An intelligent network congestion control method based on frequency domain feature analysis and dynamic weight adjustment

Abstract

The present invention discloses an intelligent network congestion control method based on frequency domain feature analysis and dynamic weight adjustment, which collects the network round-trip delay RTT and queue length and calculates the jitter value. The jitter value is converted into the frequency domain by performing DFT, and then the high-frequency energy ratio is calculated. The state vector is then constructed, and then the state vector is estimated in real time based on the Kalman filter model to obtain the network state estimation value. According to the high-frequency energy ratio, the weights of RTT, queue length and frequency domain features are dynamically adjusted, and then the parameters of the PID controller are adaptively adjusted according to the weights. Finally, the PID control signal generated by the PID controller is converted into a sending rate adjustment action or an explicit congestion notification, thereby suppressing network congestion in real time. The present invention can significantly improve the performance of data center networks in high-load and complex network environments, reduce delays, increase throughput, avoid data packet loss, and enhance the network's adaptability.

Description

Technical Field

[0001] The present invention relates to the technical field of computer network communications, and in particular to an intelligent network congestion control method based on frequency domain feature analysis and dynamic weight adjustment. Background Art

[0002] With the rapid development of cloud computing, big data, the Internet of Things, and 5G communications technologies, data center networks are facing an ever-increasing volume of traffic, significantly increasing network complexity and dynamism. The demand for transmitting large amounts of real-time data streams has led to increasing network congestion, severely restricting quality of service (QoS), reducing user experience and network performance. Therefore, timely and accurate monitoring and mitigation of network congestion has become a major challenge in the data communications field.

[0003] Traditional data center network congestion control methods rely primarily on single or simple status monitoring metrics (such as simple RTT or queue length monitoring) and employ fixed-parameter control strategies. These strategies struggle to achieve efficient, real-time congestion control responses in rapidly changing network environments. These fixed control mechanisms are prone to delayed responses, overreactions, or insufficient control when network loads fluctuate or experience transient congestion, leading to severe network performance fluctuations such as dramatically increased latency, significantly reduced throughput, and even significant packet loss.

[0004] In recent years, the Kalman filter algorithm has been widely used in network state estimation. It effectively estimates and predicts state changes such as network delay and queue length in real time through recursive methods. However, traditional Kalman filter applications in network congestion control are mostly limited to time-domain data and lack effective consideration of the frequency-domain characteristics of network traffic variations. Consequently, it struggles to accurately identify and respond to short-term network jitter and high-frequency congestion events.

[0005] On the other hand, Fourier transforms (especially discrete Fourier transforms (DFTs)) can effectively capture the frequency domain characteristics of network state changes, especially with a clear advantage in identifying short-term, high-frequency network fluctuations. However, existing technologies typically only use Fourier transforms for spectral analysis and feature identification of network traffic, thus failing to realize their full potential for real-time network control and congestion relief.

[0006] In summary, current data center network congestion control methods still have the following major issues: first, they lack a refined, real-time prediction and response mechanism for network status; second, they lack the integration of frequency-domain features with time-domain control methods; and third, their fixed control parameters cannot adaptively respond to dynamically changing network environments. These shortcomings severely restrict the control performance of data center networks in complex and ever-changing real-world application scenarios. There is an urgent need for an innovative solution that can more efficiently and accurately respond to network status changes in real time and comprehensively improve network performance and stability. Summary of the Invention

[0007] To address the problems of current traditional network congestion control methods in complex network environments, such as insufficient real-time response, low accuracy, and fixed control strategies, this invention provides an intelligent network congestion control method based on frequency domain feature analysis and dynamic weight adjustment. This method dynamically optimizes control strategies, accurately responds to various network environment changes, and thus improves network performance.

[0008] In order to achieve the above technical purpose, the technical solution of the present invention is:

[0009] An intelligent network congestion control method based on frequency domain feature analysis and dynamic weight adjustment includes the following steps:

[0010] Step 1: Collect network round-trip delay (RTT) and queue length data based on time, then segment the data into time domain segments according to a fixed window size and calculate the jitter value.

[0011] Step 2: Perform discrete Fourier transform (DFT) on the jitter values ​​of each segment of RTT and queue length data to convert them into the frequency domain, and then calculate the proportion of high-frequency energy in the frequency domain.

[0012] Step 3: Construct a state vector containing RTT deviation, queue length, and high-frequency energy ratio. Then, use the Kalman filter model to perform real-time network state estimation on the state vector to obtain the estimated RTT, queue length, and frequency domain characteristics of the network.

[0013] Step 4: Based on the proportion of high-frequency energy, the weights of RTT, queue length, and frequency domain features are dynamically adjusted using the Sigmoid function. The PID controller parameters are then adaptively adjusted based on the weights.

[0014] Step 5: Convert the PID control signal generated by the PID controller into a sending rate adjustment action or an explicit congestion notification (ECN), thereby suppressing network congestion in real time.

[0015] Furthermore, in the method, step 1 includes:

[0016] Collect RTT and queue length data in the form of time series to obtain the specific values ​​of RTT and queue length corresponding to each time point;

[0017] The data is segmented according to a fixed time length, i.e., a window. The difference between adjacent data in each segment is then calculated, and the calculated differences are used as the jitter value of this segment of data.

[0018] Furthermore, in the method, step 1 further includes the step of filtering out abnormal values ​​of the jitter value of each segment of data:

[0019] Calculate the mean of the data in each segment, and then calculate the standard deviation based on the mean. Then determine the threshold of the jitter value of this segment of data based on the standard deviation. Jitter values ​​greater than the threshold are judged as outliers and deleted.

[0020] Furthermore, in the method, step 2 includes:

[0021] Perform discrete Fourier transform (DFT) on the jitter value of each segment of data according to the following formula:

[0022] ;

[0023] in is the complex frequency domain coefficient corresponding to the kth point of the spectrum, is the fixed window length, Indicates the jitter value of each segment of data, is the natural base, represents the imaginary unit, represents pi, Indicates the first Sampling moments;

[0024] Then calculate the proportion of high-frequency energy in the frequency domain according to the following formula:

[0025] ;

[0026] in It is the set frequency threshold, which determines the starting frequency of the high frequency part. represents the energy at the kth frequency point, It is the ratio of high-frequency component energy to the total signal energy, that is, the proportion of high-frequency energy.

[0027] Furthermore, in the method, in step 3, the state vector is expressed as:

[0028] ;

[0029] in, Indicates The state vector at time t, Indicates Network delay deviation at each moment; Indicates The queue length at the time; Indicates The proportion of high-frequency energy in the frequency domain characteristics of the moment;

[0030] The state of the network is then treated as a linear dynamic system and represented by the following state evolution equation and observation equation:

[0031] ;

[0032] ;

[0033] Where F is the state transfer matrix, express The state vector of the previous moment, B is the control input matrix, express The control input at the previous moment, express The process noise at the previous moment;

[0034] express The actual observed value of RTT or queue length at time, is the observation matrix, represents the observation noise.

[0035] Furthermore, in the method described above, in the state evolution equation, the state transfer matrix F is expressed as:

[0036] ;

[0037] in Indicates the influence coefficient of queue length on RTT deviation, Indicates the influence coefficient of RTT deviation on the proportion of high-frequency energy, Indicates the influence coefficient of RTT deviation on queue length, Represents the coefficient of the relationship between network load and short-term fluctuations, Indicates the influence coefficient of high-frequency energy ratio on RTT deviation, Indicates the influence coefficient of high-frequency energy ratio on queue length;

[0038] Control Input Matrix Expressed as:

[0039] ;

[0040] in Indicates the influence coefficient of control input on RTT deviation and the influence coefficient of rate on RTT deviation. Indicates the adjustment coefficient of rate to queue length, It represents the suppression coefficient of the rate to high frequency energy, represents the rate constraint coefficient on queue growth, represents the coupling coefficient of rate to frequency domain characteristics, Indicates the smoothing coefficient of rate to delay fluctuation;

[0041] In the observation equation, the observation matrix Expressed as:

[0042] ;

[0043] in 、 and They represent the influence coefficients of RTT deviation, queue length, and high-frequency energy ratio on RTT measurement value; 、 and They represent the influence coefficients of RTT deviation, queue length, and high-frequency energy ratio on the queue length measurement value.

[0044] Furthermore, in the method, in step 3, performing real-time network state estimation on the state vector based on the Kalman filter model includes the following steps:

[0045] Real-time network status estimation is performed through two stages: prediction and correction:

[0046] The prediction phase includes the prediction state and prediction error covariance matrix obtained by the following formula:

[0047] ;

[0048] ;

[0049] in Express The state prediction value at the moment, for The optimal state estimate at the previous moment;

[0050] Express The error covariance matrix of the moment prediction, express The error covariance matrix of the previous moment prediction, Superscript represents the matrix transpose, represents the covariance matrix of the process noise;

[0051] The correction phase involves first calculating the Kalman gain using the following formula, and then combining the observed data and the predicted state based on the Kalman gain to obtain the optimal state estimate:

[0052] ;

[0053] ;

[0054] in represents the Kalman gain, Superscript represents the matrix transpose, represents the covariance matrix of the observation noise;

[0055] Express The optimal state estimate of the network state at any moment, is the innovation quantity, which represents the difference between the observed value and the predicted value;

[0056] Then the error covariance matrix is ​​updated using the following formula:

[0057] ;

[0058] in Represents the updated error covariance matrix and serves as the next moment at time k To use, Represents the identity matrix.

[0059] Furthermore, in the method, in step 4, the process of dynamically adjusting the weights of RTT, queue length, and frequency domain features by using a Sigmoid function includes:

[0060] First, use the Sigmoid function to dynamically map the high-frequency energy ratio to the weight adjustment coefficient, thereby dynamically adjusting the RTT weight. Then, based on the RTT weight, further set the queue length weight and high-frequency energy ratio weight:

[0061] ;

[0062] ;

[0063] ;

[0064] in Indicates the weight of RTT, To adjust the sensitivity coefficient, is the threshold of high-frequency energy proportion, is the natural base;

[0065] Indicates the queue length weight;

[0066] Represents the weight of high-frequency energy.

[0067] Furthermore, in the method, in step 4, adaptively adjusting the parameters of the PID controller according to the weights includes:

[0068] The proportional gain of the PID controller is calculated based on the following equations: , integral gain and differential gain Three parameters are adjusted separately:

[0069] ;

[0070] ;

[0071] ;

[0072] in is the initial proportional gain, is the adjustment factor of the proportional gain;

[0073] is the initial integral gain, is the adjustment factor of the integral gain, represents the control error of the PID controller, is the error threshold, Indicates other situations;

[0074] is the initial differential gain, is the adjustment factor for the differential gain.

[0075] Furthermore, in the method, step 5 includes:

[0076] First, the control error is calculated based on the network estimated state :

[0077] ;

[0078] in, is the current real-time estimated round-trip delay, and are the target network delay and The queue length at the time is included in middle, is the target queue length;

[0079] Then based on To calculate the control output of the PID controller :

[0080] ;

[0081] in Represents the mixed error value at the historical moment, represents the control period, express The mixed error value of the previous moment;

[0082] Based on To adjust the sending rate :

[0083] ;

[0084] ;

[0085] in Indicates the adjustment value of the sending rate. It is the conversion factor that maps the PID control signal to the actual network control quantity. express The sending rate of the next moment, express The sending rate at the moment;

[0086] At the same time, real-time monitoring of the sending rate , when the sending rate Reduced to a preset threshold When , the ECN marking operation is triggered, marking the ECN field in the IP header of the data packet, so that the source node will actively reduce the sending rate after receiving the data packet with ECN marking, thereby completing the explicit congestion notification ECN.

[0087] The technical effect of the present invention is that the present invention combines frequency domain features with Kalman filtering, proposes a dynamic weight distribution method based on the proportion of high-frequency energy, and combines it with adaptive PID parameter adjustment to achieve accurate and real-time control of network status. By capturing instantaneous fluctuations and short-term changes in the network through frequency domain feature analysis, the system can dynamically adjust the control strategy according to the real-time network status, making the network control response more sensitive and accurate. The present invention can significantly improve the performance of data center networks in high-load and complex network environments, reduce delays, increase throughput, avoid data packet loss, and enhance the network's adaptability. By estimating the network status in real time and dynamically adjusting the PID controller parameters, the system can accurately respond to changes in various network states and ensure the stability and efficiency of network performance in complex environments. BRIEF DESCRIPTION OF THE DRAWINGS

[0088] Figure 1 Schematic diagram of the functions of each step of an embodiment of the present invention;

[0089] Figure 2 This is a schematic diagram of the frequency domain energy spectrum corresponding to the RTT jitter time domain signal according to an embodiment of the present invention, where (a) is the RTT jitter time domain signal diagram, and (b) is the corresponding frequency domain energy spectrum diagram.

[0090] Figure 3 This is a schematic diagram of the frequency domain energy spectrum corresponding to the queue length jitter time domain signal according to an embodiment of the present invention, where (a) is a diagram of the queue length jitter time domain signal, and (b) is a diagram of the corresponding frequency domain energy spectrum.

[0091] Figure 4 This is a flow chart of step 5 of an embodiment of the present invention. DETAILED DESCRIPTION

[0092] See also Figure 1 This embodiment provides an intelligent network congestion control method based on frequency domain feature analysis and dynamic weight adjustment, comprising the following steps:

[0093] Step 1: Collect network round-trip delay (RTT) and queue length data based on time, then segment the data into time domains according to a fixed window size and calculate the jitter value.

[0094] This step first obtains real-time time series data of RTT and queue length from the network. Specifically, the first signal collected is the change data of RTT and queue length. The data collection will be stored in the form of time series, that is, each data point corresponds to a timestamp. For example, when storing RTT data in time series, we get a sequence:

[0095] ;

[0096] Among them That is, the corresponding timestamp The specific network round-trip delay value.

[0097] For the queue length data, the same method is used to obtain a similar sequence:

[0098] ;

[0099] Assume that each sampling point The corresponding RTT is , then the RTT jitter is:

[0100] ;

[0101] Collect queue length data from network nodes (such as switches or routers) and calculate the fluctuation of queue length. Assume that the queue length at each sampling time is , then the queue length jitter is:

[0102] .

[0103] The collected time series data is then segmented into fixed time length windows. In the following steps, the data in each window will be Fourier transformed separately, which helps to analyze the frequency domain characteristics of the signal in different time periods.

[0104] Next, the RTT and queue length data are smoothed, that is, the mean and standard deviation are calculated to filter out outliers. Assume that the data window size is , the mean and standard deviation of RTT data are calculated as:

[0105] ;

[0106] ;

[0107] Similarly, calculate the mean and standard deviation of the queue length:

[0108] ;

[0109] ;

[0110] Based on the standard deviation, outliers can be filtered out:

[0111] ;

[0112] ;

[0113] For the above inequality or It is regarded as an abnormal point and removed from the current data window. This method can effectively suppress the interference of abnormally high jitter on frequency domain feature extraction.

[0114] Step 2: Perform discrete Fourier transform (DFT) on the jitter values ​​of each segment of RTT and queue length data to convert them into the frequency domain, and then calculate the proportion of high-frequency energy in the frequency domain.

[0115] In order to extract the frequency domain features of RTT / queue jitter, this embodiment uses discrete Fourier transform (DFT) to convert the time domain signal into the frequency domain.

[0116] For RTT data, its jitter sequence can be constructed for:

[0117] ;

[0118] For queue length data, its jitter sequence can be constructed for:

[0119] ;

[0120] in, Indicates the The end-to-end delay value (original RTT data) obtained at the sampling time, Represents the discrete sampling time point index (integer), Indicates the time domain jitter sequence (differential result) calculated based on RTT.

[0121] According to the collected jitter time series, its DFT can be expressed as:

[0122] ;

[0123] in is the complex frequency domain coefficient corresponding to the kth point of the spectrum, is the fixed window length, Indicates the jitter value of each data segment, that is, the jitter sequence of the RTT or queue length calculated above. is the natural base, represents the imaginary unit, represents pi, Indicates the first Sampling moments.

[0124] Through the spectrum The energy of the signal in different frequency bands can be calculated. Through Fourier transform, the jitter signal in the time domain is converted to the frequency domain, and the amplitude and phase information of each frequency component are obtained, thereby identifying the frequency characteristics of the signal.

[0125] After Fourier transform, the spectrum The energy of the signal in different frequency bands can be further calculated. In order to measure the proportion of high-frequency fluctuations in the signal, this embodiment proposes the proportion of high-frequency energy The definition of

[0126] ;

[0127] in , represents the actual frequency value corresponding to the frequency component, is the sampling frequency. This is the frequency threshold that determines the starting frequency of the high-frequency portion. Generally speaking, this value can be set based on the typical periodic characteristics of network traffic. Represents the energy at the kth frequency point.

[0128] This is the ratio of high-frequency component energy to total signal energy, reflecting the intensity of short-term fluctuations in the network. This metric characterizes the intensity of short-term, sudden fluctuations in network status. A larger value indicates more significant high-frequency jitter and a higher risk of transient network congestion.

[0129] In this step, the output high frequency energy ratio As a key feature, it can effectively reflect the dynamic change of network jitter. A high value indicates that there are more instantaneous changes in the network, and there may be sudden traffic or temporary congestion. A lower value indicates that the network is more stable and may be in a more stable working state.

[0130] Frequency domain features (such as high frequency energy ratio) are the input for the subsequent adaptive state estimation and dynamic weight control in this embodiment. Figure 2 and Figure 3 ,Under different RTT jitter time domain signals and queue length jitter time domain signals, there will be different frequency domain energy spectra, where Figure 2 The high-frequency energy in (b) accounts for 23.07%, Figure 3 The high-frequency energy in (b) accounts for 10.30%.

[0131] When performing adaptive state estimation, these frequency domain features are used as an important dimension of the system state. State estimation is performed through Kalman filtering and combined with other network states (such as RTT, queue length, etc.) to analyze and predict network behavior.

[0132] When implementing dynamic weight control, frequency domain characteristics (particularly the proportion of high-frequency energy) are used to adjust the control strategy. For example, when high-frequency energy is high, a stronger control response can be selected to cope with rapid fluctuations; while when low-frequency energy is high, the control strategy can be adjusted to better adapt to long-term network changes.

[0133] Traditional network performance monitoring methods mostly rely on time-domain analysis, such as directly monitoring statistical information like the average, maximum, and variance of RTT and queue length. This approach can reflect the general state of the network, but struggles to capture sudden, high-frequency transient fluctuations (such as sudden load increases and traffic floods). This embodiment, however, uses Fourier transform to convert time-domain signals into frequency-domain signals, effectively capturing these frequency characteristics, particularly high-frequency noise in the network.

[0134] Fourier transform can extract frequency domain features from time domain signals, thereby effectively identifying instantaneous fluctuations (high-frequency noise) and long-term trends (low-frequency changes) in the network.

[0135] RTT and queue length data can reflect key performance indicators in the network. Combined with Fourier transform, they can more accurately capture the frequency characteristics of fluctuations in the network.

[0136] Using high-frequency energy proportion as a feature can clearly classify high-frequency noise from low-frequency trends, which helps the network control system adopt appropriate strategies based on the actual fluctuations of the network.

[0137] Step 3: Construct a state vector containing RTT deviation, queue length, and high-frequency energy ratio. Then, perform real-time network state estimation on the state vector based on the Kalman filter model to obtain the network's RTT estimate, queue length estimate, and frequency domain feature estimate.

[0138] In this step, the state vector is expressed as:

[0139] ;

[0140] in, Indicates The state vector at time t, Indicates Network delay deviation at each moment; Indicates The queue length at the time; Indicates The proportion of high-frequency energy in the frequency domain characteristics at the moment. Reflect network delay fluctuations in real time and show the changing trend of network delay. Reflects the network load and congestion level. Describing the severity of short-term fluctuations and emergencies in the network, frequency domain characteristics clearly reflect the potential risk of instantaneous network congestion.

[0141] The network state estimation is then modeled as a linear dynamic system, and Kalman filtering is used to estimate and predict the system state in real time. The state space model consists of two major equations:

[0142] State evolution equation (time update):

[0143] ;

[0144] Where F is the state transfer matrix, express The state vector of the previous moment, B is the control input matrix, express The control input at the moment before the moment. express The process noise at the previous moment is assumed to be zero-mean Gaussian noise with a covariance of Q, which is used to describe the uncertainty in network state changes. For example, sudden traffic or device failure may cause drastic fluctuations in RTT or queue length. These changes cannot be directly predicted by the model, so process noise is introduced. .

[0145] In the state evolution equation, the state transfer matrix F represents the transition from Time has come The network status change at each moment is used to describe the natural change trend of the network status under uncontrolled conditions. Its elements can be based on simulation data fitting or empirical settings to reflect the coupling and inertia between RTT jitter, queue length and frequency domain energy.

[0146] The state transfer matrix F is expressed as:

[0147] ;

[0148] in, : The coefficient of influence of queue length on RTT deviation. Because network congestion usually increases queue length, which in turn leads to an increase in RTT, this term is a positive number. The value depends on the network topology, traffic characteristics and control strategy. Usually, this value can be estimated through experiments or simulation data. Assuming that the experimental data shows that for every 10 packets increase in queue length, the RTT deviation increases by about 20ms, then It can be expressed as 0.02 (that is, for every unit increase in queue length, the RTT deviation increases by 0.02).

[0149] : The influence coefficient of RTT deviation on high-frequency energy ratio. The change of RTT is closely related to the instantaneous fluctuation and frequency domain characteristics (high-frequency components) in the network. The high-frequency energy ratio is usually used to measure the short-term fluctuation in the network. When the high-frequency component is strong, the RTT deviation may fluctuate greatly. It can be assumed that there is a certain positive correlation between RTT deviation and high-frequency energy ratio. If the high-frequency energy increases by 10%, the RTT deviation increases by 5ms. A possible value is 0.05.

[0150] : The coefficient of RTT deviation on queue length. Generally, an increase in network load (increased RTT) will lead to an increase in queue length, so this is a positive number. According to the actual network load, if the RTT increases by 10ms, the queue length may increase by 5 packets, then A possible value is 0.5.

[0151] An increase in queue length typically indicates increased network load, which can lead to increased short-term fluctuations in the network, thus affecting high-frequency energy in the frequency domain. That is, when the queue length is large, the network may experience more transient changes. Reflects the relationship coefficient between network load and short-term fluctuation. It should be a positive value, indicating that an increase in queue length is usually accompanied by an increase in the proportion of high-frequency energy. If the queue length increases by 10 packets, the proportion of high-frequency energy in the network increases by 5%, then A possible value is 0.05.

[0152] : The coefficient of influence of high-frequency energy ratio on RTT deviation. Although the relationship between high-frequency energy and RTT is not as direct as queue length, there may still be some feedback effect. This factor can be determined based on actual data. It may be smaller because high-frequency fluctuations do not affect latency as directly as queue length. In some cases, It can be a small positive value, such as 0.02, which means that the impact of high-frequency energy on RTT is weak.

[0153] : The coefficient of influence of high-frequency energy ratio on queue length. Short-term load fluctuations are usually accompanied by changes in queue length, so this term is a positive number. Indicates the impact of high-frequency energy on queue length. It should be a positive value, indicating that high-frequency fluctuations may cause the queue length to increase. If the high-frequency energy increases by 10%, the queue length increases by 5 packets, then A possible value is 0.5.

[0154] Control Input Matrix , describes the impact of control input on the state, and represents the modulation effect of control input (such as sending rate adjustment) on the system state. It usually acts mainly on the queue state and indirectly affects the delay and jitter characteristics.

[0155] Control Input Matrix Expressed as:

[0156] ;

[0157] in, : The influence coefficient of rate on RTT deviation.

[0158] Control input (send rate adjustment) directly affects the deviation of network transmission delay. Increasing the send rate can reduce link congestion and thus reduce the RTT fluctuation. It should be a negative value, indicating that the rate increase is inversely related to the RTT deviation. According to experimental data: when the sending rate increases by 10Gbps, the RTT deviation decreases by an average of 2ms, then A possible value is -0.02. In a high-speed network environment, when the rate approaches the bandwidth limit, the adjustment effect of this coefficient increases by about 30%.

[0159] : The adjustment coefficient of rate to queue length.

[0160] Quantify the regulatory effect of the send rate on the switch buffer load. A reasonable increase in the send rate can speed up queue emptying, but excessive increase may cause transient congestion. This coefficient reflects the efficiency of rate adjustment in controlling queue length. Under normal circumstances, this coefficient is positive. Measured data shows that in a 25Gbps link, for every 1Gbps increase in the sending rate, the queue length decreases by an average of 8 packets (approximately 12%). A possible value is 0.08. When the queue utilization exceeds 85%, the regulation effect increases by 50%.

[0161] : The suppression coefficient of rate to high-frequency energy.

[0162] The effect of increasing the transmission rate on the high-frequency components of the network spectrum characteristics. A moderate increase in the transmission rate can smooth network traffic, reduce sudden fluctuations, and thus reduce the proportion of high-frequency energy. This coefficient should be a negative value, indicating that the increase in rate and high-frequency energy are inversely proportional. Experimental statistics show that for every 1Gbps increase in the transmission rate, the proportion of high-frequency energy decreases by an average of 0.15%. A possible value is -0.15. In high network load scenarios (>70% bandwidth usage), the suppression effect is improved by 40%.

[0163] : The rate constraint coefficient on queue growth.

[0164] The sending rate adjusts the ability to control the expansion of the buffer queue. Properly increasing the rate can suppress excessive queue growth and prevent buffer overflow. This coefficient should be negative, indicating that the rate increase can effectively control the queue growth. Verification data shows that: for every 1Gbps increase in rate, the queue growth rate decreases by an average of 7%. A possible value is -0.07. When the queue is close to full load (>90% capacity), the constraint effectiveness increases to 3 times the baseline value.

[0165] : Coupling coefficient of rate to frequency domain characteristics.

[0166] The dynamic correlation effect between the transmission rate adjustment and the change of network frequency domain characteristics. Although the rate increase can alleviate the congestion of the main frequency band, it will indirectly affect the high-frequency energy distribution through the change of traffic characteristics. This coefficient is usually positive, indicating that the increase in rate will slightly increase the high-frequency energy. The measured results show that: for every 1Gbps increase in the transmission rate, the proportion of high-frequency energy increases by an average of 0.03%, then A possible value is 0.03. Through joint optimization, the compensation efficiency can reach 92% in a typical network environment.

[0167] : Smoothing coefficient of rate to delay fluctuation.

[0168] The stabilizing effect of transmission rate optimization on transmission delay jitter. Reasonable adjustment of the transmission rate can suppress irregular fluctuations in network transmission, thereby smoothing RTT changes. This coefficient should be a negative value, indicating that the rate increase and delay fluctuation are inversely related. Experimental analysis shows that: for every 1Gbps increase in transmission rate, the RTT standard deviation decreases by an average of 0.04ms, then A possible value is -0.04. In bursty traffic scenarios, this coefficient can improve the smoothing effect by 60%; when the queue length reaches above 50% of the capacity, the smoothing effect increases linearly.

[0169] The state-space model also includes observation equations (measurement updates). The main task of the observation equation is to link the system's internal state (RTT deviation, queue length, and high-frequency energy ratio) with measurable network parameters (such as the actual measured RTT and queue length). This process effectively reduces prediction errors, allowing the system to better adapt to environmental changes. The observation equation is expressed as:

[0170] ;

[0171] in, express The actual observed value of RTT or queue length at time, is the observation matrix, represents the observation noise.

[0172] Observation matrix Expressed as:

[0173] ;

[0174] in, : The influence coefficient of RTT deviation on RTT measurement value.

[0175] The change of RTT deviation directly affects the measurement results of end-to-end delay. When network congestion increases, the increase of RTT deviation will significantly increase the RTT measurement value. Should be a positive value, indicating that an increase in RTT deviation generally leads to an increase in RTT measurement. If the RTT measurement increases by 4.9ms for every 5ms increase in RTT deviation, then A possible value is 0.98.

[0176] : The influence coefficient of queue length on RTT measurement value.

[0177] The increase in queue length indirectly affects RTT measurements through the buffer delay effect. When the switch buffer is backed up, the packet queuing delay increases, causing the RTT measurement to increase. Typically, Should be a positive value, indicating that an increase in queue length will be accompanied by an increase in RTT measurement. If the RTT measurement increases by 0.5ms for every 10 packets added to the queue length, then A possible value is 0.05.

[0178] : The influence coefficient of high-frequency energy ratio on RTT measurement value.

[0179] High-frequency energy fluctuations reflect network bursts and react to delay measurements through congestion control mechanisms. Short-term traffic bursts trigger rapid congestion responses, which may suppress RTT growth. It should be a negative value, indicating that the increase in high-frequency energy will suppress the RTT fluctuation amplitude. If the RTT measurement value decreases by 0.3ms for every 10% increase in high-frequency energy, then A possible value is -0.03.

[0180] : The coefficient of influence of RTT deviation on queue measurement value.

[0181] The increase in RTT deviation indicates an increase in network load, which is indirectly reflected in the queue accumulation trend. The increase in transmission delay will cause the data packet to stay in the buffer for a longer time. Should be a positive value, indicating that an increase in RTT deviation is often accompanied by an increase in queue length. If the queue measurement value increases by 0.6 packets for every 5ms increase in RTT deviation, then A possible value is 0.12.

[0182] : Queue status to queue measurement value mapping coefficient.

[0183] The actual state of the switch buffer directly determines the queue measurement value. This coefficient includes correction factors such as measurement device error and system response delay. Slightly greater than 1.0 indicates that the queue measurement slightly exaggerates the actual state. If the queue state is 100 packets and the measurement shows 102 packets, then A possible value is 1.02.

[0184] : Correlation coefficient of high-frequency energy proportion to queue measurement value.

[0185] Traffic bursts caused by high-frequency energy surges can cause instantaneous accumulation in the grid buffer. Short-term impacts of burst packets can cause instantaneous fluctuations in queue depth. Should be a positive value, indicating that an increase in high-frequency energy is accompanied by a temporary increase in the queue measurement value. If the queue measurement value temporarily increases by 1.2 packets for every 15% increase in high-frequency energy, then A possible value is 0.08.

[0186] The measurement matrix is ​​used to describe how the state vector is mapped to the observation space. For network state estimation, the measurement matrix It can reflect the impact of different states on the observed quantity. For example, assuming the observed quantity includes: the actual measured RTT ( ) and the actual measured queue length ( ), in this case, the observation equation can be written as:

[0187] ;

[0188] In this step, real-time network state estimation is performed based on the Kalman filter model through two stages: prediction and correction:

[0189] 1. Forecast (time update).

[0190] In the prediction phase, the current state and error covariance are predicted based on the system's dynamic model (i.e., the state transition equation). This process does not rely on observed data, but rather relies solely on the system's previous state estimate and control inputs.

[0191] Prediction status:

[0192] ;

[0193] in, It is predicted based on the state estimation and control input at the previous moment The state prediction value at the moment is the prior estimation of the state by the system model before the current observation data is obtained. for The optimal state estimate at the previous moment.

[0194] Forecast error covariance matrix:

[0195] ;

[0196] in, Express The error covariance matrix of the moment prediction, express The error covariance matrix of the previous moment prediction, Superscript represents the matrix transpose, Represents the covariance matrix of the process noise.

[0197] 2. Calibration (measurement update).

[0198] The goal of the correction phase is to correct the prediction results by integrating the observed data. The Kalman filter compares the difference between the predicted network state and the actual observed data and uses the Kalman gain to adjust the state estimate to obtain a more accurate network state.

[0199] ;

[0200] in: .

[0201] Kalman gain Used to balance forecast uncertainty and the weights of the covariance matrix R of the observation noise, where R is a statistical description of the observation noise. If the prediction error is large and the observation noise is small, the Kalman gain will be higher, indicating that the system is more dependent on the observations.

[0202] Through the Kalman gain, the Kalman filter combines the observed data and the predicted state to update the state estimate. The updated state estimate yes The optimal estimate of the network state before time, combining predicted and observed information.

[0203] ;

[0204] in, Is the innovation, which represents the difference between the observed value and the predicted value. The Kalman filter updates the state estimate by correcting the innovation.

[0205] When updating the state estimate, the error covariance matrix also needs to be updated , to reflect the impact of observation data on the accuracy of state estimation. The updated error covariance matrix is ​​usually smaller, indicating that the uncertainty of state estimation is reduced after the introduction of observation data.

[0206] ;

[0207] is the updated error covariance matrix, which represents the uncertainty of the updated state estimate and serves as The next moment of the moment to use.

[0208] By combining prediction and correction steps, the Kalman filter effectively and dynamically estimates RTT deviation, queue length, and high-frequency energy content in network state estimation. Using the Kalman gain, the Kalman filter adjusts state estimates based on prediction error and observation noise, ensuring that network control systems can adapt to changes in network state in real time and maintain efficient performance in uncertain environments. This recursive estimation process enables congestion control schemes to precisely control network state, thereby optimizing network throughput, latency, and stability.

[0209] Step 4: Based on the proportion of high-frequency energy, the weights of RTT, queue length, and frequency domain features are dynamically adjusted through the Sigmoid function, and then the parameters of the PID controller are adaptively adjusted based on the weights.

[0210] The main function of this step is to dynamically adjust the weights of the various state variables in the network control strategy based on the network state estimation results output by the previous step, thereby achieving refined congestion control and network state optimization. This step comprehensively assesses the network congestion status by analyzing RTT deviation, queue length, and the proportion of high-frequency energy in the frequency domain. Based on this assessment, the weights of the various state variables in the control strategy are dynamically assigned. This step also includes a PID parameter adaptive adjustment mechanism, which uses dynamic weight information to optimize the parameters of the PID controller in real time, effectively reducing network latency, reducing queue backlogs, and improving the overall stability and performance of the network.

[0211] In this step, the weights are first dynamically adjusted based on the real-time relationship between the high-frequency energy ratio and the network status:

[0212] First calculate the dynamic weight distribution coefficient:

[0213] Use the sigmoid function to dynamically map the high-frequency energy ratio to the weight adjustment coefficient:

[0214] ;

[0215] in, Indicates the weight of RTT, is the natural base, To adjust the sensitivity coefficient, control the steepness of the sigmoid function (for example, a value of 5). It is the threshold of the high-frequency energy ratio. When the high-frequency energy ratio exceeds this threshold, the RTT weight will increase significantly, indicating that the instantaneous fluctuation in the network is large and the RTT needs to be adjusted faster to reduce the delay.

[0216] Correspondingly, the queue length weight and high-frequency energy proportion weight are set as:

[0217] ;

[0218] ;

[0219] This setting ensures that the sum of the RTT weight and the queue length weight remains 1, and the high-frequency energy weight reaches its maximum when the two weights are multiplied, which enhances the rapid response to instantaneous network congestion. When the network is stable ( When the queue length is low, the queue length weight increases, focusing more on long-term congestion control and load balancing. This dynamic weight adjustment mechanism enables network control policies to intelligently adjust response strength in different congestion scenarios.

[0220] The smooth adjustment characteristics of the Sigmoid function ensure that the changes in weights are not too drastic, thus avoiding drastic fluctuations in the control strategy when the network changes rapidly, and improving the stability of the system. This smoothness also improves the robustness of the control system, and it can operate efficiently in different network environments.

[0221] Then, according to the results of dynamic weight distribution, the PID controller (proportional-integral-derivative controller) parameters ( ) of the adaptive adjustment:

[0222] Proportional gain ( )Adjustment:

[0223] ;

[0224] in, is the initial proportional gain. is the adjustment factor for the proportional gain. It is the RTT control weight that is dynamically adjusted based on the high-frequency energy ratio (frequency domain characteristics), reflecting the severity of the RTT deviation in the network.

[0225] Proportional gain ( ) reflects the impact of the current error on the output control signal. For network control, when the network delay (RTT) deviation is large, the proportional gain needs to be adjusted quickly to quickly respond to network fluctuations. The adaptive adjustment of the proportional gain is based on the RTT control weight. As the network RTT changes, Adjustments are made based on the adjusted weights to enhance responsiveness to latency fluctuations.

[0226] Integral gain ( )Adjustment:

[0227] Integral gain ( ) Controlling the cumulative effect of past errors helps eliminate the system's steady-state errors. Over-reliance on integral control can lead to overshoot and system oscillation. Therefore, the integral gain needs to be adjusted appropriately when the network load is stable.

[0228] Adaptive adjustment of the integral gain is achieved by controlling the weight of the queue length. When the queue length (i.e., network load) is high, the system requires a stronger integral response to eliminate steady-state errors. When the queue is small or the system is stable, the integral gain should be appropriately reduced.

[0229] When the network error exceeds a certain threshold, the integral term becomes effective:

[0230] ;

[0231] is the initial integral gain. is the adjustment factor. It is the control weight of the queue length, reflecting the load and network stability. Is an error threshold, and the integral term will only take effect when the error exceeds this threshold.

[0232] Differential Gain ( )Adjustment:

[0233] Differential Gain ( ) controls the rate of change of the control error, used to reduce system overshoot and oscillation. In a network environment, the high-frequency energy percentage reflects instantaneous fluctuations in the network. Adjusting the differential gain can help the system suppress the impact of these short-term fluctuations on the control output. Adaptive adjustment of the differential gain is achieved based on the high-frequency energy percentage. When the network experiences large instantaneous fluctuations (for example, a high high-frequency energy percentage), the differential gain needs to be increased to reduce the impact of short-term fluctuations on network control.

[0234] The adjustment formula for the differential gain is:

[0235] ;

[0236] is the initial differential gain, is the adjustment factor for the differential gain.

[0237] The key innovation of this step lies in the real-time dynamic adjustment of the weights of various state variables in the network congestion control strategy based on frequency domain characteristics (the proportion of high-frequency energy). This overcomes the problem of traditional fixed-weight control strategies being insufficiently responsive to network changes and achieves highly adaptable congestion control. Furthermore, the adaptive adjustment mechanism of PID parameters utilizes dynamic weights to optimize PID parameters, enabling the controller to adjust control strength in real time within complex network environments. This effectively reduces network latency, balances network load, and improves overall network throughput and performance stability. This high adaptability and real-time optimization capability significantly enhances the robustness, real-time performance, and accuracy of the network congestion control system.

[0238] Step 5: Convert the PID control signal generated by the PID controller into a sending rate adjustment action or an explicit congestion notification (ECN), thereby suppressing network congestion in real time.

[0239] See also Figure 4 In this step, the PID control signal It is the core data directly input to the network execution layer, which is generated by the PID controller of the dynamic weight control layer. The specific calculation method of the PID control signal is as follows:

[0240] First, define the control error , based on the estimated state of the network, is defined as follows:

[0241] ;

[0242] in, is the current real-time estimated round-trip delay (in milliseconds), calculated from network status data using Kalman filtering. .

[0243] The target latency expected by the network, set according to the network service level agreement (SLA). The current real-time estimated queue length (unit: number of packets) is obtained by monitoring the switch buffer status. The target queue length (unit: number of packets) is dynamically configured based on the switch buffer capacity and is usually set to the queue length under ideal network load (such as the queue length under low load). , , is the dynamic weight coefficient, which is dynamically determined by the Sigmoid function driven by the frequency domain features.

[0244] The nonlinear PID controller specifically calculates the control output The formula is as follows:

[0245] ;

[0246] in Represents the mixed error value at the historical moment, represents the control period, express Mixed error value at the previous moment; PID parameters , , Real-time optimization and adjustment are achieved through the adaptive PID regulator to ensure stability and rapid responsiveness when network status changes. Denotes the mixed error value at the historical moment, denotes the mixed error value at the jth control cycle. Δt denotes the control cycle, which is a fixed time interval of the discrete control system.

[0247] In this step, the actual action of sending rate adjustment is to control the output of PID Mapped to the change of the sending rate of a specific network device. The actual adjustment rules of the sending rate are:

[0248] Assume that the network device is The sending rate at this moment is (unit: Mbps), the rate adjustment strategy at the next moment is:

[0249] ;

[0250] and According to the PID control signal To determine:

[0251] ;

[0252] Among them: coefficient It is the conversion factor that maps the PID control signal to the actual network control quantity (the conversion factor is generally determined based on the network characteristics and device bandwidth range).

[0253] The negative sign indicates that when the PID output is positive, it means that the current network condition is good, the error is reduced, and the network can increase the sending rate; when the PID output is negative, it means that the network condition is deteriorating and the sending rate needs to be reduced to alleviate congestion.

[0254] For example:

[0255] like =5, γ=0.1, the sending rate decreases by 0.5Mbps, that is, the rate becomes slower;

[0256] like =−4, γ=0.1, the sending rate increases by 0.4 Mbps, that is, the rate is improved.

[0257] In the actual control process, if the sending rate is found to be lower than a certain threshold (such as lower than the set threshold , which means that the network is in a relatively serious congested state. At this time, the network execution layer will trigger the ECN marking operation.

[0258] ECN mark trigger conditions:

[0259] ;

[0260] When ECN marking is triggered, the network device marks the ECN field in the IP header of the data packet (usually set to 11).

[0261] After receiving a data packet with ECN marking, the source node will actively reduce the sending rate (such as the CWR mechanism in TCP) to alleviate network congestion at the source.

[0262] Those skilled in the art will appreciate that the embodiments of the present application can be provided as methods, systems, or computer program products. Therefore, the present application can adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Moreover, the present application can adopt the form of a computer program product implemented on one or more computer-readable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code. The present application is a flow chart according to the method, device (system), and computer program product of the embodiment of the present application and / or the instructions executed by the processor to generate a device for realizing the function specified in one flow chart or multiple flows and / or one box or multiple boxes of the block diagram. These computer program instructions can also be stored in a computer-readable memory that can guide a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer-readable memory generate a product comprising an instruction device, which realizes the function specified in one flow chart or multiple flows and / or one box or multiple boxes of the block diagram. These computer program instructions may also be loaded onto a computer or other programmable data processing device so that a series of operational steps are executed on the computer or other programmable device to produce a computer-implemented process, whereby the instructions executed on the computer or other programmable device provide steps for implementing the functions specified in one or more processes in the flowchart and / or one or more boxes in the block diagram.

[0263] It should be emphasized that the examples of the present invention are illustrative rather than restrictive, and therefore the present invention is not limited to the examples in the specific embodiments. Any other embodiments derived by those skilled in the art based on the technical solution of the present invention that do not depart from the purpose and scope of the present invention, whether modified or replaced, also fall within the scope of protection of the present invention.

Claims

1. An intelligent network congestion control method based on frequency domain feature analysis and dynamic weight adjustment, characterized in that: The following steps are involved: Step 1: Collect network round-trip delay (RTT) and queue length data based on time, then segment the data into time domain segments according to a fixed window size and calculate the jitter value. Step 2: Perform discrete Fourier transform (DFT) on the jitter values ​​of each segment of RTT and queue length data to convert them into the frequency domain, and then calculate the proportion of high-frequency energy in the frequency domain. Step 3: Construct a state vector containing RTT deviation, queue length, and high-frequency energy ratio. Then, use the Kalman filter model to perform real-time network state estimation on the state vector to obtain the estimated RTT, queue length, and frequency domain characteristics of the network. Step 4: Based on the proportion of high-frequency energy, the weights of RTT, queue length, and frequency domain features are dynamically adjusted using the Sigmoid function. The PID controller parameters are then adaptively adjusted based on the weights. Step 5: Convert the PID control signal generated by the PID controller into a sending rate adjustment action or an explicit congestion notification (ECN), thereby suppressing network congestion in real time. The step 2 includes: Perform discrete Fourier transform (DFT) on the jitter value of each segment of data according to the following formula: ； in is the complex frequency domain coefficient corresponding to the kth point of the spectrum, is the fixed window length, Indicates the jitter value of each segment of data, is the natural base, represents the imaginary unit, represents pi, Indicates the first Sampling moments; Then calculate the proportion of high-frequency energy in the frequency domain according to the following formula: ； in It is the set frequency threshold, which determines the starting frequency of the high frequency part. represents the energy at the kth frequency point, It is the ratio of high-frequency component energy to the total signal energy, that is, the proportion of high-frequency energy; In step 4, the process of dynamically adjusting the weights of RTT, queue length, and frequency domain features using the Sigmoid function includes: First, use the Sigmoid function to dynamically map the high-frequency energy ratio to the weight adjustment coefficient, thereby dynamically adjusting the RTT weight. Then, based on the RTT weight, further set the queue length weight and high-frequency energy ratio weight: ； ； ； in Indicates the weight of RTT, To adjust the sensitivity coefficient, is the threshold of high-frequency energy proportion, is the natural base, Indicates The proportion of high-frequency energy in the frequency domain characteristics of the moment; Indicates the queue length weight; Indicates the weight of high-frequency energy proportion; In step 4, adaptively adjusting the parameters of the PID controller according to the weights includes: The proportional gain of the PID controller is calculated based on the following equations: , integral gain and differential gain Three parameters are adjusted separately: ； ； ； in is the initial proportional gain, is the adjustment factor of the proportional gain; is the initial integral gain, is the adjustment factor of the integral gain, represents the control error of the PID controller, is the error threshold, Indicates other situations; is the initial differential gain, is the adjustment factor for the differential gain.

2. The method according to claim 1, characterized in that The step 1 comprises: Collect RTT and queue length data in the form of time series to obtain the specific values ​​of RTT and queue length corresponding to each time point; The data is segmented according to a fixed time length, i.e., a window. The difference between adjacent data in each segment is then calculated, and the calculated differences are used as the jitter value of this segment of data.

3. The method according to claim 1, characterized in that The step 1 further includes the step of filtering outliers from the jitter value of each segment of data: Calculate the mean of the data in each segment, and then calculate the standard deviation based on the mean. Then determine the threshold of the jitter value of this segment of data based on the standard deviation. Jitter values ​​greater than the threshold are judged as outliers and deleted.

4. The method according to claim 1, wherein In step 3, the state vector is expressed as: ； in, Indicates The state vector at time t, Indicates Network delay deviation at each moment; Indicates The queue length at the time; Indicates The proportion of high-frequency energy in the frequency domain characteristics of the moment; The state of the network is then treated as a linear dynamic system and represented by the following state evolution equation and observation equation: ； ； Where F is the state transfer matrix, express The state vector of the previous moment, B is the control input matrix, express The control input at the previous moment, express The process noise at the previous moment; express The actual observed value of RTT or queue length at time, is the observation matrix, represents the observation noise.

5. The method according to claim 4, characterized in that In the state evolution equation, the state transfer matrix F is expressed as: ； in Indicates the influence coefficient of queue length on RTT deviation, Indicates the influence coefficient of RTT deviation on the proportion of high-frequency energy, Indicates the influence coefficient of RTT deviation on queue length, Represents the coefficient of the relationship between network load and short-term fluctuations, Indicates the influence coefficient of high-frequency energy ratio on RTT deviation, Indicates the influence coefficient of high-frequency energy ratio on queue length; Control Input Matrix Expressed as: ； in Indicates the influence coefficient of control input on RTT deviation and the influence coefficient of rate on RTT deviation. Indicates the adjustment coefficient of rate to queue length, It represents the suppression coefficient of the rate to high frequency energy, represents the rate constraint coefficient on queue growth, represents the coupling coefficient of rate to frequency domain characteristics, Indicates the smoothing coefficient of rate to delay fluctuation; In the observation equation, the observation matrix Expressed as: ； in 、 and They represent the influence coefficients of RTT deviation, queue length, and high-frequency energy ratio on RTT measurement value; 、 and They represent the influence coefficients of RTT deviation, queue length, and high-frequency energy ratio on the queue length measurement value.

6. The method according to claim 4, characterized in that In step 3, performing real-time network state estimation on the state vector based on the Kalman filter model includes the following steps: Real-time network status estimation is performed through two stages: prediction and correction: The prediction phase includes the prediction state and prediction error covariance matrix obtained by the following formula: ； ； in Express The predicted state value at the moment, for The optimal state estimate at the previous moment; Express The error covariance matrix of the moment prediction, express The error covariance matrix of the previous moment prediction, Superscript represents the matrix transpose, represents the covariance matrix of the process noise; The correction phase involves first calculating the Kalman gain using the following formula, and then combining the observed data and the predicted state based on the Kalman gain to obtain the optimal state estimate: ； ； in represents the Kalman gain, Superscript represents the matrix transpose, represents the covariance matrix of the observation noise; Express The optimal state estimate of the network state at any moment, is the innovation quantity, which represents the difference between the observed value and the predicted value; Then the error covariance matrix is ​​updated using the following formula: ； in Represents the updated error covariance matrix and serves as the next moment at time k To use, Represents the identity matrix.

7. The method according to claim 1, characterized in that The step 5 comprises: First, the control error is calculated based on the network estimated state : ； in, is the current real-time estimated round-trip delay, and are the target network delay and The queue length at the time is included in middle, is the target queue length; Then based on To calculate the control output of the PID controller : ； in Represents the mixed error value at the historical moment, represents the control period, express The mixed error value of the previous moment; Based on To adjust the sending rate : ； ； in Indicates the adjustment value of the sending rate. It is the conversion factor that maps the PID control signal to the actual network control quantity. express The sending rate of the next moment, express The sending rate at the moment; At the same time, real-time monitoring of the sending rate , when the sending rate Reduced to a preset threshold When , the ECN marking operation is triggered, marking the ECN field in the IP header of the data packet, so that the source node will actively reduce the sending rate after receiving the data packet with ECN marking, thereby completing the explicit congestion notification ECN.

Images