Multi-level network congestion control method for intelligent computing center

Abstract

The present invention discloses a multi-level network congestion control method for an intelligent computing center, and relates to the field of network communication technology. The method comprises the following steps: real-time monitoring of the traffic state parameters of each level network of the intelligent computing center; calculating the congestion coefficient of each level respectively based on the traffic state parameters of each level network; setting the dynamic congestion threshold of each level network; when the single-layer congestion coefficient exceeds the dynamic congestion threshold, performing congestion control on each level network respectively; when the congestion coefficient of two or more levels exceeds the dynamic congestion threshold, performing cross-level collaborative optimization. The method of the present invention uses single-layer independent congestion control and multi-level joint congestion control to quickly resolve the congestion of a single-layer network, and when all multi-level networks are congested, the overall network performance is optimized to a higher level through joint optimization. The present invention uses cross-level collaborative optimization to better coordinate the network conditions of the remaining levels when the intelligent computing center network is globally congested, and effectively balance sensitivity and stability.

Description

Technical Field

[0001] The present invention relates to the field of network communication technology, and in particular to a multi-level network congestion control method for an intelligent computing center. Background Art

[0002] With the rapid development of intelligent computing and cloud computing technologies, intelligent computing centers, as a key platform for centralized management and scheduling of various resources, play a key role in data processing, task scheduling, and resource utilization efficiency. With the rapid development of information technology and the deepening of digital transformation, computing network scenarios are becoming increasingly diverse and segmented, covering cloud interconnection, large-scale model applications, intelligent applications, the Internet of Things, and other fields. These diverse application scenarios place varying demands on computing network infrastructure, leading to increasingly differentiated capabilities. These diverse application scenarios place varying and complex demands on computing network infrastructure, leading to increasingly differentiated capabilities. This, in turn, requires that the design, deployment, and optimization of these infrastructures fully consider the specific application scenario requirements to achieve more efficient, flexible, and customized service support.

[0003] The congestion control service continuously monitors the congestion of the network and dynamically adjusts the sender's data transmission rate based on the overall performance of the network, including throughput, latency, packet loss rate and other indicators, so that the network load remains within a reasonable range. However, the network traffic of the intelligent computing center is highly time-varying. When the traffic suddenly stops transmitting or a large amount of new traffic is added to the transmission queue, the existing rule-based congestion control algorithm lacks the ability to learn from historical experience and is difficult to accurately adjust the congestion control strategy. Existing technologies lack the ability to dynamically adjust paths from a global perspective. For example, multi-path load balancing (such as ECMP) is prone to flow conflicts when the hashing is uneven, exacerbating congestion.

[0004] For example, the Chinese patent with publication number CN118413489A discloses a network congestion scheduling method and system for an AI intelligent computing center, the method comprising: step 1: copying messages and distinguishing between service messages and control messages, step 2: selecting a queue, step 3: updating the queue length and calculating the average queue length, step 4: matching the drop probability with the average queue length, step 5: storing the drop probability in the inlet pipe, step 6: distinguishing between admitted traffic and non-admitted traffic, and step 7: discarding non-admitted traffic according to probability. This invention decouples the congestion detection and feedback mechanisms of queue management, generates control messages by copying, implements congestion detection by service messages, and implements feedback on drop probability by control messages, thus ensuring the normal forwarding of service messages and achieving fine-grained feedback.

[0005] For example, Chinese patent application CN114760252B discloses a data center network congestion control method and system. The system comprises: periodically obtaining the current total length of data packets sent by each terminal connected to the data center network to the same destination terminal; if the total length of the data packets sent by the destination terminal exceeds the congestion threshold corresponding to the destination terminal, generating a congestion notification message for the destination terminal; and sending the congestion notification message to each terminal connected to the data center network, so that the terminal obtains the corresponding pause transmission time and pauses the transmission of data packets within the timing period of the pause transmission time. This patent can significantly shorten the feedback time of congestion signals, allowing the terminal as the data packet sender to respond to network congestion more quickly, effectively reduce the transmission of additional data packets during the convergence process, reduce the queue accumulation of switch ports, and thus improve the efficiency and reliability of the data center network congestion control process.

[0006] The above patents all have the problems raised by this background technology: the network traffic of the intelligent computing center is highly time-varying, and the number of nodes in each layer of the network is large. When encountering global network congestion and single-layer network congestion, it cannot be well hierarchically controlled, resulting in poor timeliness of congestion control and difficulty in coping with the large-scale network architecture of the intelligent computing center. Summary of the Invention

[0007] The technical problem to be solved by the present invention is to address the deficiencies of the existing technology and provide a multi-level network congestion control method for intelligent computing centers.

[0008] In order to achieve the above object, the technical solution adopted by the present invention is:

[0009] A multi-level network congestion control method for intelligent computing centers includes the following steps:

[0010] Real-time monitoring of traffic status parameters at all levels of the intelligent computing center network, including the edge access layer, aggregation layer, and core layer;

[0011] Calculate the congestion coefficient of each layer based on the network traffic status parameters of each layer;

[0012] Set dynamic congestion thresholds for each layer of the network;

[0013] When the congestion coefficient of a single layer exceeds the dynamic congestion threshold, congestion control is performed on each layer of the network separately;

[0014] When the congestion coefficients of two or more layers exceed the dynamic congestion threshold, cross-layer collaborative optimization is performed.

[0015] Furthermore, the congestion coefficient of each layer is calculated using the following formula:

[0016]

[0017] Among them, E represents the congestion index of the edge access layer, A represents the congestion index of the aggregation layer, C represents the congestion index of the core layer, and BW m and BW c They represent the peak bandwidth in the last 10ms window and the currently allocated link bandwidth, Q d and Q m Represents the current queue depth and maximum queue depth respectively, ▽P l Indicates the gradient of packet loss rate change, RTT n Indicates the normalized round-trip time RTT value, N Q Indicates queue imbalance, CE indicates the proportion of ENC congestion markings, ΔDelay and Delay indicate the delay change in the last second and the current delay respectively, μ indicates delay variance, and BU indicates bandwidth utilization. Indicates the average bandwidth utilization, F c Indicates the flow collision rate of ECMP hash collisions.

[0018] Furthermore, the dynamic congestion threshold of each layer of the network is specifically formulated as follows:

[0019]

[0020] Among them, T E represents the dynamic congestion threshold of the edge access layer, T A represents the dynamic congestion threshold of the aggregation layer, T C Indicates the core layer dynamic congestion threshold, E h represents the historical ECI mean for the same period, t represents hourly time, i represents the i-th hour, σ E Indicates the ECI standard deviation in the last hour, W indicates the sliding window size, represents the queue depth and σ in the sliding window Q represents the standard deviation of the queue depth in the sliding window, BW represents the total bandwidth, λ(o) represents the load factor during the period, σ D represents the latency variance in the last hour, and ▽BW represents the bandwidth fluctuation rate in the last hour.

[0021] Furthermore, the congestion control is performed on each layer of the network, specifically including: dynamically allocating bandwidth on the edge access layer; adjusting weights on priority queues on the convergence layer; and performing multi-path optimization on the core layer.

[0022] Furthermore, the dynamic allocation of bandwidth of the edge access layer specifically includes the following steps:

[0023] Prioritize traffic based on service type;

[0024] Bandwidth is dynamically allocated based on priority marking. The specific formula is:

[0025]

[0026] Among them, BW x Indicates the bandwidth allocated to the current service type, N indicates the total number of service types, P x Indicates the priority of the current service type, P y represents the priority of the yth service type, G represents the GPU utilization, and α=0.8 represents the adjustment factor.

[0027] Furthermore, the weight of the priority queue of the convergence layer is adjusted, and the specific formula is:

[0028]

[0029] Among them, M m Indicates the weight of the current queue, M indicates the total number of queues, Indicates the priority of the current queue. Indicates the priority of the nth queue, Q m Indicates the current queue delay, Q n represents the delay of the nth queue, and k=0.1 represents the delay sensitivity coefficient.

[0030] Furthermore, the multipath weight optimization of the core layer specifically includes:

[0031] Calculate the path weight. The specific formula is:

[0032]

[0033] Among them, R d Indicates the weight of the current path, R indicates the total number of queues, BWR d Indicates the bandwidth of the current path, BWR g represents the bandwidth of the g-th path, D d Indicates the current path delay, D g represents the j-th path delay;

[0034] The path weight is updated every 10ms.

[0035] Furthermore, the cross-level collaborative optimization specifically includes the following steps:

[0036] Obtain the network status characteristics of the edge access layer, the queue timing characteristics of the aggregation layer, and the topology characteristics of the core layer respectively;

[0037] The network status features of the edge access layer, the queue timing features of the aggregation layer, and the topology features of the core layer are embedded into multimodal features. The specific formula is:

[0038] Z=Concat(MLP(X E ),TCN(X A ),GAT(X C ))

[0039] Among them, Z represents multimodal features, Concat represents connection function, MLP represents fully connected layer, TCN represents dilated convolution layer, GAT represents graph attention, X E 、X A and X C They represent the network status feature matrix of the edge access layer, the queue timing feature matrix of the aggregation layer, and the topology feature matrix of the core layer respectively;

[0040] Input multimodal features into the spatiotemporal graph convolutional network;

[0041] The target loss function of the spatiotemporal graph convolutional network is optimized. The specific formula is:

[0042] L=0.5L Q +0.3L f +0.2L c

[0043] Among them, L represents the target loss function, L Q Denotes the service quality loss function, L f represents the fairness loss function, L c represents the consistency loss function;

[0044] Output the edge access layer bandwidth allocation matrix, aggregation layer queue weight matrix and core layer path weight matrix respectively.

[0045] Furthermore, the service quality loss function L Q The specific formula is:

[0046]

[0047] Among them, D pred and D real Denote the predicted delay and actual delay respectively, B pred and B real They represent the predicted bandwidth utilization and the actual bandwidth utilization, respectively, ||·||2 represents the L2 norm, and cs represents the cosine similarity function;

[0048] The fairness loss function L f The specific formula is:

[0049]

[0050] Among them, BW a and BWb Represents the bandwidth of the ath and bth service types respectively, P a and P b Respectively represent the priorities of the ath and bth service types;

[0051] The consistency loss function L c The specific formula is:

[0052]

[0053] Among them, ▽C and ▽E represent the congestion coefficient gradient of the edge access layer and the core layer, respectively. represents the time derivative of the congestion coefficient of the aggregation layer, and ||·||1 represents the L1 norm.

[0054] Compared with the prior art, the present invention has the following beneficial effects:

[0055] 1. The present invention provides a dynamic parameter adjustment mechanism. During the congestion coefficient calculation process, adaptive parameters are set according to historical traffic data, which can make the congestion calculation more sensitive to burst traffic.

[0056] 2. The present invention implements hierarchical control over the intelligent computing center network. When a single-layer network is congested, it can quickly perform scheduling to prevent the congestion from spreading to other layers.

[0057] 3. Through cross-level collaborative optimization, the present invention can better coordinate with the network conditions of other levels when the intelligent computing center network is globally congested, and obtain collaborative optimization results through neural networks; at the same time, the loss function of the spatiotemporal graph convolutional network is targetedly optimized based on the network characteristics of the intelligent computing center, making the output results more accurate and stable.

[0058] 4. The present invention adopts a triple design of basic value hierarchical calculation, dynamic real-time correction and cross-level collaborative optimization, which effectively balances sensitivity and stability while ensuring evaluation accuracy, and can accurately adapt to the complex scenarios of heterogeneous traffic concurrency in the intelligent computing center. BRIEF DESCRIPTION OF THE DRAWINGS

[0059] Other features, objects and advantages of the present invention will become more apparent upon reading the detailed description of non-limiting embodiments made with reference to the following drawings:

[0060] Figure 1 A schematic diagram of a flow chart of an embodiment of the present invention;

[0061] Figure 2 This is a diagram of a cross-level collaborative optimization structure according to an embodiment of the present invention;

[0062] Figure 3 This is a network topology diagram of the intelligent computing center according to an embodiment of the present invention;

[0063] Figure 4 This is a structural diagram of the spatiotemporal convolutional network according to an embodiment of the present invention. DETAILED DESCRIPTION

[0064] In order to make the objectives, technical solutions and advantages of the present invention more clear, the present invention is described in detail below with reference to the accompanying drawings and specific embodiments.

[0065] like Figure 1 As shown in FIG, a multi-level network congestion control method for an intelligent computing center includes the following steps:

[0066] Real-time monitoring of traffic status parameters at all levels of the intelligent computing center network, including the edge access layer, aggregation layer, and core layer;

[0067] Calculate the congestion coefficient of each layer based on the network traffic status parameters of each layer;

[0068] Set dynamic congestion thresholds for each layer of the network;

[0069] When the congestion coefficient of a single layer exceeds the dynamic congestion threshold, congestion control is performed on each layer of the network separately;

[0070] When the congestion coefficients of two or more layers exceed the dynamic congestion threshold, cross-layer collaborative optimization is performed.

[0071] The specific formula for calculating the congestion coefficient of each layer is as follows:

[0072]

[0073] Among them, E represents the congestion index of the edge access layer, A represents the congestion index of the aggregation layer, C represents the congestion index of the core layer, and BW m and BW c They represent the peak bandwidth in the last 10ms window and the currently allocated link bandwidth, Q d and Q m Represents the current queue depth and maximum queue depth respectively, ▽P l Indicates the gradient of packet loss rate change, RTT n Indicates the normalized round-trip time RTT value, N Q Indicates queue imbalance, CE indicates the proportion of ENC congestion markings, ΔDelay and Delay indicate the delay change in the last second and the current delay respectively, μ indicates delay variance, and BU indicates bandwidth utilization. Indicates the average bandwidth utilization, F c Indicates the flow collision rate of ECMP hash collisions.

[0074] The specific formula for the dynamic congestion threshold of each layer of network is:

[0075]

[0076] Among them, T E represents the dynamic congestion threshold of the edge access layer, T A represents the dynamic congestion threshold of the aggregation layer, T C Indicates the core layer dynamic congestion threshold, E h represents the historical ECI mean for the same period, t represents hourly time, i represents the i-th hour, σ E Indicates the ECI standard deviation in the last hour, W indicates the sliding window size, represents the queue depth and σ in the sliding window Q represents the standard deviation of the queue depth in the sliding window, BW represents the total bandwidth, λ(o) represents the load factor during the period, σ D represents the latency variance in the last hour, and ▽BW represents the bandwidth fluctuation rate in the last hour.

[0077] The load factor λ(o) is set according to the load period. When the business is at peak hours, λ(o) = 1.2 is set. When the load is low at night, λ(o) = 0.8 is set. When the load is low at night, the core layer link delay variance σ D =0.3, bandwidth fluctuation ▽BW=0.15, F c Flow conflict rate = 5%, then the core layer dynamic congestion threshold When the congestion coefficient exceeds the threshold, core layer congestion control is triggered.

[0078] The congestion control is performed on each layer of the network, specifically including: dynamically allocating bandwidth to the edge access layer; adjusting weights of priority queues at the convergence layer; and performing multi-path optimization on the core layer.

[0079] The dynamic allocation of bandwidth of the edge access layer specifically includes the following steps:

[0080] Prioritize traffic based on service type;

[0081] Bandwidth is dynamically allocated based on priority marking. The specific formula is:

[0082]

[0083] Among them, BW x Indicates the bandwidth allocated to the current service type, N indicates the total number of service types, P x Indicates the priority of the current service type, P y represents the priority of the yth service type, G represents the GPU utilization, and α=0.8 represents the adjustment factor.

[0084] Intelligent computing center services can be categorized by computing task type into:

[0085] AI model training: requires large-scale distributed computing and high-bandwidth network requirements; typical loads include NLP large model training (such as the GPT series) and computer vision model training.

[0086] AI inference services: low latency requirements, flexible resource scheduling; typical scenarios: real-time image recognition, intelligent customer service.

[0087] Scientific computing (HPC): Double-precision floating-point computing capability, MPI communication intensive; typical applications: meteorological simulation, molecular dynamics calculations.

[0088] Big data analysis: storage and computing separation architecture, high I / O throughput requirements; typical scenarios: user behavior analysis, log processing.

[0089] For current services including real-time image recognition, user behavior analysis, NLP large-model training, and molecular dynamics calculations, set the priority of bandwidth-intensive user behavior analysis and NLP large-model training to 2, and real-time image recognition and molecular dynamics calculations to 1. When GPU utilization is 50%, the total bandwidth of 1000 GBps is 1000 / 1 + 1 + 0.574 + 0.574 ≈ 317.7 GBps.

[0090] The weight adjustment of the priority queue of the convergence layer is performed using the following formula:

[0091]

[0092] Among them, M m Indicates the weight of the current queue, M indicates the total number of queues, Indicates the priority of the current queue. Indicates the priority of the nth queue, Q m Indicates the current queue delay, Q n represents the delay of the nth queue, and k=0.1 represents the delay sensitivity coefficient.

[0093] The multipath weight optimization of the core layer specifically includes:

[0094] Calculate the path weight. The specific formula is:

[0095]

[0096] Among them, R d Indicates the weight of the current path, R indicates the total number of queues, BWR d Indicates the bandwidth of the current path, BWR g represents the bandwidth of the g-th path, D dIndicates the current path delay, D g represents the j-th path delay;

[0097] The path weight is updated every 10ms.

[0098] like Figure 2 As shown, the cross-level collaborative optimization specifically includes the following steps:

[0099] Obtain the network status characteristics of the edge access layer, the queue timing characteristics of the aggregation layer, and the topology characteristics of the core layer respectively;

[0100] The network status features of the edge access layer, the queue timing features of the aggregation layer, and the topology features of the core layer are embedded into multimodal features. The specific formula is:

[0101] Z=Concat(MLP(X E ),TCN(X A ),GAT(X C ))

[0102] Among them, Z represents multimodal features, Concat represents connection function, MLP represents fully connected layer, TCN represents dilated convolution layer, GAT represents graph attention, X E 、X A and X C They represent the network status feature matrix of the edge access layer, the queue timing feature matrix of the aggregation layer, and the topology feature matrix of the core layer respectively;

[0103] Input multimodal features into the spatiotemporal graph convolutional network;

[0104] The target loss function of the spatiotemporal graph convolutional network is optimized. The specific formula is:

[0105] L=0.5L Q +0.3L f +0.2L c

[0106] Among them, L represents the target loss function, L Q Denotes the service quality loss function, L f represents the fairness loss function, L c represents the consistency loss function;

[0107] Output the edge access layer bandwidth allocation matrix, aggregation layer queue weight matrix and core layer path weight matrix respectively.

[0108] The service quality loss function L Q The specific formula is:

[0109]

[0110] Among them, D pred and D real Denote the predicted delay and actual delay respectively, B pred and B real They represent the predicted bandwidth utilization and the actual bandwidth utilization, respectively, ||·||2 represents the L2 norm, and cs represents the cosine similarity function;

[0111] The fairness loss function L f The specific formula is:

[0112]

[0113] Among them, BW a and BW b Represents the bandwidth of the ath and bth service types respectively, P a and P b Respectively represent the priorities of the ath and bth service types;

[0114] The consistency loss function L c The specific formula is:

[0115]

[0116] Among them, ▽C and ▽E represent the congestion coefficient gradient of the edge access layer and the core layer, respectively. represents the time derivative of the congestion coefficient of the aggregation layer, and ||·||1 represents the L1 norm.

[0117] like Figure 3 As shown in the figure, the three-layer network architecture of the intelligent computing center includes a convergence switch group, a core switch group, and an access switch group. The access switch is connected to the server.

[0118] Among them, the core switch group: as a network traffic hub, it undertakes the horizontal high-speed interconnection needs of cross-regional data centers, cloud computing resource pools and external networks, and supports high-bandwidth services such as AI training and inference. The aggregation switch group serves as the policy execution center: it implements inter-VLAN routing, ACL access control, and quality of service (QoS) marking, and divides independent logical channels for different AI services (such as isolation of training clusters and inference clusters). The access switch group is directly connected to the server: it connects to GPU servers, storage nodes and other devices through high-speed ports, providing single-port access capabilities of ≥200Gbps to meet the high-concurrency communication needs of AI computing nodes. In addition, each access switch is connected to two aggregation switches through two independent uplinks, forming a topology without single point failure.

[0119] like Figure 4As shown in Figure 1, the simple network structure of a spatiotemporal graph convolutional network consists of three parts: Normalization: Normalizes the input data. Spatiotemporal variation: Through multiple STGCN blocks, each block alternates between GCN and TCN. Output: Classifies features using average pooling and fully connected layers. Finally, outputs the scheduling policy for the aggregation layer, core layer, and edge access layer.

[0120] The computer-readable storage medium of this embodiment may be an internal storage unit of the terminal, such as a hard disk or memory of the terminal; the computer-readable storage medium of this embodiment may also be an external storage device of the terminal, such as a plug-in hard disk, a smart memory card, a secure digital card, a flash memory card, etc. equipped on the terminal; further, the computer-readable storage medium may also include both an internal storage unit of the terminal and an external storage device.

[0121] The computer-readable storage medium of this embodiment is used to store computer programs and other programs and data required by the terminal. The computer-readable storage medium can also be used to temporarily store data that has been output or is to be output.

[0122] Those skilled in the art will appreciate that all or part of the processes in the above-described method embodiments can be implemented by instructing related hardware through a computer program. The program can be stored in a computer-readable storage medium, and when executed, the program can include the processes in the above-described method embodiments. The storage medium can be a magnetic disk, an optical disk, a read-only memory (ROM), or a random access memory (RAM).

[0123] The examples described in the present invention are merely descriptions of the preferred embodiments of the present invention and are not intended to limit the concept and scope of the present invention. Without departing from the design concept of the present invention, various modifications and improvements made to the technical solutions of the present invention by engineers and technicians in this field should fall within the scope of protection of the present invention.

Claims

1. A multi-level network congestion control method for intelligent computing centers, characterized by: The steps include: Real-time monitoring of traffic status parameters at all levels of the intelligent computing center network, including the edge access layer, aggregation layer, and core layer; The congestion coefficient of each layer is calculated based on the network traffic status parameters of each layer. The specific formula is: Among them, E represents the congestion index of the edge access layer, A represents the congestion index of the aggregation layer, C represents the congestion index of the core layer, and BW m and BW c They represent the peak bandwidth in the last 10ms window and the currently allocated link bandwidth, Q d and Q m Respectively represent the current queue depth and maximum queue depth, Indicates the gradient of packet loss rate change, RTT n Indicates the normalized round-trip time RTT value, N Q Indicates queue imbalance, CE indicates the proportion of ENC congestion markings, ΔDelay and Delay indicate the delay change in the last second and the current delay respectively, μ indicates delay variance, and BU indicates bandwidth utilization. Indicates the average bandwidth utilization, F c Indicates the flow collision rate of ECMP hash collision; Set the dynamic congestion threshold of each layer of the network. The specific formula is: Among them, T E represents the dynamic congestion threshold of the edge access layer, T A represents the dynamic congestion threshold of the aggregation layer, T C Indicates the core layer dynamic congestion threshold, E h represents the historical ECI mean for the same period, t represents hourly time, i represents the i-th hour, σ E Indicates the ECI standard deviation in the last hour, W indicates the sliding window size, represents the queue depth and σ in the sliding window Q represents the standard deviation of the queue depth in the sliding window, BW represents the total bandwidth, λ(o) represents the load factor during the period, σ D Indicates the latency variance in the last hour. Indicates the bandwidth fluctuation rate in the last hour; When the congestion coefficient of a single layer exceeds the dynamic congestion threshold, congestion control is performed on each layer of the network separately; When the congestion coefficients of two or more layers exceed the dynamic congestion threshold, cross-layer collaborative optimization is performed; The cross-level collaborative optimization specifically includes the following steps: Obtain the network status characteristics of the edge access layer, the queue timing characteristics of the aggregation layer, and the topology characteristics of the core layer respectively; The network status features of the edge access layer, the queue timing features of the aggregation layer, and the topology features of the core layer are embedded into multimodal features. The specific formula is: Z=Concat(MLP(X E ),TCN(X A ),GAT(X C )) Among them, Z represents multimodal features, Concat represents connection function, MLP represents fully connected layer, TCN represents dilated convolution layer, GAT represents graph attention, X E 、X A and X C They represent the network status feature matrix of the edge access layer, the queue timing feature matrix of the aggregation layer, and the topology feature matrix of the core layer respectively; Input multimodal features into the spatiotemporal graph convolutional network; The target loss function of the spatiotemporal graph convolutional network is optimized. The specific formula is: <h2 style=";text-align:left;direction:ltr">L=0.5L<h2 style=";text-align:left;direction:ltr"> Q <h2 style=";text-align:left;direction:ltr"> +0.3L<h2 style=";text-align:left;direction:ltr"> f <h2 style=";text-align:left;direction:ltr"> +0.2L<h2 style=";text-align:left;direction:ltr"> c Among them, L represents the target loss function, L Q Denotes the service quality loss function, L f represents the fairness loss function, L c represents the consistency loss function; Output the edge access layer bandwidth allocation matrix, aggregation layer queue weight matrix and core layer path weight matrix respectively.

2. The method according to claim 1, characterized in that The congestion control is performed on each layer of the network, specifically including: dynamically allocating bandwidth to the edge access layer; adjusting weights of priority queues at the convergence layer; and performing multi-path optimization on the core layer.

3. The method according to claim 2, characterized in that The dynamic allocation of bandwidth of the edge access layer specifically includes the following steps: Prioritize traffic based on service type; Bandwidth is dynamically allocated based on priority marking. The specific formula is: Among them, BW x Indicates the bandwidth allocated to the current service type, N indicates the total number of service types, P x Indicates the priority of the current service type, P y represents the priority of the yth service type, G represents the GPU utilization, and α=0.8 represents the adjustment factor.

4. The method according to claim 3, characterized in that The weight adjustment of the priority queue of the convergence layer is performed using the following formula: Among them, M m Indicates the weight of the current queue, M indicates the total number of queues, Indicates the priority of the current queue. Indicates the priority of the nth queue, Q m Indicates the current queue delay, Q n represents the delay of the nth queue, and k=0.1 represents the delay sensitivity coefficient.

5. The method according to claim 4, characterized in that The multipath weight optimization of the core layer specifically includes: Calculate the path weight. The specific formula is: Among them, R d Indicates the weight of the current path, R indicates the total number of queues, BWR d Indicates the bandwidth of the current path, BWR g represents the bandwidth of the g-th path, D d Indicates the current path delay, D g represents the j-th path delay; The path weight is updated every 10ms.

6. The method according to claim 5, characterized in that The service quality loss function L Q The specific formula is: Among them, D pred and D real Denote the predicted delay and actual delay respectively, B pred and B real They represent the predicted bandwidth utilization and the actual bandwidth utilization, respectively, ||·||2 represents the L2 norm, and cs represents the cosine similarity function; The fairness loss function L f The specific formula is: Among them, BW a and BW b Represents the bandwidth of the ath and bth service types respectively, P a and P b Respectively represent the priorities of the ath and bth service types; The consistency loss function L c The specific formula is: in, and They represent the congestion coefficient gradient of the edge access layer and the core layer respectively, represents the time derivative of the congestion coefficient of the aggregation layer, and ||·||1 represents the L1 norm.

Images