A computing power cluster SDN congestion control and data flow intelligent scheduling method

Abstract

The application discloses a kind of computing power cluster SDN congestion control and data stream intelligent scheduling method, it is related to data transmission control technical field, specific steps are as follows: S1: SDN controller is periodically collected by south interface link bandwidth of virtual switch, port queue length, data packet forwarding delay and data stream five tuple characteristics, while collecting the computing resource utilization rate and task queue state of computing power node, constructs the global perception matrix of fusion network state and computing power state;S2: based on the global perception matrix, the congestion trend of each link is predicted using time series prediction model.The present application can identify potential congestion risk in future window period in advance, get rid of the hysteresis problem of traditional fixed threshold post-response, without relying on artificial configuration static threshold, can be self-adapted to adapt to sudden computing power traffic, tidal traffic complex scene, effectively suppresses traffic avalanche, data packet loss, service time delay jitter problem.

Description

Technical Field

[0001] This invention relates to the field of data transmission control technology, specifically to a method for SDN congestion control and intelligent data flow scheduling in computing power clusters. Background Technology

[0002] With the large-scale deployment of services such as AI large-scale model training, massive offline big data analysis, high-performance scientific computing, and real-time edge computing, centralized computing clusters have become the core carrier of digital infrastructure. A computing cluster contains a large number of GPU computing nodes, CPU general-purpose computing nodes, and distributed storage nodes. These nodes rely on data center networks to complete the real-time transmission of massive amounts of computational interaction data, storage read / write data, and device management and control data. The network transmission quality and computing resource scheduling efficiency directly determine the overall throughput, task execution latency, and resource utilization of the entire computing cluster. Software-defined networking (SDN), with its core advantages of decoupling the control plane and data plane, centralized network-wide management and control, and programmable flow table distribution, is widely used in the transformation of computing cluster network architectures. It enables unified management and flexible path orchestration of network traffic, effectively compensating for the rigid management and inflexible traffic scheduling of traditional Ethernet distributed networks.

[0003] However, existing SDN network congestion control and data flow scheduling solutions for computing power clusters still suffer from several core technical deficiencies, failing to adapt to the current operating scenarios of heterogeneous traffic, dynamic fluctuations in computing load, and sudden network congestion. Specific deficiencies include: a single dimension of state awareness, separating network link status from computing node operating status, failing to construct a unified global view of network and computing, and congestion judgment and scheduling decisions not conforming to actual cluster operating conditions; congestion detection is a reactive, post-event response, lacking time-series prediction capabilities, making it difficult to proactively avoid congestion problems caused by sudden traffic spikes, and exhibiting poor adaptability to fixed thresholds; traffic scheduling fails to prioritize services or link to computing resource availability, making it difficult to guarantee the transmission quality of critical services, easily leading to a double waste of network and computing resources, and failing to adapt to the complex and dynamic business operating scenarios of clusters. To address these issues, we propose a new SDN congestion control and intelligent data flow scheduling method for computing power clusters. Summary of the Invention

[0004] To address the aforementioned technical issues, this paper presents a method for SDN congestion control and intelligent data flow scheduling in computing power clusters. This technical solution solves the problems of limited state perception dimensions, difficulty in preventing congestion caused by sudden traffic surges, and double waste of network computing resources.

[0005] To achieve the above objectives, the technical solution adopted by this invention is: a method for SDN congestion control and intelligent data flow scheduling in computing power clusters, the specific steps of which are as follows:

[0006] S1: The SDN controller periodically collects the link bandwidth, port queue length, packet forwarding latency and data flow five-tuple characteristics of the virtual switch through the southbound interface, and at the same time collects the computing resource utilization and task queue status of the computing power nodes to build a global awareness matrix that integrates network status and computing power status.

[0007] S2: Based on the global perception matrix, a time-series prediction model is used to predict the congestion trend of each link. Combining the current congestion index with the predicted congestion trend, when the current congestion index exceeds a preset threshold or the predicted congestion trend will exceed the threshold within a preset time window, the link is determined to enter a congestion warning state.

[0008] S3: Divide the data flow within the computing power cluster into high-priority computing flow, medium-priority storage flow and low-priority management flow according to business type, extract the bandwidth requirements, latency tolerance and transmission duration characteristics of each data flow, and associate them with the computing resource availability of the target computing power node of each data flow.

[0009] S4: Using the global perception matrix as the state space, the link allocation and bandwidth allocation of the data flow as the action space, and the link congestion mitigation degree and the load balancing degree of computing nodes as reward signals, a data flow scheduling strategy is dynamically generated through reinforcement learning algorithm. High-priority computing flows are prioritized to links with sufficient computing resources and no congestion warnings, while bandwidth limiting and path rerouting are implemented for medium and low-priority data flows.

[0010] S5: The SDN controller sends flow table update commands to the virtual switch to complete real-time forwarding control of data flows, continuously monitors the scheduling effect, and iteratively optimizes the reinforcement learning model parameters and scheduling strategies.

[0011] Preferably, the global perception matrix in S1 includes a network state sub-matrix and a computing power state sub-matrix; the elements of the network state sub-matrix include the bandwidth utilization of each link, the queue occupancy rate of each port, the packet forwarding delay of each link, and the five-tuple statistics of each data stream; the elements of the computing power state sub-matrix include the CPU utilization, GPU utilization, memory occupancy rate, and task queue length of each computing power node.

[0012] Preferably, the congestion index in S2 is calculated by normalizing the queue occupancy rate, packet forwarding delay ratio, and bandwidth utilization rate of each link and then summing them by weight, wherein the packet forwarding delay ratio is the ratio of the current packet forwarding delay to the baseline delay.

[0013] Preferably, the time-series prediction model in S2 is a long short-term memory network, which takes the historical global perception matrix sequence as input and outputs the predicted congestion index sequence of each link within the preset time window.

[0014] Preferably, the availability of computing resources in S3 is determined based on the CPU idle rate, GPU idle rate, and estimated completion time of the task queue of the target computing node; the CPU idle rate and GPU idle rate are the difference between the current utilization rate and the full load utilization rate, and the estimated completion time of the task queue is estimated based on the computational amount of each task in the queue and the computing power of the node.

[0015] Preferably, the reinforcement learning algorithm in S4 adopts an Actor-Critic architecture. The Actor network takes the global perception matrix as input and outputs the link allocation probability distribution and bandwidth allocation ratio of each data stream. The Critic network takes the global perception matrix and the current action as input and outputs a state-action value estimate, which is used to evaluate the merits of the scheduling strategy.

[0016] Preferably, the reward signal in S4 is calculated using the following formula:

[0017] R = α·ΔCI norm + β·(1 - CV compute )

[0018] Where ΔCI norm CV represents the normalized decrease in the overall network congestion index after scheduling. compute The coefficient of variation of resource utilization is calculated for each computing node, where α and β are preset weight coefficients and the sum of α and β is 1, and R is the reward signal value.

[0019] Preferably, the training process of the reinforcement learning algorithm includes an experience replay mechanism, which stores the state transition data of each scheduling step into the experience replay buffer, and randomly samples small batches of data from the experience replay buffer to update the parameters of the Actor network and the Critic network.

[0020] Preferably, the continuous monitoring of the scheduling effect in S5 includes real-time collection of the congestion index change of each link after scheduling, the load variance of each computing node, and the data stream transmission completion rate; when the congestion index change is lower than the preset improvement threshold or the load variance exceeds the preset equilibrium threshold, the online update of the reinforcement learning model parameters is triggered.

[0021] Preferably, the network state submatrix and the computing power state submatrix in S1 adopt an asynchronous update mechanism, the update frequency of the network state submatrix is ​​higher than the update frequency of the computing power state submatrix, and the state data collected most recently is used to fill the gap between two updates.

[0022] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0023] This invention achieves dual-dimensional, full-domain integrated perception of network and computing power, constructs an integrated global state matrix, eliminates blind spots in state perception, and relies on the SDN controller to periodically and synchronously collect network link transmission parameters and computing power node resource load and task queue status. It breaks through the shortcomings of the traditional solution of separating network and computing power status, accurately depicts the coupled correlation between network congestion and computing power overload. Combining real-time congestion index and link congestion time series prediction results, it identifies potential congestion risks in the future window period in advance, gets rid of the lag problem of traditional fixed threshold post-event response, does not rely on manual configuration of static thresholds, and can adaptively adapt to complex scenarios of sudden computing power traffic and tidal business traffic, effectively suppressing traffic avalanche, packet loss, and service latency jitter problems.

[0024] This invention divides data streams into high, medium, and low priorities, and applies differentiated scheduling and bandwidth control strategies accordingly. It prioritizes the transmission needs of core computing streams, preventing low-priority traffic from preempting critical link resources. Simultaneously, the scheduling path is correlated with the availability of target computing node resources, eliminating the problem of ineffective scheduling where network links are idle but computing nodes are overloaded, and reducing the waste of network bandwidth and computing resources. It can dynamically adjust forwarding paths and bandwidth quotas in real time according to network fluctuations and changes in computing load, avoiding secondary congestion problems. The model also has online iterative optimization capabilities, maintaining excellent scheduling performance even under long-term operation, and comprehensively improving the overall throughput and resource utilization of the computing cluster. Attached Figure Description

[0025] Figure 1 This is a flowchart of the mobilization steps of the present invention. Detailed Implementation

[0026] The following description is intended to disclose the invention and enable those skilled in the art to implement it. The preferred embodiments described below are merely examples, and other obvious variations will occur to those skilled in the art.

[0027] System Architecture Overview

[0028] The SDN congestion control and intelligent data flow scheduling method for computing power clusters described in this invention is deployed in a computing power cluster environment using an SDN architecture. This environment includes an SDN controller, multiple virtual switches, and multiple computing power nodes. The SDN controller, acting as the global decision-making center, communicates with each virtual switch through a southbound interface to obtain network status data; simultaneously, it communicates with each computing power node through a management interface to obtain computing resource status data. The virtual switches are responsible for data forwarding, executing forwarding actions according to the flow tables issued by the SDN controller. The computing power nodes provide computing services, running various computing tasks on them.

[0029] This invention introduces a closed-loop mechanism that integrates sensing, predictive warning, and reinforcement learning scheduling in the SDN controller, enabling congestion predictive control and intelligent data flow scheduling under the condition of joint sensing of computing power and network.

[0030] S1: Construction of the Global Awareness Matrix

[0031] The SDN controller periodically collects link status data and data flow characteristic data of virtual switches through the southbound interface, and at the same time collects computing resource status data of computing nodes through the management interface, and merges the two types of data to construct a global awareness matrix.

[0032] The global awareness matrix includes a network state submatrix. and computing power state submatrix Both adopt an asynchronous update mechanism. Since the frequency of network state changes is typically higher than the frequency of changes in computing node resource states, the update frequency of the network state sub-matrix is ​​higher than that of the computing power state sub-matrix. Between two updates, the most recently collected state data is used to fill the gap, ensuring that the global perception matrix is ​​always available for downstream modules to read and use.

[0033] Network state submatrix The elements include bandwidth utilization of each link, queue occupancy of each port, packet forwarding latency of each link, and five-tuple statistics of each data stream. Assume the cluster has... Link Ports, If there are 1 active data stream, the network state submatrix is ​​represented as:

[0034]

[0035] in, for The link bandwidth utilization vector, its 3D link bandwidth utilization vector, element For link The ratio of current traffic rate to link capacity; for The port queue occupancy vector, its 3rd dimension element For port The ratio of the current queue length to the maximum queue length; for The dimensional link packet forwarding delay vector, its 3D link packet forwarding delay vector, is as follows: element For link The average forwarding delay of data packets from the ingress port to the egress port; for A five-tuple statistical matrix for data streams, where each row corresponds to the statistical feature encoding of the source IP, destination IP, source port, destination port, and protocol type of a data stream.

[0036] Computing power state submatrix The elements include the CPU utilization, GPU utilization, memory usage, and task queue length of each computing node. Assume the cluster has a total of... If there are 10 computing power nodes, then the computing power state submatrix is ​​represented as:

[0037]

[0038] in, for 3D CPU utilization vector for 3D GPU utilization vector for 3D memory occupancy vector for A vector of task queue lengths.

[0039] Global perception matrix It consists of two sub-matrices joined together:

[0040]

[0041] This matrix serves as a unified input for subsequent prediction and scheduling modules, enabling joint perception of computing power status and network status.

[0042] S2: Congestion Trend Prediction and Early Warning

[0043] Based on the global awareness matrix, a time-series prediction model is used to predict the congestion trend of each link. Combining the current congestion index with the predicted congestion trend, it is determined whether a link has entered a congestion warning state. The congestion index is calculated for each link. Congestion Index The results were obtained by weighted summation after normalizing the queue occupancy rate, packet forwarding latency ratio, and bandwidth utilization rate respectively.

[0044]

[0045] in, For link The normalized value of the port queue occupancy rate. This is the normalized value of the packet forwarding latency ratio. This is the normalized value of link bandwidth utilization; packet forwarding latency ratio. Defined as the ratio of the current packet forwarding delay to the baseline delay:

[0046]

[0047] in, For link The baseline forwarding delay under congestion-free conditions is determined through the initialization phase or by statistical analysis of historical minimum values. Weighting coefficients. , , satisfy The specific value can be adjusted according to the characteristics of the cluster business.

[0048] The time-series prediction model employs a Long Short-Term Memory (LSTM) network, taking the historical global perception matrix sequence as input and outputting a predicted congestion index sequence for each link within a preset time window. Specifically, it uses the most recent... Global sensing matrix for each acquisition cycle The data is sequentially input into an LSTM network. After temporal feature extraction by multiple LSTM units, the future data is output through a fully connected layer. Congestion index forecast for each period LSTM networks can capture the temporal dependencies of congestion indices and predict congestion trends in advance.

[0049] Congestion warning determination, for links The link is considered to be in a congestion warning state when any of the following conditions are met:

[0050] Condition 1: Current congestion index Exceeding the preset congestion threshold ;

[0051] Condition 2: The predicted congestion trend indicator is within a preset time window. Within each cycle, the predicted congestion index is... Will exceed the threshold .

[0052] Condition one is used to capture the current congestion state, and condition two is used to provide predictive warnings before congestion actually occurs. The combination of the two enables the system to have both reactive and predictive congestion awareness capabilities, providing more response time for subsequent scheduling decisions.

[0053] S3: Data Flow Classification and Resource Association

[0054] The data streams within the computing power cluster are prioritized according to business type, and the service quality requirement characteristics of each data stream are extracted. At the same time, the computing resource availability of the target computing power nodes of each data stream is associated.

[0055] Data flow priority is assigned to three levels based on business type: high-priority computing flow, medium-priority storage flow, and low-priority management flow. Computing flow carries traffic for core business processes sensitive to latency, such as training and inference, and scientific computing; storage flow carries traffic with high bandwidth requirements but high latency tolerance, such as data read / write and model synchronization; and management flow carries operational traffic with low bandwidth and latency requirements, such as heartbeat detection and configuration distribution.

[0056] Data stream feature extraction, for each data stream Extracting 3D feature vectors ,in The bandwidth requirement is calculated based on historical data stream rates. The latency tolerance is determined based on the SLA requirements corresponding to the service type; The transmission duration characteristic is calculated based on the data stream's ongoing transmission time and the estimated remaining transmission time.

[0057] Computational resource availability calculation, for data streams Target computing nodes Computational resource availability Determined based on the CPU idle rate, GPU idle rate, and estimated task queue completion time of the target computing node:

[0058]

[0059] in, For nodes CPU idle rate is defined as the difference between CPU full-load utilization and current utilization. For nodes GPU idle rate is defined similarly to CPU idle rate; The estimated completion time for the task queue is derived from the computational load of each task in the queue and the computing power of the nodes. This is a normalized upper limit for the estimated completion time; , , These are weighting coefficients, and the sum of the three is 1; calculate resource availability. The larger the value, the stronger the node's ability to accept new tasks.

[0060] The feature vector of each data stream The availability of its target computing nodes The associations form flow-resource association features, which serve as one of the inputs to the reinforcement learning scheduling module.

[0061] S4: Reinforcement Learning Scheduling Policy Generation

[0062] Using a global perception matrix as the state space, link allocation and bandwidth allocation of data flow as the action space, and link congestion mitigation degree and computing node load balancing degree as reward signals, a data flow scheduling strategy is dynamically generated through reinforcement learning algorithm.

[0063] State space, state Take the global perception matrix at the current time. It contains all elements of the network state submatrix and the computing power state submatrix, enabling the reinforcement learning agent to simultaneously perceive network congestion and computing power node load.

[0064] Action space, action For the joint decision-making of link allocation and bandwidth allocation for each data stream, for each data stream The action includes two dimensions: link selection. Represents data stream The allocated forwarding path and bandwidth allocation Represents data stream The allocated bandwidth ratio; the overall action is represented as:

[0065]

[0066] The reinforcement learning algorithm architecture adopts an Actor-Critic architecture, which includes an Actor network and a Critic network. The Actor network uses a global perception matrix. As input, output the link allocation probability distribution and bandwidth allocation ratio for each data stream. The link allocation probability distribution is normalized using Softmax, and the bandwidth allocation ratio is constrained using Sigmoid. Interval. The Critic network uses a global perception matrix. and current action Input: State-Action Value Estimate It is used to evaluate the merits of the current scheduling strategy and guide the parameter update direction of the Actor network.

[0067] Reward signal design, reward signal Calculated using the following formula:

[0068]

[0069] in, This represents the normalized decrease in the network congestion index after scheduling, reflecting the degree to which scheduling alleviates network congestion. The coefficient of variation of resource utilization is calculated for each computing node, reflecting the degree of load balancing among computing nodes; and For preset weighting coefficients and , This is to award a signal value.

[0070] Normalized decrease in the overall network congestion index The calculation method is as follows:

[0071]

[0072] in, and These are the average congestion indices of all links in the entire network before and after the scheduling.

[0073] coefficient of variation The calculation method is as follows:

[0074]

[0075] in, Calculate the average resource utilization rate for each computing node. Calculate the standard deviation of utilization rate for each computing node. A smaller value indicates a more balanced load across nodes. The larger the value, the higher the balance.

[0076] The design of the reward function unifies the two objectives of congestion mitigation and load balancing into the same optimization framework, enabling the reinforcement learning agent to take into account both network performance and computing resource utilization efficiency when generating scheduling strategies.

[0077] The scheduling strategy is executed by the reinforcement learning agent, which dynamically generates a data flow scheduling strategy based on the current state and the trained Actor network: high-priority computing flows are prioritized for scheduling to links with sufficient computing resources and no congestion warnings, ensuring that core business traffic receives low-latency and high-bandwidth transmission guarantees; bandwidth limiting and path rerouting are implemented for medium and low-priority data flows, migrating them from congested or soon-to-be-congested links to less-loaded backup paths, thus releasing core link resources.

[0078] The experience replay mechanism is included in the training process of reinforcement learning algorithms. After each scheduling step, the state transition data is transferred. The data is stored in the experience replay buffer. During training, small batches of data are randomly sampled from the experience replay buffer to update the parameters of the Actor and Critic networks. The experience replay mechanism breaks the temporal correlation of data, improving training stability and sample utilization efficiency.

[0079] The parameters of the Actor and Critic networks are updated using gradient descent. The loss function of the Critic network is the mean square value of the temporal difference error.

[0080]

[0081] in, The discount factor is used; the update gradient of the Actor network is guided by the value estimate of the Critic network.

[0082]

[0083] in, The policy distribution output by the Actor network. These are the parameters of the Actor network.

[0084] S5: Flow Table Distribution and Online Optimization

[0085] Based on the scheduling strategy generated by the reinforcement learning scheduling module, the SDN controller issues flow table update instructions to the virtual switch to complete real-time forwarding control of data flows.

[0086] Flow table distribution transforms the link allocation results in the scheduling policy into flow table rules: each flow table rule matches the five-tuple characteristics of the data flow, and the action field specifies the forwarding port and queue priority. Bandwidth allocation results are transformed into traffic shaping rules: token bucket rate limiting parameters are set for low- and medium-priority data flows to ensure that their bandwidth consumption does not exceed the allocated ratio. The SDN controller distributes the flow table rules and traffic shaping rules to each virtual switch through the southbound interface, and the virtual switches execute data forwarding according to the rules.

[0087] To continuously monitor the scheduling effect, after the flow table is issued, the SDN controller continuously collects the following monitoring indicators: the change in congestion index of each link, the load variance of each computing node, and the data flow transmission completion rate.

[0088] The change in congestion index is the difference in the congestion index of link $l$ before and after scheduling. A positive value indicates that congestion has been alleviated.

[0089] Load variance is the variance of the computing resource utilization of each computing node, reflecting the load balancing effect;

[0090] Data stream transmission completion rate: The percentage of data streams that are transmitted within a specified latency threshold.

[0091] Online updates are triggered when the change in the congestion index falls below a preset improvement threshold. When the load variance exceeds the preset balancing threshold, it indicates that the current scheduling strategy has failed to effectively alleviate congestion. When this occurs, it indicates that the load imbalance between computing nodes is intensifying. When any of the above conditions are met, the online update of the reinforcement learning model parameters is triggered. The parameters of the Actor network and Critic network are fine-tuned using the latest collected state data, so that the scheduling strategy can adapt to the changes in the current network and computing environment.

[0092] Online updates employ gradient updates with a small learning rate, making incremental adjustments based on existing model parameters to avoid policy oscillations caused by full training and ensure the continuity and stability of the scheduling service.

[0093] Example 1

[0094] This embodiment takes a computing cluster containing 8 computing nodes, 12 virtual switches, and 20 links as an example to illustrate the specific implementation process of the present invention.

[0095] Phase S1: The SDN controller collects link status data from each virtual switch via the OpenFlow protocol at 500ms intervals and resource status data from each computing node via the REST API at 2s intervals. The dimensions of the network state sub-matrix are... (Link bandwidth utilization) (Port queue utilization, 2 ports per switch) (Link forwarding delay) (Data flow quintuple statistics), the dimension of the computing power state submatrix is... (CPU utilization, GPU utilization, memory usage, task queue length). The network state submatrix is ​​updated every 500ms, and the computing power state submatrix is ​​updated every 2s. The interval between two updates is filled with the most recently collected data.

[0096] S2 Phase: Congestion Index Weighting , , Reference latency of each link The congestion threshold is determined during the initialization phase by collecting forwarding latency under no-load conditions, with a value ranging from 50μs to 200μs. Set to 0.7; the LSTM network uses 2 layers of LSTM units, with a hidden layer dimension of 128, and the input is the nearest neighbor. The global sensing matrix sequence of each acquisition cycle outputs the future... The congestion index prediction value for each cycle. When the link... satisfy or (Within a prediction window of 5 periods) determine the link Entering congestion warning status.

[0097] S3 Phase: The cluster currently has 50 active data streams, including 15 compute streams, 20 storage streams, and 15 management streams; extract data from each data stream. Feature: Computational stream delay tolerance Set the ms timer to 10ms, the storage flow to 50ms, and the management flow to 200ms; the computing resource availability weight is set to... , , . Take 300 seconds.

[0098] Phase S4: Both the Actor and Critic networks employ a 3-layer fully connected network with hidden layer dimensions of 256, 128, and 64 respectively, using ReLU as the activation function. The Actor network output layer: the link allocation part outputs the probability distribution of 20 links via Softmax, and the bandwidth allocation part outputs the bandwidth ratio via Sigmoid. The reward weights are... , Discount factor The experience replay buffer has a capacity of 10,000 records, and the training batch size is 64. The initial learning rate for the Actor network is... Critic network is .

[0099] After the scheduling policy is generated, the 15 computational flows are prioritized for allocation to nodes with no congestion warnings and high target node availability. The 20 storage streams were allocated to links with ample remaining bandwidth, with bandwidth ratios limited to 60%-80%; the 15 management streams were allocated to the lightest-loaded links, with bandwidth ratios limited to 20%-40%.

[0100] Phase S5: Improving the Threshold Set the threshold to 0.05. Use 0.15. Monitoring collects the change in congestion index for each link and the load variance of each node at 1-second intervals. The decrease in congestion index after a link is scheduled is also recorded. For three consecutive cycles, or node load variance At that time, an online update is triggered to... The learning rate is used to perform a 5-step gradient update on the Actor and Critic networks.

[0101] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the claimed invention.

Claims

1. A method for SDN congestion control and intelligent data flow scheduling in a computing power cluster, characterized in that, The specific steps are as follows: S1: The SDN controller periodically collects the link bandwidth, port queue length, packet forwarding latency and data flow five-tuple characteristics of the virtual switch through the southbound interface, and at the same time collects the computing resource utilization and task queue status of the computing power nodes to build a global awareness matrix that integrates network status and computing power status. S2: Based on the global perception matrix, a time-series prediction model is used to predict the congestion trend of each link. Combining the current congestion index with the predicted congestion trend, when the current congestion index exceeds a preset threshold or the predicted congestion trend will exceed the threshold within a preset time window, the link is determined to enter a congestion warning state. S3: Divide the data flow within the computing power cluster into high-priority computing flow, medium-priority storage flow and low-priority management flow according to business type, extract the bandwidth requirements, latency tolerance and transmission duration characteristics of each data flow, and associate them with the computing resource availability of the target computing power node of each data flow. S4: Using the global perception matrix as the state space, the link allocation and bandwidth allocation of the data flow as the action space, and the link congestion mitigation degree and the load balancing degree of computing nodes as reward signals, a data flow scheduling strategy is dynamically generated through reinforcement learning algorithm. High-priority computing flows are prioritized to links with sufficient computing resources and no congestion warnings, while bandwidth limiting and path rerouting are implemented for medium and low-priority data flows. S5: The SDN controller sends flow table update commands to the virtual switch to complete real-time forwarding control of data flows, continuously monitors the scheduling effect, and iteratively optimizes the reinforcement learning model parameters and scheduling strategies.

2. The method for SDN congestion control and intelligent data flow scheduling of a computing power cluster according to claim 1, characterized in that: The global awareness matrix in S1 includes a network state sub-matrix and a computing power state sub-matrix; the elements of the network state sub-matrix include the bandwidth utilization of each link, the queue occupancy rate of each port, the packet forwarding latency of each link, and the five-tuple statistics of each data stream; the elements of the computing power state sub-matrix include the CPU utilization, GPU utilization, memory occupancy rate, and task queue length of each computing power node.

3. The method for SDN congestion control and intelligent data flow scheduling of a computing power cluster according to claim 1, characterized in that: The congestion index mentioned in S2 is calculated as follows: the queue occupancy rate, packet forwarding delay ratio and bandwidth utilization of each link are normalized and then weighted and summed, wherein the packet forwarding delay ratio is the ratio of the current packet forwarding delay to the baseline delay.

4. The method for SDN congestion control and intelligent data flow scheduling of a computing power cluster according to claim 1, characterized in that: The time-series prediction model described in S2 is a long short-term memory network that takes the historical global perception matrix sequence as input and outputs the predicted congestion index sequence of each link within the preset time window.

5. The method for SDN congestion control and intelligent data flow scheduling of a computing power cluster according to claim 1, characterized in that: The availability of computing resources mentioned in S3 is determined based on the CPU idle rate, GPU idle rate, and estimated completion time of the task queue of the target computing power node; the CPU idle rate and GPU idle rate are the difference between the current utilization rate and the full load utilization rate, and the estimated completion time of the task queue is estimated based on the computational amount of each task in the queue and the computing power of the node.

6. The method for SDN congestion control and intelligent data flow scheduling of a computing power cluster according to claim 1, characterized in that: The reinforcement learning algorithm described in S4 adopts an Actor-Critic architecture. The Actor network takes the global perception matrix as input and outputs the link allocation probability distribution and bandwidth allocation ratio of each data stream. The Critic network takes the global perception matrix and the current action as input and outputs the state-action value estimate, which is used to evaluate the merits of the scheduling strategy.

7. The method for SDN congestion control and intelligent data flow scheduling of a computing power cluster according to claim 1, characterized in that, The reward signal mentioned in S4 is calculated using the following formula: R = α·ΔCI norm + β·(1 - CV compute ) Where ΔCI norm CV represents the normalized decrease in the overall network congestion index after scheduling. compute The coefficient of variation of resource utilization is calculated for each computing node, where α and β are preset weight coefficients and the sum of α and β is 1, and R is the reward signal value.

8. The method for SDN congestion control and intelligent data flow scheduling of a computing power cluster according to claim 6, characterized in that: The training process of the reinforcement learning algorithm includes an experience replay mechanism, which stores the state transition data of each scheduling step into the experience replay buffer, and randomly samples small batches of data from the experience replay buffer to update the parameters of the Actor network and the Critic network.

9. The method for SDN congestion control and intelligent data flow scheduling of a computing power cluster according to claim 1, characterized in that: The continuous monitoring of scheduling effect described in S5 includes real-time collection of the congestion index change of each link after scheduling, the load variance of each computing node, and the data stream transmission completion rate; when the congestion index change is lower than the preset improvement threshold or the load variance exceeds the preset equilibrium threshold, the online update of the reinforcement learning model parameters is triggered.

10. The method for SDN congestion control and intelligent data flow scheduling of a computing power cluster according to claim 1, characterized in that: The network state submatrix and computing power state submatrix mentioned in S1 adopt an asynchronous update mechanism. The update frequency of the network state submatrix is ​​higher than that of the computing power state submatrix, and the state data collected most recently is used to fill the gap between two updates.

Images