Automatic parameter adjustment method, device and equipment for controlling network congestion and medium

Abstract

The invention relates to the field of computer networks, and discloses an automatic parameter adjustment method and device for controlling network congestion, equipment and a medium. According to the method disclosed by the invention, a sliding window sample maintenance strategy is designed, random forest proxy modeling and an acquisition function based on a predicted mean value and uncertainty are combined, global-local candidate generation and a random exploration mechanism with a certain probability are adopted, and congestion control parameters exposed by a kernel are automatically adjusted by taking a retransmission rate as a main optimization target; the problem that low retransmission rate and high throughput are difficult to guarantee at the same time in a dynamic network environment by traditional static empirical parameter adjustment and single-target optimization is solved, so that the transmission efficiency and the system stability are improved under changing network conditions.

Description

Technical Field

[0001] This invention relates to the field of computer networks and discloses a method, apparatus, device, and medium for automatically adjusting parameters to control network congestion. Background Technology

[0002] Congestion control is a core mechanism in network transmission protocols. Its main goal is to coordinate the data transmission rate of senders to avoid network congestion and ensure end-to-end performance under conditions of limited and dynamically changing network resources. Common congestion control algorithms include several configurable parameters, such as the initial congestion window, slow start gain, congestion window reduction factor, and bandwidth estimation smoothing factor. The values ​​of these parameters directly affect the throughput, latency, retransmission rate, and stability of the link. Under different network conditions (such as high latency, packet loss rate fluctuations, bandwidth jitter, and mobile network handover), the optimal settings for these parameters often vary significantly.

[0003] However, most current congestion control algorithms employ static or quasi-static parameter configurations during deployment. These fixed parameters are often unsuitable under varying network conditions. Furthermore, traditional parameter adjustments typically rely on manual experience, making it difficult to adapt to fluctuating network conditions in real-time. The complex tuning process also leads to low network resource utilization. Network environments are influenced by a combination of factors, including latency, packet loss rate, bandwidth fluctuations, link jitter, user traffic distribution, and mobility. Static parameters cannot adaptively adjust to changes in these conditions, resulting in poor performance in scenarios with high-latency links, sudden increases in packet loss, or drastic bandwidth fluctuations. For example, parameter mismatch can cause the sender to continue sending with an aggressive strategy even when network conditions are deteriorating, leading to high retransmission rates and significant bandwidth waste. Conversely, conservative parameters may fail to fully utilize bandwidth even when network conditions are good, resulting in reduced throughput. Summary of the Invention

[0004] Given the shortcomings of existing technologies in dealing with dynamic network environments and limited samples, this application provides an automatic parameter tuning method, apparatus, device, and medium for controlling network congestion. The technical solution is as follows: An automatic parameter tuning method for controlling network congestion, running on an electronic device with a configurable congestion control algorithm, includes the following steps: S1. Select the congestion control algorithm parameters to be tuned, first perform initial sampling on the parameters to obtain a preset number of candidate parameter groups, wherein the candidate parameter groups are the values ​​of each parameter to be applied to the system; S2. For each candidate parameter group, modify the kernel congestion control algorithm parameters to the parameter values ​​corresponding to the candidate parameter group, so that the system runs under the parameter configuration; after the system stabilizes, sample the system's performance indicators for multiple cycles to form an initial dataset, the performance indicators including download speed and retransmission rate; S3. Construct an objective function based on performance metrics, and then train a surrogate regression model to fit the objective function based on the initial dataset; S4. Use iterative optimization methods to iteratively optimize the objective function to obtain the optimal congestion control algorithm parameters.

[0005] Preferably, the congestion control algorithm parameters to be tuned include the initial window, slow start gain coefficient, minimum bandwidth, congestion window reduction coefficient after packet loss, or other parameters related to the congestion control algorithm exposed by the operating system kernel module, and upper and lower boundaries need to be defined for each parameter to constrain the search space; The initial sampling method uses Latin hypercube sampling to ensure uniform coverage of the initial samples in the high-dimensional parameter space.

[0006] Preferably, step S2 specifically includes the following steps: S21. Round the candidate parameter set to the nearest integer; S22. Write the candidate parameter set into the operating system kernel and modify the corresponding parameters of the congestion control algorithm; S23. Wait for D seconds to ensure that the system state tends to be balanced. In the subsequent K sampling cycles, call the monitoring interface every t seconds to obtain the download speed and retransmission rate of each sampling. K, D and t are all configurable parameters. S24. Calculate average download speed and retransmission rate The calculation method is as follows: ; ; in, and These are the download speed and retransmission rate for the j-th sample, respectively.

[0007] Preferably, the monitoring interface can collect the number of requests, download speed, and retransmission rate of each network card on the local machine in each sampling. The download speed T and retransmission rate R for each sampling are calculated as follows: ; ; in, , , ...

[0008] Preferably, the objective function in step S3 is: ; Where x is a parameter set, and The average download speed and retransmission rate are obtained from the sampling of parameter group x. This is a penalty factor for the retransmission rate.

[0009] Specifically, this application preferably uses retransmission rate and download speed as core performance indicators because retransmission rate can most directly reflect the reliability of network links and the extent of resource waste. Based on this, those skilled in the art will understand that, to further refine the network profile, indicators such as bandwidth utilization and round-trip time (RTT) fluctuations can also be used as supplementary references. These indicators are highly correlated with retransmission rate in reflecting network congestion trends and can all be included in the input dimensions or objective function evaluation scope of the proxy model described in this application.

[0010] Preferably, the surrogate regression model in step S3 is a random forest regressor, which consists of several independent decision trees and can predict the objective function value based on the parameter set x. When predicting the candidate parameter set x, the random forest regressor calls each decision tree to output the predicted objective function value and the prediction standard deviation, and calculates the arithmetic mean of the objective function values ​​and prediction standard deviations of all decision trees as the final predicted mean of the parameter set. and the predicted standard deviation .

[0011] Preferably, the Bayesian optimization method in step S4 includes the following steps: S41. Generate candidate parameter sets for the next evaluation using a surrogate regression model and a data collection function; S42. Apply the candidate parameter set to the system and collect performance metrics to add to the dataset; S43. Use a sliding window to maintain the dataset. The window length is W. When the number of samples exceeds W due to the addition of new data, delete the oldest sample data from the top of the window to ensure that the dataset can reflect the dynamic changes in the recent network environment. S44. Update the surrogate regression model using the dataset. Repeat steps S41 to S44. The candidate parameter set will gradually approach the optimal value.

[0012] Preferably, the acquisition function in S41 is: ; in, and These are the mean and standard deviation of the objective function values ​​predicted by the surrogate regression model for parameter set x. These are configurable parameters.

[0013] The preferred procedure for generating candidate parameter sets for the next evaluation is as follows: First, N_CAND parameter groups are uniformly and randomly generated within the defined parameter boundaries, and the acquisition function value U of each parameter group is calculated using a surrogate regression model. The parameter group corresponding to the largest U is selected as the global optimal point x_best. Then, Gaussian perturbations are applied to each parameter dimension with x_best as the center to generate n_noisy local parameter groups. The acquisition function value is recalculated for these n_noisy local parameter groups, and the parameter group corresponding to the locally optimal acquisition function value is selected as the candidate parameter group for the next evaluation. N_CAND and n_noisy are both configurable parameters. The generation of candidate parameter sets for the next evaluation also incorporates a stochastic exploration strategy. In each iteration, purely random candidate sampling is performed with probability explore_p to maintain global exploration capability. Use a strategy based on the collected function values.

[0014] An automatic parameter tuning device for controlling network congestion includes a parameter sampling unit, a calculation unit, a proxy model unit, and a candidate generation unit. Parameter sampling unit: Used to write candidate parameter sets into the operating system kernel, modify the corresponding parameters of the congestion control algorithm through the system-permitted interface, and sample performance indicators in multiple cycles after stabilization to obtain the average download speed and retransmission rate; Calculation unit: Based on the download speed and retransmission rate obtained by the parameter sampling unit, it calculates the corresponding objective function value and establishes a correspondence between the parameter set and the objective function value, adding it to the dataset; Proxy model unit: used to train a proxy regression model using a sliding window-based dataset. The proxy regression model employs a random forest regressor and is able to predict the objective function value and standard deviation based on the parameter set. Candidate generation unit: It is used to calculate the acquisition function value based on the prediction results of the surrogate model unit, and at the same time, it combines a certain probability random exploration strategy to generate candidate parameter sets for the next evaluation, so as to ensure local optimization while maintaining global exploration capability. The generated candidate parameter sets are output to the parameter sampling unit for the next round of evaluation.

[0015] A computer device includes at least one processor and a computer-readable medium storing a computer program, which is read and executed by the processor to implement the methods described in this application.

[0016] A computer-readable medium storing a computer program that is read and executed by a processor to implement the methods described in this application.

[0017] Compared with the prior art, the beneficial effects of this application are as follows: (1) Taking retransmission rate as the core objective to directly reduce the loss caused by retransmission: This invention applies a weighted penalty to the retransmission rate in the performance score, making the optimization direction significantly biased towards reducing the retransmission rate. Compared with the tuning strategy that only targets throughput, this invention can maintain throughput as much as possible while suppressing retransmission, reducing bandwidth waste and delay and resource overhead caused by repeated transmission.

[0018] (2) Sliding window sample and random forest agent modeling to improve online robustness and response speed: The sliding window is used to retain only recent samples to train the agent model, so that the model can quickly reflect the dynamic changes of the network environment; the random forest as an agent is not only robust to noisy observations, but also provides uncertainty estimates, reducing the mistuning caused by observation noise or scarce samples, thereby reducing the increase in retransmissions caused by mistuning.

[0019] (3) This solution has stronger deployment independence and compatibility. Since the optimization process is entirely based on modifying the local operating system kernel parameters and calling the local monitoring interface, there is no need to change the communication protocol message format or coordinate with the network peer devices. This allows this solution to achieve adaptive optimization of transmission performance without changing the existing network topology and peer configuration, significantly reducing the system upgrade and maintenance costs.

[0020] (4) Efficient global-local candidate generation combined with random exploration reduces trial and error costs and accelerates convergence: First, a global coarse screening is performed, then a local fine search is conducted around the advantages, and a pure random exploration is carried out with a certain probability, which maintains the global exploration capability and improves the efficiency of local optimization. Attached Figure Description

[0021] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below; Figure 1 This is a schematic diagram of the process disclosed in the embodiments of the present invention; Figure 2 This is a comparison chart of the retransmission rate and throughput before and after parameter tuning when applying the present invention. Detailed Implementation

[0022] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention.

[0023] This invention provides an automatic parameter tuning method for congestion control algorithms based on iterative optimization. This method is executed on an operating system running a configurable congestion control algorithm module, such as... Figure 1 As shown, it includes the following steps: S1. Select the congestion control algorithm parameters to be tuned. First, perform initial sampling on the parameters to obtain a preset number of candidate parameter groups. The candidate parameter groups are the values ​​of each parameter to be applied to the system. Specifically, for ease of subsequent calculation and program implementation, this preset number can be denoted as the initial sampling number N. inital The candidate parameter set consists of the values ​​of each parameter to be applied to the system, N. inital N is the initial number of samples. inital The specific value can be flexibly configured according to the complexity of the network environment and the requirements for parameter tuning accuracy. For example, it can be set to 10 groups in the experiment to ensure sufficient initial coverage in the parameter space.

[0024] The congestion control algorithm parameters to be tuned in step S1 include the initial window, slow start gain coefficient, minimum bandwidth, congestion window reduction coefficient after packet loss, or other parameters related to the congestion control algorithm exposed by the operating system kernel module. Furthermore, upper and lower boundaries need to be defined for each parameter to constrain the search space.

[0025] The initial sampling in step S1 uses the Latin hypercube sampling method to ensure that the initial samples achieve uniform coverage in the high-dimensional parameter space.

[0026] S2. For each candidate parameter group, modify the kernel congestion control algorithm parameters to the corresponding parameter values ​​for that candidate parameter group, enabling the system to operate under that parameter configuration. After the system stabilizes, perform performance metric sampling on the system over multiple cycles to form an initial dataset. These performance metrics include download speed and retransmission rate.

[0027] Step S2 specifically includes the following steps: (1) Round the candidate parameter set to the nearest integer; (2) Use the write interface allowed by the operating system to write the candidate parameter group into the operating system kernel and modify the corresponding parameters of the congestion control algorithm; (3) Wait for D seconds to ensure that the system state tends to be balanced, and then call the monitoring interface every t seconds in the subsequent K sampling cycles to obtain the download speed and retransmission rate of each sampling. K, D and t are all configurable parameters. (4) Calculate the average download speed and retransmission rate The calculation method is as follows: ; ; in, and These are the download speed and retransmission rate for the j-th sample, respectively.

[0028] Furthermore, the monitoring interface can collect the number of requests, download speed, and retransmission rate of each network card on the local machine in each sampling. The download speed T and retransmission rate R for each sampling are calculated as follows: ; ; in, , , ...

[0029] The objective function in step S3 is: ; Where x is a parameter set, and The average download speed and retransmission rate are obtained from the sampling of parameter group x. This is a penalty factor for the retransmission rate.

[0030] S3. Construct an objective function based on performance metrics, and then train a proxy model to fit the objective function based on the initial dataset.

[0031] The surrogate model aims to fit a nonlinear mapping relationship between candidate parameter sets and objective function values. In an optional implementation, the surrogate model can be implemented using a regression model, such as a random forest regressor. The random forest regressor consists of several independent decision trees and can predict objective function values ​​based on the parameter set x. When predicting candidate parameter set x, the random forest regressor calls each decision tree to output the predicted objective function value and the prediction standard deviation, and calculates the arithmetic mean of the objective function values ​​and prediction standard deviations of all decision trees as the final predicted mean for the parameter set. and the predicted standard deviation The prediction results not only output the predicted mean, but also reflect the uncertainty of the prediction, which is of great significance for the subsequent optimization process.

[0032] S4. Use iterative optimization methods to iteratively optimize the objective function to obtain the optimal congestion control algorithm parameters; This application does not strictly limit the specific algorithm type for iterative optimization. As an efficient implementation scheme, this embodiment preferably adopts the Bayesian optimization method. By utilizing the prior information provided by the surrogate model, Bayesian optimization can balance exploration and exploitation, quickly approximating the optimal congestion control algorithm parameters in a smaller number of iterations, thereby significantly improving the transmission efficiency of the system. In a specific embodiment, the Bayesian optimization method in step S4 includes the following steps: (1) Generate candidate parameter sets for the next evaluation using a surrogate regression model and a data collection function; (2) Apply the candidate parameter set to the system and collect performance metrics to add to the dataset; (3) Use a sliding window to maintain the dataset. The window length is W. When the number of samples exceeds W due to the addition of new data, delete the oldest sample data from the top of the window to ensure that the dataset can reflect the dynamic changes of the network environment in the near future.

[0033] Furthermore, the acquisition function mentioned in (1) is: ; in, and These are the mean and standard deviation of the objective function values ​​predicted by the surrogate regression model for parameter set x. These are configurable parameters.

[0034] The process for generating candidate parameter sets for the next evaluation is as follows: First, N_CAND parameter sets are uniformly and randomly generated within the defined parameter boundaries, and the acquisition function value U of each parameter set is calculated using a surrogate regression model. The parameter set corresponding to the largest U is selected as the global optimal point x_best. Then, Gaussian perturbations are applied to each parameter dimension with x_best as the center to generate n_noisy local parameter sets. The acquisition function value is calculated again for these n_noisy local parameter sets, and the parameter set corresponding to the locally optimal acquisition function value is selected as the candidate parameter set for the next evaluation. N_CAND and n_noisy are both configurable parameters.

[0035] The generation of candidate parameter sets for the next evaluation also incorporates a stochastic exploration strategy. In each iteration, purely random candidate sampling is performed with probability explore_p to maintain global exploration capability. Use a strategy based on the collected function values.

[0036] (4) Update the surrogate regression model using the dataset. Repeat steps (1) to (4). The candidate parameter set will gradually approach the optimal value. Specific Implementation

[0037] This invention provides an automatic parameter tuning method for a congestion control algorithm based on Bayesian optimization. This method can be executed on an operating system with a configurable congestion control module, such as the BBR congestion control algorithm in a Linux system kernel environment. The system has a parameter writing interface ` / sys / module / bbr_adapt / parameters / ` and a performance monitoring interface ` / proc / tcp_metrics`, enabling automated parameter tuning and real-time performance acquisition.

[0038] In this embodiment, the congestion control algorithm parameters to be optimized are first selected, such as the initial congestion window size (init_cwnd), the slow start gain coefficient (high_gain), and the congestion window reduction coefficient after packet loss (beta). Upper and lower boundaries are set for each parameter; for example, the initial congestion window size is limited to [4, 100], and the slow start gain coefficient is limited to [1.0, 3.0]. These boundaries can be given by kernel protocol stack characteristics, RFC standards, or empirical values ​​to ensure that the parameters are searched within a reasonable range.

[0039] Its beneficial effects are: by setting parameter boundaries, the search space is constrained, preventing the algorithm from exploring unsafe or unstable parameter regions, thus ensuring that the optimization process is stable and controllable.

[0040] After determining the optimization parameters and boundaries, the Latin hypercube sampling (LHS) method is used to generate an initial candidate parameter set. In this embodiment, the initial sampling number N is set. inital By using SciPy's lhs() function to sample in the high-dimensional parameter space, 10 sets of candidate parameters are obtained, each set being uniformly distributed within a defined boundary range.

[0041] Its beneficial effects are as follows: Latin hypercube sampling achieves uniform coverage in multidimensional space, avoids the concentration or gap problems of random sampling, and enables the surrogate model to obtain high-quality training data from the beginning stage, thereby accelerating the optimization convergence speed.

[0042] For each set of candidate parameters, the program writes them to the kernel parameter interface using the command format `echo`. <value>> / sys / module / bbr_adapt / parameters / <param> To avoid floating-point numbers being rejected by the kernel, the system rounds or t-rounds parameter values. A wait time D (e.g., 5 seconds) is set to ensure the kernel parameters take effect and the TCP connection enters a stable phase. In the subsequent K sampling cycles, metrics provided by the monitoring interface are read every t seconds, including download speed, retransmission rate, and the number of requests per network interface card. The average of the K sampling results is calculated to obtain the performance metrics corresponding to this parameter set.

[0043] For example, in one experiment, the candidate parameter set {init_cwnd=32, ssthresh_gain=1.5, beta=0.7} was applied to the system. After waiting for 5 seconds, the system ran for 60 seconds, and samples were taken every 10 seconds to obtain 6 sets of throughput and retransmission rate data. Finally, the performance evaluation was calculated.

[0044] Its beneficial effects are: by delaying the write operation and averaging multiple samples, it avoids misjudgments caused by instantaneous network fluctuations and ensures that the sampling results can reflect the system performance under stable operating conditions.

[0045] During the sampling process, if the system has multiple network cards (e.g., eth0 and eth1), the number of requests, download speed, and retransmission rate of each network card are read at each sampling time, and the overall performance is calculated using the following weighted formula: ; in, , , ...

[0046] Its beneficial effects are: the weighted calculation method can take into account the load of different network cards, making the overall performance indicators more accurate, and is suitable for multi-link, multi-interface device environments.

[0047] Then, construct the objective function based on the performance metrics: ; Where x is a parameter set, and The average download speed and retransmission rate are obtained from the sampling of parameter group x. This is a penalty factor for the retransmission rate, used to reflect the importance of reducing the retransmission rate; in this example, it is set to 5.0. If the average download speed is 8.2 Mbps and the retransmission rate is 6.1%, then the objective function value is: .

[0048] Its beneficial effect lies in the fact that the objective function achieves a balance between the two performance indicators of "high throughput and low retransmission," and can be adjusted according to actual business needs. Preference control allows for flexible adaptation to different scenarios.

[0049] A surrogate regression model is trained based on the dataset of the parameter set and objective function values ​​described above. In this embodiment, a random forest regressor is used as the surrogate model, consisting of 100 independent decision trees. For each parameter set x, the model outputs the predicted mean μ(x) and the predicted standard deviation σ(x), that is, the overall predicted mean and variance are obtained by taking the arithmetic mean of the prediction results of each tree.

[0050] Its advantages are: random forests do not require hypothesis function form, can handle complex nonlinear relationships, and the prediction variance can naturally reflect the uncertainty of the model, which ensures both computational speed and provides reliable uncertainty information for Bayesian optimization.

[0051] Based on this, a Bayesian optimization process is performed. First, 10,000 candidate parameter sets are randomly generated within the defined parameter boundaries, and their acquisition function values ​​are calculated. ; in For the exploration coefficient, this embodiment sets it to 1.5. The parameter set with the largest acquisition function value is selected as the global optimal point x_best. Then, a Gaussian perturbation is applied to the vicinity of x_best to generate 1000 local parameter sets, and their acquisition function values ​​are calculated to select the local optimum x_next. In each round of optimization, a purely random sampling is performed with a probability of 10% (explore_p) to enhance global exploration.

[0052] Its beneficial effects are that this strategy of combining "global randomness + local perturbation" can both escape local optima and focus on searching for potential high-yield regions, significantly improving the convergence speed and the probability of global optima.

[0053] After each round of optimization, a new parameter set x_next is obtained, written into the system for execution, and performance data is collected. The results are added to the dataset, and the dataset is updated using a sliding window mechanism (window length W=200). When the number of samples exceeds the window length, the oldest sample is automatically deleted. By repeating the above steps, the system congestion control parameter values ​​will gradually approach the optimal value.

[0054] Figure 2 The paper presents a comparison of retransmission rate and throughput before and after parameter tuning of the optimized congestion control algorithm of this invention. The congestion control algorithm used here is the BBR algorithm. It can be seen that after parameter tuning using this method, the retransmission rate is reduced by 24% and the throughput is increased by 9%.

[0055] An automatic parameter tuning device for controlling network congestion includes: The parameter sampling unit is used to write the candidate parameter set into the operating system kernel, modify the corresponding parameters of the congestion control algorithm through the system-permitted interface, and sample performance indicators in multiple cycles after stabilization to obtain the average download speed and retransmission rate. The calculation unit is used to calculate the corresponding objective function value based on the download speed and retransmission rate obtained by the parameter sampling unit, and to establish a correspondence between the parameter set and the objective function value and add it to the dataset. The surrogate model unit is used to train a surrogate regression model using a sliding window-based dataset. The surrogate regression model employs a random forest regressor and is able to predict the objective function value and standard deviation based on the parameter set. The candidate generation unit is used to calculate the acquisition function value based on the prediction results of the surrogate model unit, and at the same time, combine a certain probability random exploration strategy to generate the candidate parameter set for the next evaluation, so as to ensure local optimization while maintaining global exploration capability. The generated candidate parameter set is output to the parameter sampling unit for the next round of evaluation.

[0056] A computer device includes at least one processor and a computer-readable medium storing a computer program, which is read and executed by the processor to implement the method of this application.

[0057] A computer-readable medium storing a computer program that is read and executed by a processor to implement the method of this application.

[0058] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.< / value>

Claims

1. An automatic parameter tuning method for controlling network congestion, running on an electronic device with a configurable congestion control algorithm, characterized in that, Includes the following steps: S1. Select the congestion control algorithm parameters to be tuned, first perform initial sampling on the parameters to obtain a preset number of candidate parameter groups, wherein the candidate parameter groups are the values ​​of each parameter to be applied to the system; S2. For each candidate parameter group, modify the kernel congestion control algorithm parameters to the parameter values ​​corresponding to the candidate parameter group, so that the system runs under the parameter configuration; after the system stabilizes, sample the system's performance indicators for multiple cycles to form an initial dataset, the performance indicators including download speed and retransmission rate; S3. Construct an objective function based on performance metrics, and then train a proxy model to fit the objective function based on the initial dataset; S4. Use iterative optimization methods to iteratively optimize the objective function to obtain the optimal congestion control algorithm parameters.

2. The automatic parameter tuning method for controlling network congestion according to claim 1, characterized in that, The congestion control algorithm parameters to be tuned include the initial window, slow start gain coefficient, minimum bandwidth, congestion window reduction coefficient after packet loss, or other parameters related to the congestion control algorithm exposed by the operating system kernel module. It is also necessary to define the upper and lower boundaries of each parameter to constrain the search space. The initial sampling method uses Latin hypercube sampling to ensure uniform coverage of the initial samples in the high-dimensional parameter space.

3. The automatic parameter tuning method for controlling network congestion according to claim 1, characterized in that, Step S2 specifically includes the following steps: S21. Round the candidate parameter set to the nearest integer; S22. Write the candidate parameter set into the operating system kernel and modify the corresponding parameters of the congestion control algorithm; S23. Wait for D seconds to ensure that the system state tends to be balanced. In the subsequent K sampling cycles, call the monitoring interface every t seconds to obtain the download speed and retransmission rate of each sampling. K, D and t are all configurable parameters. S24. Calculate average download speed and retransmission rate The calculation method is as follows: ； ； in, and These are the download speed and retransmission rate for the j-th sample, respectively.

4. The automatic parameter tuning method for controlling network congestion according to claim 3, characterized in that, The monitoring interface can collect the number of requests, download speed, and retransmission rate for each network card on the local machine in each sampling. The download speed T and retransmission rate R for each sampling are calculated as follows: ； in, , , ...

5. The automatic parameter tuning method for controlling network congestion according to claim 3, characterized in that, The objective function mentioned in step S3 is: ； Where x is a parameter set, and The average download speed and retransmission rate are obtained from the sampling of parameter group x. This is a penalty factor for the retransmission rate.

6. The automatic parameter tuning method for controlling network congestion according to claim 1, characterized in that, The surrogate regression model in step S3 is a random forest regressor, which consists of several independent decision trees and can predict the objective function value based on the parameter set x. When predicting the candidate parameter set x, the random forest regressor calls each decision tree to output the predicted objective function value and the prediction standard deviation, and calculates the arithmetic mean of the objective function values ​​and prediction standard deviations of all decision trees as the final predicted mean of the parameter set. and the predicted standard deviation .

7. The automatic parameter tuning method for controlling network congestion according to claim 1, characterized in that, The Bayesian optimization method described in step S4 includes the following steps: S41. Generate candidate parameter sets for the next evaluation using a surrogate regression model and a data collection function; S42. Apply the candidate parameter set to the system and collect performance metrics to add to the dataset; S43. Use a sliding window to maintain the dataset. The window length is W. When the number of samples exceeds W due to the addition of new data, delete the oldest sample data from the top of the window to ensure that the dataset can reflect the dynamic changes in the recent network environment. S44. Update the surrogate regression model using the dataset. Repeat steps S41 to S44. The candidate parameter set will gradually approach the optimal value.

8. The automatic parameter tuning method for controlling network congestion according to claim 7, characterized in that, The acquisition function described in S41 is: ； in, and These are the mean and standard deviation of the objective function values ​​predicted by the surrogate regression model for parameter set x. These are configurable parameters.

9. The automatic parameter tuning method for controlling network congestion according to claim 8, characterized in that, The process for generating candidate parameter sets for the next evaluation is as follows: First, N_CAND parameter groups are uniformly and randomly generated within the defined parameter boundaries, and the acquisition function value U of each parameter group is calculated using a surrogate regression model. The parameter group corresponding to the largest U is selected as the global optimal point x_best. Then, Gaussian perturbations are applied to each parameter dimension with x_best as the center to generate n_noisy local parameter groups. The acquisition function value is recalculated for these n_noisy local parameter groups, and the parameter group corresponding to the locally optimal acquisition function value is selected as the candidate parameter group for the next evaluation. N_CAND and n_noisy are both configurable parameters. The generation of candidate parameter sets for the next evaluation also incorporates a stochastic exploration strategy. In each iteration, purely random candidate sampling is performed with probability explore_p to maintain global exploration capability. Use a strategy based on the collected function values.

10. An automatic parameter adjustment device for controlling network congestion, characterized in that, It includes a parameter sampling unit, a calculation unit, a surrogate model unit, and a candidate generation unit; Parameter sampling unit: Used to write candidate parameter sets into the operating system kernel, modify the corresponding parameters of the congestion control algorithm through the system-permitted interface, and sample performance indicators in multiple cycles after stabilization to obtain the average download speed and retransmission rate; Calculation unit: Based on the download speed and retransmission rate obtained by the parameter sampling unit, it calculates the corresponding objective function value and establishes a correspondence between the parameter set and the objective function value, adding it to the dataset; Proxy model unit: used to train a proxy regression model using a sliding window-based dataset. The proxy regression model employs a random forest regressor and is able to predict the objective function value and standard deviation based on the parameter set. Candidate generation unit: It is used to calculate the acquisition function value based on the prediction results of the surrogate model unit, and at the same time, it combines a certain probability random exploration strategy to generate candidate parameter sets for the next evaluation, so as to ensure local optimization while maintaining global exploration capability. The generated candidate parameter sets are output to the parameter sampling unit for the next round of evaluation.

11. A computer device, characterized in that, It includes at least one processor and a computer-readable medium storing a computer program, which is read and executed by the processor to implement the method as described in any one of claims 1-9.

12. A computer-readable medium, characterized in that, The computer-readable medium stores a computer program that is read and executed by a processor to implement the method as described in any one of claims 1-9.

Images