A CDN edge node cooperative optimization method and system based on a lightweight AI model

Abstract

The application discloses a CDN edge node cooperative optimization method and system based on a lightweight AI model, and relates to the technical field of content distribution networks. In view of the problems of response lag and low resource utilization caused by the traditional CDN static strategy, the scheme comprises the following steps: collecting specified data, aligning the time sequence, and then performing real-time analysis by a lightweight TCN model to output a pre-cache priority list and a routing target node score table; aggregating the global state information reported by each node, combining the analysis results, and generating a global balanced strategy by a MARL framework through node demand similar matching and resource negotiation; based on the global balanced strategy, the following actions are performed in parallel: intelligent pre-caching based on the pre-cache priority list, dynamic request routing realized in combination with the routing target node score table and the QUIC protocol, and containerized elastic resource allocation completed through Kubernetes according to quantitative indicators; collecting the execution effect data of the foregoing actions, and optimizing the MARL framework and the lightweight TCN model. The application is used in a high-concurrency dynamic traffic scene.

Description

Technical Field

[0001] This invention relates to the field of content delivery network technology, specifically a method and system for collaborative optimization of CDN edge nodes based on a lightweight AI model. Background Technology

[0002] In existing CDN (Content Delivery Network) systems, static preset caching strategies are commonly used, with each edge node independently storing and distributing content, which has several significant technical drawbacks.

[0003] First, the system response is severely lagging. During periods of sudden traffic surges, the edge nodes fail to identify and cache trending content in real time, causing massive user requests to directly penetrate to the origin server. Actual test data shows that the average request latency in this scenario spikes to over 2.5 seconds, severely impacting user experience.

[0004] Secondly, the system resource utilization is low. The existing static caching strategy is severely out of sync with real-time user needs, and the lack of an effective coordination mechanism between nodes results in storage space utilization of less than 45% and bandwidth resource idle rate as high as 50%. At the same time, isolated decision-making by each node also causes content duplication caching, with a duplication rate exceeding 35%, further reducing the overall resource utilization efficiency.

[0005] Furthermore, the existing system has high maintenance costs. Faced with dynamically changing network traffic, the existing solution mainly relies on manual strategy tuning, and the average response time for sudden failures exceeds 10 minutes, indicating a serious lack of automated management capabilities.

[0006] The technical limitations of existing improvement solutions can be summarized into the following three categories:

[0007] (1) Rule-based scheduling systems rely on fixed thresholds (such as triggering traffic splitting when CPU utilization is >80%), lack the ability to perceive network conditions and content hotspots in real time, and cannot adapt to dynamically changing environments.

[0008] (2) Centralized AI decision-making solutions require data to be uploaded to the cloud for unified calculation, and the strategy generation latency is as high as 800ms, which cannot meet the millisecond-level response requirements of edge scenarios;

[0009] (3) In the isolated node optimization scheme, each node makes independent decisions and lacks a global perspective, resulting in redundant caching of content and waste of resources. Summary of the Invention

[0010] This invention addresses the issues of response delays (hot requests penetrating to the origin server) and low resource utilization (duplication due to lack of inter-node collaboration) caused by static caching strategies in traditional CDNs (Content Delivery Networks). It provides a CDN edge node collaborative optimization method and system based on a lightweight AI model, thereby effectively reducing content distribution latency, improving global cache hit rate, and enhancing the network's adaptability to dynamic load.

[0011] Firstly, this invention provides a CDN edge node collaborative optimization method based on a lightweight AI model. The technical solution adopted to solve the above-mentioned technical problems is as follows:

[0012] A collaborative optimization method for CDN edge nodes based on a lightweight AI model includes the following steps:

[0013] S1. Collect HTTP / QUIC traffic characteristics, user request logs and node resource monitoring data, and transmit them to the edge node after time-series alignment. The lightweight TCN model deployed on the edge node analyzes them in real time and outputs a pre-caching priority list and a route target node score table. When the preset conditions are met, the pre-caching instruction is automatically triggered, and the pre-caching priority list and route target node score table are synchronized to the next process.

[0014] S2. Aggregate the global status information reported by each node, combine it with the pre-caching and routing analysis results output in step S1, and generate a global balancing strategy with cache hit rate as the core and low latency and load balancing by the multi-agent reinforcement learning (MARL) framework through node demand similarity matching and resource negotiation. This strategy plans three major execution directions: cross-node content pre-caching, dynamic request diversion, and elastic resource allocation.

[0015] S3. Based on the global balancing strategy generated in step S2, three core actions are executed in parallel: intelligent pre-caching is performed based on the pre-caching priority list, dynamic request routing is achieved by combining the routing target node scoring table with the QUIC protocol, and containerized elastic resource allocation is completed through Kubernetes according to quantitative indicators, so as to achieve bandwidth optimization, low-latency migration and resource elastic adaptation in a coordinated manner.

[0016] S4. Collect execution effect data of the three core actions (including indicators such as cache hit rate, transmission latency, and node load), feed it back to the multi-agent reinforcement learning framework and lightweight TCN model, update the model parameters to optimize subsequent analysis and decision-making logic, and form a closed-loop optimization link of "data collection - edge analysis - global decision-making - execution feedback - model iteration".

[0017] Optionally, step S1 specifically includes the following operations:

[0018] S1.1. Deploy NetFlow traffic sensors with a sampling frequency of 100ms on edge nodes to capture HTTP / QUIC traffic characteristics in real time and collect user request logs and node resource monitoring data simultaneously. Among them, user request logs include content type distribution and device geographical location information, and node resource monitoring data uses cgroups technology to statistically analyze CPU, memory and NIC metrics.

[0019] S1.2 and S1.1 The multi-source data collected in step S1.1 are time-aligned by a sliding window processor to ensure consistency in the time dimension of the data, and then transmitted to the edge node;

[0020] S1.3 Deploy a lightweight 8-layer TCN model optimized by channel pruning and INT8 quantization at the edge nodes; the TCN model performs real-time analysis on the output data of step S1.2, and then outputs a pre-cached priority list and a route target node score table that meet preset conditions; the preset conditions for the pre-cached priority list are: TOP10 sorted by content popularity, with a confidence level ≥90%; the preset conditions for the route target node score table are: neighboring nodes within 3 hops with a load score <0.6 and an RTT <50ms.

[0021] S1.4 When the TCN model detects that the request growth rate ΔQ / Δt > 200% and the content popularity H(c) > 0.7, the pre-caching instruction generation process is automatically started.

[0022] Optionally, step S2 specifically includes the following operations:

[0023] S2.1 Aggregates the global status information reported by each node in the CDN network, specifically covering: a 1024-node adjacency list (in the form of a topology matrix stored in binary compression), the distribution of the top 1000 content requests (presented as popular data compressed by Bloom Filter), and a normalized load vector (a CPU / bandwidth value with float16 precision).

[0024] S2.2 Based on the aggregated global state information and the pre-cached priority list and route target node scoring table output by step S1, the Locality-Sensitive Hashing algorithm is adopted, with a Jaccard coefficient greater than 0.7 as the matching threshold, to quickly filter out nodes with similar needs, and a node demand association model is constructed based on the improved collaborative filtering mechanism to clarify the demand association relationship between nodes.

[0025] S2.3. Based on the node demand association model and the real-time load of nodes, initiate the resource negotiation process: Nodes whose load exceeds the preset threshold publish task tenders (including key information such as the ID of the content to be distributed, the request volume, and the maximum latency constraint) based on the Contract Net protocol. Idle nodes respond with a list of available resources (clearly specifying resource status parameters such as bandwidth margin and real-time load rate), thus completing the resource supply and demand matching between nodes.

[0026] S2.4 Based on the supply and demand matching results of the resource negotiation phase, and combined with the pre-caching priority list and routing target node scoring table output in step S1, a global balancing strategy is generated by integrating three core indicators through the Multi-Agent Reinforcement Learning (MARL) framework. The strategy focuses on improving cache hit rate, reducing latency, and load balancing as core objectives, and plans three execution directions: cross-node content pre-caching, dynamic request routing, and elastic resource allocation. The specific implementation process is as follows:

[0027] S2.4.1 Based on the supply and demand matching results of the resource negotiation phase, combined with the pre-cached priority list and route target node scoring table output in step S1, as well as the three core indicators of cache hit rate, P99 request latency and node load standard deviation collected in real time by CDN network nodes, as the environment status input of the MARL framework.

[0028] S2.4.2. Using each edge node as an independent intelligent agent, generate three types of candidate actions based on local state and global indicators: cross-node content pre-caching, dynamic request routing, and elastic resource allocation.

[0029] S2.4.3 The central server aggregates all agent actions and calculates the global reward value according to a preset formula:

[0030] R t =α×H-β×max(0,L) P99 -L target +δ×(1-σ)),

[0031] Where: α is the cache hit rate weight, the core objective of which is to maximize content reuse; H is the content reuse degree, used to measure the degree of overlap in content sharing between nodes; β is the latency constraint weight coefficient, used to constrain P99 latency from exceeding the target threshold; L P99 For the node's P99 latency; L target δ is the preset P99 delay target threshold; δ is the load balancing weight coefficient; σ is the node load standard deviation, used to measure the uniformity of load distribution across all nodes.

[0032] S2.4.4 Update the agent's policy network parameters through the gradient ascent algorithm, strengthen high-reward actions, suppress negative feedback actions, and complete policy evaluation and iterative optimization;

[0033] S2.4.5 Generate a global load balancing strategy, clearly defining the core objectives as improving cache hit rate, reducing latency, and load balancing, while determining the specific execution direction corresponding to the three types of candidate actions.

[0034] Optionally, step S3 specifically includes the following operations:

[0035] (a) The pre-caching priority list clearly marks the popularity ranking and fragment type of each content fragment (with special emphasis on H.264 encoded I-frame fragments); when performing intelligent pre-caching based on the pre-caching priority list, the top 10% of core fragments in popularity (i.e. H.264 I-frames, which serve as key benchmark fragments for video content decoding) are pushed to the edge node cache first; at the same time, the cache hit rate and bandwidth usage data of the edge nodes are synchronized in real time. If the cache hit rate of the top 10% of fragments of a certain type of content is lower than the threshold, the fragment push weight of that type of content is dynamically adjusted to ensure that the pre-caching effect of core fragments accurately matches the bandwidth optimization target and stably achieves a 40% reduction in bandwidth usage;

[0036] (b) When a user initiates a request or a network switch occurs, the optimal target node is selected based on the routing target node scoring table. At the same time, relying on the 0-RTT (Round-Trip Time) handshake feature of the QUIC protocol, the request migration can be completed without re-establishing a TCP connection. During the migration process, the latency data of node switching is monitored in real time. By dynamically adjusting the node scoring weight (such as increasing the latency index weight to 30%), the node switching latency is strictly controlled within 15ms, achieving low-latency adaptation of request routing.

[0037] (c) Using QPS (queries per second) change rate as the core quantitative indicator, a Kubernetes containerized resource scheduling system is built; QPS data of the Kubernetes container cluster is collected in real time, and the number of Pod replicas is dynamically adjusted through a preset elastic scaling formula ΔPod = k·ΔQPS (where the proportional coefficient k = 0.05, that is, one Pod replica is added for every 20 QPS increase); at the same time, a scaling threshold triggering mechanism is set (such as scaling up when QPS is above 80% of the threshold for 30 consecutive seconds, and scaling down when QPS is below 40% of the threshold) to avoid frequent scaling up and down caused by resource fluctuations; through this quantitative scheduling logic, it is ensured that resource allocation is accurately matched with the business request volume, and the elastic adaptation and efficient utilization of resources are achieved.

[0038] Secondly, this invention provides a CDN edge node collaborative optimization system based on a lightweight AI model, and the technical solution adopted to solve the above-mentioned technical problems is as follows:

[0039] A CDN edge node collaborative optimization system based on a lightweight AI model, comprising:

[0040] The edge awareness layer is used to collect HTTP / QUIC traffic characteristics, user request logs and node resource monitoring data. After time-series alignment, the data is transmitted to the edge nodes, where a lightweight TCN model deployed on the edge nodes analyzes it in real time and outputs a pre-caching priority list and a route target node score table. When preset conditions are met, the pre-caching command is automatically triggered, and the pre-caching priority list and route target node score table are synchronized to the collaborative optimization layer.

[0041] The collaborative optimization layer is used to aggregate the compressed global state information reported by each node. Combined with the pre-caching and routing analysis results output by the edge perception layer, the multi-agent reinforcement learning (MARL) framework generates a global balancing strategy with cache hit rate as the core and low latency and load balancing as the considerations. This strategy plans three major execution directions: cross-node content pre-caching, dynamic request diversion, and elastic resource allocation.

[0042] The dynamic execution layer is used to execute three core actions in parallel based on a global balancing strategy: intelligent pre-caching based on a pre-caching priority list, dynamic request routing by combining a routing target node scoring table with the QUIC protocol, and containerized elastic resource allocation through Kubernetes according to quantitative indicators, thereby achieving bandwidth optimization, low-latency migration and elastic resource adaptation.

[0043] The closed-loop feedback layer is used to collect execution effect data of the three core actions (including indicators such as cache hit rate, transmission latency, and node load), and feeds them back to the multi-agent reinforcement learning framework and the lightweight TCN model to update the model parameters to optimize subsequent analysis and decision-making logic, forming a closed-loop optimization link of "data collection - edge analysis - global decision-making - execution feedback - model iteration".

[0044] Optionally, the edge-aware layer involved specifically includes:

[0045] The data acquisition module is used to capture HTTP / QUIC traffic characteristics in real time through NetFlow traffic sensors with a sampling frequency of 100ms deployed on edge nodes, and to collect user request logs and node resource monitoring data simultaneously. The user request logs include content type distribution and device geographic location information, and the node resource monitoring data uses cgroups technology to statistically analyze CPU, memory and NIC metrics.

[0046] The timing alignment module is used to perform timing alignment on multi-source data collected by the data acquisition module through a sliding window processor to ensure the consistency of the data in the time dimension, and then transmit it to the edge node;

[0047] The model deployment module is used to deploy a lightweight 8-layer TCN model optimized by channel pruning and INT8 quantization at edge nodes. This TCN model performs real-time analysis on the output data of the time-series alignment module, and then outputs a pre-cached priority list and a route target node score table that meet preset conditions. The preset conditions for the pre-cached priority list are: TOP10 ranked by content popularity, with a confidence level ≥90%; the preset conditions for the route target node score table are: neighboring nodes within 3 hops with a load score <0.6 and an RTT <50ms.

[0048] The pre-caching trigger module is used to automatically start the pre-caching instruction generation process when the TCN model detects that the request growth rate ΔQ / Δt > 200% and the content popularity H(c) > 0.7.

[0049] Optionally, the collaborative optimization layers involved specifically include:

[0050] The data aggregation module is used to aggregate the global status information reported by each node in the CDN network, specifically covering: a 1024-node adjacency list (in the form of a topology matrix stored in binary compression), the distribution of the top 1000 content requests (presented as popular data compressed by Bloom Filter), and a normalized load vector (a CPU / bandwidth value with float16 precision).

[0051] The demand collaborative matching module is used to quickly filter out nodes with similar demands based on aggregated global state information and the pre-cached priority list and route target node scoring table output by the edge perception layer. It adopts the Locality-Sensitive Hashing algorithm with a Jaccard coefficient greater than 0.7 as the matching threshold, and builds a node demand association model based on the improved collaborative filtering mechanism to clarify the demand association relationship between nodes.

[0052] The resource negotiation module is used to initiate the resource negotiation process based on the node demand association model and the real-time load of the nodes: nodes whose load exceeds the preset threshold publish task tenders (including key information such as the ID of the content to be distributed, the request volume, and the maximum latency constraint) based on the Contract Net protocol, and idle nodes respond with a list of available resources (clearly specifying resource status parameters such as bandwidth margin and real-time load rate), thus completing the resource supply and demand matching between nodes.

[0053] The strategy generation module, based on the supply and demand matching results from the resource negotiation phase, combines the pre-cached priority list output by the edge awareness layer and the routing target node scoring table. It then uses a multi-agent reinforcement learning (MARL) framework to integrate three core metrics to generate a global balancing strategy. The strategy focuses on improving cache hit rate, reducing latency, and achieving load balancing as its core objectives. It plans three main execution directions: cross-node content pre-caching, dynamic request routing, and elastic resource allocation. The specific implementation process is as follows:

[0054] (1) Based on the supply and demand matching results in the resource negotiation phase, combined with the pre-cached priority list and route target node scoring table output by the edge awareness layer, as well as the three core indicators of cache hit rate, P99 request latency and node load standard deviation collected in real time by CDN network nodes, the environment status input of the MARL framework is used.

[0055] (2) Each edge node is treated as an independent intelligent agent. Based on the local state and global indicators, three types of candidate actions are generated: cross-node content pre-caching, dynamic request diversion, and elastic resource allocation.

[0056] (3) The central server aggregates all agent actions and calculates the global reward value according to a preset formula:

[0057] R t =α×H-β×max(0,L) P99 -L target +δ×(1-σ)),

[0058] Where: α is the cache hit rate weight, the core objective of which is to maximize content reuse; H is the content reuse degree, used to measure the degree of overlap in content sharing between nodes; β is the latency constraint weight coefficient, used to constrain P99 latency from exceeding the target threshold; L P99 For the node's P99 latency; L target δ is the preset P99 delay target threshold; δ is the load balancing weight coefficient; σ is the node load standard deviation, used to measure the uniformity of load distribution across all nodes.

[0059] (4) Update the agent's policy network parameters through the gradient ascent algorithm, strengthen high-reward actions and suppress negative feedback actions, and complete policy evaluation and iterative optimization.

[0060] (5) Generate a global balancing strategy, clearly defining the core objectives as improving cache hit rate, reducing latency and load balancing, and determining the specific execution direction corresponding to the three types of candidate actions.

[0061] Optionally, the dynamic execution layer involved executes the three core actions in parallel through the following operations:

[0062] (a) The pre-caching priority list clearly marks the popularity ranking and fragment type of each content fragment (with special emphasis on H.264 encoded I-frame fragments);

[0063] When performing intelligent pre-caching based on the pre-caching priority list, the top 10% of core segments (i.e., H.264 I frames, which serve as key benchmark segments for video content decoding) are prioritized and pushed to the edge node cache. At the same time, the cache hit rate and bandwidth usage data of the edge nodes are synchronized in real time. If the cache hit rate of the top 10% of segments of a certain type of content is lower than the threshold, the push weight of the segments of that type of content is dynamically adjusted to ensure that the pre-caching effect of core segments accurately matches the bandwidth optimization target and stably achieves a 40% reduction in bandwidth usage.

[0064] (b) When a user initiates a request or a network switch occurs, the optimal target node is selected based on the routing target node scoring table. At the same time, relying on the 0-RTT (Round-Trip Time) handshake feature of the QUIC protocol, the request migration can be completed without re-establishing a TCP connection. During the migration process, the latency data of node switching is monitored in real time. By dynamically adjusting the node scoring weight (such as increasing the latency index weight to 30%), the node switching latency is strictly controlled within 15ms, achieving low-latency adaptation of request routing.

[0065] (c) Using QPS (queries per second) change rate as the core quantitative indicator, a Kubernetes containerized resource scheduling system is built; QPS data of the Kubernetes container cluster is collected in real time, and the number of Pod replicas is dynamically adjusted through a preset elastic scaling formula ΔPod = k·ΔQPS (where the proportional coefficient k = 0.05, that is, one Pod replica is added for every 20 QPS increase); at the same time, a scaling threshold triggering mechanism is set (such as scaling up when QPS is above 80% of the threshold for 30 consecutive seconds, and scaling down when QPS is below 40% of the threshold) to avoid frequent scaling up and down caused by resource fluctuations; through this quantitative scheduling logic, it is ensured that resource allocation is accurately matched with the business request volume, and the elastic adaptation and efficient utilization of resources are achieved.

[0066] The CDN edge node collaborative optimization method and system based on a lightweight AI model of the present invention have the following advantages compared with the prior art:

[0067] 1. This invention covers time-series prediction, collaborative filtering, multi-agent reinforcement learning and dynamic resource scheduling technologies. It is suitable for high-concurrency dynamic traffic scenarios, such as large-scale live broadcasts and breaking news events. It can solve the problems of response delay and low resource utilization caused by traditional CDN static strategies, significantly reduce distribution latency, improve cache hit rate and enhance network adaptability.

[0068] 2. This invention can reduce the average latency of content distribution in real-world testing scenarios with tens of millions of concurrent users, and significantly reduce video stuttering rates in 4K live streaming scenarios. By improving edge cache hit rate, reducing bandwidth costs, reducing back-to-origin traffic, and improving hardware resource utilization, it can increase the workload that an equivalent server can handle. Attached Figure Description

[0069] Appendix Figure 1 This is a flowchart of the method according to Embodiment 1 of the present invention;

[0070] Appendix Figure 2 This is a flowchart of step S1 of Embodiment 1 of the present invention;

[0071] Appendix Figure 3 This is a flowchart of step S2 of Embodiment 1 of the present invention;

[0072] Appendix Figure 4 This is a module connection block diagram of Embodiment 2 of the present invention. Detailed Implementation

[0073] To make the technical solution, the technical problem solved, and the technical effect of the present invention clearer, the technical solution of the present invention will be clearly and completely described below in conjunction with specific embodiments.

[0074] Example 1:

[0075] Combined with appendix Figure 1 This embodiment proposes a collaborative optimization method for CDN edge nodes based on a lightweight AI model, which includes the following steps:

[0076] S1. Collect HTTP / QUIC traffic characteristics, user request logs, and node resource monitoring data. After time-series alignment, transmit the data to the edge node. The lightweight TCN model deployed on the edge node analyzes the data in real time and outputs a pre-caching priority list and a routing target node score table. When preset conditions are met, the pre-caching command is automatically triggered, and the pre-caching priority list and the routing target node score table are synchronized to the next process.

[0077] Combined with appendix Figure 2 Step S1 in this embodiment specifically includes the following operations:

[0078] S1.1. Deploy NetFlow traffic sensors with a sampling frequency of 100ms on edge nodes to capture HTTP / QUIC traffic characteristics in real time and collect user request logs and node resource monitoring data simultaneously. Among them, user request logs include content type distribution and device geographical location information, and node resource monitoring data uses cgroups technology to statistically analyze CPU, memory and NIC metrics.

[0079] S1.2 and S1.1 The multi-source data collected in step S1.1 are time-aligned by a sliding window processor to ensure consistency in the time dimension of the data, and then transmitted to the edge node;

[0080] S1.3 Deploy a lightweight 8-layer TCN model optimized by channel pruning and INT8 quantization at the edge nodes; the TCN model performs real-time analysis on the output data of step S1.2, and then outputs a pre-cached priority list and a route target node score table that meet preset conditions; the preset conditions for the pre-cached priority list are: TOP10 sorted by content popularity, with a confidence level ≥90%; the preset conditions for the route target node score table are: neighboring nodes within 3 hops with a load score <0.6 and an RTT <50ms.

[0081] S1.4 When the TCN model detects that the request growth rate ΔQ / Δt > 200% and the content popularity H(c) > 0.7, the pre-caching instruction generation process is automatically started.

[0082] S2. Aggregate the global status information reported by each node, and combine it with the pre-caching and routing analysis results output in step S1. Through node demand similarity matching and resource negotiation, the multi-agent reinforcement learning (MARL) framework generates a global balancing strategy with cache hit rate as the core and low latency and load balancing as the main consideration. This strategy plans three major execution directions: cross-node content pre-caching, dynamic request diversion, and elastic resource allocation.

[0083] Combined with appendix Figure 3 Step S2 in this embodiment specifically includes the following operations:

[0084] S2.1 Aggregate global status information reported by each node in the CDN (Content Delivery Network), specifically including: a 1024-node adjacency list (in the form of a topology matrix stored in binary compression), the distribution of the top 1000 content requests (presented as popular data compressed by Bloom Filter), and a normalized load vector (a CPU / bandwidth value with float16 precision).

[0085] S2.2 Based on the aggregated global state information and the pre-cached priority list and route target node scoring table output by step S1, the Locality-Sensitive Hashing algorithm is adopted, with a Jaccard coefficient greater than 0.7 as the matching threshold, to quickly filter out nodes with similar needs, and a node demand association model is constructed based on the improved collaborative filtering mechanism to clarify the demand association relationship between nodes.

[0086] S2.3. Based on the node demand association model and the real-time load of nodes, initiate the resource negotiation process: Nodes whose load exceeds the preset threshold publish task tenders (including key information such as the ID of the content to be distributed, the request volume, and the maximum latency constraint) based on the Contract Net protocol. Idle nodes respond with a list of available resources (clearly specifying resource status parameters such as bandwidth margin and real-time load rate), thus completing the resource supply and demand matching between nodes.

[0087] S2.4 Based on the supply and demand matching results of the resource negotiation phase, and combined with the pre-caching priority list and routing target node scoring table output in step S1, a global balancing strategy is generated by integrating three core indicators through the Multi-Agent Reinforcement Learning (MARL) framework. The strategy focuses on improving cache hit rate, reducing latency, and load balancing as core objectives, and plans three execution directions: cross-node content pre-caching, dynamic request routing, and elastic resource allocation. The specific implementation process is as follows:

[0088] S2.4.1 Based on the supply and demand matching results of the resource negotiation phase, combined with the pre-cached priority list and route target node scoring table output in step S1, as well as the three core indicators of cache hit rate, P99 request latency and node load standard deviation collected in real time by CDN network nodes, as the environment status input of the MARL framework.

[0089] S2.4.2. Using each edge node as an independent intelligent agent, generate three types of candidate actions based on local state and global indicators: cross-node content pre-caching, dynamic request routing, and elastic resource allocation.

[0090] S2.4.3 The central server aggregates all agent actions and calculates the global reward value according to a preset formula:

[0091] R t =α×H-β×max(0,L) P99 -L target +δ×(1-σ)),

[0092] Where: α is the cache hit rate weight, the core objective of which is to maximize content reuse; H is the content reuse degree, used to measure the degree of overlap in content sharing between nodes; β is the latency constraint weight coefficient, used to constrain P99 latency from exceeding the target threshold; L P99 For the node's P99 latency; L target α is the preset P99 delay target threshold; δ is the load balancing weight coefficient; σ is the node load standard deviation, used to measure the uniformity of load distribution across all nodes; specifically set α = 0.6, β = 0.3, δ = 0.1.

[0093] S2.4.4 Update the agent's policy network parameters through gradient ascent algorithms (such as Actor-Critic), strengthen high-reward actions (such as pre-caching to accurately hit hot content), suppress negative feedback actions (such as latency deterioration after traffic splitting), and complete policy evaluation and iterative optimization.

[0094] S2.4.5 Generate a global load balancing strategy, clearly defining the core objectives as improving cache hit rate, reducing latency, and load balancing, while determining the specific execution direction corresponding to the three types of candidate actions.

[0095] S3. Based on the global balancing strategy generated in step S2, three core actions are executed in parallel: intelligent pre-caching is performed based on the pre-caching priority list; dynamic request routing is achieved by combining the routing target node scoring table with the QUIC protocol; and containerized elastic resource allocation is completed through Kubernetes according to quantitative indicators, so as to achieve bandwidth optimization, low-latency migration and resource elastic adaptation in a coordinated manner.

[0096] Step S3 in this embodiment specifically includes the following operations:

[0097] (a) The pre-caching priority list clearly marks the popularity ranking and fragment type of each content fragment (with special emphasis on H.264 encoded I-frame fragments); when performing intelligent pre-caching based on the pre-caching priority list, the top 10% of core fragments in popularity (i.e. H.264 I-frames, which serve as key benchmark fragments for video content decoding) are pushed to the edge node cache first; at the same time, the cache hit rate and bandwidth usage data of the edge nodes are synchronized in real time. If the cache hit rate of the top 10% of fragments of a certain type of content is lower than the threshold, the fragment push weight of that type of content is dynamically adjusted to ensure that the pre-caching effect of core fragments accurately matches the bandwidth optimization target and stably achieves a 40% reduction in bandwidth usage;

[0098] (b) When a user initiates a request or a network switch occurs, the optimal target node is selected based on the routing target node scoring table. At the same time, relying on the 0-RTT (Round-Trip Time) handshake feature of the QUIC protocol, the request migration can be completed without re-establishing a TCP connection. During the migration process, the latency data of node switching is monitored in real time. By dynamically adjusting the node scoring weight (such as increasing the latency index weight to 30%), the node switching latency is strictly controlled within 15ms, achieving low-latency adaptation of request routing.

[0099] (c) Using QPS (queries per second) change rate as the core quantitative indicator, a Kubernetes containerized resource scheduling system is built; QPS data of the Kubernetes container cluster is collected in real time, and the number of Pod replicas is dynamically adjusted through a preset elastic scaling formula ΔPod = k·ΔQPS (where the proportional coefficient k = 0.05, that is, one Pod replica is added for every 20 QPS increase); at the same time, a scaling threshold triggering mechanism is set (such as scaling up when QPS is above 80% of the threshold for 30 consecutive seconds, and scaling down when QPS is below 40% of the threshold) to avoid frequent scaling up and down caused by resource fluctuations; through this quantitative scheduling logic, it is ensured that resource allocation is accurately matched with the business request volume, and the elastic adaptation and efficient utilization of resources are achieved.

[0100] S4. At set intervals, such as every five minutes, collect execution effect data of the three core actions (including indicators such as cache hit rate, transmission latency, and node load), and feed it back to the multi-agent reinforcement learning framework and lightweight TCN model to update model parameters to optimize subsequent analysis and decision-making logic, forming a closed-loop optimization link of "data collection - edge analysis - global decision-making - execution feedback - model iteration".

[0101] Example 2:

[0102] Combined with appendix Figure 4 This embodiment proposes a CDN edge node collaborative optimization system based on a lightweight AI model, which includes:

[0103] The edge awareness layer is used to collect HTTP / QUIC traffic characteristics, user request logs and node resource monitoring data. After time-series alignment, the data is transmitted to the edge nodes, where a lightweight TCN model deployed on the edge nodes analyzes it in real time and outputs a pre-caching priority list and a route target node score table. When preset conditions are met, the pre-caching command is automatically triggered, and the pre-caching priority list and route target node score table are synchronized to the collaborative optimization layer.

[0104] The collaborative optimization layer is used to aggregate the compressed global state information reported by each node. Combined with the pre-caching and routing analysis results output by the edge perception layer, the multi-agent reinforcement learning (MARL) framework generates a global balancing strategy with cache hit rate as the core and low latency and load balancing as the considerations. This strategy plans three major execution directions: cross-node content pre-caching, dynamic request diversion, and elastic resource allocation.

[0105] The dynamic execution layer is used to execute three core actions in parallel based on a global balancing strategy: intelligent pre-caching based on a pre-caching priority list, dynamic request routing by combining a routing target node scoring table with the QUIC protocol, and containerized elastic resource allocation through Kubernetes according to quantitative indicators, thereby achieving bandwidth optimization, low-latency migration and elastic resource adaptation.

[0106] The closed-loop feedback layer is used to collect execution effect data of the three core actions (including indicators such as cache hit rate, transmission latency, and node load), and feeds them back to the multi-agent reinforcement learning framework and the lightweight TCN model to update the model parameters to optimize subsequent analysis and decision-making logic, forming a closed-loop optimization link of "data collection - edge analysis - global decision-making - execution feedback - model iteration".

[0107] In this embodiment, the edge-aware layer specifically includes:

[0108] The data acquisition module is used to capture HTTP / QUIC traffic characteristics in real time through NetFlow traffic sensors with a sampling frequency of 100ms deployed on edge nodes, and to collect user request logs and node resource monitoring data simultaneously. The user request logs include content type distribution and device geographic location information, and the node resource monitoring data uses cgroups technology to statistically analyze CPU, memory and NIC metrics.

[0109] The timing alignment module is used to perform timing alignment on multi-source data collected by the data acquisition module through a sliding window processor to ensure the consistency of the data in the time dimension, and then transmit it to the edge node;

[0110] The model deployment module is used to deploy a lightweight 8-layer TCN model optimized by channel pruning and INT8 quantization at edge nodes. This TCN model performs real-time analysis on the output data of the time-series alignment module, and then outputs a pre-cached priority list and a route target node score table that meet preset conditions. The preset conditions for the pre-cached priority list are: TOP10 ranked by content popularity, with a confidence level ≥90%; the preset conditions for the route target node score table are: neighboring nodes within 3 hops with a load score <0.6 and an RTT <50ms.

[0111] The pre-caching trigger module is used to automatically start the pre-caching instruction generation process when the TCN model detects that the request growth rate ΔQ / Δt > 200% and the content popularity H(c) > 0.7.

[0112] In this embodiment, the collaborative optimization layer specifically includes:

[0113] The data aggregation module is used to aggregate the global status information reported by each node in the CDN network, specifically covering: a 1024-node adjacency list (in the form of a topology matrix stored in binary compression), the distribution of the top 1000 content requests (presented as popular data compressed by Bloom Filter), and a normalized load vector (a CPU / bandwidth value with float16 precision).

[0114] The demand collaborative matching module is used to quickly filter out nodes with similar demands based on aggregated global state information and the pre-cached priority list and route target node scoring table output by the edge perception layer. It adopts the Locality-Sensitive Hashing algorithm with a Jaccard coefficient greater than 0.7 as the matching threshold, and builds a node demand association model based on the improved collaborative filtering mechanism to clarify the demand association relationship between nodes.

[0115] The resource negotiation module is used to initiate the resource negotiation process based on the node demand association model and the real-time load of the nodes: nodes whose load exceeds the preset threshold publish task tenders (including key information such as the ID of the content to be distributed, the request volume, and the maximum latency constraint) based on the Contract Net protocol, and idle nodes respond with a list of available resources (clearly specifying resource status parameters such as bandwidth margin and real-time load rate), thus completing the resource supply and demand matching between nodes.

[0116] The strategy generation module, based on the supply and demand matching results from the resource negotiation phase, combines the pre-cached priority list output by the edge awareness layer and the routing target node scoring table. It then uses a multi-agent reinforcement learning (MARL) framework to integrate three core metrics to generate a global balancing strategy. The strategy focuses on improving cache hit rate, reducing latency, and achieving load balancing as its core objectives. It plans three main execution directions: cross-node content pre-caching, dynamic request routing, and elastic resource allocation. The specific implementation process is as follows:

[0117] (1) Based on the supply and demand matching results in the resource negotiation phase, combined with the pre-cached priority list and route target node scoring table output by the edge awareness layer, as well as the three core indicators of cache hit rate, P99 request latency and node load standard deviation collected in real time by CDN network nodes, the environment status input of the MARL framework is used.

[0118] (2) Each edge node is treated as an independent intelligent agent. Based on the local state and global indicators, three types of candidate actions are generated: cross-node content pre-caching, dynamic request diversion, and elastic resource allocation.

[0119] (3) The central server aggregates all agent actions and calculates the global reward value according to a preset formula:

[0120] R t =α×H-β×max(0,L) P99-L target +δ×(1-σ)),

[0121] Where: α is the cache hit rate weight, the core objective of which is to maximize content reuse; H is the content reuse degree, used to measure the degree of overlap in content sharing between nodes; β is the latency constraint weight coefficient, used to constrain P99 latency from exceeding the target threshold; L P99 For the node's P99 latency; L target α is the preset P99 delay target threshold; δ is the load balancing weight coefficient; σ is the node load standard deviation, used to measure the uniformity of load distribution across all nodes; specifically set α = 0.6, β = 0.3, δ = 0.1.

[0122] (4) Update the agent policy network parameters through gradient ascent algorithm (such as Actor-Critic), strengthen high reward actions (such as pre-caching to accurately hit hot content), suppress negative feedback actions (such as latency deterioration after diversion), and complete policy evaluation and iterative optimization.

[0123] (5) Generate a global balancing strategy, clearly defining the core objectives as improving cache hit rate, reducing latency and load balancing, and determining the specific execution direction corresponding to the three types of candidate actions.

[0124] In this embodiment, the dynamic execution layer specifically executes the three core actions in parallel through the following operations:

[0125] (a) The pre-caching priority list clearly marks the popularity ranking and fragment type of each content fragment (with special emphasis on H.264 encoded I-frame fragments);

[0126] When performing intelligent pre-caching based on the pre-caching priority list, the top 10% of core segments (i.e., H.264 I frames, which serve as key benchmark segments for video content decoding) are prioritized and pushed to the edge node cache. At the same time, the cache hit rate and bandwidth usage data of the edge nodes are synchronized in real time. If the cache hit rate of the top 10% of segments of a certain type of content is lower than the threshold, the push weight of the segments of that type of content is dynamically adjusted to ensure that the pre-caching effect of core segments accurately matches the bandwidth optimization target and stably achieves a 40% reduction in bandwidth usage.

[0127] (b) When a user initiates a request or a network switch occurs, the optimal target node is selected based on the routing target node scoring table. At the same time, relying on the 0-RTT (Round-Trip Time) handshake feature of the QUIC protocol, the request migration can be completed without re-establishing a TCP connection. During the migration process, the latency data of node switching is monitored in real time. By dynamically adjusting the node scoring weight (such as increasing the latency index weight to 30%), the node switching latency is strictly controlled within 15ms, achieving low-latency adaptation of request routing.

[0128] (c) Using QPS (queries per second) change rate as the core quantitative indicator, a Kubernetes containerized resource scheduling system is built; QPS data of the Kubernetes container cluster is collected in real time, and the number of Pod replicas is dynamically adjusted through a preset elastic scaling formula ΔPod = k·ΔQPS (where the proportional coefficient k = 0.05, that is, one Pod replica is added for every 20 QPS increase); at the same time, a scaling threshold triggering mechanism is set (such as scaling up when QPS is above 80% of the threshold for 30 consecutive seconds, and scaling down when QPS is below 40% of the threshold) to avoid frequent scaling up and down caused by resource fluctuations; through this quantitative scheduling logic, it is ensured that resource allocation is accurately matched with the business request volume, and the elastic adaptation and efficient utilization of resources are achieved.

[0129] In summary, the CDN edge node collaborative optimization method and system based on a lightweight AI model of the present invention covers time-series prediction, collaborative filtering, multi-agent reinforcement learning and dynamic resource scheduling technologies. It is suitable for high-concurrency dynamic traffic scenarios, such as large-scale live broadcasts and breaking news events. It can solve the problems of response lag and low resource utilization caused by traditional CDN static strategies, significantly reduce distribution latency, improve cache hit rate and enhance network adaptability.

[0130] The above specific examples illustrate the principles and implementation methods of the present invention in detail. These embodiments are merely for the purpose of helping to understand the core technical content of the present invention. Based on the above specific embodiments of the present invention, any improvements and modifications made to the present invention by those skilled in the art without departing from the principles of the present invention should fall within the patent protection scope of the present invention.

Claims

1. A CDN edge node collaborative optimization method based on a lightweight AI model, characterized in that, Includes the following steps: S1. Collect HTTP / QUIC traffic characteristics, user request logs and node resource monitoring data, and transmit them to the edge node after time-series alignment. The lightweight TCN model deployed on the edge node analyzes them in real time and outputs a pre-caching priority list and a route target node score table. When the preset conditions are met, the pre-caching instruction is automatically triggered, and the pre-caching priority list and route target node score table are synchronized to the next process. S2. Aggregate the global status information reported by each node, combine it with the pre-caching and routing analysis results output in step S1, and generate a global balancing strategy with cache hit rate as the core and low latency and load balancing by a multi-agent reinforcement learning framework through node demand similarity matching and resource negotiation. This strategy plans three major execution directions: cross-node content pre-caching, dynamic request diversion, and elastic resource allocation. S3. Based on the global load balancing strategy generated in step S2, three core actions are executed in parallel: intelligent pre-caching based on the pre-caching priority list, dynamic request routing by combining the routing target node scoring table and the QUIC protocol, and containerized elastic resource allocation through Kubernetes according to quantitative indicators, collaboratively achieving bandwidth optimization, low-latency migration, and elastic resource adaptation; specifically as follows: (a) The popularity ranking and type of each content shard are clearly marked in the pre-caching priority list; when performing intelligent pre-caching based on the pre-caching priority list, the top 10% of core shards in terms of popularity are pushed to the edge node cache first. At the same time, the cache hit rate and bandwidth usage data of edge nodes are synchronized in real time. If the cache hit rate of the top 10% of a certain type of content is lower than the threshold, the push weight of that type of content will be dynamically adjusted. (b) When a user initiates a request or a network switch occurs, the optimal target node is selected based on the routing target node scoring table, and the request migration is completed by relying on the 0-RTT handshake feature of the QUIC protocol. During the migration process, the latency data of node switching is monitored in real time. By dynamically adjusting the node scoring weight, the node switching latency is strictly controlled within 15ms, achieving low-latency adaptation of request routing. (c) Using QPS change rate as the core quantitative indicator, build a Kubernetes containerized resource scheduling system; collect QPS data of Kubernetes container clusters in real time, and dynamically adjust the number of Pod replicas through the preset elastic scaling formula ΔPod=k·ΔQPS; at the same time, set up a scaling threshold trigger mechanism. S4. Collect the execution effect data of the three core actions, feed it back to the multi-agent reinforcement learning framework and the lightweight TCN model, update the model parameters to optimize the subsequent analysis and decision-making logic, and form a closed-loop optimization link of "data collection - edge analysis - global decision-making - execution feedback - model iteration".

2. The CDN edge node collaborative optimization method based on a lightweight AI model according to claim 1, characterized in that, Step S1 specifically includes the following operations: S1.

1. Deploy NetFlow traffic sensors with a sampling frequency of 100ms on edge nodes to capture HTTP / QUIC traffic characteristics in real time and collect user request logs and node resource monitoring data simultaneously. Among them, user request logs include content type distribution and device geographical location information, and node resource monitoring data uses cgroups technology to statistically analyze CPU, memory and NIC metrics. S1.2 and S1.1 The multi-source data collected in step S1.1 are time-aligned by a sliding window processor to ensure consistency in the time dimension of the data, and then transmitted to the edge node; S1.3 Deploy a lightweight 8-layer TCN model optimized by channel pruning and INT8 quantization at the edge nodes; the TCN model performs real-time analysis on the output data of step S1.2, and then outputs a pre-cached priority list and a route target node score table that meet the preset conditions; S1.4 When the TCN model detects that the request growth rate ΔQ / Δt > 200% and the content popularity H(c) > 0.7, the pre-caching instruction generation process is automatically started.

3. The CDN edge node collaborative optimization method based on a lightweight AI model according to claim 2, characterized in that, The default conditions for the pre-caching priority list are: TOP 10 content ranked by popularity, with a confidence level of ≥90%; The default conditions for the routing target node rating table are: neighboring nodes within 3 hops with a load rating of <0.6 and an RTT of <50ms.

4. The CDN edge node collaborative optimization method based on a lightweight AI model according to claim 1, characterized in that, Step S2 specifically includes the following operations: S2.1 Aggregate the global status information reported by each node in the CDN network, specifically covering: 1024 node adjacency list, Top 1000 content request distribution and normalized load vector; S2.2 Based on the aggregated global state information and the pre-cached priority list and route target node scoring table output by step S1, the Locality-Sensitive Hashing algorithm is adopted, with a Jaccard coefficient greater than 0.7 as the matching threshold, to quickly filter out nodes with similar needs, and a node demand association model is constructed based on the improved collaborative filtering mechanism to clarify the demand association relationship between nodes. S2.

3. Based on the node demand association model and real-time node load, initiate the resource negotiation process: Nodes with loads exceeding a preset threshold publish task tenders based on the Contract Net protocol, and idle nodes respond with a list of available resources, thus completing the resource supply and demand matching between nodes. S2.4 Based on the supply and demand matching results of the resource negotiation phase, and combined with the pre-caching priority list and routing target node scoring table output in step S1, a global balancing strategy is generated by integrating three core indicators through a multi-agent reinforcement learning framework. The core objectives are clearly defined as improving cache hit rate, reducing latency, and load balancing. Three execution directions are planned: cross-node content pre-caching, dynamic request routing, and elastic resource allocation. The specific implementation process is as follows: S2.4.1 Based on the supply and demand matching results of the resource negotiation phase, combined with the pre-cached priority list and route target node scoring table output in step S1, as well as the three core indicators of cache hit rate, P99 request latency and node load standard deviation collected in real time by CDN network nodes, as the environment status input of the MARL framework. S2.4.

2. Using each edge node as an independent intelligent agent, generate three types of candidate actions based on local state and global indicators: cross-node content pre-caching, dynamic request routing, and elastic resource allocation. S2.4.3 The central server aggregates all agent actions and calculates the global reward value according to a preset formula: ， in: The cache hit rate weight is used to maximize content reuse; H represents content reuse degree, which measures the degree of overlap in content sharing between nodes. These are the delay constraint weighting coefficients, used to constrain the P99 delay from exceeding the target threshold; For the P99 latency of the node; The preset P99 delay target threshold; σ is the load balancing weighting coefficient; σ is the standard deviation of node load, used to measure the uniformity of load distribution across all nodes. S2.4.4 Update the agent's policy network parameters through the gradient ascent algorithm, strengthen high-reward actions, suppress negative feedback actions, and complete policy evaluation and iterative optimization; S2.4.5 Generate a global load balancing strategy, clearly defining the core objectives as improving cache hit rate, reducing latency, and load balancing, while determining the specific execution direction corresponding to the three types of candidate actions.

5. A CDN edge node collaborative optimization system based on a lightweight AI model, characterized in that, It includes: The edge awareness layer is used to collect HTTP / QUIC traffic characteristics, user request logs and node resource monitoring data. After time-series alignment, the data is transmitted to the edge nodes, where a lightweight TCN model deployed on the edge nodes analyzes it in real time and outputs a pre-caching priority list and a route target node score table. When preset conditions are met, the pre-caching command is automatically triggered, and the pre-caching priority list and route target node score table are synchronized to the collaborative optimization layer. The collaborative optimization layer is used to aggregate the compressed global state information reported by each node. Combined with the pre-caching and routing analysis results output by the edge perception layer, the multi-agent reinforcement learning framework generates a global balancing strategy with cache hit rate as the core and low latency and load balancing through node demand similarity matching and resource negotiation. This strategy plans three major execution directions: cross-node content pre-caching, dynamic request diversion, and elastic resource allocation. The dynamic execution layer is used to execute three core actions in parallel based on a global balancing strategy: intelligent pre-caching based on a pre-caching priority list, dynamic request routing by combining a routing target node scoring table with the QUIC protocol, and containerized elastic resource allocation through Kubernetes according to quantitative indicators, thereby achieving bandwidth optimization, low-latency migration and elastic resource adaptation. The closed-loop feedback layer is used to collect the execution effect data of the three core actions, which is fed back to the multi-agent reinforcement learning framework and the lightweight TCN model to update the model parameters to optimize the subsequent analysis and decision-making logic, forming a closed-loop optimization link of "data collection - edge analysis - global decision-making - execution feedback - model iteration". Specifically, the dynamic execution layer performs three core actions in parallel through the following operations: (a) The pre-caching priority list clearly indicates the popularity ranking and fragment type of each content fragment; When performing intelligent pre-caching based on the pre-caching priority list, the top 10% of core shards in terms of popularity are pushed to the edge node cache first. At the same time, the cache hit rate and bandwidth usage data of the edge nodes are synchronized in real time. If the cache hit rate of the top 10% of shards of a certain type of content is lower than the threshold, the shard push weight of that type of content is dynamically adjusted. (b) When a user initiates a request or a network switch occurs, the optimal target node is selected based on the routing target node scoring table. At the same time, relying on the 0-RTT handshake feature of the QUIC protocol, the request migration can be completed without re-establishing a TCP connection. During the migration process, the latency data of node switching is monitored in real time. By dynamically adjusting the node scoring weight, the node switching latency is strictly controlled within 15ms, thus achieving low-latency adaptation of the request route. (c) Using QPS change rate as the core quantitative indicator, build a Kubernetes containerized resource scheduling system; collect QPS data of Kubernetes container clusters in real time, and dynamically adjust the number of Pod replicas through the preset elastic scaling formula ΔPod = k·ΔQPS; at the same time, set a scaling threshold trigger mechanism.

6. The CDN edge node collaborative optimization system based on a lightweight AI model according to claim 5, characterized in that, The edge-aware layer specifically includes: The data acquisition module is used to capture HTTP / QUIC traffic characteristics in real time through NetFlow traffic sensors with a sampling frequency of 100ms deployed on edge nodes, and to collect user request logs and node resource monitoring data simultaneously. The user request logs include content type distribution and device geographic location information, and the node resource monitoring data uses cgroups technology to statistically analyze CPU, memory and NIC metrics. The timing alignment module is used to perform timing alignment on multi-source data collected by the data acquisition module through a sliding window processor to ensure the consistency of the data in the time dimension, and then transmit it to the edge node; The model deployment module is used to deploy a lightweight 8-layer TCN model optimized by channel pruning and INT8 quantization at edge nodes. The TCN model performs real-time analysis on the output data of the time alignment module and then outputs a pre-cached priority list and a route target node score table that meet preset conditions. The pre-caching trigger module is used to automatically start the pre-caching instruction generation process when the TCN model detects that the request growth rate ΔQ / Δt > 200% and the content popularity H(c) > 0.

7.

7. The CDN edge node collaborative optimization system based on a lightweight AI model according to claim 6, characterized in that, The default conditions for the pre-caching priority list are: TOP 10 content ranked by popularity, with a confidence level of ≥90%; The default conditions for the routing target node rating table are: neighboring nodes within 3 hops with a load rating of <0.6 and an RTT of <50ms.

8. A CDN edge node collaborative optimization system based on a lightweight AI model according to claim 5, characterized in that, The collaborative optimization layer specifically includes: The data aggregation module is used to aggregate the global status information reported by each node in the CDN network, specifically covering: a 1024-node adjacency list, the distribution of the top 1000 content requests, and the normalized load vector. The demand collaborative matching module is used to quickly filter out nodes with similar demands based on aggregated global state information and the pre-cached priority list and route target node scoring table output by the edge perception layer. It adopts the Locality-Sensitive Hashing algorithm with a Jaccard coefficient greater than 0.7 as the matching threshold, and builds a node demand association model based on the improved collaborative filtering mechanism to clarify the demand association relationship between nodes. The resource negotiation module is used to initiate the resource negotiation process based on the node demand association model and the real-time load of nodes: nodes whose load exceeds the preset threshold publish task tenders based on the Contract Net protocol, and idle nodes respond with a list of available resources, thus completing the resource supply and demand matching between nodes. The strategy generation module, based on the supply and demand matching results from the resource negotiation phase, combines the pre-cached priority list output by the edge awareness layer and the routing target node scoring table. It then uses a multi-agent reinforcement learning framework to integrate three core metrics to generate a global balancing strategy. The strategy focuses on improving cache hit rate, reducing latency, and achieving load balancing as its core objectives. It plans three main execution directions: cross-node content pre-caching, dynamic request routing, and elastic resource allocation. The specific implementation process is as follows: (1) Based on the supply and demand matching results of the resource negotiation stage, combined with the pre-cached priority list and route target node scoring table output by the edge awareness layer, as well as the three core indicators of cache hit rate, P99 request latency and node load standard deviation collected in real time by CDN network nodes, the environment status input of the MARL framework is used. (2) Using each edge node as an independent intelligent agent, generate three types of candidate actions based on local state and global indicators: cross-node content pre-caching, dynamic request diversion, and elastic resource allocation. (3) The central server aggregates all agent actions and calculates the global reward value according to a preset formula: ， in: The cache hit rate weight is used to maximize content reuse; H represents content reuse degree, which measures the degree of overlap in content sharing between nodes. These are the delay constraint weighting coefficients, used to constrain the P99 delay from exceeding the target threshold; For the P99 latency of the node; The preset P99 delay target threshold; σ is the load balancing weighting coefficient; σ is the standard deviation of node load, used to measure the uniformity of load distribution across all nodes. (4) Update the agent's policy network parameters through the gradient ascent algorithm, strengthen high-reward actions and suppress negative feedback actions, and complete policy evaluation and iterative optimization; (5) Generate a global balancing strategy, clearly define the core objectives as improving cache hit rate, reducing latency and load balancing, and determine the specific execution direction corresponding to the three types of candidate actions.

Images