Flow jump buffer pool optimization method, medium and system

Abstract

The invention provides a flow jump buffer pool optimization method, a medium and a system, belongs to the technical field of flow jump buffer pool optimization, and predicts a switching behavior of a user among a plurality of content distribution network nodes by constructing an embedded neural network model. And based on the switching prediction vector, establishing a multi-objective optimization model, and solving by adopting a particle algorithm to obtain the optimal buffer pool size by taking the minimization of the traffic buffer pool size and the maximization of the user experience index as objectives. And the system dynamically adjusts the traffic buffer among the network nodes according to the size of the optimal buffer pool, monitors user experience indexes in real time, and periodically updates the neural network model to realize the adaptive optimization of the size of the traffic buffer pool. According to the method, the comprehensive prediction model is constructed, the multi-objective optimization mechanism is established, and the adaptive online optimization is realized, so that the problem that the adaptive online optimization of the traffic jump buffer pool is difficult to realize in the prior art is effectively solved.

Description

Technical Field

[0001] The present invention belongs to the technical field of optimizing traffic jump buffer pools. Specifically, it relates to a method, medium, and system for optimizing traffic jump buffer pools. Background Art

[0002] Content Delivery Network (CDN) is one of the important infrastructures of current Internet services. Its main function is to deploy cache servers at multiple geographical locations to provide the required content to users nearby, thereby reducing network congestion and improving the user access experience. Typical CDN application scenarios include video on demand, software download, and web page loading. In these applications, user terminals often switch between different CDN nodes, resulting in dynamic changes and bursts in network traffic, which poses great challenges to CDN network management.

[0003] To address this problem, various optimization strategies have been proposed in the prior art. Among them, the most common method is to adopt dynamic resource scheduling based on traffic prediction, that is, according to the prediction of future network traffic, dynamically adjust the cache capacity and bandwidth resources of different nodes to adapt to traffic changes. The core of such methods lies in constructing an accurate traffic prediction model. Common models include time series analysis, machine learning, and deep learning. At the same time, some studies also focus on the prediction and analysis of user switching behavior, using the user's historical switching records to predict the future switching probability, thereby optimizing the scheduling of network resources.

[0004] However, these existing optimization strategies still have some limitations: 1) A single traffic prediction model cannot accurately capture the dynamic change characteristics of network traffic, and the prediction accuracy is limited; 2) The prediction of user switching behavior is too simplistic and cannot comprehensively reflect the user's switching decision-making mechanism; 3) The dynamic adjustment strategy of cache resources is too single and cannot take into account multiple optimization goals, such as minimizing cache costs and maximizing user experience. These problems lead to the fact that the existing CDN optimization methods cannot truly meet the increasingly complex network environment and user needs.

[0005] In summary, the prior art has the problem of being difficult to achieve adaptive online optimization of traffic jump buffer pools. Summary of the Invention

[0006] In view of this, the present invention provides a method, medium, and system for optimizing traffic jump buffer pools, which can solve the problem in the prior art that it is difficult to achieve adaptive online optimization of traffic jump buffer pools.

[0007] The present invention is implemented as follows:

[0008] The first aspect of the present invention provides a method for optimizing a traffic jump buffer pool, which includes the following steps:

[0009] S10. Construct an embedded neural network model, which includes a series of prediction equation sets and a prediction network in series;

[0010] S20. Collect historical traffic data of multiple content distribution network nodes, and obtain the switching records of the user terminal between the multiple content distribution network nodes;

[0011] S30. Preprocess the historical traffic data and extract network traffic feature parameters;

[0012] S40. Input the network traffic feature parameters into the embedded neural network model to obtain a switching prediction vector of the user terminal between the multiple content distribution network nodes;

[0013] S50. Establish a multi-objective optimization model according to the switching prediction vector. The optimization objectives of the multi-objective optimization model include minimizing the size of the traffic buffer pool and maximizing the user experience index;

[0014] S60. Use a particle algorithm to solve the multi-objective optimization model to obtain the optimal buffer pool size;

[0015] S70. Dynamically adjust the traffic buffer pool between the multiple content distribution network nodes according to the optimal buffer pool size;

[0016] S80. Real-time monitor the network experience index of the user terminal, and the network experience index includes video playback smoothness, network latency, and packet loss rate;

[0017] S90. Regularly update the embedded neural network model online according to the network experience index to realize the adaptive optimization of the traffic buffer pool size.

[0018] On the basis of the above technical solutions, a method for optimizing a traffic jump buffer pool of the present invention can be further improved as follows:

[0019] Among them, the prediction equation set includes a traffic prediction equation, a switching probability prediction equation, and a resource scheduling equation.

[0020] Further, the traffic prediction equation is used to predict the network traffic at a future moment, and is specifically expressed as follows:

[0021]

[0022] In the formula, F(t) is the predicted traffic value at time t, with the unit of Mbps; f i(t-i) is the historical traffic value at the previous i moments, with the unit of Mbps; n is the size of the historical data window, and the value range is from 10 to 100; α1, α2, α3 are weight coefficients, obtained by least squares fitting; ω is the angular frequency of the periodic change of traffic, with the unit of rad / s; φ is the phase offset, with the unit of rad; λ is the attenuation coefficient, with the unit of 1 / s; ∈ f is the prediction error term.

[0023] Parameter acquisition method:

[0024] 1. The historical traffic value f i (t-i) is obtained by deploying traffic monitoring probes at the content distribution network nodes;

[0025] 2. The periodic parameters ω and φ are obtained by performing Fourier transform on the historical traffic data:

[0026]

[0027] Furthermore, the handover probability prediction equation is used to predict the probability that the user terminal will hand over to other nodes, and is specifically expressed as follows:

[0028]

[0029] In the formula, P(s) is the handover probability; L is the current node load rate, and the value range is from 0 to 1; D is the network delay, with the unit of ms; Q is the quality of service index; β1, β2, β3 are weight coefficients; ∈ p is the prediction error term.

[0030] Parameter acquisition method:

[0031] 1. The calculation formula for the load rate L:

[0032] In the formula, C used is the used bandwidth, and C total is the total bandwidth;

[0033] 2. The network delay D is obtained by active probing:

[0034] In the formula, RTT k is the round-trip delay of the kth probe, and m is the number of probes, with the value of 100.

[0035] Furthermore, the resource scheduling equation is used to optimize the allocation of network resources, and is specifically expressed as follows:

[0036]

[0037] In the formula, R(x) is the resource scheduling index; is the flow rate change rate; is the flow acceleration; γ1, γ2, γ3 are weight coefficients; ∈ r is the scheduling error term.

[0038] Furthermore, the prediction network adopts a lightweight convolutional neural network, with the input being the output of the prediction equation set and the output being the node switching vector.

[0039] Furthermore, the prediction network contains 3 convolutional layers and 2 fully connected layers; optionally, the prediction network structure formula is described as follows:

[0040] Conv1: kernel_size = 3, channels = 16;

[0041] Conv2: kernel_size = 3, channels = 32;

[0042] Conv3: kernel_size = 3, channels = 64;

[0043] FC1: neurons = 128;

[0044] FC2: neurons = num_nodes.

[0045] Conv1~3 represent three convolutional layers; FC1~2 represent two fully connected layers.

[0046] Furthermore, the network traffic characteristic parameters include traffic volume, traffic duration, and traffic direction.

[0047] Furthermore, the optimization objectives of the multi-objective optimization model include minimizing the size of the traffic buffer pool and maximizing the user experience index, which are specifically expressed as follows:

[0048] min{Z1(B), Z2(B)};

[0049] The first optimization objective is to minimize the total size of the traffic buffer pool:

[0050] The second optimization objective is to maximize the reciprocal of the user experience index:

[0051] Constraint conditions:

[0052] 1) Buffer pool size constraint: 0 ≤ B i ≤ B max , i = 1, 2,..., N;

[0053] 2) Traffic constraint: F(t i ) ≤ Ci , i = 1, 2, ..., N;

[0054] 3) Handoff probability constraint: 0 ≤ P(s i ) ≤ 1, i = 1, 2, ..., N;

[0055] 4) Resource scheduling constraint: R min ≤ r(x i ) ≤ R max , i = 1, 2, ..., N;

[0056] Among them, the calculation formula of the user experience index:

[0057]

[0058] In the formula, B i is the buffer pool size of the i-th node, in MB; N is the total number of nodes; F(t i ) is the predicted traffic value of the i-th node by the traffic prediction equation at time t, in Mbps; P(s i ) is the predicted handoff probability value of the i-th node by the handoff probability prediction equation; R(x i ) is the scheduling index of the i-th node by the resource scheduling equation; QoE i is the user experience index; B max is the maximum buffer pool size, with a value of 1000 MB; C i is the bandwidth capacity of the i-th node, in Mbps; T i is the handoff delay, in ms; R min is the minimum resource scheduling index, with a value of 0; R max is the maximum resource scheduling index, with a value of 1; μ1, μ2, μ3 are the weight coefficients of the user experience index, satisfying μ1 + μ2 + μ3 = 1.

[0059] Solution method:

[0060] 1) Construct the Pareto optimal solution set:

[0061]

[0062] 2) Solve using the multi-objective particle swarm optimization algorithm:

[0063] Particle update formula:

[0064]

[0065] In the formula, is the velocity of the i-th particle at the k-th iteration; is the position of the i-th particle at the k-th iteration; is the historical optimal position of the i-th particle; g k is the global optimal position; ω is the inertia weight, with a value of 0.7; c1 and c2 are acceleration constants, both with a value of 2; r1 and r2 are random numbers within the interval [0, 1].

[0066] Algorithm parameter settings:

[0067] particle_number = 200;

[0068] iteration_number = 500;

[0069] ω max = 0.9, ω min = 0.4.

[0070] 3) Use the fuzzy comprehensive evaluation method to select the optimal solution:

[0071] Construct the membership function:

[0072]

[0073] Comprehensive evaluation index:

[0074] S(B) = λ1μ1(B) + λ2μ2(B);

[0075] In the formula, λ1 and λ2 are evaluation weights, satisfying λ1 + λ2 = 1.

[0076] The finally selected optimal solution is:

[0077] B * = argmax B∈Ω S(B).

[0078] The maximum number of iterations of the particle algorithm is 100 times.

[0079] Specifically, the step S10 specifically includes the following sub-steps: constructing an embedded neural network model, which includes two main parts: a prediction equation set and a prediction network. The prediction equation set is used to establish mathematical models for traffic prediction, handover probability prediction, and resource scheduling, and the prediction network is used to integrate these prediction results into a final node handover prediction vector. The prediction equation set includes three sub-equations: a traffic prediction equation, a handover probability prediction equation, and a resource scheduling equation. The traffic prediction equation comprehensively considers historical traffic data, periodic change rules, and exponential decay characteristics, and uses the least squares method to fit and obtain various parameters; the handover probability prediction equation considers the load rate, network delay, and quality of service indicators of the current node, and models the user's handover behavior through the Logistic function; the resource scheduling equation comprehensively considers the change rate, acceleration, and handover probability of traffic, and gives a resource scheduling indicator. The prediction network adopts a lightweight convolutional neural network structure, with the outputs of the above three prediction equations as the input and the node handover prediction vector as the output.

[0080] Specifically, the step S20 includes the following sub-steps: deploying traffic monitoring probes at key nodes of the content delivery network to collect network traffic data in real time and record the handover situation of user terminals between different nodes. These historical data will be used as the basis for subsequent prediction modeling and optimization decisions.

[0081] Specifically, the step S30 includes the following sub-steps: preprocessing the collected historical traffic data to extract three major characteristic parameters of network traffic, namely traffic volume, traffic duration, and traffic direction. The traffic volume can be directly obtained from the monitoring data, the traffic duration can be calculated by statistically analyzing the traffic changes between two adjacent time points, and the traffic direction needs to be judged in combination with the handover records of user terminals. These three characteristic parameters can comprehensively describe the dynamic change characteristics of network traffic and lay a foundation for subsequent prediction modeling.

[0082] Specifically, the step S40 includes the following sub-steps: inputting the extracted network traffic characteristic parameters into the previously constructed embedded neural network model to obtain a handover prediction vector of the user terminal between multiple content delivery network nodes. Specifically, first, the outputs of the traffic prediction equation, the handover probability prediction equation, and the resource scheduling equation are used as the input of the prediction network, and then after being processed by 3 convolutional layers and 2 fully connected layers, a node handover probability vector is finally output. This handover prediction vector provides a key input for subsequent multi-objective optimization.

[0083] Specifically, the step S50 includes the following sub-steps: Based on the handover prediction vector obtained in step S40, a multi-objective optimization model is established. The optimization objectives of this model include minimizing the total size of the traffic buffer pool and maximizing the user experience metric. The constraint conditions include buffer pool size constraint, traffic constraint, handover probability constraint, and resource scheduling constraint. The calculation formula of the user experience metric takes into account buffer delay, handover probability, and resource scheduling factors.

[0084] Specifically, the step S60 includes the following sub-steps: The multi-objective particle swarm optimization algorithm (MOPSO) is used to solve the multi-objective optimization model established in step S50 to obtain the optimal buffer pool size. The specific process of the MOPSO algorithm includes initializing the particle swarm, iterative optimization, constructing the Pareto optimal solution set, and finally using the fuzzy comprehensive evaluation method to select the optimal solution. This optimization algorithm can find a balance between minimizing the total buffer pool size and maximizing the user experience metric to obtain the optimal buffer pool size.

[0085] Specifically, the step S70 includes the following sub-steps: Based on the optimal buffer pool size obtained in step S60, a traffic buffer pool is set on each content delivery network node and its size is dynamically adjusted. When the user switches between different nodes, the network will dynamically allocate buffer resources according to the predicted handover probability to ensure the user experience to the greatest extent. This dynamically adjusted buffer pool mechanism can effectively avoid problems such as playback stuttering caused by insufficient buffer resources.

[0086] Specifically, the step S80 includes the following sub-steps: The network experience metrics of the user terminal are monitored in real time, including video playback smoothness, network latency, and packet loss rate, etc. These metrics can be obtained by deploying monitoring probes on the user terminal or by combining user feedback information. The monitored network experience metrics will be used as the basis for the online update of the embedded neural network model in the subsequent step S90.

[0087] Specifically, the step S90 includes the following sub-steps: Based on the network experience metrics obtained in step S80, the embedded neural network model constructed previously is periodically updated online using the method of reinforcement learning, so that the prediction results can better reflect the actual network conditions, thereby guiding the dynamic adjustment of the buffer pool size more accurately. This adaptive optimization mechanism enables the entire system to continuously learn and improve during operation, adapt to changes in the network environment, and provide a better service experience for users.

[0088] The second aspect of the present invention provides a computer-readable storage medium, wherein program instructions are stored in the computer-readable storage medium, and when the program instructions run on a computer, they are used to execute the above-mentioned traffic jump buffer pool optimization method.

[0089] The third aspect of the present invention provides a traffic jump buffer pool optimization system, which includes the above-mentioned computer-readable storage medium.

[0090] Compared with the prior art, the beneficial effects of a traffic jump buffer pool optimization method, medium and system provided by the present invention are as follows:

[0091] 1. Stronger prediction ability: By constructing an embedded neural network model including a prediction equation set and a prediction network, it can comprehensively capture the dynamic change characteristics of network traffic, including periodicity, exponential decay, and burst fluctuations, etc., and the prediction accuracy is greatly improved. At the same time, the model can also accurately predict the switching probability of users between different CDN nodes, providing a key basis for subsequent resource optimization.

[0092] 2. More comprehensive optimization objectives: The multi-objective optimization model of the present invention simultaneously considers two objectives of minimizing the total buffer pool size and maximizing the user experience index, and finds the optimal buffer pool size allocation scheme under the premise of meeting various constraint conditions. This optimization strategy that balances network costs and user satisfaction can better meet the needs of practical applications.

[0093] 3. Stronger adaptability: The solution of the present invention has the functions of real-time monitoring and online update, and can dynamically adjust the parameters of the embedded neural network model according to the observed user experience index, so that its prediction results can better reflect the actual network conditions. This adaptive optimization mechanism enables the entire system to continuously learn and improve as the environment changes, providing a more stable and reliable service for users.

[0094] In summary, the traffic jump buffer pool optimization method based on an embedded neural network proposed by the present invention effectively solves the problem in the prior art that it is difficult to achieve adaptive online optimization of the traffic jump buffer pool by constructing a comprehensive prediction model, establishing a multi-objective optimization mechanism, and implementing adaptive online optimization. BRIEF DESCRIPTION OF THE DRAWINGS

[0095] Figure 1 is a flowchart of the method provided by the present invention;

[0096] Figure 2 is a comparison chart of buffer pool sizes;

[0097] Figure 3 is an online optimization performance chart;

[0098] Figure 4 is a user network experience index chart;

[0099] Figure 5 is a network traffic prediction model performance chart. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0100] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the following will clearly and completely describe the technical solutions in the embodiments of the present invention with reference to the accompanying drawings in the embodiments of the present invention.

[0101] As Figure 1 shown, it is a flowchart of an optimized method for a traffic jump buffer pool provided by the first aspect of the present invention. This method includes the following steps:

[0102] S10. Construct an embedded neural network model, where the embedded neural network model includes a series-connected prediction equation set and a prediction network;

[0103] S20. Collect historical traffic data of multiple content distribution network nodes and obtain the switching records of the user terminal between multiple content distribution network nodes;

[0104] S30. Preprocess the historical traffic data and extract network traffic feature parameters;

[0105] S40. Input the network traffic feature parameters into the embedded neural network model to obtain the switching prediction vector of the user terminal between multiple content distribution network nodes;

[0106] S50. According to the switching prediction vector, establish a multi-objective optimization model, and the optimization objectives of the multi-objective optimization model include minimizing the size of the traffic buffer pool and maximizing the user experience index;

[0107] S60. Use the particle algorithm to solve the multi-objective optimization model to obtain the optimal buffer pool size;

[0108] S70. Dynamically adjust the traffic buffer pool between multiple content distribution network nodes according to the optimal buffer pool size;

[0109] S80. Real-time monitor the network experience index of the user terminal, and the network experience index includes video playback smoothness, network latency, and packet loss rate;

[0110] S90. Regularly update the embedded neural network model online according to the network experience index to achieve adaptive optimization of the traffic buffer pool size.

[0111] The following will describe the specific implementation manners of the above steps in detail:

[0112] The specific implementation manner of step S10 is: construct an embedded neural network model. This model includes two main parts: a prediction equation set and a prediction network. The prediction equation set is used to establish a mathematical model for traffic prediction, switching probability prediction, and resource scheduling, and the prediction network is used to integrate these prediction results into the final node switching prediction vector.

[0113] First, the prediction equation set includes three sub-equations:

[0114] 1) Flow prediction equation: Used to predict the network traffic at future moments. This equation comprehensively considers historical traffic data, periodic change patterns, and exponential decay characteristics, and uses the least squares method to fit the values of various parameters. The historical traffic data is obtained by deploying traffic monitoring probes at content distribution network nodes, and the periodic parameters are calculated by performing a Fourier transform on the historical traffic data. The specific expression of the flow prediction equation is:

[0115]

[0116] where F(t) is the predicted traffic value at time t, f i (t - i) is the historical traffic value at the previous i moments, n is the size of the historical data window, α1, α2, α3 are weight coefficients, ω is the angular frequency of the periodic traffic change, φ is the phase shift, λ is the decay coefficient, ∈ f is the prediction error term.

[0117] 2) Handoff probability prediction equation: Used to predict the probability that a user terminal will hand off to other nodes. This equation considers the load rate of the current node, network latency, and quality of service metrics, and models the handoff behavior of users through a logistic function. The load rate is calculated by the ratio of the used bandwidth to the total bandwidth, and the network latency is the average of the round-trip latencies obtained through active probing. The specific expression of the handoff probability prediction equation is:

[0118]

[0119] where P(s) is the handoff probability, L is the load rate of the current node, D is the network latency, Q is the quality of service metric, β1, β2, β3 are weight coefficients, ∈ p is the prediction error term.

[0120] 3) Resource scheduling equation: Used to optimize the allocation of network resources. This equation comprehensively considers the change rate, acceleration, and handoff probability of traffic, and gives a resource scheduling metric. The specific expression of the resource scheduling equation is:

[0121]

[0122] where R(x) is the resource scheduling metric, is the traffic change rate, is the traffic acceleration, γ1, γ2, γ3 are weight coefficients, ∈ r is the scheduling error term.

[0123] Next is the prediction network, which adopts a lightweight convolutional neural network structure. The input is the output of the above three prediction equations, and the output is the node switching prediction vector. The specific structure includes 3 convolutional layers and 2 fully connected layers. The kernel_size of the convolutional layers are 3, 3, 3 respectively, the channels are 16, 32, 64 respectively, and the neurons of the fully connected layers are 128 and num_nodes (the number of nodes) respectively. The role of this prediction network is to fuse multiple factors such as traffic prediction, handover probability prediction, and resource scheduling metrics, and give the final node switching prediction result.

[0124] Generally speaking, the embedded neural network model constructed in step S10 can comprehensively predict the future network state through the collaborative work of the prediction equation set and the prediction network, providing support for subsequent optimization decisions.

[0125] The specific implementation of step S20 is: collect the historical traffic data of multiple content delivery network nodes and obtain the switching records of user terminals between these nodes. These data will be used as the basis for training and validating the neural network model. Specifically, traffic monitoring probes can be deployed at key nodes of the content delivery network to collect network traffic data in real time and record the switching situations of user terminals between different nodes. These historical data provide the basis for subsequent prediction modeling and optimization decisions.

[0126] The specific implementation of step S30 is: preprocess the collected historical traffic data and extract three major characteristic parameters of network traffic, namely traffic volume, traffic duration, and traffic direction. Among them, the traffic volume can be directly obtained from the monitoring data, with the unit of Mbps; the traffic duration can be calculated by statistically analyzing the traffic changes between two adjacent time points, with the unit of ms; the traffic direction needs to be judged in combination with the switching records of user terminals. If the user switches from node A to node B, the traffic direction at this time point is from A to B. These three characteristic parameters can comprehensively describe the dynamic change characteristics of network traffic and lay the foundation for subsequent prediction modeling.

[0127] The specific implementation of step S40 is: input the extracted network traffic characteristic parameters into the previously constructed embedded neural network model to obtain the switching prediction vector of user terminals between multiple content delivery network nodes. Specifically, first, the outputs of the traffic prediction equation, handover probability prediction equation, and resource scheduling equation are used as the input of the prediction network, and then through the processing of 3 convolutional layers and 2 fully connected layers, finally, a node switching probability vector is output, indicating the possibility of the user switching between each node. This switching prediction vector provides the key input for subsequent multi-objective optimization.

[0128] The specific implementation of step S50 is as follows: Based on the handover prediction vector obtained in step S40, a multi-objective optimization model is established. The optimization objectives of this model include: 1) minimizing the total size of the traffic buffer pool; 2) maximizing the user experience metric. The specific expression of the multi-objective optimization model is as follows:

[0129] min{Z1(B), Z2(B)};

[0130] Where, represents the total buffer pool size, represents the reciprocal of the user experience metric.

[0131] The constraint conditions include:

[0132] 1) Buffer pool size constraint: 0 ≤ B i ≤ B max , i = 1, 2,..., N, where B max takes the value of 1000MB;

[0133] 2) Traffic constraint: F(t i ) ≤ C i , i = 1, 2,..., N, where C i is the bandwidth capacity of the i-th node;

[0134] 3) Handoff probability constraint: 0 ≤ P(s i ) ≤ 1, i = 1, 2,..., N;

[0135] 4) Resource scheduling constraint: R min ≤ R(x i ) ≤ R max , i = 1, 2,..., N, where R min = 0, R max = 1.

[0136] The calculation formula for the user experience metric QoE i is: Where T i is the handover delay, and μ1, μ2, μ3 are weight coefficients, satisfying μ1 + μ2 + μ3 = 1.

[0137] Generally speaking, this multi-objective optimization model attempts to find a Pareto optimal solution set under the premise of satisfying various constraint conditions, that is, minimizing the total buffer pool size and maximizing the user experience metric simultaneously.

[0138] The specific implementation of step S60 is as follows: The multi-objective particle swarm optimization algorithm (MOPSO) is used to solve the multi-objective optimization model established in step S50 to obtain the optimal buffer pool size.

[0139] The specific process of the MOPSO algorithm is as follows:

[0140] 1) Initialize the particle swarm: Set the number of particles to 200, and generate 200 random feasible solutions as the initial particle positions.

[0141] 2) Iterative optimization: Perform 500 iterations of optimization. In each iteration, update the velocity and position of the particles according to the historical best solutions of the particles and the global best solution:

[0142]

[0143] Among them, is the velocity of the i-th particle in the k-th iteration, is the position of the i-th particle in the k-th iteration, is the historical best position of the i-th particle, g k is the global best position, ω is the inertia weight, with a value of 0.7, c1 and c2 are acceleration constants, both with a value of 2, and r1 and r2 are random numbers in the interval [0, 1].

[0144] 3) Construct the Pareto optimal solution set:

[0145]

[0146] 4) Select the optimal solution: Through the fuzzy comprehensive evaluation method, construct the membership functions of the two objective functions, and calculate the comprehensive evaluation index S(B) = λ1μ1(B) + λ2μ2(B), where λ1 and λ2 are weight coefficients, satisfying λ1 + λ2 = 1. Finally, select B * = argmax B∈Ω S(B) as the optimal buffer pool size.

[0147] Generally speaking, step S60 uses the MOPSO algorithm to solve the multi-objective optimization model, seeking a balance between minimizing the total buffer pool size and maximizing the user experience index, and obtaining the optimal buffer pool size.

[0148] The specific implementation method of step S70 is: According to the optimal buffer pool size obtained in step S60, dynamically adjust the traffic buffer pools among multiple content distribution network nodes. Specifically, a traffic buffer pool can be set on each content distribution network node, and the size of the buffer pool is adjusted according to the optimization result of step S60. When the user switches between different nodes, the network will dynamically allocate buffer resources according to the predicted switching probability to maximize the user experience. This dynamically adjusted buffer pool mechanism can effectively avoid problems such as playback jams caused by insufficient buffer resources.

[0149] The specific implementation of step S80 is as follows: Real-time monitor the network experience metrics of the user terminal, including video playback smoothness, network latency, packet loss rate, etc. These metrics can be obtained by deploying monitoring probes on the user terminal or by combining user feedback information. The monitored network experience metrics will serve as the basis for the online update of the embedded neural network model in the subsequent step S90.

[0150] The specific implementation of step S90 is as follows: According to the network experience metrics obtained in step S80, regularly update the previously constructed embedded neural network model online to achieve adaptive optimization of the size of the traffic buffer pool. Specifically, the method of reinforcement learning can be adopted, taking the monitored network experience metrics as reward and punishment signals, and by continuously adjusting the parameters of the neural network model, the prediction results can better reflect the actual network conditions, so as to more accurately guide the dynamic adjustment of the buffer pool size. This adaptive optimization mechanism enables the entire system to continuously learn and improve during operation, adapt to changes in the network environment, and provide a better service experience for users.

[0151] In summary, this traffic jump buffer pool optimization method based on an embedded neural network constructs a prediction equation set and a prediction network, establishes a multi-objective optimization model, and uses the MOPSO algorithm to solve it to obtain the optimal buffer pool size. At the same time, this method also has the ability of real-time monitoring and adaptive optimization, can dynamically adjust the buffer pool size, and maximize the user experience. This comprehensive optimization method can effectively alleviate the congestion problem in the content delivery network and improve the utilization efficiency of network resources.

[0152] The second aspect of the present invention provides a computer-readable storage medium, wherein program instructions are stored in the computer-readable storage medium, and when the program instructions run on a computer, they are used to execute the above-mentioned traffic jump buffer pool optimization method.

[0153] The third aspect of the present invention provides a traffic jump buffer pool optimization system, which includes the above-mentioned computer-readable storage medium.

[0154] Specifically, the principle of the present invention is: construct an embedded neural network model, which includes two main parts: a prediction equation set and a prediction network.

[0155] The prediction equation set includes three sub-equations, which are respectively used for traffic prediction, handover probability prediction, and resource scheduling. These three equations form a closed-loop prediction system, where traffic prediction is the input, handover probability prediction and resource scheduling are the outputs, and there is a coupling relationship between them.

[0156] The traffic prediction equation adopts the form of a hybrid model, comprehensively considering historical traffic data, periodic change patterns, and exponential decay characteristics. Among them, historical traffic data reflects the short-term dynamic changes of traffic, periodic change patterns capture long-term seasonal fluctuations, and exponential decay characteristics characterize the suddenness of traffic. Concentrating these three factors in a prediction equation can better adapt to the complex change characteristics of network traffic.

[0157] The handover probability prediction equation uses the Logistic function to model the node handover behavior of users, considering factors such as the load rate of the current node, network delay, and quality of service. These factors directly affect the handover decision of users, so it is reasonable to use them as input variables for handover probability prediction.

[0158] The resource scheduling equation comprehensively considers the change rate, acceleration of traffic, and handover probability, and gives an index for optimizing the allocation of network resources. This index can be used as the basis for subsequent optimization of the buffer pool size to balance network costs and user experience.

[0159] The neural network part acts as a bridge between the prediction equation set and the optimization decision. It adopts a lightweight convolutional neural network structure, fuses the outputs of traffic prediction, handover probability prediction, and resource scheduling, and gives the final node handover prediction vector. This prediction vector provides a key basis for subsequent multi-objective optimization.

[0160] On this basis, the present invention constructs a multi-objective optimization model, with the objectives including minimizing the total buffer pool size and maximizing the user experience index. Among them, the user experience index comprehensively considers buffer delay, handover probability, and resource scheduling factors, reflecting the quality of service perceived by users. The MOPSO algorithm is used to solve this multi-objective optimization problem to obtain the Pareto optimal solution set, and then the final optimal solution is selected through the fuzzy comprehensive evaluation method, which not only ensures the minimization of network costs but also maximizes user satisfaction.

[0161] In addition, the solution of the present invention also has the ability of adaptive optimization. It can real-time monitor network experience indicators of user terminals, such as video playback smoothness, network delay, and packet loss rate, etc., and perform online updates and parameter adjustments on the embedded neural network model according to these indicators. In this way, the prediction model can continuously learn and improve, enabling its prediction results to better reflect the actual network conditions, thereby guiding the dynamic optimization of the buffer pool size and ultimately improving the user's service experience.

[0162] Generally speaking, the traffic jump buffer pool optimization method based on an embedded neural network proposed by the present invention gives full play to the synergistic effect of the prediction equation set and the neural network, achieving a comprehensive prediction of network traffic, user switching behavior, and resource scheduling requirements. On this basis, a multi-objective optimization model is constructed to seek a balance between minimizing the total buffer pool size and maximizing the user experience index. At the same time, this solution also has the ability of adaptive optimization, which can dynamically adjust the prediction model parameters and continuously optimize the buffer pool size allocation to meet the content distribution requirements in a complex network environment.

[0163] A specific embodiment 1 of the method of the present invention is given below. The specific implementation manners of each step in this embodiment 1 are described in detail as follows: The specific implementation manner of step S10 is: Construct an embedded neural network model. This model includes two main parts: a prediction equation set and a prediction network.

[0164] The prediction equation set includes the following 3 sub-equations:

[0165] 1) The traffic prediction equation F(t): Used to predict the network traffic at a future moment. This equation adopts the form of a hybrid model, comprehensively considering historical traffic data, periodic change rules, and exponential decay characteristics. The specific expression is where F(t) is the predicted traffic value at time t, with the unit of Mbps; f i (t - i) is the historical traffic value at the previous i moments, with the unit of Mbps; n is the size of the historical data window, and the value range is from 10 to 100; α1, α2, α3 are weight coefficients obtained by least squares fitting; ω is the angular frequency of traffic periodic change, with the unit of rad / s; φ is the phase shift, with the unit of rad; λ is the decay coefficient, with the unit of 1 / s; ∈ f is the prediction error term. The historical traffic data f i (t - i) is obtained by deploying traffic monitoring probes at content distribution network nodes. The periodic parameters ω and φ are calculated by performing a Fourier transform on the historical traffic data. The specific formulas are and

[0166] 2) The switching probability prediction equation P(s): Used to predict the probability that the user terminal switches to other nodes. This equation adopts the form of a Logistic function, considering the load rate, network delay, and service quality index of the current node. The specific expression is where P(s) is the switching probability; L is the load rate of the current node, and the value range is from 0 to 1. The calculation formula is where C used is the used bandwidth, and C total is the total bandwidth; D is the network delay, with the unit of ms, obtained by active probing. The calculation formula is Among them, RTT k is the round-trip delay of the k-th detection, m is the number of detections, and its value is 100; Q is the quality of service index; β1, β2, β3 are weight coefficients; ∈ p is the prediction error term.

[0167] 3) Resource scheduling equation R(x): used to optimize the allocation of network resources. This equation comprehensively considers the change rate, acceleration, and handover probability of traffic. The specific expression is where R(x) is the resource scheduling index; is the traffic change rate; is the traffic acceleration; γ1, γ2, γ3 are weight coefficients; ∈ r is the scheduling error term.

[0168] The prediction network adopts a lightweight convolutional neural network structure. The input is the output of the above three prediction equations, and the output is the node handover prediction vector. The specific structure includes 3 convolutional layers and 2 fully connected layers. The kernel_size of the convolutional layers are 3, 3, 3 respectively, the channels are 16, 32, 64 respectively, and the neurons of the fully connected layers are 128 and num_nodes (the number of nodes) respectively.

[0169] Generally speaking, the embedded neural network model constructed in step S10 can comprehensively predict the future network state through the collaborative work of the prediction equation set and the prediction network, providing support for subsequent optimization decisions.

[0170] The specific implementation method of step S20 is: collect the historical traffic data of multiple content distribution network nodes and obtain the handover records of user terminals between these nodes. These data will be used as the basis for training and validating the neural network model. Specifically, traffic monitoring probes can be deployed at key nodes of the content distribution network to collect network traffic data in real time and record the handover situations of user terminals between different nodes. These historical data provide the basis for subsequent prediction modeling and optimization decisions.

[0171] The specific implementation method of step S30 is: preprocess the collected historical traffic data and extract three major characteristic parameters of network traffic, namely traffic size F, traffic duration D, and traffic direction d. Among them, the traffic size F can be directly obtained from the monitoring data, with the unit of Mbps; the traffic duration D can be calculated by statistically analyzing the traffic changes between two adjacent time points, with the unit of ms; the traffic direction d needs to be judged in combination with the handover records of user terminals. If the user switches from node A to node B, the traffic direction at this time point is from A to B. These three characteristic parameters can comprehensively describe the dynamic change characteristics of network traffic and lay the foundation for subsequent prediction modeling.

[0172] The specific implementation of step S40 is as follows: Input the extracted network traffic characteristic parameters F, D, and d into the previously constructed embedded neural network model to obtain the switching prediction vector of the user terminal among multiple content distribution network nodes. Specifically, first, take the outputs of the traffic prediction equation F(t), the switching probability prediction equation P(s), and the resource scheduling equation R(x) as the inputs of the prediction network. Then, after being processed by 3 convolutional layers and 2 fully connected layers, finally output a node switching probability vector. It represents the possibility of the user switching among various nodes. This switching prediction vector provides a key input for subsequent multi-objective optimization.

[0173] The specific implementation of step S50 is as follows: Based on the switching prediction vector obtained in step S40 Establish a multi-objective optimization model. The optimization objectives of this model include: 1) Minimize the total traffic buffer pool size Z1(B); 2) Maximize the reciprocal of the user experience metric Z2(B). The specific expression of the multi-objective optimization model is as follows:

[0174] min{Z1(B), Z2(B)};

[0175] Among them, the total buffer pool size The reciprocal of the user experience metric

[0176] The constraint conditions include:

[0177] 1) Buffer pool size constraint: 0 ≤ B i ≤ B max , i = 1, 2,..., N, where B max takes a value of 1000MB;

[0178] 2) Traffic constraint: F(t i ) ≤ C i , i = 1, 2,..., N, where C i is the bandwidth capacity of the i-th node;

[0179] 3) Switching probability constraint: 0 ≤ P(s i ) ≤ 1, i = 1, 2,..., N;

[0180] 4) Resource scheduling constraint: R min ≤ R(x i ) ≤ R max , i = 1, 2,..., N, where R min = 0, R max = 1.

[0181] User experience metric QoEi The calculation formula is: where T i is the handover delay, and μ1, μ2, μ3 are weight coefficients, satisfying μ1 + μ2 + μ3 = 1.

[0182] Generally speaking, under the premise of meeting various constraint conditions, this multi-objective optimization model attempts to find a Pareto optimal solution set, that is, to minimize the total buffer pool size and maximize the user experience index simultaneously.

[0183] The specific implementation of step S60 is: using the multi-objective particle swarm optimization algorithm (MOPSO) to solve the multi-objective optimization model established in step S50 to obtain the optimal buffer pool size

[0184] The specific process of the MOPSO algorithm is as follows:

[0185] 1) Initialize the particle swarm: Set the number of particles to 200, and generate 200 random feasible solutions as the initial particle positions

[0186] 2) Iterative optimization: Perform 500 iterations of optimization. In each iteration, update the velocity and position of the particle according to the historical best solution of the particle and the global best solution :

[0187]

[0188] where, is the velocity of the i-th particle at the k-th iteration, is the position of the i-th particle at the k-th iteration, is the historical best position of the i-th particle, is the global best position, ω is the inertia weight, with a value of 0.7, c1, c2 are acceleration constants, both with a value of 2, and r1, r2 are random numbers in the interval [0, 1].

[0189] 3) Construct the Pareto optimal solution set:

[0190]

[0191] 4) Select the optimal solution: Through the fuzzy comprehensive evaluation method, construct the membership functions of the two objective functions and Calculate the comprehensive evaluation index S(B) = λ1μ1(B) + λ2μ2(B), where λ1, λ2 are weight coefficients, satisfying λ1 + λ2 = 1. Finally, select B * = argmax B∈ΩLet S(B) be the optimal buffer pool size.

[0192] Generally speaking, step S60 uses the MOPSO algorithm to solve the multi-objective optimization model, seeking a balance between minimizing the total buffer pool size and maximizing the user experience metric, and obtaining the optimal buffer pool size.

[0193] The specific implementation of step S70 is as follows: According to the optimal buffer pool size obtained in step S60 dynamically adjust the traffic buffer pool among multiple content distribution network nodes. Specifically, a traffic buffer pool can be set on each content distribution network node, and the size of the buffer pool is adjusted according to the optimization result of step S60 When a user switches between different nodes, the network dynamically allocates buffer resources according to the predicted switching probability P(s i ) to maximize the user experience to the greatest extent. This dynamically adjusted buffer pool mechanism can effectively avoid problems such as playback stuttering caused by insufficient buffer resources.

[0194] The specific implementation of step S80 is as follows: Real-time monitor the network experience metrics of the user terminal, including video playback smoothness q, network latency D, packet loss rate ρ, etc. These metrics can be obtained by deploying monitoring probes on the user terminal or by combining user feedback information. The monitored network experience metrics will be used as the basis for the online update of the embedded neural network model in the subsequent step S90.

[0195] The specific implementation of step S90 is as follows: According to the network experience metrics q, D, and ρ obtained in step S80, use the method of reinforcement learning to perform regular online updates on the previously constructed embedded neural network model, so that the prediction results can better reflect the actual network conditions, and thus more accurately guide the dynamic adjustment of the buffer pool size. Specifically, the monitored network experience metrics can be used as reward and punishment signals, and by continuously adjusting the parameters of the neural network model, the output switching prediction vector can better match the actually observed user switching behavior. This adaptive optimization mechanism enables the entire system to continuously learn and improve during operation, adapt to changes in the network environment, and provide users with a better service experience.

[0196] To better understand and implement the present invention, the following provides Example 2 of a specific application scenario of the present invention: A content distribution network operator ABC enterprise has deployed several content distribution nodes in multiple geographical regions to provide services such as video on demand and software download for users. In order to optimize the user service experience and reduce network costs, ABC enterprise decides to adopt the traffic jump buffer pool optimization method based on an embedded neural network proposed by the present invention. The following is the specific implementation situation of this enterprise.

[0197] First, traffic monitoring probes are deployed at key nodes of the content delivery network to collect network traffic data of the nodes in real time. Through the analysis of this historical traffic data, it is found that there are obvious periodic change rules, such as larger traffic during the morning, noon, and evening peak hours every day, and a slight decrease in traffic on weekends. At the same time, some sudden traffic fluctuations are also observed, such as during the launch of new products or major events. In addition, the switching situations of user terminals between different nodes are recorded, and it is found that the switching behavior of users is related to factors such as the load rate of the nodes, network latency, and service quality. Based on these observations, Company ABC begins to construct the embedded neural network model proposed in the present invention.

[0198] Specifically, the prediction equation set includes the following three sub - equations:

[0199] 1) Traffic prediction equation F(t):

[0200]

[0201] Among them, F(t) is the predicted traffic value at time t, with the unit of Mbps; f i (t - i) are the historical traffic values at the previous i time points, with the unit of Mbps; n is the size of the historical data window, taking the value of 50; α1 = 0.6, α2 = 0.3, α3 = 0.1 are weight coefficients; ω = 2π / 86400 is the angular frequency of the periodic change of traffic, corresponding to a 24 - hour cycle; φ = π / 4 is the phase shift; λ = 0.01 is the exponential decay coefficient; ∈ f is the prediction error term.

[0202] By analyzing the historical traffic data through Fourier transform, the periodic parameters ω and φ are obtained. At the same time, the least - squares method is used to fit the traffic values at the first 50 time points to estimate each weight coefficient.

[0203] 2) Switching probability prediction equation P(s):

[0204]

[0205] Among them, P(s) is the switching probability; L is the current node load rate, calculated by the ratio of the used bandwidth C used and the total bandwidth C total , C used and C total are 100 Mbps and 500 Mbps respectively, so L = 0.2; D is the network latency, and the average round - trip latency obtained by active probing is 30 ms; Q is the service quality index, taking the value of 0.8; β1 = 0.5, β2 = 0.3, β3 = 0.2 are weight coefficients; ∈ p is the prediction error term.

[0206] 3) Resource Scheduling Equation R(x):

[0207]

[0208] Among them, R(x) is the resource scheduling index; is the flow change rate; is the flow acceleration; ∈ r is the scheduling error term.

[0209] The prediction network adopts the following structure:

[0210] Convolutional layer 1: kernel_size = 3, channels = 16;

[0211] Convolutional layer 2: kernel_size = 3, channels = 32;

[0212] Convolutional layer 3: kernel_size = 3, channels = 64;

[0213] Fully connected layer 1: neurons = 128;

[0214] Fully connected layer 2: neurons = 3 (number of nodes);

[0215] The input of this network is the output of the above 3 prediction equations, and the output is the node switching prediction vector indicating the switching probability of the user among 3 nodes.

[0216] With the above prediction model, ABC Enterprise then constructed a multi-objective optimization model, and the objectives include minimizing the total buffer pool size and maximizing the user experience index:

[0217] min{Z1(B), Z2(B)};

[0218] Among them, the total buffer pool size the reciprocal of the user experience index

[0219] The constraint conditions include:

[0220] 1) Buffer pool size constraint: 0 ≤ B i ≤ 300MB, i = 1, 2, 3;

[0221] 2) Flow constraint: F(t i ) ≤ 300Mbps, i = 1, 2, 3;

[0222] 3) Switching probability constraint: 0 ≤ P(s i ) ≤ 1, i = 1, 2, 3;

[0223] 4) Resource scheduling constraint: 0 ≤ R(x i ) ≤ 1, i = 1, 2, 3.

[0224] The user experience metric QoE i is calculated as follows:

[0225]

[0226] The MOPSO algorithm is used to solve the multi-objective optimization model, obtaining the Pareto optimal solution set, and finally selecting B * = (200, 150, 100) MB as the optimal buffer pool size allocation scheme.

[0227] To verify the optimization effect, ABC enterprise conducted a 3-month trial run on this scheme. During this period, the enterprise monitored the network experience metrics of user terminals in real time, including video playback smoothness, network latency, and packet loss rate, etc. The data statistical results are shown in Table 1:

[0228] Table 1 User network experience metrics

[0229] Index Average value Fluctuation range Video playback smoothness 95％ 92％-98％ Network latency 35ms 30ms - 45ms Packet loss rate 2％ 1％-3％

[0230] It can be seen from the table that after adopting the scheme of the present invention, the network experience metrics of users remain stable at a relatively high level. The average video playback smoothness reaches 95%, the network latency is controlled within 35 ms, and the packet loss rate is less than 3%, showing obvious improvement compared with before.

[0231] The following provides multiple charts to describe Embodiment 2 in detail:

[0232] Figure 2 Shows the comparison of the total buffer pool size before and after optimization. The figure shows that the buffer pool size before optimization is 500 MB, while it is reduced to 450 MB after optimization. This intuitively demonstrates the resource saving effect brought by the scheme of the present invention. The buffer pool size is reduced by 10%, which not only reduces the investment cost of network equipment but also reduces the daily operation and maintenance expenses.

[0233] Figure 3 Shows the change trends of various metrics during the online optimization process. The figure contains three lines: red represents the user experience metric (QoE), blue represents the network latency, and green represents the packet loss rate. The horizontal axis represents time (month), and the left side of the vertical axis represents the QoE metric, while the right side represents the latency (ms) and packet loss rate (%). It can be clearly seen from the figure that as time goes by, QoE continuously improves, rising from 0.85 to 0.93; at the same time, the network latency is reduced from 40 ms to 33 ms, and the packet loss rate is reduced from 3% to 1.5%. This figure intuitively demonstrates the adaptive optimization ability of the scheme of the present invention, proving the continuous improvement of system performance.

[0234] Figure 4 Shows three key user network experience metrics: video playback smoothness, network latency, and packet loss rate. Each metric shows the average value and the fluctuation range. The average video playback smoothness reaches 95%, and the fluctuation range is between 92% - 98%; the average network latency is 35 ms, and the fluctuation range is between 30 ms - 45 ms; the average packet loss rate is 2%, and the fluctuation range is between 1% - 3%. This figure visually demonstrates the excellent performance of the solution of the present invention in ensuring user experience, and each metric is maintained at a high level and remains stable.

[0235] Figure 5 Shows the performance of the network traffic prediction model. The blue solid line represents the actual traffic, the orange dashed line represents the predicted traffic, and the light blue area represents the prediction interval. The horizontal axis represents the time change within 24 hours, and the vertical axis represents the traffic size (Mbps). It can be seen from the figure that the trend of the predicted traffic basically coincides with the actual traffic, and it can better capture the periodic changes of the traffic. The prediction interval also reasonably encloses most of the actual traffic points, indicating that the prediction model of the present invention has good accuracy and robustness. This provides a reliable basis for subsequent resource scheduling and optimization.

[0236] As described above, it is only the specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art within the technical scope disclosed by the present invention can easily think of changes or substitutions, and all should be covered within the protection scope of the present invention.

Claims

1. An optimization method for a traffic jump buffer pool, characterized in that, Including: Construct an embedded neural network model including a series of prediction equation sets and a prediction network; collect historical traffic data of multiple content distribution network nodes, and obtain the switching records of user terminals among the multiple content distribution network nodes; preprocess the historical traffic data, extract network traffic feature parameters and input them into the embedded neural network model to obtain a switching prediction vector of user terminals among the multiple content distribution network nodes; establish a multi-objective optimization model including minimizing the size of the traffic buffer pool and maximizing the user experience index according to the switching prediction vector, obtain the optimal buffer pool size, and dynamically adjust the traffic buffer pool among the multiple content distribution network nodes; monitor the network experience index of user terminals in real time, where the network experience index includes video playback smoothness, network latency, and packet loss rate; regularly update the embedded neural network model online according to the network experience index to achieve adaptive optimization of the traffic buffer pool size.

2. The optimized method for a traffic jump buffer pool according to claim 1, characterized in that The prediction equation set includes a traffic prediction equation, a switching probability prediction equation, and a resource scheduling equation.

3. The optimization method of a traffic jump buffer pool according to claim 2, wherein The traffic prediction equation is used to predict the network traffic at a future moment. The switching probability prediction equation is used to predict the probability that a user terminal switches to other nodes. The resource scheduling equation is used to output parameters for optimizing the allocation of network resources.

4. The optimization method of a traffic jump buffer pool according to claim 3, characterized in that, The prediction network adopts a lightweight convolutional neural network, with the input being the output of the prediction equation set and the output being a node switching vector.

5. The optimization method of a traffic jump buffer pool according to claim 4, characterized in that The prediction network includes 3 convolutional layers and 2 fully connected layers.

6. The optimized method for a traffic jump buffer pool according to claim 5, characterized in that, The network traffic feature parameters include traffic size, traffic duration, and traffic direction.

7. A method for optimizing a flow jump buffer pool according to claim 6, characterized in that, The optimization objectives of the multi-objective optimization model include minimizing the size of the traffic buffer pool and maximizing the user experience index.

8. A method for optimizing a traffic jump buffer pool according to claim 7, characterized in that, The constraint conditions of the multi-objective optimization model at least include: buffer pool size constraint, traffic constraint.

9. A computer-readable storage medium, characterized in that, Program instructions are stored in the computer-readable storage medium, and when the program instructions run on a computer, they are used to execute a traffic jump buffer pool optimization method according to any one of claims 1-8.

10. A flow jump buffer pool optimization system, characterized in that Including the computer-readable storage medium according to claim 9.

Images