Content distribution network intelligent scheduling system based on machine learning and edge nodes

Abstract

The invention relates to the technical field of computer network communication and edge computing, in particular to a content distribution network intelligent scheduling system based on machine learning and edge nodes, which comprises an edge data acquisition module for acquiring and sending real-time performance data and content access logs of the edge nodes; the intelligent scheduling server is used for constructing a virtual space-time potential energy field data model representing a network topology state; performing evolution prediction on the future distribution state to generate a dynamic potential energy distribution map; resolving an accumulated potential energy minimum path and a potential energy low-lying area, and sending a routing scheduling and cache preheating instruction; the routing execution node responds to the instruction and prefetches specified content from the source station to the local; the user request is redirected to the target edge node; according to the method, intelligent prediction and accurate scheduling of the network traffic are realized by constructing the virtual potential energy field model.

Description

Technical Field

[0001] This invention relates to the fields of computer network communication and edge computing technology, specifically to an intelligent scheduling system for content delivery networks based on machine learning and edge nodes. Background Technology

[0002] Content Delivery Network (CDN) technology is a traditional means of improving internet access speed and stability. It typically uses the Domain Name System (DNS) to direct access requests to the nearest edge node based on the user's geographical location or network operator attributes. It can also remove faulty nodes through a simple health check mechanism, and is characterized by mature deployment and significant basic acceleration effect.

[0003] With the evolution of cloud computing and edge computing technologies, content delivery networks have gradually evolved from static resource distribution to dynamic intelligent routing scheduling. However, when faced with massive concurrent requests and complex network fluctuations, existing scheduling strategies mainly rely on instantaneous load states or preset static rules, lacking the ability to predict the spatiotemporal evolution of network topology and a multi-dimensional global optimization mechanism. This leads to problems such as routing oscillations, delayed scheduling responses, and uneven allocation of node resources during peak traffic periods, making it difficult to meet the service quality requirements of high real-time services. Summary of the Invention

[0004] The purpose of this invention is to provide an intelligent scheduling system for content delivery networks based on machine learning and edge nodes, which solves the following technical problems: it can construct a virtual spatiotemporal potential energy field data model representing the network topology state and perform evolution prediction, effectively suppress network routing oscillations and avoid getting trapped in local optima, achieve multi-objective dynamic balance scheduling of latency, load and bandwidth costs, and accurately perform content cache preheating based on low potential energy areas.

[0005] The objective of this invention can be achieved through the following technical solutions:

[0006] Includes: edge data acquisition module, intelligent scheduling server, and routing execution node;

[0007] The edge data acquisition module is used to collect real-time performance data and content access logs of edge nodes, and send the real-time performance data and content access logs to the intelligent scheduling server.

[0008] The intelligent scheduling server is used to construct a virtual spatiotemporal potential energy field data model representing the network topology state based on the real-time performance data and the content access log, and to predict the evolution of the distribution state of the virtual spatiotemporal potential energy field data model in future time steps to generate a dynamic potential energy distribution map.

[0009] The intelligent scheduling server is also used to calculate the route path with the minimum accumulated potential energy and the potential energy low-lying area with potential energy value below a preset threshold based on the dynamic potential energy distribution map, and send route scheduling instructions and cache preheating instructions to the route execution node accordingly.

[0010] The routing execution node is used to respond to the cache preheating instruction by pre-fetching specified content data from the origin server to the local storage space;

[0011] The routing execution node is also used to respond to the routing scheduling instruction by redirecting the received user request to the target edge node.

[0012] As a further aspect of the present invention: the intelligent scheduling server is also used for:

[0013] Extract content popularity features from the content access logs and map the content popularity features to gravitational potential energy parameters in the virtual spatiotemporal potential energy field;

[0014] Extract node load features from the real-time performance data and map the node load features to the repulsive potential energy parameters in the virtual spatiotemporal potential energy field;

[0015] Network latency features are extracted from the real-time performance data and mapped to the transmission damping coefficient in the virtual spatiotemporal potential field.

[0016] As a further aspect of the present invention: the intelligent scheduling server is also used for:

[0017] Based on the dynamic potential energy distribution map, potential energy depression areas where the potential energy value is lower than the preset potential energy threshold are identified.

[0018] The target edge node is determined from the potential energy depression area, and it is determined whether the target edge node has stored the specified hotspot content;

[0019] In response to determining that the target edge node does not store the specified hotspot content, the cache preheating instruction is generated;

[0020] In response to the determination that the target edge node has already stored the specified hotspot content, the cache preheating instruction is not generated and the current storage state is maintained.

[0021] As a further aspect of the present invention: the intelligent scheduling server is also used for:

[0022] Define a multi-objective energy function that includes delay, load, and bandwidth cost terms;

[0023] The gradient of the multi-objective energy function is calculated in the virtual spatiotemporal potential energy field data model to obtain the flow direction vector requested by the user in the field;

[0024] The flow vector is mapped to the identification information of the specific target edge node to generate the routing scheduling instruction.

[0025] As a further aspect of the present invention: the intelligent scheduling server is also used for:

[0026] Obtain the scheduling gradient vector for historical time steps;

[0027] The scheduling gradient vector of the historical time step and the currently calculated flow direction vector are weighted and smoothed using the momentum factor.

[0028] The routing scheduling instructions are generated based on the smoothed vectors to suppress routing oscillations.

[0029] As a further aspect of the present invention: the intelligent scheduling server is also used for:

[0030] Monitor the service type identifier of the current network environment;

[0031] In response to the service type being identified as latency-sensitive, the weighting coefficients in the multi-objective energy function are adjusted so that the weighting coefficient of the latency term is greater than the weighting coefficients of the load term and the bandwidth cost term.

[0032] In response to the service type being identified as cost-sensitive, the weighting coefficients in the multi-objective energy function are adjusted so that the weighting coefficient of the bandwidth cost item is greater than the weighting coefficients of the latency item and the load item.

[0033] As a further aspect of the present invention: the intelligent scheduling server includes: a graph neural network module and a multi-stream attention prediction module;

[0034] The graph neural network module is used to construct the spatial structure of the virtual spatiotemporal potential energy field data model based on the physical connection relationship of the edge nodes.

[0035] The multi-stream attention prediction module is used to process the time series data of the gravitational potential energy parameter, the repulsive potential energy parameter and the transmission damping coefficient respectively, and outputs the dynamic potential energy distribution map after feature fusion.

[0036] As a further aspect of the present invention: the routing execution node includes: a smart network interface card and a storage control unit;

[0037] The smart network interface card is used to parse the routing scheduling instruction and perform traffic redirection by modifying DNS resolution records or sending HTTP redirection messages.

[0038] The storage control unit is used to parse the cache warm-up instruction and control the solid-state storage medium to perform data writing operations.

[0039] The beneficial effects of this invention are:

[0040] 1) This invention constructs a virtual spatiotemporal potential energy field data model representing the network topology state, maps real-time performance data and content access logs into field parameters, and predicts the evolution of future distribution states. This mechanism breaks through the limitations of traditional scheduling that relies on instantaneous states or static rules, and can globally perceive the network potential energy distribution, identify traffic flow and hotspot areas in advance, thereby significantly improving the foresight and global optimization capabilities of scheduling.

[0041] 2) This invention calculates low-potential areas based on dynamic potential energy distribution maps, enabling precise location of future traffic convergence nodes and triggering cache preheating commands. By driving routing execution nodes to pre-fetch specified content to local storage, it achieves a shift in content distribution from passive response to proactive awareness, effectively reducing the latency of the first packet accessed by users, avoiding the cold start problem of nodes under sudden traffic surges, and optimizing the resource allocation balance of edge nodes.

[0042] 3) This invention introduces a multi-objective energy function that incorporates latency, load, and bandwidth costs, and determines request flow direction by calculating gradients. Specifically, when generating routing instructions, a momentum factor is used to weight and smooth historical scheduling gradients, simulating physical inertia to suppress data jitter. This design achieves multi-dimensional routing decision optimization and effectively prevents routing oscillations common during traffic peaks, ensuring system stability.

[0043] 4) This invention can monitor service type identifiers in real time and dynamically adjust the weighting coefficients of latency, load, and bandwidth cost terms in the multi-objective energy function accordingly. For latency-sensitive services, transmission damping is reduced first; for cost-sensitive services, bandwidth overhead is optimized first. This differentiated scheduling strategy not only improves the user experience of critical services but also effectively controls the overall network's operating bandwidth costs while meeting different service quality requirements.

[0044] 5) This invention employs a hardware execution architecture comprised of a smart network interface card (NIC) and a storage control unit, offloading routing redirection and cache write operations to dedicated hardware. The smart NIC directly parses instructions and modifies packets, while the storage unit controls direct writing to the control medium, significantly reducing the occupancy rate of the main general-purpose processor on the edge server. This hardware-software co-design ensures that scheduling execution maintains microsecond-level low latency and high response speed in scenarios with massive concurrent throughput. Attached Figure Description

[0045] The invention will now be further described with reference to the accompanying drawings.

[0046] Figure 1 This is a schematic diagram of the module structure of the intelligent scheduling system for content delivery networks based on machine learning and edge nodes in this application embodiment. Detailed Implementation

[0047] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0048] Please see Figure 1 As shown, the intelligent scheduling system for content delivery networks based on machine learning and edge nodes includes: an edge data acquisition module, an intelligent scheduling server, and a routing execution node; wherein, the edge data acquisition module is used to collect real-time performance data and content access logs of edge nodes, and send the real-time performance data and content access logs to the intelligent scheduling server;

[0049] The intelligent scheduling server is used to construct a virtual spatiotemporal potential energy field data model representing the network topology state based on real-time performance data and content access logs, and to predict the evolution of the distribution state of the virtual spatiotemporal potential energy field data model in future time steps to generate a dynamic potential energy distribution map. The intelligent scheduling server is also used to calculate the routing path with the minimum accumulated potential energy and the potential energy low-lying area with potential energy value below a preset threshold based on the dynamic potential energy distribution map, and send routing scheduling instructions and cache preheating instructions to the routing execution node accordingly.

[0050] The routing execution node is used to respond to cache preheating instructions by pre-fetching specified content data from the origin server to the local storage space; the routing execution node is also used to respond to routing scheduling instructions by redirecting received user requests to the target edge node.

[0051] In this embodiment, the edge data acquisition module is deployed in the Linux kernel state and captures network packets at the tc receive and send functions through eBPF hooks; the process of the intelligent scheduling server constructing the virtual spatiotemporal potential energy field data model includes two steps: coordinate mapping and field function fitting; the Vivaldi network coordinate algorithm is used to map the RTT delay of physical nodes into three-dimensional delay space coordinates. The Euclidean distance directly represents the communication delay in milliseconds (ms), and the coordinate update follows the spring relaxation equation:

[0052]

[0053] Convergence factor in the equation Set to 0.25. To prevent numerical calculation anomalies when the vector magnitude is zero, the server uses radial basis function (RBF) interpolation to fit the scalar potential energy values ​​of discrete nodes to a continuous field function. The calculation formula is as follows:

[0054]

[0055] The basis functions adopt the form Shape parameters Set as the average latency distance of all nodes in the network The reciprocal of; weight By solving the system of linear equations Obtain, where the coefficient matrix for A symmetric matrix, whose elements are defined as Represents the spatial correlation between nodes; vector The gravitational potential energy of each node With repulsive potential energy After maximum value normalization, the components are linearly superimposed to eliminate dimensional differences; in the evolution prediction stage, the server defines a spatial bounding box covering all network nodes, and its calculation formula is as follows:

[0056]

[0057] Among them, coordinate boundary margin Set to 50ms; bounding box Divided into The voxel grid is used to calculate the field strength at the center of each voxel using trilinear interpolation; and the field strength is mapped to a value using Min-Max normalization. Interval as input tensor Input the grid data into the ConvLSTM network, based on the network state transition equation:

[0058]

[0059] The kernel size is set to... The hidden layer has 64 channels, the stacked layers are 2, and the activation function is Tanh to adapt to the nonlinear changes in the potential field and output the future. Voxel potential energy prediction at time t.

[0060] In a preferred embodiment of the present invention, the intelligent scheduling server is further configured to: extract content popularity features from content access logs and map the content popularity features to gravitational potential energy parameters in a virtual spatiotemporal potential energy field; extract node load features from real-time performance data and map the node load features to repulsive potential energy parameters in a virtual spatiotemporal potential energy field; and extract network latency features from real-time performance data and map the network latency features to transmission damping coefficients in a virtual spatiotemporal potential energy field.

[0061] This embodiment illustrates the engineering implementation of mapping physical network features to virtual field parameters, ensuring that the field function is differentiable everywhere in three-dimensional space; and extracting content popularity features in units of req / s. It is defined as a sliding window. The request count is performed within the function; the gravitational potential energy function is constructed, and its calculation formula is as follows:

[0062]

[0063] in, Let the coordinates be any test point in the field. For storage nodes Vivaldi coordinates, For the set of nodes holding this content; Gravity gain The value was set to 100.0, which was determined through gradient sensitivity experiments to ensure that the gravitational term accounts for approximately 40% of the total potential energy. The normalized reference heat value is set to the highest historical heat record across the entire network (e.g., 10000req / s) to eliminate the impact of absolute value fluctuations on the potential energy field. Used for numerical stability; extraction of node load characteristics The calculation formula is:

[0064]

[0065] in, , , These represent the node's CPU utilization, memory usage, and disk I / O wait rate, respectively, all of which have been normalized to [value missing]. Interval; weight vector Based on historical root cause analysis statistics, this reflects the contribution rate of resource depletion to service interruption.

[0066] Weight vector Based on historical root cause analysis statistics, the contribution rate of each resource depletion to service interruption is reflected; the repulsive potential energy function is constructed as follows:

[0067]

[0068] Wherein, the tangent function Used to construct a hard threshold potential barrier when the load approaches a critical value. At that time, the potential energy will increase gradually and rapidly, thus physically simulating an insurmountable repulsive force field; through The function limits the load ratio to an upper limit of 0.99 to prevent the tangent function from generating infinite values ​​that could cause computational overflow.

[0069] Repulsion coefficient Set as Critical load Set to 85% of the node's theoretical throughput limit, obtained through stress testing using the stress-ng tool; Gaussian decay term. Setting it to 15.0 corresponds to an equivalent distance of 20msRTT in the physical network, limiting the repulsive force to act only on the local neighborhood;

[0070] Extracting network latency features in milliseconds The Kriging interpolation method is used to map the RTT data of discrete nodes into a transmission damping coefficient field in continuous space. The calculation formula is as follows:

[0071]

[0072] Among them, the normalized damping factor The value is 1.0, used to balance the order-of-magnitude difference between the delay term and the potential energy term; This represents the current average RTT value for all nodes in the network, with the statistical window being the most recent 5 minutes, used to analyze the interpolated latency field. Dimensionless processing is performed; all the above potential energy parameters are normalized to dimensionless field strengths before superposition.

[0073] In a preferred embodiment of the present invention, the intelligent scheduling server is further configured to: identify potential energy depression areas where the potential energy value is lower than a preset potential energy threshold based on a dynamic potential energy distribution map; determine target edge nodes from the potential energy depression areas and determine whether the target edge nodes have stored specified hot content; generate a cache preheating instruction in response to determining that the target edge nodes have not stored the specified hot content; and not generate a cache preheating instruction and maintain the current storage state in response to determining that the target edge nodes have stored the specified hot content.

[0074] This embodiment details the node localization and preheating path calculation based on the potential energy field gradient; during the identification stage of low potential energy areas, the system statistically analyzes the virtual field. The potential energy values ​​of all voxels are taken as the threshold at the 15th percentile. Extracting the required information Connected regions And calculate its centroid. During the virtual-physical mapping phase, the system searches for distance in the 3D KD-Tree. Recent Calculate the overall score for each physical node. The calculation formula is as follows:

[0075]

[0076] in, The average latency distance across all network nodes is used to perform dimensionless processing on the distance term, making it comparable to the normalized load term. Within the same order of magnitude; select The smallest value is the target node; during the deadline calculation phase, the system uses the source station coordinates in the frozen potential energy field snapshot. The gradient descent simulation is performed starting from this point, and its calculation formula is as follows:

[0077]

[0078] in, The negative gradient direction of the continuous potential field obtained by the aforementioned RBF fitting is given by the step size. The resolution is set to 0.5 times the grid resolution to ensure iterative convergence, generating a discrete pathpoint sequence. The estimated transmission time is calculated using the following formula:

[0079]

[0080] in, Updated using an exponentially weighted moving average:

[0081]

[0082] Smoothing factor ; The instantaneous link bandwidth value measured by sending probe packets at the current time step; set the deadline. Boundary margin The value is set to three times the standard deviation of the link RTT, which is used to cover network jitter risks and ensure that content is ready before traffic peaks arrive.

[0083] In a preferred embodiment of the present invention, the intelligent scheduling server is further configured to: define a multi-objective energy function including a delay term, a load term, and a bandwidth cost term; calculate the gradient of the multi-objective energy function in a virtual spatiotemporal potential energy field data model to obtain the flow direction vector of the user request in the field; map the flow direction vector to the identification information of specific target edge nodes, and generate routing scheduling instructions.

[0084] In this embodiment, the multi-objective energy function Construct a dimensionless scalar field: Weight coefficient vector Read from the SLA policy table in the configuration center and normalize it. Bandwidth cost item The Gaussian mixture model is defined as follows:

[0085]

[0086] in, For nodes The unit cost of data traffic (unit: yuan / GB), which is obtained by calling the operator's billing API; Set to 20.0, corresponding to the average service radius coverage of city-level edge nodes; calculated... All other terms are mapped to the maximum value through normalization. Interval obtained Dimensionless terms are used to eliminate unit differences such as yuan / GB and milliseconds; during the flow vector calculation phase, the server uses an automatic differential operator to calculate the current coordinates. The negative gradient at point A is calculated using the following formula:

[0087]

[0088] The node mapping logic is as follows: Traverse the set of physical neighbor nodes. The projection score is calculated using the following formula:

[0089]

[0090] If the candidate set is empty, i.e., it is trapped in a local potential trap, then the escape strategy is triggered: based on probability... Choose the one with the closest Euclidean distance and the load. Forced redirection to the node, or by probability Perform a random walk to avoid routing deadlock; the above probability values and The empirically set value can be dynamically adjusted in the configuration file according to the size and density of network nodes to balance convergence speed and escape success rate; the final selected value... The highest node generates routing instructions;

[0091] In a preferred embodiment of the present invention, the intelligent scheduling server is further configured to: obtain the scheduling gradient vector of the historical time step; perform weighted smoothing processing on the scheduling gradient vector of the historical time step and the currently calculated flow direction vector using a momentum factor; and generate routing scheduling instructions based on the smoothed vector to suppress routing oscillations.

[0092] This embodiment describes a mechanism for suppressing routing oscillations using a momentum factor; to address the ping-pong effect that may result from pure gradient descent, the intelligent scheduling server introduces a momentum gradient update algorithm; the calculation formula is:

[0093]

[0094] in, Characterize the current time step Used to generate a smoothed vector of instructions; For the previous time step Historical scheduling gradient vector; This is the currently calculated original flow direction vector; The momentum factor is derived from the system configuration, and its value range is... The momentum factor is used to characterize the inertia of the system. In this embodiment, the momentum factor is introduced to simulate the mass inertia of the physical entity. Even if the repulsive field of a node fluctuates instantaneously at the current moment, the historical momentum will maintain the original flow direction, thereby smoothing the routing decision and effectively avoiding routing oscillations caused by data jitter.

[0095] In a preferred embodiment of the present invention, the intelligent scheduling server is further configured to: monitor the service type identifier of the current network environment; and, in response to the service type identifier being latency-sensitive, adjust the weight coefficients in the multi-objective energy function so that the weight coefficient of the latency term is greater than the weight coefficients of the load term and the bandwidth cost term.

[0096] In response to the service type being identified as cost-sensitive, the weighting coefficients in the multi-objective energy function are adjusted so that the weighting coefficient of the bandwidth cost item is greater than that of the delay item and the load item.

[0097] This embodiment illustrates a dynamic adjustment strategy for weight coefficients based on business type; the intelligent scheduling server maintains a pre-set weight vector lookup table, where the weight vector is defined as:

[0098]

[0099] These correspond to the delay term, load term, bandwidth cost term, and gravitational potential energy term, respectively, and satisfy the following conditions: Real-time parsing of QoS flags or URL suffixes in business message headers; In response to messages identified as latency-sensitive, such as cloud gaming, a table lookup is used to select the weight vector:

[0100]

[0101] This makes the energy function dominated by the transmission damping term, forcing the route to flow to the node with the closest physical distance;

[0102] In response to tasks identified as cost-sensitive, such as log archiving, a table lookup is used to select the weight vector. This makes the energy function dominated by the bandwidth cost term, forcing traffic to flow to low-rate edge nodes; the lookup table mechanism ensures that the parameters of the multi-objective function are reconstructed within a microsecond-level scheduling cycle.

[0103] In a preferred embodiment of the present invention, the intelligent scheduling server includes: a graph neural network module and a multi-stream attention prediction module;

[0104] The graph neural network module is used to construct the spatial structure of the virtual spatiotemporal potential energy field data model based on the physical connection relationship of the edge nodes.

[0105] The multi-stream attention prediction module is used to process the time series data of the gravitational potential energy parameter, the repulsive potential energy parameter and the transmission damping coefficient respectively, and outputs the dynamic potential energy distribution map after feature fusion.

[0106] This embodiment constructs an ST-GNN model to predict the evolution of the potential energy field; in the graph structure construction stage, the system calculates the Euclidean distance between nodes, if... Then set the adjacency matrix. The connection radius threshold The equivalent Euclidean distance is set to 40ms to ensure that only strongly correlated neighbor nodes are retained; during the feature input stage, node feature tensors are constructed. Among them, historical time steps Set to 12, feature dimensions Corresponding to attraction, repulsion, and damping;

[0107] Entering the multi-flow prediction phase, gravitational flow input Features are output after passing through a 4-HeadTemporalAttention layer. Repulsive flow input Features output after 3 layers of TCN Damped flow input Features output after 3 layers of TCN In the graph neural network fusion stage, the splicing features are calculated using the following formula:

[0108]

[0109] Inputting a two-layer GraphSAGE network, the computation is as follows:

[0110]

[0111] in, The weight transformation matrix is ​​learnable and updated during offline training by minimizing the mean squared error loss function; the next-time node potential vector is output via regression through a fully connected layer. Perform field reconstruction based on the current coordinates of the nodes. Constructing the radial basis matrix Its construction method is the same as that defined in the aforementioned coordinate update stage. The matrix remains consistent, and its elements are:

[0112]

[0113] Among them, the reconstruction smoothing parameter The value was set to 10.0 to ensure the local detail fitting ability of the field function; Cholesky decomposition was used to solve the problem. Obtain the weight vector Thus, the continuous field function is determined. .

[0114] In a preferred embodiment of the present invention, the routing execution node includes: a smart network interface card (NIC) and a storage control unit; wherein, the smart NIC is used to parse routing scheduling instructions and perform traffic redirection by modifying DNS resolution records or sending HTTP redirection messages; the storage control unit is used to parse cache warm-up instructions and control the solid-state storage medium to perform data writing operations.

[0115] This embodiment describes the hardware-level implementation of the routing execution node; the smart network card adopts a programmable network card based on FPGA or SoC architecture, and runs an eBPF program or P4 pipeline inside the network card, directly parsing the routing scheduling instructions from the server at the data link layer; when a target user request is matched, the network card directly modifies the MAC / IP header of the data packet, or injects RST / SYN-ACK packets during the TCP handshake phase to achieve zero-copy traffic redirection;

[0116] Meanwhile, the storage control unit refers to the main controller chip of the NVMeSSD or the RAID card connected to the storage. This unit receives cache warm-up instructions through the PCIe bus and uses direct memory access technology to directly move the prefetched data blocks from the network card buffer to the storage particles without going through the general CPU processing path.

[0117] This embodiment significantly reduces the main CPU utilization of the edge server by offloading routing and storage operations to dedicated hardware, ensuring that the scheduling execution latency remains constant at the microsecond level in high-concurrency throughput scenarios.

[0118] The foregoing has provided a detailed description of one embodiment of the present invention, but this description is merely a preferred embodiment and should not be construed as limiting the scope of the invention. All equivalent variations and modifications made within the scope of the claims of this invention should still fall within the patent coverage of this invention.

Claims

1. A content delivery network intelligent scheduling system based on machine learning and edge nodes, characterized in that, include: Edge data acquisition module, intelligent scheduling server, and routing execution node; The edge data acquisition module is used to collect real-time performance data and content access logs of edge nodes, and send the real-time performance data and content access logs to the intelligent scheduling server. The intelligent scheduling server is used to construct a virtual spatiotemporal potential energy field data model representing the network topology state based on the real-time performance data and the content access log, and to predict the evolution of the distribution state of the virtual spatiotemporal potential energy field data model in future time steps to generate a dynamic potential energy distribution map. The intelligent scheduling server is also used to calculate the route path with the minimum accumulated potential energy and the potential energy low-lying area with potential energy value below a preset threshold based on the dynamic potential energy distribution map, and send route scheduling instructions and cache preheating instructions to the route execution node accordingly. The routing execution node is used to respond to the cache preheating instruction by pre-fetching specified content data from the origin server to the local storage space; The routing execution node is also used to respond to the routing scheduling instruction by redirecting the received user request to the target edge node.

2. The intelligent scheduling system for content delivery networks based on machine learning and edge nodes according to claim 1, characterized in that, The intelligent scheduling server is also used for: Extract content popularity features from the content access logs and map the content popularity features to gravitational potential energy parameters in the virtual spatiotemporal potential energy field; Extract node load features from the real-time performance data and map the node load features to the repulsive potential energy parameters in the virtual spatiotemporal potential energy field; Network latency features are extracted from the real-time performance data and mapped to the transmission damping coefficient in the virtual spatiotemporal potential field.

3. The intelligent scheduling system for content delivery networks based on machine learning and edge nodes according to claim 1, characterized in that, The intelligent scheduling server is also used for: Based on the dynamic potential energy distribution map, potential energy depression areas where the potential energy value is lower than the preset potential energy threshold are identified. The target edge node is determined from the potential energy depression area, and it is determined whether the target edge node has stored the specified hotspot content; In response to determining that the target edge node does not store the specified hotspot content, the cache preheating instruction is generated; In response to the determination that the target edge node has already stored the specified hotspot content, the cache preheating instruction is not generated and the current storage state is maintained.

4. The intelligent scheduling system for content delivery networks based on machine learning and edge nodes according to claim 1, characterized in that, The intelligent scheduling server is also used for: Define a multi-objective energy function that includes delay, load, and bandwidth cost terms; The gradient of the multi-objective energy function is calculated in the virtual spatiotemporal potential energy field data model to obtain the flow direction vector requested by the user in the field; The flow vector is mapped to the identification information of the specific target edge node to generate the routing scheduling instruction.

5. The intelligent scheduling system for content delivery networks based on machine learning and edge nodes according to claim 4, characterized in that, The intelligent scheduling server is also used for: Obtain the scheduling gradient vector for historical time steps; The scheduling gradient vector of the historical time step and the currently calculated flow direction vector are weighted and smoothed using the momentum factor. The routing scheduling instructions are generated based on the smoothed vectors to suppress routing oscillations.

6. The intelligent scheduling system for content delivery networks based on machine learning and edge nodes according to claim 4, characterized in that, The intelligent scheduling server is also used for: Monitor the service type identifier of the current network environment; In response to the service type being identified as latency-sensitive, the weighting coefficients in the multi-objective energy function are adjusted so that the weighting coefficient of the latency term is greater than the weighting coefficients of the load term and the bandwidth cost term. In response to the service type being identified as cost-sensitive, the weighting coefficients in the multi-objective energy function are adjusted so that the weighting coefficient of the bandwidth cost item is greater than the weighting coefficients of the latency item and the load item.

7. The intelligent scheduling system for content delivery networks based on machine learning and edge nodes according to claim 2, characterized in that, The intelligent scheduling server includes: a graph neural network module and a multi-stream attention prediction module; The graph neural network module is used to construct the spatial structure of the virtual spatiotemporal potential energy field data model based on the physical connection relationship of the edge nodes. The multi-stream attention prediction module is used to process the time series data of the gravitational potential energy parameter, the repulsive potential energy parameter and the transmission damping coefficient respectively, and outputs the dynamic potential energy distribution map after feature fusion.

8. The intelligent scheduling system for content delivery networks based on machine learning and edge nodes according to any one of claims 1-7, characterized in that, The routing execution node includes: a smart network interface card and a storage control unit; The smart network interface card is used to parse the routing scheduling instruction and perform traffic redirection by modifying DNS resolution records or sending HTTP redirection messages. The storage control unit is used to parse the cache warm-up instruction and control the solid-state storage medium to perform data writing operations.

Images