance and in the graph structure after the insertion of the virtual node of the content instance. Read the set of neighboring nodes of the current node, obtain the link status parameters of the corresponding link and the node status parameters of the neighboring nodes, and record the cumulative path delay of the current path; The link latency, link bandwidth, and link congestion level of adjacent nodes are normalized, and the node cache status, node remaining cache capacity, and node load value are also normalized. Based on the normalized link state parameters, node state parameters, and cumulative path delay, the link delay and link congestion level are processed with reverse weights, the link bandwidth, node buffer status and node remaining buffer capacity are processed with forward weights, and the node load value is processed with reverse weights. The weighted summation is then performed to obtain the transfer weight of each adjacent node. The transition weights of each adjacent node are normalized to form a transition probability distribution. The next node is selected based on the transition probability distribution, the selected node is added to the path sequence, and the current link delay is accumulated to the cumulative path delay. Starting from the current node, the node transfer process is executed cyclically, including determining adjacent nodes, processing state parameters, calculating transfer weights, and selecting nodes, until the cumulative path delay reaches the delay threshold or the target edge node or source node is reached, generating a single path sequence. Perform N random walks in the graph structure before and after insertion respectively to obtain the path sequence before deployment and the path sequence after deployment.

6. The CDN access acceleration method based on graph neural networks according to claim 1, characterized in that, Step five specifically involves: Input the path feature sequences corresponding to the pre-deployment path sequence and the post-deployment path sequence into the improved RWNN model; The improved RWNN model includes a pre-deployment path coding branch, a post-deployment path coding branch, a differential path coding module, a multi-state path memory module, and a path representation output module; At the input end, the features of each path step in the pre-deployment path sequence and the post-deployment path sequence are vectorized and encoded respectively. The path step features include node cache status, node remaining cache capacity, node load value, link latency, link bandwidth, link congestion degree and cumulative path latency, forming the pre-deployment path input sequence and the post-deployment path input sequence. Input the pre-deployment path input sequence into the pre-deployment path encoding branch, and input the post-deployment path input sequence into the post-deployment path encoding branch. Perform sequence encoding calculation step by step in the two path encoding branches according to the path step order to generate the pre-deployment path hidden state sequence and the post-deployment path hidden state sequence. In the multi-state path memory module, the following processing is performed for each path step: Extract the link latency, link bandwidth, and link congestion level corresponding to the current path step, and construct a local link state memory vector; Based on the type of the node corresponding to the current path step, the region migration state memory vector is constructed by encoding the category among user access node, regional center node, edge cache node and source station node. Calculate the number of graph hops from the node corresponding to the current path step to the virtual node of the content instance and the source node, and construct the content proximity state memory vector; The local link state memory vector, the region migration state memory vector, and the content proximity state memory vector are concatenated with the current path step hidden state to generate the path step joint encoding result. The following processing is performed in the differential path encoding module: Extract the terminal hidden states of the pre-deployment path hidden state sequence and the post-deployment path hidden state sequence; calculate the element-wise difference between the post-deployment path terminal hidden state and the pre-deployment path terminal hidden state to generate a difference vector; concatenate the pre-deployment path terminal hidden state, the post-deployment path terminal hidden state, and the difference vector in sequence; perform a linear transformation on the concatenation result to generate the differential path encoding result. The path step joint encoding result is used as the path step level input, and the differential path encoding result is used as the path level feature input to the path representation output module. The weight coefficients corresponding to each path step are calculated based on the path step joint encoding result in the path step dimension, and a weighted summation calculation is performed according to the weight coefficients to generate the path representation result of the corresponding candidate insertion graph.

7. The CDN access acceleration method based on graph neural networks according to claim 1, characterized in that, Step six specifically involves: Read the path representation results corresponding to each candidate insertion graph, and normalize the path representation results according to the path sequence dimension; The path delay estimate is calculated based on the normalized path representation results. The path delay estimate is calculated by linearly combining the path representation results with the mapping coefficients calculated based on the distribution of the path representation results. The access delay value corresponding to the candidate insertion graph is obtained by weighted summation of multiple path delay estimates corresponding to the same candidate insertion graph. The weight coefficients are determined based on the vector norm of the path representation result. The baseline access latency value is calculated by weighted summation of the path latency estimates for the corresponding path sequence when no content instance virtual nodes are inserted. The content replication cost is calculated by using the ratio of the content size to the remaining cache capacity of the target node, and weighted by combining the current load value of the node. Deployment evaluation metrics are constructed based on access latency, baseline access latency, content replication cost, and node load. The deployment evaluation metrics are obtained by calculating the difference between the baseline access latency and the access latency and subtracting the weighted result of the content replication cost and node load. The deployment evaluation metrics corresponding to each candidate edge node are compared, and the candidate edge node with the largest deployment evaluation metric value is selected as the target deployment node.

8. The CDN access acceleration method based on graph neural networks according to claim 1, characterized in that, Step seven specifically involves: Receive the target deployment node identifier, and read the current cache usage, total cache capacity, and list of cached content of the target deployment node; Determine the relationship between the remaining cache capacity of the target deployment node and the size of the target content. When the remaining cache capacity is less than the size of the target content, sort the list of cached content in ascending order of content popularity value, select the content with the lowest popularity value and delete it, and update the cache usage at the same time until the remaining cache capacity is not less than the size of the target content. Write the target content to the cache space of the target deployment node, and update the cache usage and the list of cached content; Update the routing table entries in the user access node to map the target content address to the target deployment node address, and forward the access request to the target deployment node according to the updated routing table; Read the content index table in the cache space and perform a matching operation based on the target content identifier; When a matching result exists, the corresponding content data is read from the cache and returned. When no matching result is found, a back-to-origin connection is established between the target deployment node and the origin node to retrieve content data, and the retrieved content data is written to the cache space. During the access request processing, the request initiation time and response return time are recorded. The access latency is calculated by performing a difference between the response return time and the request initiation time, and the number of back-to-origin requests is also recorded. The access latency and number of times the data returns to the origin are associated with the corresponding target deployment node identifier and stored, and then written into the access record data table.

Images