An 8K high code rate panoramic live broadcast hierarchical distribution method for sports event private network

Abstract

The application provides an 8K high code rate panoramic live broadcast grading distribution method for a sports event private network, and belongs to the technical field of live broadcast distribution.The original 8K panoramic video is compressed and encoded by a high-efficiency video coding standard to generate a high code rate video stream in a live source node, and the high code rate video stream is pushed to a first forwarding node through a real-time transmission protocol.The maximum transmission unit of the private network full link is uniformly configured as 9000 bytes in the first forwarding node, and a data plane development kit user state protocol stack is enabled.A second distribution node concurrently pushes the high code rate video stream to terminal equipment through a lock-free ring buffer.A fault detection module combines active heartbeat detection with passive response time monitoring to monitor the node state in real time, and migrates the affected terminal to a backup node within a second, thereby solving the technical problems of high transmission delay and poor concurrent stability of the high code rate video stream during the grading distribution of the 8K high code rate panoramic video of the sports event private network.

Description

Technical Field

[0001] This invention belongs to the field of live streaming distribution technology, and more specifically, it relates to a hierarchical distribution method for 8K high bitrate panoramic live streaming for sports event dedicated networks. Background Technology

[0002] With the large-scale application of 8K panoramic video technology in live sports broadcasts, a hierarchical distribution architecture for private networks has become the mainstream technical path for achieving real-time transmission of ultra-high-definition video. In existing technologies, the live source node compresses the original video through an encoding module, then pushes the video stream tier by tier to multiple secondary distribution nodes via a primary forwarding node. Terminal devices then pull the stream from the secondary distribution nodes for decoding and playback. Currently, sports event private networks generally use standard maximum transmission unit (MTB) configurations and operating system kernel protocol stacks for data forwarding. Load balancing relies on static routing rules or simple round-robin strategies to allocate distribution paths, and the system control module only periodically collects single-node metrics for threshold alarms and manual intervention. In current high-bitrate video distribution on private networks, due to the extremely high bitrate of 8K video and the massive number of data packets, the standard MTB leads to frequent fragmentation and reassembly, introduces significant context switching overhead into the kernel protocol stack, and consumes substantial CPU resources at the primary forwarding nodes. Simultaneously, static routing and simple round-robin strategies cannot perceive the global load distribution and dynamic changes in the private network topology. When local nodes experience congestion or failure, the distribution path cannot adaptively adjust, leading to accumulated high-bitrate video stream transmission delays and a sharp decline in system stability as the number of concurrent terminals increases. In other words, existing technologies have technical problems such as high CPU utilization of primary forwarding nodes during the hierarchical distribution of 8K high bitrate panoramic video in sports event private networks, and the inability of load balancing strategies to perceive dynamic changes in the global topology, resulting in high latency and poor concurrency stability of high bitrate video stream transmission. Summary of the Invention

[0003] In view of this, the present invention provides a hierarchical distribution method for 8K high bitrate panoramic live streaming for sports event private networks, which can solve the technical problems in the prior art where the central processing unit occupancy rate of the first-level forwarding node is high and the load balancing strategy cannot perceive the dynamic changes in the global topology, resulting in high transmission latency and poor concurrency stability of high bitrate video streams.

[0004] This invention is implemented as follows: This invention provides a hierarchical distribution method for 8K high bitrate panoramic live streaming on dedicated networks for sports events, including the following steps:

[0005] The live source node receives raw 8K panoramic video data, compresses and encodes it using an efficient video encoding standard through the encoding module, generates a high bitrate video stream, and pushes the high bitrate video stream to the first-level forwarding node through the stream push module via a real-time transmission protocol.

[0006] The first-level forwarding node receives high-bitrate video streams through the stream receiving module, stores the high-bitrate video streams in an unlocked circular buffer, and uniformly configures the maximum transmission unit of the entire private network link to 9000 bytes. The user-space protocol stack of the data plane development kit is enabled on the first-level forwarding node, and the high-bitrate video streams are forwarded to the second-level distribution node through zero-copy large packet forwarding.

[0007] The system control and management module periodically collects the CPU load rate, network bandwidth utilization rate, number of connected terminals, and buffer queue depth of each node. It constructs a directed weighted graph based on the connection relationship between the nodes of the private network, using the CPU load rate, network bandwidth utilization rate, number of connected terminals, and buffer queue depth as node feature vectors, and link latency and link bandwidth as edge features. This is input into the private network topology-aware load balancing network, and outputs the probability distribution of distribution paths for each node. Simultaneously, it executes a multi-level node importance assessment and distribution priority ranking algorithm based on random walk restart using the directed weighted graph, outputting the importance score of each node. Based on the distribution path probability distribution and the importance score of each node, the distribution paths are prioritized, and high-bitrate video streams are preferentially distributed to secondary distribution nodes with lower overall load values ​​via paths with high importance scores.

[0008] The system control and management module calculates the comprehensive load value of each node based on the collected central processing unit load rate, network bandwidth utilization rate, and number of connected terminals. It then inputs the comprehensive load value into the distribution path adjustment function and adjusts the path allocation strategy parameters of the private network topology-aware load balancing network according to the range to which the comprehensive load value belongs, thereby achieving dynamic load balancing.

[0009] The secondary distribution node obtains the high bitrate video stream from the primary forwarding node and pushes the high bitrate video stream concurrently to the terminal devices in the area using an unlocked circular buffer. The terminal devices maintain a connection with the secondary distribution node through a heartbeat mechanism, receive the high bitrate video stream, and decode and play it.

[0010] The fault detection module monitors the status of each node in real time by combining active heartbeat detection with passive response time monitoring. When an abnormality is detected in a secondary distribution node or the overall load value exceeds the preset fault threshold, the system control and management module will migrate the affected terminal devices to the backup node within seconds to ensure the continuity of high bitrate video stream playback.

[0011] Specifically, storing high-bitrate video streams into a lock-free circular buffer refers to using a circular memory queue structure that implements multi-threaded concurrent read and write operations with atomic comparison and swap instructions. The advancement of read and write pointers is completed through atomic operations. The read and write pointers of each thread are aligned and filled with CPU cache lines, so that each pointer occupies a different cache line, thereby eliminating the additional cache synchronization overhead caused by false sharing.

[0012] Specifically, configuring the maximum transmission unit of the entire private network link to 9000 bytes means enabling jumbo frames to eliminate the network layer fragmentation requirement caused by the inconsistency of the maximum transmission unit when transmitting large data packets of 8K high bitrate video streams between private network links, thus avoiding the overhead of central processing unit for fragmentation and reassembly at the receiving end; the user-space protocol stack of the data plane development kit bypasses the network protocol stack of the operating system kernel and directly drives the network card to complete data transmission and reception in user space.

[0013] Specifically, the private network topology-aware load balancing network refers to a network that uses an improved GraphSAGE structure as a state encoder. It extracts the node embedding representation of each node through a three-layer neighborhood aggregation operation. During each aggregation layer, the node feature vectors of neighboring nodes are averaged and pooled, and then concatenated with the node feature vector of the current node. After linear transformation and nonlinear activation, a new node embedding representation is obtained.

[0014] The dedicated network topology-aware load balancing network encodes link latency and link bandwidth as attention weights during the message propagation phase. Through the bidirectional attention propagation mechanism of edge features, each node's node embedding representation simultaneously perceives upstream distribution pressure and downstream pull pressure. When congestion features are detected in a subgraph, the message propagation depth is dynamically increased by 2 layers to expand the perception range.

[0015] The dedicated network topology-aware load balancing network inputs node embedding representations as state vectors into the actor-critic reinforcement learning framework. The actor network outputs the distribution path probability distribution of each candidate distribution path, and the critic network estimates the long-term benefit of the current strategy within a future time window. The long-term benefit is defined as the weighted negative of the total system latency and packet loss rate. An additional physical constraint layer is fused, embedding the upper limit of each link bandwidth as a hard constraint into the normalized exponential function mask of the actor network strategy output.

[0016] Specifically, the training of the dedicated network topology-aware load balancing network refers to performing stress injection in a simulated network environment with different numbers of concurrent terminals, different node load combinations, and different link bandwidth conditions, collecting a state-action-reward triplet sample set, collecting no less than 50,000 samples, dividing the training set, validation set, and test set in an 8:1:1 ratio, updating the actor network parameters with a near-end policy optimization algorithm, and optimizing the long-term revenue estimation of the critic network using the mean squared error loss function.

[0017] The formula for calculating the comprehensive load value is as follows: ,in For the overall load value, The current CPU load rate, This is the baseline value for the central processing unit at full load. This represents the current network bandwidth utilization. This is the baseline value for full-load link bandwidth utilization. This represents the number of currently connected terminals. This is the baseline value for the maximum number of connected terminals per node. , , The weighting coefficients are satisfied. .

[0018] Specifically, the distribution path adjustment function refers to when When the probability distribution of the current distribution path remains unchanged; when When the probability weights of low-importance scoring paths are shifted to high-importance scoring paths; when When the secondary distribution node requests a new high-bitrate video stream replica connection from the upstream primary forwarding node; The failover process is triggered in time, forcibly migrating some terminal devices to the backup node.

[0019] Specifically, the multi-level node importance assessment and distribution priority ranking algorithm based on random walk restart refers to using the live stream source nodes as the initial distribution of personalized random walks, and ranking them on a directed weighted graph with probability... Transition from the current node to its neighboring node along the outgoing edge, with probability. Returning to the live stream source node, the algorithm iterates until it reaches a stationary convergence distribution. The components of the convergence vector represent the importance scores of each node. The overall time complexity of the algorithm is O(n log n). ,in The number of iterations required for convergence, Let be the number of edges in a directed weighted graph.

[0020] The link weights of the directed weighted graph comprehensively encode the link bandwidth, link delay and comprehensive load value between nodes. Every fixed time window, the link weights of the directed weighted graph are updated according to the latest link bandwidth measurement value and the multi-level node importance assessment and distribution priority sorting algorithm based on random walk restart is re-executed to achieve dynamic priority adaptive adjustment.

[0021] The fault detection module monitors the status of each node in real time by combining active heartbeat detection with passive response time monitoring. The preset fault threshold is 0.9. When the comprehensive load value... The system control and management module triggers a fault switching process, migrating the affected terminal devices to the backup node within seconds.

[0022] Specifically, the near-end policy optimization algorithm refers to introducing a truncation ratio constraint during policy updates to limit the offset range of the probability ratio between the old and new policies during each parameter update, thereby preventing the policy update step from being too large and causing training instability.

[0023] Specifically, the normalized exponential function mask refers to applying a very large negative bias to the component corresponding to the non-action in the normalized exponential function layer of the actor network strategy output, so that the component approaches zero after being transformed by the normalized exponential function, thereby forcibly excluding the distribution action that violates the physical bandwidth constraint at the probability level.

[0024] The secondary distribution node concurrently pushes high-bitrate video streams to terminal devices in its area using an unlocked circular buffer. The terminal devices maintain a connection with the secondary distribution node through a heartbeat mechanism, receive the high-bitrate video streams, and decode and play them. The CPU load rate, network bandwidth utilization rate, and number of connected terminals periodically collected by the system control and management module are directly used in the comprehensive load value calculation formula. After the comprehensive load value is output, it is used for the interval judgment of the distribution path adjustment function and the trigger judgment of fault switching.

[0025] This invention combines jumbo frame configuration, zero-copy forwarding of user-space protocol stacks from the data plane development kit, and a topology-aware load balancing network for private networks to construct a hierarchical distribution method for 8K high bitrate panoramic live streaming for sports events. It solves the technical problems of high CPU occupancy in primary forwarding nodes and the inability of load balancing strategies to perceive dynamic changes in the global topology, which lead to high latency and poor concurrency stability in high bitrate video stream transmission. This invention uniformly configures the maximum transmission unit to 9000 bytes to eliminate network layer fragmentation and reassembly overhead. Simultaneously, it enables the user-space protocol stack of the data plane development kit to bypass the operating system kernel protocol stack, and combines lock-free ring buffer line alignment padding to eliminate false sharing, fundamentally reducing the CPU consumption of first-level forwarding nodes in high-concurrency scenarios. Furthermore, the dedicated network topology-aware load balancing network models the entire network topology using an improved GraphSAGE structure. It combines immediate load balancing decisions with long-term system stability goals through an actor-commentator reinforcement learning framework. A multi-level node importance evaluation algorithm based on random walk restarts globally prioritizes distribution paths, ensuring that high-bitrate video streams are always transmitted along the optimal path. The distribution path adjustment function dynamically adjusts the path allocation strategy based on the comprehensive load value, and the fault detection module completes the migration of backup nodes for affected terminals within seconds. In summary, this invention solves the technical problems mentioned in the background art, such as high CPU occupancy of first-level forwarding nodes and the inability of the load balancing strategy to perceive dynamic changes in the global topology during the hierarchical distribution of 8K high-bitrate panoramic video in dedicated sports event networks, leading to high transmission latency and poor concurrency stability of high-bitrate video streams. Attached Figure Description

[0026] Figure 1 This is a flowchart of the method of the present invention.

[0027] Figure 2 A graph showing the change in video stream reception delay of affected terminal devices before and after node fault injection.

[0028] Figure 3 A dynamic adjustment curve showing the importance score of each secondary distribution node as the number of concurrent terminals changes. Detailed Implementation

[0029] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below.

[0030] like Figure 1 The diagram shown is a flowchart of a hierarchical distribution method for 8K high bitrate panoramic live streaming on a dedicated sports event network provided by this invention. This method includes the following steps:

[0031] S01. The live source node receives the raw 8K panoramic video data, compresses and encodes it using an efficient video encoding standard through the encoding module to generate a high bitrate video stream, and pushes the high bitrate video stream to the first-level forwarding node through the stream push module via a real-time transmission protocol.

[0032] S02. The first-level forwarding node receives the high bitrate video stream through the stream receiving module, stores the high bitrate video stream in the lockless ring buffer, and uniformly configures the maximum transmission unit of the private network full link to 9000 bytes. The first-level forwarding node enables the user-space protocol stack of the data plane development kit and forwards the high bitrate video stream to the second-level distribution node through zero-copy large packet forwarding.

[0033] S03. The system control and management module periodically collects the CPU load rate, network bandwidth utilization rate, number of connected terminals, and buffer queue depth of each node. It constructs a directed weighted graph based on the connection relationship of each node in the private network, uses the CPU load rate, network bandwidth utilization rate, number of connected terminals, and buffer queue depth as node feature vectors, and uses link latency and link bandwidth as edge features. It inputs the graph into the private network topology-aware load balancing network and outputs the probability distribution of distribution paths for each node. At the same time, it executes a multi-level node importance assessment and distribution priority ranking algorithm based on random walk restart using the directed weighted graph, outputs the importance score of each node, and prioritizes the distribution paths according to the probability distribution of distribution paths and the importance scores of each node. It prioritizes the distribution paths and distributes the high bitrate video stream to the secondary distribution nodes with lower overall load values ​​via paths with high importance scores.

[0034] S04. The system control and management module calculates the comprehensive load value of each node based on the central processing unit load rate, network bandwidth utilization rate and network bandwidth utilization rate collected in step S03. The comprehensive load value is input into the distribution path adjustment function. The path allocation strategy parameters of the private network topology-aware load balancing network are adjusted according to the interval to which the comprehensive load value belongs, so as to achieve dynamic load balancing.

[0035] S05. The secondary distribution node obtains the high bitrate video stream from the primary forwarding node and concurrently pushes the high bitrate video stream to the terminal devices in the area using an unlocked circular buffer. The terminal devices maintain a connection with the secondary distribution node through a heartbeat mechanism, receive the high bitrate video stream, and decode and play it.

[0036] S06. The fault detection module monitors the status of each node in real time by combining active heartbeat detection and passive response time monitoring. When an abnormality is detected in the secondary distribution node or the comprehensive load value exceeds the preset fault threshold, the system control and management module will migrate the affected terminal devices to the backup node within seconds to ensure the continuity of the high bitrate video stream playback.

[0037] The lock-free circular buffer refers to a circular memory queue structure that uses compare-and-swap atomic instructions to achieve concurrent read and write operations by multiple threads without the need for mutexes. The advancement of read and write pointers is completed through atomic operations, eliminating mutex contention for the shared buffer among multiple terminal read threads. The lock-free circular buffer eliminates the additional cache synchronization overhead caused by false sharing by aligning and filling the CPU cache lines of each thread's read and write pointers, ensuring that each pointer occupies a different cache line. This significantly reduces thread context switching overhead and improves the computational efficiency of forwarding nodes in high-concurrency scenarios. False sharing refers to the phenomenon where multiple threads' read and write pointers, although belonging to different logical variables, frequently invalidate the cache line because they share the same CPU cache line. CPU cache line alignment and filling involves filling the memory addresses of each thread's read and write pointers, ensuring that each pointer occupies a different CPU cache line, thereby eliminating false sharing.

[0038] The maximum transmission unit (MTB) is uniformly configured to 9000 bytes, enabling jumbo frames. This eliminates the network layer fragmentation requirement caused by inconsistent MTBs when transmitting large data packets of 8K high-bitrate video streams between dedicated network links, avoiding the CPU overhead of fragment reassembly at the receiving end. The user-space protocol stack of the data plane development kit bypasses the operating system kernel network protocol stack, directly driving the network card to complete data transmission and reception in user space. Combined with jumbo frames, it achieves zero-copy large packet forwarding, solving the problem of high CPU utilization at the first-level forwarding node in high-concurrency scenarios. The 9000-byte MTB value was obtained through experiments: a multi-switch network similar to the sports event dedicated network was built in a laboratory environment, with MTBs configured as 1500 bytes, 4500 bytes, and 9000 bytes respectively. Stress tests were conducted with 8K high-bitrate video streams, and the CPU utilization and packet loss rate of the first-level forwarding node were collected. Through multiple rounds of iterative experiments and data analysis, it was confirmed that 9000 bytes has the best compatibility on mainstream 10 Gigabit Ethernet switches and has a significant fragmentation elimination effect.

[0039] The specific structure of the dedicated network topology-aware load balancing network is as follows: the node connection relationship of the dedicated network hierarchical distribution system is modeled as a directed weighted graph. The node feature vector of each node consists of four physical parameters: current central processing unit load rate, network bandwidth utilization rate, number of connected terminals, and buffer queue depth. The edge feature of each directed edge consists of two parameters: link delay and link bandwidth. An improved GraphSAGE structure is used as a state encoder. A three-layer neighborhood aggregation operation is used to extract the node embedding representation of each node in the topological context. During each aggregation layer, the node feature vectors of neighboring nodes are averaged and pooled, then concatenated with the node feature vector of the current node, and then subjected to linear transformation and nonlinear activation to obtain a new node embedding representation. During the message propagation stage, the link delay and link bandwidth are encoded as attention weights and distributed through edges. The bidirectional attention propagation mechanism enables each node's node embedding representation to simultaneously perceive upstream distribution pressure and downstream pull pressure. When congestion is detected in a subgraph, the message propagation depth dynamically increases by two layers to expand the perception range. The node embedding representation is input as a state vector into the actor-critic reinforcement learning framework. The actor network consists of two fully connected layers, outputting the distribution path probability distribution of each candidate distribution path. The critic network consists of two fully connected layers, estimating the long-term benefit of the current strategy within the future time window. The long-term benefit is defined as the weighted negative of the total system latency and packet loss rate. An additional physical constraint layer is fused, embedding the upper limit of each link bandwidth as a hard constraint into the normalized exponential function mask of the actor network strategy output, forcing the probability of impossible actions to zero, and ensuring that the output distribution action meets the physical bandwidth feasibility. The dedicated network topology-aware load balancing network uses graph neural networks to structurally model the dedicated network topology, enabling load balancing decisions to simultaneously perceive the local node status and the global topology pressure distribution. The actor-commentator reinforcement learning framework combines real-time load balancing decisions with long-term system stability goals, and the physical constraint layer ensures the engineering feasibility of the decisions. Overall, it achieves adaptive optimization of the distribution path under dynamic load changes in the dedicated network, effectively reducing overall system latency and improving concurrent processing stability.

[0040] The steps for establishing the training dataset for the dedicated network topology-aware load balancing network specifically include: building a simulated network environment in the laboratory that is similar to the topology of the sports event dedicated network; performing stress injection with different numbers of concurrent terminals, different node load combinations, and different link bandwidth conditions; collecting node feature vectors, edge features, distribution path probability distributions corresponding to distribution actions, and the total system latency and packet loss rate at each time point to form a state-action-reward triplet sample set; simultaneously collecting samples under node fault injection and burst traffic scenarios to enhance the dataset coverage; collecting no less than 50,000 samples in total, and dividing them into training set, validation set, and test set in an 8:1:1 ratio. The specific steps for training the dedicated network topology-aware load balancing network include: online training using the actor-critic algorithm framework; updating the actor network parameters using a near-end policy optimization algorithm to constrain the policy update magnitude; optimizing the long-term revenue estimation of the critic network using the mean squared error loss function; incrementally fine-tuning the network parameters with the latest link state data at fixed time windows; using the weighted average of the system's total latency and packet loss rate on the validation set as the early stopping criterion during training; and using the revenue improvement rate being lower than a preset convergence threshold within 10 consecutive validation periods as the convergence criterion, which is determined through multiple rounds of experimental iterations.

[0041] Among them, the near-end policy optimization algorithm introduces a truncation ratio constraint during policy updates to limit the offset range of the probability ratio between the old and new policies during each parameter update. This prevents excessively large policy update steps from causing training instability, thereby improving the robustness of the training process while ensuring learning efficiency. The normalized exponential function mask applies a large negative bias to the components corresponding to inactive actions in the normalized exponential function layer of the actor network policy output. This causes the components to approach zero after the normalized exponential function transformation, thus forcibly excluding actions that violate physical bandwidth constraints at the probabilistic level.

[0042] The formula for calculating the comprehensive load value is as follows: ;in This is the overall load value, with a dimension of 1. The current CPU load rate, This is the baseline value for the central processing unit at full load; both have the same dimensions. This represents the current network bandwidth utilization. This is the baseline value for full-load link bandwidth utilization; both have the same unit of measurement. This represents the number of currently connected terminals. This is a baseline value for the maximum number of connected terminals per node; both have the same unit of measurement. , , The weighting coefficients are satisfied. Weighting coefficients , , The specific values ​​were obtained through experiments: In a simulated private network environment, multiple rounds of stress tests were conducted using 8K high bitrate video streams, and the values ​​were enumerated respectively. , , The optimal weight combination is determined through at least 30 rounds of iterative experiments, using a weighted average of total system latency and packet loss rate as the optimization objective, on a grid with a step size of 0.1. The distribution path adjustment function is based on the overall load value. The path allocation strategy parameters of the private network topology-aware load balancing network within the specified interval are adjusted as follows: when When <0.4, the private network topology-aware load balancing network maintains the current distribution path probability distribution unchanged; when When, the probability weights of low-importance score paths in the actor network output distribution path probability distribution are shifted to high-importance score paths, and the shift ratio is calculated by linear interpolation of the distribution path adjustment function; when When this occurs, the secondary distribution node is triggered to request a new high-bitrate video stream replica connection from the upstream primary forwarding node, while simultaneously compressing the normalized exponential function mask weights of low-bandwidth paths in the physical constraint layer; when At this time, the system control and management module triggers a fault switching process, forcibly migrating some terminal devices to the backup node and notifying the actor network to block all distribution actions corresponding to the node with the current high overall load value. The interval boundary values ​​of 0.4, 0.7, and 0.9 were obtained experimentally: stress tests were conducted in a simulation environment at different concurrency scales, collecting the overall load value and system service quality degradation curves. The boundaries of each interval were determined through piecewise linear fitting and inflection point analysis, and the boundary stability was verified through no less than 20 rounds of iterative experiments. Preset fault thresholds and overall load values... The interval boundary is consistent, and the value is 0.9. The method of obtaining the value is the same as the interval boundary value of 0.9.

[0043] In step S04, the parameters used in calculating the comprehensive load value are: the CPU load rate and network bandwidth utilization rate, which are periodically collected by the system control and management module in step S03; and the number of connected terminals, which are collected in step S03. All three parameters are directly used in the comprehensive load value calculation formula. The comprehensive load value is then used for the interval judgment of the distribution path adjustment function in step S04 and the trigger judgment of fault switching in step S06.

[0044] The multi-level node importance assessment and distribution priority ranking algorithm based on random walk restart models the node connection relationship of the private network hierarchical distribution system as a directed weighted graph. The link weights of the directed weighted graph comprehensively encode the link bandwidth, link delay, and overall load value between nodes. The algorithm uses the live stream source nodes as the initial distribution for personalized random walks and performs a random walk process with restart on the directed weighted graph. At each step, a probability is applied... Transition from the current node to its neighboring node along the outgoing edge, with probability. Returning to the live stream source node, the algorithm iterates until a stable distribution converges, and the components of the convergence vector represent the importance score of each node. Nodes with high importance scores are physically close to the live stream source node and have strong link bandwidth transmission capabilities. Based on this, the algorithm prioritizes the distribution paths, prioritizing the transmission of the high-bitrate video stream through high-importance-score paths, while low-importance-score paths serve as backups. Every fixed time window, the link weights of the directed weighted graph are updated based on the latest link bandwidth measurements, and the multi-level node importance evaluation and distribution priority ranking algorithm based on random walk restarts is re-executed to achieve dynamic priority adaptive adjustment. The overall time complexity of the algorithm is O(log n). ,in The number of iterations required for convergence, The number of edges in the directed weighted graph typically converges within 20–50 steps. Restart probability. The specific values ​​were obtained through experiments: in a simulated private network environment, the values ​​were taken as follows: The optimal values ​​were determined through multiple rounds of iterative experiments, using convergence speed and the stability of the importance score ranking of each node as evaluation indicators, with values ​​of 0.10, 0.15, 0.20, and 0.25. The multi-level node importance assessment and distribution priority ranking algorithm based on random walk restart modeled the global structure of the private network topology, allowing the priority ranking of distribution paths to simultaneously reflect the physical distance between nodes, link bandwidth capacity, and current comprehensive load value. This avoids the limitations of traditional static routing, which relies solely on local information, enabling high-bitrate video streams to be transmitted along the optimal path within the private network, effectively improving bandwidth utilization efficiency and distribution stability.

[0045] Optionally, the present invention also provides a computer-based approach to form an 8K high bitrate panoramic live streaming hierarchical distribution system for sports event private networks. The computer is equipped with a readable storage medium that stores program instructions, which execute the above-described method when the computer is run. Furthermore, the above-described method is executed through multiple modules, including an encoding module, a stream pushing module, a stream receiving module, a system control and management module, and a fault detection module.

[0046] The specific implementation of step S01 is as follows: The live source node continuously receives raw 8K panoramic video data from the camera array. The raw video data is processed by the encoding module using an efficient video coding standard for inter-frame prediction, transform quantization, and entropy coding to generate a high-bitrate video stream. Compared with traditional coding standards, the efficient video coding standard can achieve higher compression efficiency at the same visual quality, which is suitable for the bitrate control requirements of 8K ultra-high-definition panoramic video. After encoding, the stream push module encapsulates the high-bitrate video stream into real-time transmission protocol data packets and continuously pushes them to the primary forwarding node through the private network link. The real-time transmission protocol provides end-to-end media transmission capabilities, ensuring the orderly transmission of the video stream within the private network.

[0047] The specific implementation of step S02 is as follows: After the stream receiving module of the first-level forwarding node receives the high-bitrate video stream from the live source node, it writes the data into a lock-free circular buffer. The lock-free circular buffer uses comparison and swap atomic instructions to advance the read and write pointers, eliminating the overhead of mutex lock contention in multi-threaded concurrent scenarios; the memory addresses where the read and write pointers of each thread are located are aligned and filled with CPU cache lines, so that each pointer occupies a different cache line, thereby eliminating the cache invalidation overhead caused by false sharing, and significantly reducing the frequency of thread context switching in high-concurrency scenarios. At the same time, the maximum transmission unit of all switches and nodes in the private network is uniformly configured to 9000 bytes, that is, jumbo frames are enabled, so that the large data packets of 8K high-bitrate video streams do not need to undergo network layer fragmentation and receiver reassembly when transmitted between private network links, completely eliminating the CPU overhead introduced by fragmentation and reassembly. The user-space protocol stack of the data plane development kit is enabled at the first-level forwarding node. This protocol stack bypasses the operating system kernel network protocol stack and the user-space program directly polls and drives the network card to complete the transmission and reception of data packets. Combined with jumbo frames, it realizes zero-copy large packet forwarding, that is, the data is directly mapped and transmitted between the network card memory and the user-space buffer without going through the kernel space copy. This fundamentally solves the problem of high CPU utilization of the first-level forwarding node in high-concurrency scenarios.

[0048] The specific implementation of step S03 is as follows: The system control and management module collects four physical parameters from each node at a fixed period (reference collection period is 500 milliseconds): central processing unit load rate, network bandwidth utilization rate, number of connected terminals, and buffer queue depth, forming the node feature vector of each node; simultaneously, it collects the link delay and link bandwidth of each link, forming the edge features of a directed weighted graph. The aforementioned directed weighted graph uses each node of the private network as vertices and the physical connection relationships between nodes as directed edges, comprehensively describing the topology of the private network hierarchical distribution system. The network uses node feature vectors and edge features as inputs to a dedicated network topology-aware load balancing network. This network employs an improved GraphSAGE structure as a state encoder. Through a three-layer neighborhood aggregation operation, it extracts the node embedding representation for each node within the topological context. Each aggregation layer performs mean pooling on the feature vectors of neighboring nodes and concatenates them with the current node's feature vector. After linear transformation and nonlinear activation, a new node embedding is output. During the message propagation phase, link latency and link bandwidth are encoded as attention weights. Through a bidirectional attention propagation mechanism based on edge features, the node embedding simultaneously perceives upstream distribution pressure and downstream pull pressure. When congestion is detected in a subgraph, the message propagation depth is dynamically increased by two layers to expand the perception range. Node embedding serves as the input state vector to the actor-critic reinforcement learning framework. The actor network consists of two fully connected layers, outputting the probability distribution of each candidate distribution path. The critic network also consists of two fully connected layers, estimating the long-term reward (i.e., the weighted negative of the total system latency and packet loss rate). The physical constraint layer embeds the upper limit of each link bandwidth into the actor network output in the form of a normalized exponential function mask, applying a large negative bias to distribution actions that violate the physical bandwidth constraints, making their probability approach zero, thus ensuring the engineering feasibility of the output distribution actions. Simultaneously, a multi-level node importance evaluation and distribution priority ranking algorithm based on random walk restart uses the live source nodes as the initial distribution and performs a random walk with restart on the directed weighted graph: each step is based on probability... Transition along the outgoing edge to the adjacent node with probability. (Reference value is 0.15) Returning to the live stream source node, iterate until the distribution converges smoothly (usually 20-50 steps). The components of the convergence vector are the importance scores of each node. The link weights of the directed weighted graph comprehensively encode the link bandwidth, link delay, and overall load value. Nodes with high importance scores are topologically closer to the live stream source node and have strong link transmission capabilities. Based on the probability distribution of the distribution paths and the importance scores of each node, the distribution paths are prioritized. High-bitrate video streams are preferentially distributed to secondary distribution nodes with lower overall load values ​​via high-importance-score paths, while low-importance-score paths serve as backups.

[0049] The specific implementation of step S04 is as follows: The system control and management module utilizes the central processing unit load rate collected in step S03. Network bandwidth utilization and the number of connected terminals According to the comprehensive load value calculation formula Calculate the overall load value of each node The weighting coefficient , , satisfy The specific value was determined through no fewer than 30 rounds of stress testing experiments in a simulation environment. (Comprehensive load value) Input distribution path adjustment function: when When the private network topology-aware load balancing network maintains the current distribution path probability distribution unchanged; when When the probability weights of low-importance scoring paths in the actor network output are shifted to high-importance scoring paths, the shift ratio is calculated by linear interpolation of the distribution path adjustment function; when When this occurs, the secondary distribution node is triggered to request a new high-bitrate video stream replica connection from the upstream primary forwarding node, while simultaneously compressing the normalized exponential function mask weights of low-bandwidth paths in the physical constraint layer; when At this time, the system control and management module triggers the fault switching process, forcibly migrating some terminal devices to the backup node, and notifying the actor network to block all distribution actions corresponding to the node with the current high comprehensive load value. The interval boundary values ​​of 0.4, 0.7, and 0.9 are determined through no less than 20 rounds of simulation stress test experiments. The comprehensive load value of each node is updated once per collection cycle, and the distribution path adjustment function continuously and dynamically adjusts the path allocation strategy parameters to achieve dynamic load balancing across the entire network.

[0050] The specific implementation of step S05 is as follows: After establishing a persistent connection with the primary forwarding node, the secondary distribution node continuously pulls high-bitrate video streams and writes the received video data into a local lock-free circular buffer. The lock-free circular buffer also employs a compare-and-swap atomic instruction and cache line alignment filling mechanism, supporting multiple terminal reading threads to concurrently read video data from the buffer. Each terminal reading thread maintains an independent read pointer, ensuring no interference between them. The terminal device maintains a long connection with the secondary distribution node through a heartbeat mechanism (reference heartbeat interval is 3 seconds). If no heartbeat response is received within a timeout, a reconnection process is triggered to ensure the continuous stability of the connection. After receiving the high-bitrate video stream, the terminal device's local decoding module completes video decoding and rendering playback, achieving smooth real-time display of 8K panoramic video.

[0051] The specific implementation of step S06 is as follows: The fault detection module works in conjunction with two mechanisms: active heartbeat detection and passive response time monitoring. Active heartbeat detection sends probe messages to each node at fixed intervals and waits for responses, while passive response time monitoring continuously calculates the response delay of each node to service requests. When it is detected that the secondary distribution node has experienced multiple consecutive heartbeat timeouts (reference threshold is 3 times) or the response time exceeds the preset upper limit (reference value is 2 seconds), or the comprehensive load value calculated by step S04 is reached, the fault detection module will detect the fault. When the system control and management module is activated, it immediately initiates the failover process: based on the current comprehensive load value of the backup node, the affected terminal devices are redirected to the backup node with the lowest comprehensive load value within seconds. The terminal devices establish a connection with the new node through the heartbeat mechanism and resume continuous reception and decoding playback of high bitrate video streams, ensuring uninterrupted playback.

[0052] It should be noted that the key technologies of this invention include three core technologies: collaborative optimization of jumbo frames and zero-copy forwarding of the user-space protocol stack of the data plane development kit; a private network topology-aware load balancing network (combining GraphSAGE with the actor-critic reinforcement learning framework); and a multi-level node importance evaluation algorithm based on random walk restart. Jumbo frames increase the maximum transmission unit to 9000 bytes, eliminating the CPU overhead introduced by fragmentation and reassembly at the data link layer; the user-space protocol stack of the data plane development kit bypasses kernel-mode switching, and combined with cache line alignment filling of the lock-free circular buffer, further eliminates false sharing, minimizing the CPU resource consumption of the primary forwarding nodes in high-concurrency scenarios, and ensuring stable forwarding of high-bitrate video streams from the physical transmission level. The dedicated network topology-aware load balancing network uses GraphSAGE's three-layer neighborhood aggregation to structurally encode the entire network topology into node embeddings. A bidirectional attention propagation mechanism based on edge features enables decision-making to simultaneously perceive local node states and global topology pressure. The actor-critic framework combines immediate load balancing decisions with long-term system stability goals, while a physical constraint layer ensures the feasibility of distribution actions. An algorithm based on random walk restart prioritizes distribution paths from a global topology perspective, avoiding the limitations of local information in traditional static routing. These three technologies form a collaborative closed loop in eliminating overhead at the transport layer, perceiving the global topology at the scheduling layer, and prioritizing paths at the path layer, systematically improving the end-to-end transmission latency and concurrency stability of high-bitrate video streams within the dedicated network.

[0053] It should be noted that this invention also solves the following technical problem: In the scenario of high bitrate video distribution on a sports event dedicated network, when a secondary distribution node experiences a sudden failure or a sharp increase in the overall load value, traditional systems rely on manual inspection or coarse-grained timed alarms, with fault response times typically on the order of minutes. During this period, affected terminal devices will experience video stream interruptions or severe stuttering, failing to guarantee the continuity of 8K panoramic video playback. This invention combines active heartbeat detection with passive response time monitoring through a fault detection module. It achieves millisecond-to-second-level perception of node anomalies using both continuous heartbeat timeout counts and response time as dual criteria. Combined with real-time calculation of the overall load value, it triggers the fault switching process in advance when the overall load value reaches a preset fault threshold of 0.9. The system control and management module selects the optimal migration target based on the current overall load value of the backup node, redirecting the affected terminal devices within seconds. The terminal devices quickly establish a stable connection with the new node and resume video stream reception using the heartbeat mechanism, thereby ensuring uninterrupted 8K high bitrate panoramic video playback in node anomaly scenarios. This solves the technical problem of traditional methods failing to restore terminal video streams within seconds when a node fails.

[0054] Specifically, the principle of this invention is as follows: This invention can solve the aforementioned core technical problems because the following two levels of technical logic form a mutually supportive closed loop. At the forwarding layer, 8K high bitrate video stream data packets are large in size and numerous in number per unit time. The standard maximum transmission unit forces the network layer to split large data packets into multiple smaller fragments, and the receiving end needs to consume central processing unit resources to reassemble the fragments. The kernel protocol stack needs to switch between user mode and kernel mode every time it sends or receives data, generating a large amount of system call overhead. After uniformly configuring the maximum transmission unit to 9000 bytes, a single data packet can carry more video data, the number of fragments is greatly reduced, and the packet reassembly overhead is eliminated accordingly. The user mode protocol stack of the data plane development kit directly drives the network card, bypassing kernel mode switching. Combined with the cache line alignment and filling of the lockless ring buffer, the cache synchronization overhead caused by false sharing is eliminated, so that forwarding efficiency is guaranteed at the physical level. At the scheduling layer, traditional static routing relies only on local information and cannot perceive the pressure on the entire network topology. The private network topology-aware load balancing network uses an improved GraphSAGE to encode the local node state and the global topology structure into a unified node embedding through three-layer neighborhood aggregation. The bidirectional attention propagation mechanism of edge features enables decision-making to perceive the pressure on upstream and downstream simultaneously. The actor-critic framework incorporates long-term system stability into the optimization objective. The physical constraint layer ensures the engineering feasibility of the distribution action. The importance evaluation algorithm based on random walk restart prioritizes the path from the perspective of global topology. The two-layer mechanism works together to ensure that high bitrate video streams are always transmitted along the optimal path under dynamic load conditions, realizing the organic unity of forwarding efficiency and scheduling intelligence.

[0055] The following provides a specific embodiment 1 of the present invention, and the specific implementation of each step in this embodiment 1 is described in detail below.

[0056] The specific implementation of step S01 is as follows: The live source node receives the original 8K panoramic video data, compresses and encodes it using an efficient video encoding standard through the encoding module, generates a high bitrate video stream, and then pushes it to the first-level forwarding node through the real-time transmission protocol. This step is a conventional encoding and streaming operation.

[0057] The specific implementation of step S02 is as follows: After the first-level forwarding node receives the high-bitrate video stream through the stream receiving module, it stores it in a lock-free circular buffer. The lock-free circular buffer uses compare-and-swap atomic instructions to implement multi-threaded concurrent read and write operations. The advancement of the read and write pointers is completed through atomic operations, eliminating mutex lock contention. To eliminate false sharing, the memory addresses where the read and write pointers of each thread are located are aligned and filled with CPU cache lines, so that each pointer exclusively occupies a different cache line. Let the first... The read pointer of each thread is Write pointer as The cache line size is If the bytes are padded, then the pointers satisfy the following:

[0058] ;

[0059] In the formula, and The first The thread and the first The read pointer address value of each thread. The CPU cache line size, in bytes, ensures that no two threads' read / write pointers reside on the same cache line. Simultaneously, the maximum transmission unit (MTB) across the entire private network link is uniformly configured to 9000 bytes to enable jumbo frames. Furthermore, the user-space protocol stack of the data plane development kit is enabled at the primary forwarding nodes. By bypassing the operating system kernel network protocol stack and directly driving the network card in user space, combined with jumbo frames, zero-copy large packet forwarding is achieved, eliminating the network layer fragmentation requirement caused by inconsistent MTBs when transmitting 8K high-bitrate video streams between private network links. The 9000-byte MTB was determined experimentally: a multi-switch network similar to a sports event private network was built in a laboratory environment, with MTBs configured at 1500 bytes, 4500 bytes, and 9000 bytes respectively. Stress tests were conducted with 8K high-bitrate video streams, and the CPU utilization and packet loss rate of the primary forwarding nodes were collected. Through multiple rounds of iterative experiments and data analysis, it was confirmed that 9000 bytes offers the best compatibility on mainstream 10 Gigabit Ethernet switches and significantly reduces fragmentation.

[0060] The specific implementation of step S03 is as follows: The system control and management module periodically collects the CPU load rate, network bandwidth utilization rate, number of connected terminals, and buffer queue depth of each node, and constructs a directed weighted graph based on the connection relationship between each node in the private network. ,in For a set of nodes, It is a set of directed edges. Each node Node feature vectors It consists of the following four physical parameters:

[0061] ;

[0062] In the formula, For nodes Current CPU load rate For nodes Current network bandwidth utilization For nodes Number of connected terminals For nodes Buffer queue depth, This represents the transpose operation. Each directed edge... The edge feature vector is ,in For the edge The link delay corresponding to the link, For the edge The corresponding link bandwidth. The private network topology-aware load balancing network uses an improved graph sampling and aggregation structure as a state encoder. Layer nodes The embedding representation update formula is expressed as follows:

[0063] ;

[0064] In the formula, For the first Layer nodes The embedding representation vector, For nodes The initial feature vector, For nodes The set of neighbor node indices The number of neighboring nodes. This represents a vector concatenation operation. For the first The layer can learn linear transformation matrices. For the first Layer bias vector, It is a non-linear activation function that performs three aggregation operations, namely... During the message propagation phase, link latency and link bandwidth are encoded as attention weights, and the edges... Attention weights The formula is calculated using the bidirectional attention propagation mechanism based on edge features, and is expressed as follows:

[0065] ;

[0066] In the formula, The edge feature scoring function is given by a linear weighted form. ,in , For learnable parameters, This serves as the reference value for link delay normalization. This serves as the reference value for normalized link bandwidth. and All are dimensionless quantities with a dimension of 1. The dimensionless score value has a dimension of 1. The attention weights are of dimensionless 1. When congestion is detected in a subgraph, the message propagation depth is dynamically increased by 2 layers to expand the perception range. Node embeddings are input as state vectors into the actor-critic reinforcement learning framework, and the actor network outputs the probability distribution of each candidate distribution path. ,in To distribute action vectors, The current system state observation vector is constructed by concatenating the embedding representations of each node; the commentator network estimates the long-term returns. Long-term returns are defined as:

[0067] ;

[0068] In the formula, For long-term returns with a dimension of 1 This represents the average total system latency within the future time window. This is the normalized reference value for the total system delay; both have the same dimensions. The average packet loss rate. This serves as a normalized baseline for packet loss rate; both values ​​have the same dimensions. , The weighting coefficients and The negative sign indicates that the lower the latency and packet loss rate, the greater the benefit. The physical constraint layer embeds the upper limit of bandwidth for each link as a hard constraint into a normalized exponential function mask, and applies it to inactive actions. The corresponding component is subjected to a maximum negative bias. After being transformed by the normalized exponential function, it approaches zero, as expressed by the following formula:

[0069] ;

[0070] In the formula, For actors online The raw output components of each action. The set of possible actions to satisfy physical bandwidth constraints. This is an indicator function; it takes the value 1 when the condition within the parentheses is true, and 0 otherwise. For maximum negative bias, the empirical value is This forces the probability of non-action to zero. For the first Each candidate distribution action is performed. Simultaneously, a multi-level node importance evaluation and distribution priority ranking algorithm based on random walk restart is executed using a directed weighted graph. The link weights of the directed weighted graph comprehensively encode the inter-node link bandwidth. Link delay Compared with the overall load value ,side Comprehensive link weight The formula is expressed as follows:

[0071] ;

[0072] In the formula, The dimensionless composite link weight has a dimension of 1. and All are dimensionless quantities with a dimension of 1. For nodes The overall load value, with a dimension of 1. This is a dimensionless comprehensive load adjustment factor with a dimension of 1. The algorithm uses the live source nodes as the initial distribution and the initial time setting... ,in for In a dimensional column vector, the component corresponding to the live source node is 1, and the other components are 0. The value at the corresponding position of the live stream source node is 1. For directed weighted graphs The total number of nodes. Perform a random walk with restart on the directed weighted graph, iteratively updating the importance score vector of each node. The formula is expressed as follows:

[0073] ;

[0074] In the formula, For the first The importance score vector of each node in the next iteration. For the probability of restarting, for The dimensional transition probability matrix is ​​expressed as follows:

[0075] ;

[0076] In the formula, When node With nodes When there is no directed edge between them , The sum of each row is 1. Iterate until... Convergence The convergence threshold is, empirically, [value]. , The algorithm represents the L2 norm of a vector, and its overall time complexity is O(n log n). ,in The number of iterations required for convergence, The number of edges in a directed weighted graph typically converges within 20–50 steps. Restart probability. Through experiments, it was obtained that in a simulated private network environment, the following were taken respectively: The optimal values ​​were set to 0.10, 0.15, 0.20, and 0.25, respectively, and were determined through multiple rounds of iterative experiments, using convergence speed and the stability of node importance score ranking as evaluation metrics. This was based on the probability distribution of the distribution path. Importance score vector of each node The convergence results are used to prioritize the distribution paths, and high bitrate video streams are preferentially distributed to secondary distribution nodes with lower overall load values ​​via paths with high importance scores.

[0077] The specific implementation of step S04 is as follows: The system control and management module calculates the comprehensive load value of each node based on the collected central processing unit load rate, network bandwidth utilization rate, and number of connected terminals. The formula is expressed as follows:

[0078] ;

[0079] In the formula, This is the overall load value, with a dimension of 1. The current CPU load rate, This is the baseline value for the central processing unit at full load; both have the same dimensions. This represents the current network bandwidth utilization. This is the baseline value for full-load link bandwidth utilization; both have the same unit of measurement. This represents the number of currently connected terminals. This is a baseline value for the maximum number of connected terminals per node; both have the same unit of measurement. , , The weighting coefficients are dimensionless and The weighting coefficients were obtained experimentally: multiple rounds of stress testing were conducted in a simulated private network environment using 8K high-bitrate video streams, with enumeration steps of 0.1. , , The optimal weight combination is determined through at least 30 rounds of iterative experiments, using a weighted average of total system latency and packet loss rate as the optimization objective. The distribution path adjustment function is based on... The parameters for adjusting the path allocation strategy within the specified interval: when When the probability distribution of the current distribution path remains unchanged; when When this happens, the probability weights of low-importance scoring paths are shifted to high-importance scoring paths, with a shift ratio of [percentage missing]. The calculation is performed using linear interpolation, and the formula is expressed as follows:

[0080] ;

[0081] In the formula, The probability weight transfer ratio has a dimension of 1. The molecule is the overall loading value with dimensionless weight of 1. With denominator Both have the same dimension, which is 1. When When the secondary distribution node requests a new high-bitrate video stream replica connection, it simultaneously compresses the normalized exponential function mask weights of low-bandwidth paths in the physical constraint layer; when The fault switching process is triggered in time. The boundary values ​​of 0.4, 0.7, and 0.9 for each interval were obtained experimentally: stress tests were conducted in a simulation environment with different concurrency scales, and the comprehensive load value and system service quality degradation curve were collected. The boundaries of each interval were determined by piecewise linear fitting and inflection point analysis, and the stability of the boundaries was verified through no less than 20 rounds of iterative experiments.

[0082] The specific implementation of step S05 is as follows: the secondary distribution node obtains a high bitrate video stream from the primary forwarding node and pushes it concurrently to the terminal devices in the area using an unlocked circular buffer. The terminal devices maintain a connection with the secondary distribution node through a heartbeat mechanism, receive the video stream, and decode and play it. This step is a routine concurrent streaming and heartbeat maintenance operation.

[0083] The specific implementation of step S06 is as follows: The fault detection module monitors the status of each node in real time by combining active heartbeat detection with passive response time monitoring. When an abnormality is detected in the secondary distribution node or the overall load value... In this case, the system control and management module will migrate the affected terminal devices to the backup node within seconds. The preset fault threshold is consistent with the boundary of the comprehensive load value range, with a value of 0.9. The acquisition method is the same as the range boundary value of 0.9, ensuring the continuity of high bitrate video stream playback.

[0084] The steps for establishing the training dataset for a dedicated network topology-aware load balancing network are as follows: A simulated network environment with a topology similar to that of a sports event dedicated network is built in the laboratory. Pressure is injected with different numbers of concurrent terminals, different node load combinations, and different link bandwidth conditions, and node feature vectors are collected at each time point. Edge feature vectors Distribution action The system's total latency and packet loss rate constitute a state-action-reward triplet sample set. Samples from node fault injection and sudden traffic scenarios are also collected, totaling no fewer than 50,000 samples. These samples are divided into training, validation, and test sets in an 8:1:1 ratio. The training steps are as follows: online training is performed using the actor-critic algorithm framework. The near-end policy optimization algorithm updates the actor network parameters. The near-end policy optimization algorithm introduces a truncation ratio constraint to limit the offset range of the probability ratio between the old and new policies. Its objective function formula is as follows:

[0085] ;

[0086] In the formula, Optimize the objective function value for a proximal policy with a dimension of 1. The ratio of the probabilities of the new and old strategies, with a dimension of 1. For the current policy in state Next action The probability, Perform actions for the old strategy under the same conditions. The probability, For the first System state observation vector at any given time. For the first Action vectors are distributed at each time step. The value is the estimate of the dominance function with dimension 1. For the first Real-time benefits For the commentator network on the state Long-term return estimates, For the current actor's network parameters, These are the actor network parameters from before the last update. To truncate the hyperparameter, an empirical value of 0.2 is used. This is a truncation function. Let be the expectation operator. The commentator network uses the mean squared error loss function to optimize the long-term return estimation, as expressed in the following formula:

[0087] ;

[0088] In the formula, The value of the commentator network loss function is 1 in dimension. and All have the same dimension, namely 1. The network parameters are used for commentators. The network parameters are incrementally fine-tuned at fixed time windows using the latest link state data. The weighted value of the total system latency and packet loss rate on the validation set is used as the early stopping criterion. The convergence criterion is that the increase in revenue is less than a preset convergence threshold within 10 consecutive validation periods. The preset convergence threshold is determined through multiple rounds of experimental iterations.

[0089] To better understand and implement this invention, the following is a specific application scenario example 2: Technicians built a hierarchical distribution test environment for 8K high-bitrate panoramic live streaming on a dedicated sports event network. This environment adopted a three-layer hierarchical topology structure consisting of one live streaming source node, two primary forwarding nodes, six secondary distribution nodes, and 120 terminal devices. All nodes were interconnected via 10 Gigabit Ethernet switches. The maximum transmission unit across the entire dedicated network link was uniformly configured to 9000 bytes, and each node enabled the user-space protocol stack of the data plane development kit. The live streaming source node inputs raw 8K panoramic video with a resolution of 7680×4320. After encoding using a high-efficiency video coding standard, a high-bitrate video stream with a bitrate of 320Mbps is generated and pushed to the two primary forwarding nodes via a real-time transmission protocol.

[0090] In step S01, the encoding module performs inter-frame prediction and entropy encoding on the original 8K panoramic video to generate a high bitrate video stream with a stable bitrate of 320Mbps. The stream push module encapsulates the video stream into real-time transmission protocol data packets and pushes them to the first-level forwarding node via a private network link. The end-to-end push latency is stable within 8 milliseconds.

[0091] In step S02, the lock-free circular buffer of the first-level forwarding node uses compare-and-swap atomic instructions to advance the read and write pointers, and performs CPU cache line alignment and filling on the read and write pointers of each thread to eliminate false sharing; the jumbo frame configuration enables the data packets of the 320Mbps video stream to be forwarded between private network links without fragmentation, and the user-space protocol stack of the data plane development kit completes data forwarding in a zero-copy large packet forwarding mode. Under the condition of 120 terminals concurrently pulling the stream, the CPU utilization of the first-level forwarding node is shown in Table 1:

[0092] Table 1. Comparison of CPU utilization rates of first-level forwarding nodes under different maximum transmission units and forwarding methods.

[0093]

[0094] As shown in Table 1, the collaborative configuration of jumbo frames and the user-space protocol stack of the data plane development kit plays a role in eliminating fragmentation and reassembly overhead and kernel-mode switching overhead. The CPU utilization rate is significantly reduced compared to the traditional configuration of standard frames plus kernel protocol stack, thus verifying the effectiveness of the forwarding optimization mechanism in step S02.

[0095] In step S03, the system control and management module continuously collects node feature vectors (CPU load rate, network bandwidth utilization, number of connected terminals, and buffer queue depth) and link latency and bandwidth of six secondary distribution nodes at a 500-millisecond acquisition cycle. A directed weighted graph is constructed and input into the private network topology-aware load balancing network. The GraphSAGE state encoder extracts the embeddings of each node through three-layer neighborhood aggregation, and the bidirectional attention propagation mechanism of edge features enables each node embedding to simultaneously perceive upstream and downstream pressure. The actor network outputs the probability distribution of each candidate distribution path, and the physical constraint layer forces the probability of distribution actions exceeding the link bandwidth limit to zero through a normalized exponential function mask. A multi-level node importance assessment and distribution priority ranking algorithm based on random walk restart uses restart probability... A random walk was performed on the directed weighted graph, and after 32 iterations, the graph converged. The importance scores of each secondary distribution node are shown in Table 2.

[0096] Table 2 Comparison of Importance Score and Overall Load Value of Each Secondary Distribution Node

[0097]

[0098] As shown in Table 2, there is a clear negative correlation between importance score and overall load value. Nodes with high importance scores are closer to the live stream source node, have strong link transmission capabilities, and currently have low overall load values. Based on this, the algorithm prioritizes the allocation of high bitrate video streams to nodes 1 and 2, demonstrating the rationality of the priority ranking mechanism based on random walk restart. Figure 3 As shown, the importance score of each node under different numbers of concurrent terminals adaptively adjusts with the dynamic update of the weights of the directed weighted graph links, verifying the dynamic adaptability of the algorithm.

[0099] In step S04, using weighting coefficients , , (Determined through no fewer than 30 rounds of simulated stress testing experiments) Calculate the comprehensive load value of each node. Taking node 6 as an example, when the number of concurrent terminals reaches its peak, The value rose to 0.91, triggering the distribution path adjustment function. During the interval, the system control and management module immediately initiated the fault switching process, redirecting the terminal devices under the responsibility of node 6 to node 1, which has the lowest overall load value. The entire migration process was completed within 1.2 seconds, and the terminal devices did not experience any video stream interruption.

[0100] In step S05, each of the six secondary distribution nodes concurrently pushes high-bitrate video streams to the terminal devices in its respective area through an unlocked circular buffer. Each terminal device maintains a long connection with its respective secondary distribution node with a heartbeat interval of 3 seconds. After receiving the video stream, the local decoding module completes the real-time decoding and playback of the 8K panoramic video.

[0101] In step S06, the fault detection module actively detects the heartbeats of the six secondary distribution nodes at 1-second intervals, while continuously monitoring the response latency of each node (with a preset upper limit of 2 seconds). During the test, a sudden failure of node 3 is simulated. After three consecutive heartbeat timeouts (i.e., within 3 seconds), the fault detection module triggers a failover. The system control and management module redirects the 19 terminal devices managed by node 3 to node 1, which has a comprehensive load value of 0.28. The total migration time is 1.6 seconds, and the affected terminal devices resume high-bitrate video streaming playback. There is no record of video stream interruption throughout the process. Figure 2 As shown, before and after node fault injection, the video stream reception delay curve of the affected terminal device quickly returned to normal level after the migration was completed, demonstrating the second-level response capability of the fault switching mechanism.

[0102] Compared to traditional kernel protocol stack forwarding and static routing load balancing schemes, this invention achieves fundamental progress in the following technical aspects: At the forwarding layer, jumbo frames eliminate the computational overhead of fragmentation and reassembly at the data link layer; the user-space protocol stack of the data plane development kit bypasses kernel-space switching; and the cache line alignment and filling of the lockless ring buffer eliminates false sharing. The synergy of these three features frees up the central processing unit resources of the primary forwarding node from heavy protocol processing work, enabling it to support higher-concurrency video stream forwarding requirements. At the scheduling layer, GraphSAGE expands load balancing decisions from local information to global topology awareness through topology-structured modeling; the actor-critic framework incorporates long-term system stability into the optimization objective; the physical constraint layer ensures the engineering feasibility of distribution decisions; and the importance assessment algorithm based on random walk restart dynamically prioritizes distribution paths from a global perspective, ensuring that high-bitrate video streams are always transmitted along the optimal path. At the fault recovery layer, the dual mechanism of active heartbeat and passive response time monitoring compresses fault perception time to the second level; and the real-time calculation of comprehensive load values ​​enables the system to trigger preventive migration in advance before node overload, fundamentally eliminating the disruption to video playback continuity caused by minute-level fault response in the traditional manual inspection mode.

[0103] It should be noted that the variables involved in this invention are explained in detail in Tables 3 and 4.

[0104] Table 3. Variable Explanation Table (Part 1)

[0105]

[0106] Table 4. Variable Explanation Table (Part Two)

[0107]

[0108] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A hierarchical distribution method for 8K high bitrate panoramic live streaming on a dedicated sports event network, characterized in that, Includes the following steps: The live source node receives raw 8K panoramic video data, compresses and encodes it using an efficient video encoding standard through the encoding module, generates a high bitrate video stream, and pushes the high bitrate video stream to the first-level forwarding node through the stream push module via a real-time transmission protocol. The first-level forwarding node receives high-bitrate video streams through the stream receiving module, stores the high-bitrate video streams in an unlocked circular buffer, and uniformly configures the maximum transmission unit of the entire private network link to 9000 bytes. The user-space protocol stack of the data plane development kit is enabled on the first-level forwarding node, and the high-bitrate video streams are forwarded to the second-level distribution node through zero-copy large packet forwarding. The system control and management module periodically collects the CPU load rate, network bandwidth utilization rate, number of connected terminals, and buffer queue depth of each node. It constructs a directed weighted graph based on the connection relationship between the nodes of the private network, using the CPU load rate, network bandwidth utilization rate, number of connected terminals, and buffer queue depth as node feature vectors, and link latency and link bandwidth as edge features. This is input into the private network topology-aware load balancing network, and outputs the probability distribution of distribution paths for each node. Simultaneously, it executes a multi-level node importance assessment and distribution priority ranking algorithm based on random walk restart using the directed weighted graph, outputting the importance score of each node. Based on the distribution path probability distribution and the importance score of each node, the distribution paths are prioritized, and high-bitrate video streams are preferentially distributed to secondary distribution nodes with lower overall load values ​​via paths with high importance scores. The system control and management module calculates the comprehensive load value of each node based on the collected central processing unit load rate, network bandwidth utilization rate, and number of connected terminals. It then inputs the comprehensive load value into the distribution path adjustment function and adjusts the path allocation strategy parameters of the private network topology-aware load balancing network according to the range to which the comprehensive load value belongs, thereby achieving dynamic load balancing. The secondary distribution node obtains the high bitrate video stream from the primary forwarding node and pushes the high bitrate video stream concurrently to the terminal devices in the area using an unlocked circular buffer. The terminal devices maintain a connection with the secondary distribution node through a heartbeat mechanism, receive the high bitrate video stream, and decode and play it. The fault detection module monitors the status of each node in real time by combining active heartbeat detection with passive response time monitoring. When an abnormality is detected in a secondary distribution node or the overall load value exceeds the preset fault threshold, the system control and management module will migrate the affected terminal devices to the backup node within seconds to ensure the continuity of high bitrate video stream playback.

2. The tiered distribution method for 8K high bitrate panoramic live streaming for sports event private networks according to claim 1, characterized in that, The method of storing high-bitrate video streams into a lock-free circular buffer specifically refers to using a circular memory queue structure that implements multi-threaded concurrent read and write operations by comparing and swapping atomic instructions. The advancement of read and write pointers is completed through atomic operations, and the read and write pointers of each thread are aligned and filled with CPU cache lines, so that each pointer occupies a different cache line, thereby eliminating the additional cache synchronization overhead caused by false sharing.

3. The tiered distribution method for 8K high bitrate panoramic live streaming for sports event private networks according to claim 2, characterized in that, The aforementioned configuration of the maximum transmission unit (MTU) across the entire private network link to 9000 bytes specifically refers to enabling jumbo frames to eliminate the network layer fragmentation requirement caused by inconsistent maximum MTUs when transmitting large data packets of 8K high bitrate video streams between private network links, thus avoiding the overhead of central processing unit (CPU) for fragmentation and reassembly at the receiving end. The user-space protocol stack of the data plane development kit bypasses the operating system kernel network protocol stack and directly drives the network card in user space to complete data transmission and reception.

4. The tiered distribution method for 8K high bitrate panoramic live streaming for sports event private networks according to claim 3, characterized in that, The dedicated network topology-aware load balancing network specifically refers to the use of an improved GraphSAGE structure as a state encoder, which extracts the node embedding representation of each node through a three-layer neighborhood aggregation operation. During each aggregation layer, the node feature vectors of neighboring nodes are averaged and pooled, and then concatenated with the node feature vector of the current node. After linear transformation and nonlinear activation, a new node embedding representation is obtained.

5. The tiered distribution method for 8K high bitrate panoramic live streaming for sports event private networks according to claim 4, characterized in that, The dedicated network topology-aware load balancing network encodes link latency and link bandwidth as attention weights during the message propagation phase. Through the edge feature bidirectional attention propagation mechanism, each node's node embedding representation simultaneously perceives upstream distribution pressure and downstream pull pressure. When congestion features are detected in a subgraph, the message propagation depth is dynamically increased by 2 layers to expand the perception range.

6. The tiered distribution method for 8K high bitrate panoramic live streaming for sports event private networks according to claim 5, characterized in that, The dedicated network topology-aware load balancing network inputs node embedding representations as state vectors into the actor-critic reinforcement learning framework. The actor network outputs the distribution path probability distribution of each candidate distribution path, and the critic network estimates the long-term benefit of the current strategy within the future time window. The long-term benefit is defined as the weighted negative of the total system latency and packet loss rate. An additional physical constraint layer is incorporated, embedding the upper limit of bandwidth for each link as a hard constraint into the normalized exponential function mask of the actor's network strategy output.

7. The method for hierarchical distribution of 8K high bitrate panoramic live streaming for sports event private networks according to claim 6, characterized in that, The training of the dedicated network topology-aware load balancing network specifically refers to injecting pressure in a simulated network environment with different numbers of concurrent terminals, different node load combinations, and different link bandwidth conditions, collecting a state-action-reward triplet sample set, collecting no less than 50,000 samples, dividing the training set, validation set, and test set in an 8:1:1 ratio, updating the actor network parameters with a near-end policy optimization algorithm, and optimizing the long-term revenue estimation of the critic network using the mean squared error loss function.

8. The hierarchical distribution method for 8K high bitrate panoramic live streaming for sports event private networks according to claim 7, characterized in that, The link weights of the directed weighted graph comprehensively encode the link bandwidth, link delay, and comprehensive load value between nodes. Every fixed time window, the link weights of the directed weighted graph are updated based on the latest link bandwidth measurement value, and the multi-level node importance assessment and distribution priority sorting algorithm based on random walk restart is re-executed to achieve dynamic priority adaptive adjustment.

9. The hierarchical distribution method for 8K high bitrate panoramic live streaming for sports event private networks according to claim 8, characterized in that, The fault detection module monitors the status of each node in real time by combining active heartbeat detection with passive response time monitoring. The preset fault threshold is 0.

9. When the comprehensive load value... The system control and management module triggers a fault switching process, migrating the affected terminal devices to the backup node within seconds.

10. The tiered distribution method for 8K high bitrate panoramic live streaming for sports event private networks according to claim 9, characterized in that, The aforementioned proximal policy optimization algorithm specifically refers to introducing a truncation ratio constraint during policy updates to limit the offset range of the probability ratio between the old and new policies during each parameter update, thereby preventing the policy update step from being too large and causing training instability.

Images