SDN data center load balancing method based on SRv6

Abstract

The invention relates to the technical field of software defined networks, in particular to an SRv6-based SDN (Software Defined Network) data center load balancing method and system, and the method comprises the steps: obtaining node real-time load and SRv6 routing available resource information through an SDN controller, constructing a load balancing model, and recognizing an overload node and a congestion link; generating a flow scheduling path in combination with segmented routing characteristics; in combination with network state information, monitoring the load condition in real time when the node is unbalanced or the link packet loss rate is gt; when 5%, a load balancing adjustment mechanism is triggered, unbalance types are distinguished, a related flow set is determined, optimization is performed by using a DQN algorithm, and a flow scheduling strategy is generated and dynamically adjusted by taking throughput maximization and unbalance degree minimization as targets; and issuing the scheduling strategy to the data center network equipment, updating the flow table through the Netconf protocol, and redistributing the flow. Therefore, the problems of low load balancing efficiency, insufficient dynamic adaptability, low resource utilization rate and the like in the prior art are solved.

Description

Technical Field

[0001] The present application relates to the field of software-defined network technology, and in particular to an SRv6-based SDN data center load balancing method and system. Background Art

[0002] With the rapid development of cloud computing, big data, and artificial intelligence technologies, data centers, as core infrastructure for carrying massive amounts of business, are experiencing explosive growth in network traffic. Business types are becoming increasingly complex and diverse, placing higher demands on the load balancing capabilities of data center networks. Data center networks must achieve efficient traffic distribution, intelligent path optimization, and full utilization of network resources in a dynamically changing traffic environment to ensure low latency, high reliability, and high throughput for services.

[0003] Current data center load balancing technology has the following shortcomings: the traditional ECMP method distributes traffic based on a simple hash algorithm, lacks real-time status awareness and business adaptation, and can easily lead to uneven traffic and low resource utilization; traditional SDN controller solutions have high communication latency and poor real-time response, and the centralized architecture has single points of failure and insufficient scalability; existing solutions do not fully utilize SRv6 technology and fail to effectively utilize its path orchestration and control capabilities, making it difficult to meet the fine-grained scheduling needs of diverse businesses; when dealing with complex scenarios such as multi-path transmission and service chain orchestration, there is a lack of efficient global optimization mechanisms, resulting in reduced network performance. Summary of the Invention

[0004] The present application provides an SRv6-based SDN data center load balancing method and system to solve the problems of low load balancing efficiency, insufficient dynamic adaptability, and low resource utilization in the prior art.

[0005] The first embodiment of the present application provides an SRv6-based SDN data center load balancing method, comprising the following steps: Step A1: obtaining the real-time load status of each node in the data center and the available resource information of each segment route in the SRv6 network through the SDN controller; Step A2: constructing a load balancing model based on the real-time load status and the available resource information to identify overloaded nodes and congested links in the network; Step A3: generating a traffic scheduling path based on the overloaded nodes and congested links, combined with the segmented routing characteristics of SRv6, wherein the path is an ordered list containing multiple SRv6 segment identifiers, and each SID corresponds to a Logical or physical path segment; Step A4: Based on the traffic scheduling path and combined with network status information, the load situation of the data center network is monitored in real time. When a node load imbalance or a link packet loss rate greater than 5% is detected, a load balancing adjustment mechanism is triggered, the imbalance type is distinguished, and the relevant traffic set is determined. The load balancing adjustment mechanism is optimized using the DQN algorithm to maximize network throughput and minimize load imbalance, thereby generating and dynamically adjusting the traffic scheduling strategy; Step A5: The adjusted SRv6 traffic scheduling strategy is sent to the data center network device, the device flow table information is updated through the Netconf protocol, and the traffic is redistributed.

[0006] In a preferred embodiment, a load balancing model is constructed based on the real-time load status and the available resource information to identify overloaded nodes and congested links in the network. The specific steps are as follows: Step B1, extracting the real-time load status parameters of the node, including CPU utilization , memory usage , network throughput , and the link's available resource information, including bandwidth utilization , delay , packet loss rate ; Step B2, construct a multi-dimensional dynamic load weight function , ;in, ; ; is the weight coefficient of each node parameter; is the multi-dimensional dynamic load weight function of the link, and the weight coefficient is optimized through historical data training; Step B3, if the node satisfies ( =0.8) or the link satisfies ( =0.75), To set a threshold, it is marked as an overloaded node or congested link.

[0007] In a preferred embodiment, the weight coefficient training method in step B2 includes: step B201, collecting historical load data to build a training set, and defining the loss function as mean square error (MSE): ,in, is the normalization factor; ; is the predicted value of node load; is the true value of the node load; is the square of the prediction error of a single sample; Step B202, optimize by gradient descent method .

[0008] In a preferred embodiment, according to the load balancing model, combined with the segment routing characteristics of SRv6, a traffic scheduling path is generated, and the specific steps are as follows: Step C1, abstract the data center network into a graph structure G (N, L, W), where N is a node set, L is a link set, and W is a network load matrix; Step C2, for the source node S and the destination node D, use the Dijkstra algorithm to find an initial path in the graph structure G, and the selection of the initial path aims to minimize the sum of the comprehensive load coefficients on the path, that is, ,in, To find the minimum value; Summation symbol, path Perform sum operation on each element i in ; is the weight of the link or node related to node i in the path; step C3, converting the initial path into an SRv6 segment route list, and configuring corresponding SRv6 header information according to the endpoint behavior (EndPointBehavior) of the segment route.

[0009] In a preferred embodiment, the traffic scheduling strategy is dynamically adjusted, and the specific steps are as follows: Step D1, constructing an environment model for reinforcement learning, wherein the environment model includes nodes, links, and their load status and resource information of the data center network; Step D2, initializing the reinforcement learning agent based on the environment model, setting the agent's state space, action space, and reward function, wherein the state space is a combination of information such as load balance, congestion, and remaining bandwidth of all nodes and links in the network, and the action space includes operations such as adding, deleting, or modifying SRv6 segment routes; Step D3, the agent selects an action to adjust the traffic scheduling path based on the current network status, observes the new network status and the reward value obtained after executing the action, and uses the Q-learning algorithm to update the agent's strategy, and the formula is: , where s is the current state; a is the current action; is the next state; R is the reward value; is the learning rate; is the discount factor; is the expected long-term reward of performing action a in state s; is the maximum Q value among all possible actions a′ after transferring to the new state s′; step D4, after multiple trainings, the intelligent agent can autonomously select the target SRv6 segment routing adjustment strategy according to the changes in the network status and dynamically optimize the traffic scheduling path.

[0010] In a preferred embodiment, when a node load imbalance or a link packet loss rate > 5% is detected, a load balancing adjustment mechanism is triggered, the imbalance type is distinguished, the relevant traffic set is determined, and resources are allocated. The specific steps are as follows: Step E1, when a sudden increase in traffic at a node or link is detected, exceeding a preset burst traffic threshold, the load balancing adjustment mechanism is triggered; Step E2, the SDN controller immediately queries whether there is a backup segment routing path in the SRv6 network, and the backup path must meet the requirement that the remaining bandwidth is greater than the peak bandwidth of the burst traffic; Step E3, if there is a backup path, the burst traffic is directed to the backup path for transmission; if there is no backup path, a new SRv6 segment routing path is temporarily created, and the required network resources are allocated to the burst traffic through the load balancing adjustment mechanism.

[0011] In a preferred embodiment, in step E1, when a sudden increase in traffic of a node or link is detected, exceeding a preset burst traffic threshold, a load balancing adjustment mechanism is triggered. The specific rules for triggering the load balancing adjustment mechanism are as follows: Step E101: If the node load imbalance continues to exceed =5s or link packet loss rate >5%, it is determined to be short-term congestion, triggering local path adjustment; Step E102, if the node load imbalance is accompanied by >90% and If the traffic volume exceeds 85%, it is considered as a long-term overload and triggers global traffic redistribution.

[0012] In a preferred embodiment, the load balancing adjustment mechanism is optimized using the DQN algorithm. The specific design of the DQN algorithm is as follows: Step F1, define the state space S including the node load weights , link weight , SRv6 segment list availability; Step F2, define the action space A as the insertion, deletion or replacement operation of the SRv6 segment list; Step F3, design the reward function ,in, + + =1; is the total path weight; ; is the load imbalance.

[0013] In a preferred embodiment, the training process of the DQN algorithm includes: step G1, using a dual network architecture (target network + online network), synchronizing parameters every K = 1000 steps; step G2, storing transfer samples (s, a, r, s') through the experience replay pool, and randomly sampling to break data correlation; step G3, using the Huber loss function to optimize Q value updates: in, is the true value; is the predicted value of the model, where s is the state and a is the action; hour, When , a form similar to L2 loss is adopted, the gradient changes smoothly and is more sensitive to smaller errors; , adopts a form similar to L1 loss, is insensitive to large errors (outliers), and avoids gradient explosion.

[0014] The third embodiment of the present application provides an SRv6-based SDN data center load balancing system, including: an acquisition module for obtaining the real-time load status of each node in the data center and the available resource information of each segment route in the SRv6 network through the SDN controller; a construction module for building a load balancing model based on the real-time load status and the available resource information, and identifying overloaded nodes and congested links in the network; a generation module for generating a traffic scheduling path based on the overloaded nodes and congested links, combined with the segmented routing characteristics of SRv6, wherein the path is an ordered list containing multiple SRv6 segment identifiers, and each SID corresponds to a logical a monitoring module for monitoring the load of the data center network in real time based on the traffic scheduling path and in combination with network status information. When a node load imbalance or a link packet loss rate greater than 5% is detected, a load balancing adjustment mechanism is triggered to distinguish the imbalance type, determine the relevant traffic set, and optimize the load balancing adjustment mechanism using the DQN algorithm to maximize network throughput and minimize load imbalance, thereby generating and dynamically adjusting the traffic scheduling strategy. An updating module is used to send the adjusted SRv6 traffic scheduling strategy to the data center network device, update the device's flow table information through the Netconf protocol, and redistribute traffic.

[0015] The beneficial effects of the present invention are: through the SDN controller to collect the load status of data center nodes and SRv6 routing resource information in real time, based on the dynamic load balancing model to accurately identify overloaded nodes and congested links, breaking through the blindness of traditional ECMP hash distribution and the limitations of the SDN controller's response lag, generating an orderly scheduling path with multiple segment identifiers (SIDs) according to the SRv6 segment routing characteristics, realizing fine-grained path orchestration of traffic, and providing differentiated path selection for diverse scenarios such as real-time services and high-bandwidth demand services. The DQN algorithm is introduced to construct an intelligent optimization mechanism, which triggers dynamic policy adjustment by real-time monitoring of node load imbalance or link packet loss rate (threshold > 5%), with the goal of maximizing network throughput and minimizing load imbalance, dynamically optimizing traffic distribution strategies, and improving the global optimization capabilities for complex scenarios such as multi-path transmission and service chain orchestration. By automatically issuing SRv6 scheduling policies and updating device flow tables through the Netconf protocol, traffic can be quickly redirected, reducing manual intervention costs, improving network response efficiency, and dynamically adapting to changes in data center traffic tidal patterns. This provides efficient and stable network support for emerging services such as cloud computing and AI computing power scheduling, enhancing the intelligent and flexible operational capabilities of data center networks. This solves the problems of low load balancing efficiency, insufficient dynamic adaptability, and low resource utilization in existing technologies.

[0016] Additional aspects and advantages of the present application will be given in part in the description below, and in part will become apparent from the description below, or will be learned through practice of the present application. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] The above and / or additional aspects and advantages of the present application will become apparent and easily understood from the following description of the embodiments in conjunction with the accompanying drawings, in which: Figure 1 This is a flow chart of an SRv6-based SDN data center load balancing method provided according to an embodiment of the present application; Figure 2 A schematic diagram of an SDN network optimization system for a cloud computing data center provided according to one embodiment of the present application; Figure 3 This is an example diagram of a data center network path system provided according to one embodiment of the present application; Figure 4 This is an example diagram of a video service peak period monitoring system provided according to one embodiment of the present application; Figure 5 This is an example diagram of a real-time trading system for a financial data center provided according to one embodiment of the present application; Figure 6 This is an example diagram of a system for scheduling sudden virtual machine migrations in a cloud computing data center according to one embodiment of the present application; Figure 7This is a flow chart of an SRv6-based SDN data center load balancing method provided according to one embodiment of the present application; Figure 8 This is a structural diagram of an SRv6-based SDN data center load balancing system provided according to an embodiment of the present application. DETAILED DESCRIPTION

[0018] The following describes in detail embodiments of the present application. Examples of the embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are intended to be used to explain the present application, and should not be construed as limiting the present application.

[0019] The following describes an SRv6-based SDN data center load balancing method and system according to an embodiment of the present application, with reference to the accompanying drawings. To address the low load balancing efficiency mentioned in the background art, the present application provides an SRv6-based SDN data center load balancing method. In this method, an SDN controller collects data center node load status and SRv6 routing resource information in real time, accurately identifies overloaded nodes and congested links based on a dynamic load balancing model, and overcomes the limitations of traditional ECMP hash distribution and SDN controller response lag. Based on the SRv6 segment routing feature, an ordered scheduling path containing multiple segment identifiers (SIDs) is generated to achieve fine-grained path orchestration for traffic. This provides differentiated path selection for diverse scenarios such as real-time services and high-bandwidth services. The DQN algorithm is introduced to construct an intelligent optimization mechanism. Dynamic policy adjustments are triggered by real-time monitoring of node load imbalance or link packet loss rate (threshold > 5%). With the goal of maximizing network throughput and minimizing load imbalance, the method dynamically optimizes traffic distribution strategies, improving global optimization capabilities for complex scenarios such as multi-path transmission and service chain orchestration. By automatically issuing SRv6 scheduling policies and updating device flow tables through the Netconf protocol, traffic can be quickly redirected, reducing manual intervention costs, improving network response efficiency, and dynamically adapting to changes in data center traffic tidal patterns. This provides efficient and stable network support for emerging services such as cloud computing and AI computing power scheduling, enhancing the intelligent and flexible operational capabilities of data center networks. This solves the problems of low load balancing efficiency, insufficient dynamic adaptability, and low resource utilization in existing technologies.

[0020] Specifically, Figure 1 This is a flow chart of an SRv6-based SDN data center load balancing method provided in an embodiment of the present application.

[0021] like Figure 1 As shown in FIG, the SRv6-based SDN data center load balancing method includes the following steps: In step A1, the real-time load status of each node in the data center and the available resource information of each route segment in the SRv6 network are obtained through the SDN controller.

[0022] Among them, the SDN controller is the core control plane component in the software-defined network architecture that is responsible for centrally managing network devices, monitoring network status in real time, and dynamically controlling data traffic paths by issuing policies.

[0023] It can be understood that the embodiments of the present application can centrally manage data center network equipment through the SDN controller, accurately monitor the load status of each node and the available resource information of SRv6 routing in real time, flexibly adjust the traffic path through dynamic delivery strategies, perform real-time perception of network status and intelligent scheduling of traffic, improve network resource utilization, reduce the probability of link congestion, and at the same time simplify network management complexity and enhance the adaptability of the data center network to diversified business needs and dynamic changes in traffic.

[0024] For example, Figure 2 As shown in the figure, during peak hours in a cloud computing data center, the SDN controller monitored in real time that a link between spine nodes in the leaf-spine architecture had a bandwidth utilization of 90% and a packet loss rate exceeding 8%, identifying it as a congested link. Based on the global network topology and node load data, the controller generated an SRv6 segment routing policy through dynamic path calculation, diverting the video streaming service traffic carried by this link to two lower-loaded paths (with bandwidth utilizations of 30% and 40%, respectively). The controller then updated the device flow table within 50ms using the NETCONF protocol. After this adjustment, the utilization of the previously congested link dropped to 55%, overall network throughput increased by 18%, and video service latency decreased from 25ms to 12ms. This effectively validated the SDN controller's core role in real-time network status awareness, intelligent traffic routing, and optimized resource allocation, significantly improving the dynamic adaptability and service carrying capacity of the data center network.

[0025] In step A2, a load balancing model is constructed based on the real-time load status and available resource information to identify overloaded nodes and congested links in the network.

[0026] Among them, a congested link refers to a communication link where data traffic exceeds its carrying capacity, resulting in increased transmission delay, increased packet loss rate and other performance degradations.

[0027] It can be understood that the embodiments of the present application analyze real-time load and resource information to accurately locate links with excessive traffic and degraded performance, provide targets for traffic scheduling, avoid increased congestion caused by blind traffic allocation, and rationally redistribute network resources by dynamically adjusting paths, thereby reducing link packet loss rate and transmission delay, and improving overall network throughput and stability.

[0028] In an embodiment of the present application, a load balancing model is constructed based on real-time load status and available resource information to identify overloaded nodes and congested links in the network. The specific steps are as follows: Step B1: Extract the real-time load status parameters of the node, including CPU utilization , memory usage , network throughput , and the link's available resource information, including bandwidth utilization , delay , packet loss rate ; Step B2: Constructing a multi-dimensional dynamic load weight function , ;in, ; ; is the weight coefficient of each node parameter; It is the multi-dimensional dynamic load weight function of the link, and the weight coefficient is optimized through historical data training; Step B3: If the node satisfies ( =0.8) or the link satisfies ( =0.75), To set a threshold, the node is marked as an overloaded node or a congested link.

[0029] The multi-dimensional dynamic load weight function is a mathematical model that dynamically calculates the load weight of each node or link based on multi-dimensional real-time status parameters such as network node load, link bandwidth, and latency, providing a quantitative decision-making basis for load balancing strategies. The formula for the multi-dimensional dynamic load weight function is: , ;in, ; ; is the weight coefficient of each node parameter; is the multi-dimensional dynamic load weight function of the link.

[0030] It can be understood that the embodiments of the present application use a multi-dimensional dynamic load weight function to optimize the weight coefficient based on real-time parameters such as node CPU utilization, memory occupancy, link bandwidth utilization, and latency through historical data training, dynamically calculate the load weight values of nodes and links, and quantitatively identify overloaded nodes and congested links, breaking through the limitations of single indicator evaluation, accurately capturing network load anomalies, providing a scientific and quantitative decision-making basis for load balancing strategies, avoiding the mechanical defects of static thresholds, improving the dynamic adaptability and accuracy of network status identification, dynamically adjusting traffic paths, optimizing resource allocation, reducing the probability of link congestion, and improving network throughput and stability.

[0031] For example, in real-time data center monitoring, the multidimensional dynamic load weight function collects information about node A's CPU utilization of 75% (weight 0.4), memory usage of 82% (weight 0.3), and network throughput of 90 Mbps (weight 0.3). The calculated load weight is (0.75 × 0.4 + 0.82 × 0.3 + 0.9 × 0.3 = 0.801), exceeding the threshold of 0.8 and marking it as overloaded. Meanwhile, link B has a bandwidth utilization of 85% (weight 0.5), a latency of 20 ms (weight 0.3), and a packet loss rate of 4% (weight 0.2). The calculated weight is (0.85 × 0.5 + 0.2 × 0.3 + 0.04 × 0.2 = 0.425 + 0.06 + 0.008 = 0.493), which does not reach the link threshold of 0.75 and is not marked. Based on this result, the system dispatched part of the database query traffic carried by node A to the low-load node B. After the adjustment, the load weight of node A dropped to 0.68, and the utilization rate of link B remained at 45%. This effectively verified the accuracy of the function in quantitatively evaluating multi-dimensional parameters and dynamically identifying overloaded nodes, providing a scientific basis for intelligent traffic scheduling and improving the level of refinement of network resource allocation.

[0032] In the embodiment of the present application, the training method of the weight coefficient in step B2 includes: step B201, collecting historical load data to build a training set, and defining the loss function as mean square error (MSE): ,in, is the normalization factor; ; is the predicted value of node load; is the true value of the node load; is the square of the prediction error of a single sample; Step B202, optimize by gradient descent method .

[0033] Among them, the loss function is the mean square error, which means that the average value of the square of the difference between the predicted value and the true value is used as a metric to measure the model prediction error. The loss function is the mean square error formula: ,in, is the normalization factor; ; is the predicted value of node load; is the true value of the node load; is the square of the prediction error for a single sample.

[0034] It can be understood that the embodiment of the present application quantifies the model prediction error by calculating the average of the square of the difference between the predicted value and the true value, provides an error metric for model optimization, uses the square operation to amplify the impact of large errors, and makes the model more focused on reducing errors. At the same time, it cooperates with optimization algorithms such as gradient descent to quickly adjust the weight coefficient so that the predicted value approaches the true value, thereby improving the model's fitting accuracy to historical load data.

[0035] In step A3, based on the overloaded nodes and congested links, combined with the segment routing feature of SRv6, a traffic scheduling path is generated. The path is an ordered list of multiple SRv6 segment identifiers, where each SID corresponds to a logical or physical path segment in the data center network.

[0036] The traffic scheduling path refers to an ordered forwarding path containing multiple SRv6 segment identifiers (SIDs) generated based on the SRv6 segment routing feature in a data center network to achieve load balancing or optimize transmission efficiency.

[0037] It can be understood that the embodiments of the present application, by combining the SRv6 segment routing characteristics, generate an ordered forwarding path containing multiple SRv6 segment identifiers (SIDs) based on overloaded node and congested link information, plan precise transmission paths for data traffic, avoid network bottlenecks, and use the SID's ability to identify logical / physical path segments to perform fine-grained path orchestration and flexible forwarding control of traffic, dynamically avoiding congested areas; through the combination of ordered lists of paths, support customized routing for differentiated services and improve service quality; with the help of SRv6's source routing mechanism, reduce the processing complexity of intermediate nodes, improve the response speed of traffic scheduling, adapt to network load changes, balance node and link loads, improve network resource utilization, and reduce congestion rates.

[0038] For example, during peak hours for a video platform, link X (95% bandwidth utilization, 40ms latency) between spine nodes in the data center's leaf-spine architecture is congested, while node Y (88% CPU utilization, 92% memory usage) is overloaded. The system generates two traffic scheduling paths based on SRv6 segment routing. Path one, consisting of SID_A (Locator: M-Function: Forwarding-Args: Low Latency) and SID_B (Locator: N-Function: Load Balancing), directs the HD live stream along the logical path M→N→P (18ms latency). Path two, consisting of SID_C (Locator: O-Function: Diversion-Args: High Bandwidth) and SID_D (Locator: P-Function: Forwarding), directs video cache download traffic along the physical path O→P→Q (30% bandwidth utilization). Through the NETCONF protocol, policy delivery was completed within 100ms, reducing link X utilization to 58%, node Y load to 70%, live streaming latency from 45ms to 22ms, and download speeds by 50%. This effectively demonstrated the ability of traffic scheduling paths to achieve refined traffic diversion and dynamically avoid network bottlenecks through orderly SID combinations, significantly enhancing the data center's traffic processing efficiency and business stability in high-concurrency scenarios.

[0039] In the embodiment of the present application, based on the load balancing model and combined with the segment routing characteristics of SRv6, a traffic scheduling path is generated. The specific steps are as follows: Step C1: Abstract the data center network into a graph structure G(N, L, W), where N is the node set, L is the link set, and W is the network load matrix; Step C2: For the source node S and the destination node D, use the Dijkstra algorithm to find the initial path in the graph structure G. The selection of the initial path aims to minimize the sum of the comprehensive load coefficients on the path, that is, ,in, To find the minimum value; Summation symbol, path Perform sum operation on each element i in; is the weight of the link or node related to node i in the path; Step C3: Convert the initial path into an SRv6 segment route list, and configure the corresponding SRv6 header information according to the endpoint behavior (EndPointBehavior) of the segment route.

[0040] Among them, the Dijkstra algorithm is a greedy algorithm for solving the single-source shortest path in a weighted graph with non-negative weights.

[0041] It can be understood that the embodiment of the present application abstracts the data center network into a graph structure with load weights, with the goal of minimizing the sum of the comprehensive load coefficients of the paths, quickly calculates the initial optimized path for the source node and the destination node, and converts multi-dimensional parameters such as node load and link bandwidth utilization into path weights, breaking through the limitations of the traditional shortest path that only relies on the number of hops or delay, avoiding overloaded nodes and congested links, and using greedy strategies to efficiently iterate and solve, ensuring the real-time nature of path calculation and adapting to the dynamic load changes of the data center network; by quantifying the comprehensive load, global optimization of path selection is achieved, reducing the congestion risk caused by blind traffic distribution; the generated initial path can be directly mapped to an SRv6 segment routing list, providing a data basis for fine-grained path orchestration, matching the optimal transmission path for differentiated services, improving the rationality of network resource allocation and the intelligence of traffic scheduling, and enhancing the ability of data centers to cope with complex load scenarios.

[0042] For example, Figure 3 As shown in Figure 1, in data center network modeling, node and link loads are abstracted as a weighted graph (for example, the link load from source node R to node M is weighted 4, the link load to node N is weighted 6, the load from M to target node T is weighted 3, the load from M to node P is weighted 5, and the load from N to P is weighted 2). When R needs to transmit real-time service traffic to T, the Dijkstra algorithm is used to calculate the path with the goal of minimizing the overall path load. Starting from R, the algorithm first selects M, the node with the lowest load among its neighbors (total load of 4), then evaluates the path load from M to T as 4 + 3 = 7. Simultaneously, the path R→N→P→T is calculated, with a load of 6 from R to N, a load of 2 from N to P, and an assumed load of 3 from P to T, for a total load of 6 + 2 + 3 = 11. In comparison, the R→M→T path is superior. This path is converted into an SRv6 segment routing list (including SID_M and SID_T) and then issued, allowing real-time traffic to avoid the highly loaded N→P link. The utilization rate of the original congested N→P link is reduced from 80% to 50%, and the end-to-end latency is reduced from 40ms to 22ms. This demonstrates the Dijkstra algorithm's ability to quickly generate low-load paths by quantifying load weights and ensure the quality of real-time service transmission. It also provides an efficient initial path planning solution for SRv6 traffic scheduling.

[0043] In step A4, based on the traffic scheduling path and combined with network status information, the load situation of the data center network is monitored in real time. When a node load imbalance or a link packet loss rate greater than 5% is detected, the load balancing adjustment mechanism is triggered, the imbalance type is distinguished, the relevant traffic set is determined, and the load balancing adjustment mechanism is optimized using the DQN algorithm. With the goal of maximizing network throughput and minimizing load imbalance, the traffic scheduling strategy is generated and dynamically adjusted.

[0044] Among them, the link packet loss rate is an indicator to measure the link quality during data transmission, which refers to the ratio of the number of data packets lost in a specific link to the total number of transmitted data packets.

[0045] It can be understood that the embodiments of the present application can quantify the link data transmission quality in real time by monitoring the link packet loss rate, accurately identify data packet loss anomalies caused by traffic overload, link failure, etc. (such as packet loss rate > 5%), and provide key trigger signals for the load balancing adjustment mechanism. By dynamically tracking changes in packet loss rate, congested links or performance-degraded nodes can be located in a timely manner, reducing the additional network load increased by data retransmission; combined with the node load imbalance state, the imbalance type can be accurately distinguished and the relevant traffic set can be determined, providing real-time status input for the DQN algorithm to optimize the load balancing strategy, and generating dynamic traffic scheduling with the goal of maximizing network throughput and minimizing load imbalance; at the same time, the transmission stability is improved and the impact of packet loss on delay-sensitive applications is reduced.

[0046] For example, Figure 4 As shown in the figure, during peak video service hours in the data center, the monitoring system detected that the packet loss rate on link L3 between spine nodes in the leaf-spine architecture suddenly increased from a normal 1.2% to 6.5% (exceeding the 5% threshold). Simultaneously, the CPU utilization of the associated node S5 reached 88% and the memory usage reached 92%, indicating link congestion caused by overload. The system immediately triggered the load balancing mechanism. By analyzing traffic characteristics, it identified HD video streams as the primary imbalanced traffic. Using the DQN algorithm, it optimized and generated a new SRv6 traffic scheduling policy, diverting 30% of the video stream traffic from link L3 to two low-loss paths, L1 (with a 0.8% packet loss rate) and L2 (with a 1.1% packet loss rate). After the policy adjustment, the packet loss rate on link L3 dropped to 2.1%, the load on node S5 decreased to 70%, and the overall network video service throughput increased by 22%, with latency reduced from 35ms to 18ms. This effectively demonstrated the effectiveness of monitoring link packet loss rates in real-time congestion identification and driving intelligent traffic scheduling to ensure service transmission quality.

[0047] In the embodiment of the present application, the traffic scheduling strategy is dynamically adjusted. The specific steps are as follows: Step D1: Build an environment model for reinforcement learning. The environment model includes nodes and links in the data center network, as well as their load status and resource information. Step D2: Initialize the reinforcement learning agent based on the environment model and set the agent's state space, action space, and reward function. The state space is a combination of information such as load balancing, congestion, and remaining bandwidth for all nodes and links in the network. The action space includes operations such as adding, deleting, or modifying SRv6 segment routes. Step D3: The agent selects an action to adjust the traffic scheduling path based on the current network status. After executing the action, it observes the new network status and the reward value obtained, and uses the Q-learning algorithm to update the agent's strategy. The formula is: , where s is the current state; a is the current action; is the next state; R is the reward value; is the learning rate; is the discount factor; is the expected long-term reward of performing action a in state s; is the maximum Q value of all possible actions a′ after transitioning to the new state s′; Step D4: After multiple training sessions, the agent is able to autonomously select the target SRv6 segment routing adjustment strategy based on changes in network status and dynamically optimize the traffic scheduling path.

[0048] Among them, the Q-learning algorithm is a model-free offline strategy algorithm in reinforcement learning that optimizes the strategy by iteratively updating the action value function Q(s,a). The Q-learning algorithm formula is: , where s is the current state; a is the current action; is the next state; R is the reward value; is the learning rate; is the discount factor; is the expected long-term reward of performing action a in state s; is the maximum Q value of all possible actions a′ after transitioning to the new state s′.

[0049] It can be understood that the embodiment of the present application constructs a data center network environment model, takes multi-dimensional status information such as node load and link congestion as the input of the intelligent agent, adjusts the SRv6 segment routing strategy into the action space, and iteratively updates the action value function Q(s,a) through the Bellman equation, so that the intelligent agent can autonomously learn the optimal traffic scheduling strategy under dynamic network conditions without a network prior model, with the goal of maximizing network throughput and minimizing load imbalance, accurately capture the real-time status changes of nodes and links, dynamically optimize the traffic scheduling strategy monitoring, adapt to the uncertainty and real-time variability of data center network load, reduce excessive dependence on current strategies through offline strategy learning, and use reward mechanisms to guide the intelligent agent to explore long-term optimal decisions, improve the adaptability of traffic scheduling to complex business needs, reduce link packet loss rate, balance node load, and enhance the autonomous optimization capability of network resource utilization efficiency and business service quality.

[0050] For example, Figure 5As shown in the figure, during peak real-time trading hours in a financial data center, core node B experienced a burst of high-frequency trading traffic, causing its memory utilization to reach 90% and its L7 packet loss rate to climb to 6.8%. A Q-learning agent uses node load balance, link latency (L5 latency 15ms, L6 latency 18ms), and remaining bandwidth as its state space. Its action space consists of dynamically replacing SRv6 segment identifiers on highly congested paths (e.g., replacing SID_L7 with SID_L5-20 and SID_L6-25). The reward function is defined as a reward of +50 for every 10% reduction in load imbalance and a penalty of -20 for every 1% increase in packet loss above a threshold. Under the initial strategy, the agent attempts to reroute 40% of transaction traffic from L7 to parallel paths on L5 and L6. This strategy reduces node B's memory utilization to 75% and its L7 packet loss rate to 3.2%, resulting in a reward of +35. The agent then updates its strategy using a Q-learning formula (learning rate α = 0.25, discount factor γ = 0.85). After 800 training cycles, the agent developed an optimization strategy: when node memory usage is ≥85% and link packet loss rate is >5%, it prioritizes the top three links that match latency and traffic reliability requirements to form an SRv6 segment routing combination. In practice, this reduced transaction order processing latency from 45ms to 20ms, and reduced link L7 congestion by 65%. This demonstrates Q-learning's ability to autonomously optimize traffic scheduling strategies through state-value iteration, dynamically adapting to network load fluctuations in high-frequency trading scenarios, and effectively ensuring the stable transmission of low-latency, high-reliability services.

[0051] In the embodiment of the present application, when a node load imbalance or a link packet loss rate greater than 5% is detected, a load balancing adjustment mechanism is triggered to distinguish the type of imbalance and determine the relevant traffic set. The specific steps are as follows: Step E1: When a sudden increase in traffic on a node or link is detected, exceeding a preset burst traffic threshold, a load balancing adjustment mechanism is triggered; Step E2: The SDN controller immediately queries whether there is a backup segment routing path in the SRv6 network. The backup path must have a remaining bandwidth greater than the peak bandwidth of the burst traffic. Step E3: If a backup path exists, direct the burst traffic to the backup path for transmission; if no backup path exists, temporarily create a new SRv6 segment routing path and allocate the required network resources to the burst traffic through the load balancing adjustment mechanism.

[0052] Among them, the load balancing adjustment mechanism is an automated control mechanism that monitors the load status of network nodes and links in real time, and dynamically adjusts traffic paths, task allocation or resource configuration based on preset strategies or intelligent algorithms to balance system load, improve resource utilization, reduce latency and ensure business stability.

[0053] It can be understood that the embodiments of the present application trigger an automated control process by real-time monitoring of abnormal conditions such as node load imbalance or link packet loss rate exceeding the threshold: with the help of the SDN controller, quickly query or create SRv6 backup segment routing paths that meet bandwidth requirements, dynamically guide burst traffic to avoid congested areas, respond to network load fluctuations in a timely manner, accurately distinguish the imbalance type and match the traffic scheduling strategy, quickly divert burst traffic and dynamically allocate resources, use the flexible path control capability of SRv6 segment routing, respond to traffic burst scenarios in milliseconds, and reduce network paralysis caused by single-point overload; by prioritizing the use of existing backup paths or temporarily creating new paths, maximize the use of remaining network resources, improve link bandwidth utilization and node load balancing, and reduce data packet retransmission and delay jitter caused by congestion.

[0054] For example, during prime time at night on a live video streaming platform, the memory utilization of edge node Edge-09 suddenly reached 89% (exceeding the preset 80% threshold). The packet loss rate of its connected aggregation link, L-09-22, rose to 5.8% (>5%), triggering the load balancing mechanism. The SDN controller quickly searched the SRv6 network and discovered that the temporarily created backup path needed to meet the peak bandwidth of the burst live stream (18 Gbps). It then invoked the dynamically generated SRv6 segment routing combination SID_Live1 (Locator: Edge-11 - Function: Divert) and SID_Live2 (Locator: Core-03 - Function: Forward), rerouting 40% of the HD live streaming traffic to the path from Edge-11 to Core-03 to the CDN node cluster. After the adjustment, the Edge-09 memory usage dropped to 72%, the L-09-22 packet loss rate dropped to 1.9%, the live stream freeze rate dropped from 12% to 3%, and the end-to-end latency dropped from 60ms to 35ms. This effectively demonstrated the load balancing adjustment mechanism's ability to intelligently divert live traffic through real-time detection and dynamic path creation, ensuring user viewing experience and system stability in high-concurrency live broadcast scenarios.

[0055] In the embodiment of the present application, in step E1, when a sudden increase in traffic on a node or link is detected, exceeding a preset burst traffic threshold, a load balancing adjustment mechanism is triggered. The specific rules for triggering the load balancing adjustment mechanism are as follows: Step E101: If the node load imbalance continues to exceed =5s or link packet loss rate If the traffic congestion rate is >5%, it is considered short-term congestion and triggers local path adjustment. Step E102: If the node load is unbalanced, >90% and If the traffic volume exceeds 85%, it is considered as a long-term overload and triggers global traffic redistribution.

[0056] It can be understood that the embodiments of the present application set differentiated judgment conditions, identify network anomaly types, provide data for hierarchical regulation, trigger local path adjustments for short-term congestion, and respond to sudden traffic fluctuations in seconds to prevent mild overloads from spreading into global failures; implement global traffic redistribution for long-term overloads, optimize resource allocation from the underlying network architecture level, and prevent key nodes from continuously running at high loads and causing system crashes; reduce misjudgments and excessive adjustments through hierarchical threshold design, ensure network response agility, and at the same time, improve the accuracy of regulation strategies, conduct fine-grained management of data center network load status, balance traffic scheduling efficiency and system resource consumption, and enhance network resilience and service quality stability in complex business scenarios.

[0057] In step A5, the adjusted SRv6 traffic scheduling policy is sent to the data center network device, the device's flow table information is updated through the Netconf protocol, and the traffic is redistributed.

[0058] Among them, the Netconf protocol is an XML-based network configuration protocol.

[0059] It can be understood that the embodiments of the present application accurately convert abstract traffic scheduling policies into device-executable flow table configurations through standardized XML modeling and remote procedure call mechanisms, and use the transactional operations of the protocol to ensure the atomicity and reliability of configuration updates, thereby preventing network anomalies caused by partial configuration failures; the XML-based structured data model improves the clarity and scalability of policy descriptions, and accurately transmits complex SRv6 path parameters; the remote real-time operation capability meets the dynamic requirements of traffic scheduling, can complete policy issuance and device configuration refresh, and adapt to the data center network's rapid response requirements to load changes. At the same time, the standardized protocol eliminates the compatibility barriers of multi-vendor equipment, ensures the unified execution of SRv6 policies in a cross-vendor environment, and enhances the standardization and operation and maintenance efficiency of network configuration management.

[0060] For example, Figure 6 As shown in the figure, when a sudden wave of virtual machine migrations occurred in the cloud computing data center, the bandwidth utilization of the spine link L23-45 connected to the leaf node Leaf-23 soared to 92% (threshold 80%), triggering a traffic scheduling policy adjustment. The SDN controller encoded the optimized SRv6 segment routing policy (including SID_Migrate1 (Locator: Leaf-18-Function: Traffic Diversion-Args: Bandwidth Threshold 5Gbps) and SID_Migrate2 (Locator: Spine-09-Function: Load Balancing)) into an XML configuration file using the NETCONF protocol. <lock>and <commit>The RPC operation performed an atomic configuration update on the target switch, S45: first locking the configuration database, then issuing two path rules in parallel, redirecting 45% of the migration traffic along the path from Leaf-18 to Spine-09 to the target cluster. The protocol ensured that the configuration took effect within 120ms, reducing L23-45 utilization to 68%, shortening the average VM migration time from 28 minutes to 19 minutes, and reducing the migration failure rate from 15% to 4%. This demonstrates NETCONF's ability to achieve rapid and reliable policy deployment in dynamic cloud resource scheduling through standardized XML modeling and transactional operations, effectively ensuring network stability and service efficiency in large-scale data migration scenarios.

[0061] In the embodiment of the present application, the DQN algorithm is used to optimize the load balancing adjustment mechanism. The specific design of the DQN algorithm is as follows: Step F1: Define the state space S including the node load weights , link weight ,SRv6 segment list availability; Step F2: define action space A as insert, delete, or replace operations of the SRv6 segment list; Step F3: Design reward function ,in, + + =1; is the total weight of the path; ; is the load imbalance.

[0062] Among them, the reward function is a function used to quantify the effectiveness of the agent's behavior in reinforcement learning.

[0063] It is understandable that the embodiment of the present application quantifies the total weight of the path and load imbalance To evaluate the quality of the intelligent agent's actions, with the help of reinforcement learning mechanism, the intelligent agent can learn through continuous trial and error, automatically explore and select operation strategies that reduce the total weight of the path and reduce the load imbalance, optimize the load balancing adjustment mechanism, improve network resource utilization and operation efficiency, and perform intelligent traffic scheduling and resource allocation.

[0064] In this embodiment of the present application, the training process of the DQN algorithm includes: Step G1: Use a dual network architecture (target network + online network) and synchronize parameters every K = 1000 steps; Step G2: Store the transferred samples (s, a, r, s') in the experience replay pool and randomly sample to break the data correlation; Step G3: Use Huber loss function to optimize Q value update: in, is the true value; is the predicted value of the model, where s is the state and a is the action; hour, When , a form similar to L2 loss is adopted, the gradient changes smoothly and is more sensitive to smaller errors; , adopts a form similar to L1 loss, is insensitive to large errors (outliers), and avoids gradient explosion.

[0065] Among them, the experience replay pool is a buffer used in reinforcement learning to store the experience generated by the interaction between the agent and the environment.

[0066] It can be understood that the embodiment of the present application stores the transfer samples (s, a, r, s') generated by the interaction between the intelligent agent and the environment, and randomly samples in DQN training to break the time correlation of the data, satisfy the independent and identically distributed assumption, and prevent the instability of model training caused by the strong correlation of consecutive samples. By reusing historical experience, the sample utilization efficiency is improved, the dependence on real-time interaction data is reduced, the training cost is reduced, the gradient variance problem is alleviated, and the robustness of the Huber loss function to outliers is combined to enhance the stability of the Q value update, prevent gradient explosion, and improve the convergence speed and generalization ability of the DQN algorithm in optimizing the load balancing strategy in a dynamic network environment.

[0067] For example, in an intelligent transportation system (ITS) intersection traffic control scenario, a DQN agent is responsible for optimizing signal timing and vehicle routing strategies at the intersection. An experience replay pool stores 150,000 experience samples of the agent's interactions with the traffic environment, including the length of vehicle queues in each direction at the intersection (state s), signal phase switching or routing actions (action a), delay reduction (reward r), and the traffic distribution at the next moment (state s'). During training, 64 batches of samples are randomly drawn from the pool at each step (e.g., experience with emergency dispatch during the morning rush hour). This prevents the model from being affected by the strong correlation of traffic tidal effects across consecutive samples. When the target network synchronizes parameters every 800 steps, the experience replay pool enables the agent to repeatedly learn from historically effective policies. For example, experiences with "preferring to extend the green light duration for left turns and redirecting some vehicles to secondary roads during east-west congestion" are sampled repeatedly, preventing the risk of intersection paralysis caused by excessive exploration of new policies. After 3,000 training sessions, compared with traditional reinforcement learning algorithms without an experience replay mechanism, the average delay time at intersections was reduced by 38%, the standard deviation of queue lengths was reduced by 52%, and the number of strategy oscillations during training was reduced by 70%. This significantly improved the stability and real-time performance of the DQN algorithm in optimizing signal timing and path guidance in dynamic traffic flow scenarios, effectively alleviating congestion problems at complex traffic intersections.

[0068] According to an SRv6-based SDN data center load balancing method proposed in an embodiment of the present application, the SDN controller collects data center node load status and SRv6 routing resource information in real time, accurately identifies overloaded nodes and congested links based on a dynamic load balancing model, breaks through the blindness of traditional ECMP hash distribution and the limitations of delayed response of the SDN controller, generates an ordered scheduling path containing multiple segment identifiers (SIDs) based on the SRv6 segment routing characteristics, realizes fine-grained path orchestration of traffic, and can provide differentiated path selection for diverse scenarios such as real-time services and high-bandwidth demand services. The DQN algorithm is introduced to construct an intelligent optimization mechanism, which triggers dynamic policy adjustment by real-time monitoring of node load imbalance or link packet loss rate (threshold > 5%). With the goal of maximizing network throughput and minimizing load imbalance, it dynamically optimizes traffic distribution strategies and improves the global optimization capabilities for complex scenarios such as multi-path transmission and service chain orchestration. By automatically issuing SRv6 scheduling policies and updating device flow tables through the Netconf protocol, traffic can be quickly redirected, reducing manual intervention costs, improving network response efficiency, and dynamically adapting to changes in data center traffic tidal patterns. This provides efficient and stable network support for emerging services such as cloud computing and AI computing power scheduling, enhancing the intelligent and flexible operational capabilities of data center networks. This solves the problems of low load balancing efficiency, insufficient dynamic adaptability, and low resource utilization in existing technologies.

[0069] The following will describe an SRv6-based SDN data center load balancing method through a specific embodiment. Figure 7 Shown, including: During high-frequency trading hours in the financial data center, the SDN controller collects real-time data on the operating status of 200 server nodes and links within the data center at a 5-second interval. At the node level, it monitors CPU utilization, memory usage, and network throughput. For example, at 9:15 a.m., the CPU utilization of node Server-037 reached 82% and the memory usage reached 75%, far exceeding the average level. At the link level, it monitors bandwidth utilization, latency, and packet loss rate. For example, during monitoring, the bandwidth utilization of the core link Leaf1-Spine2 reached 88%, and the latency increased from 2ms to 8ms, indicating mild congestion. This provides real-time data support for load balancing.

[0070] A multi-dimensional weighted model is used to quantify node and link loads: Node load is calculated as 0.5 × CPU utilization + 0.3 × memory usage + 0.2 × normalized network throughput, while link load is calculated as 0.6 × bandwidth utilization + 0.3 × normalized latency + 0.1 × packet loss rate. Weights are dynamically trained based on historical trading period data (e.g., weight differences between early morning low-peak and morning peak hours). Thresholds are set for node load ≥ 0.85 and link load ≥ 0.8. For example, if the calculated node load value for Server-037 is 0.89 (exceeding the threshold), and the calculated link load value for Leaf1-Spine2 is 0.86 (exceeding the threshold), these nodes are marked as overloaded and congested, respectively.

[0071] The data center network is abstracted as a "node-link" graph structure with real-time load weights. When the source node Server-010 transmits transaction data to the destination node Server-150, the SDN controller uses the Dijkstra algorithm to calculate the initial path with the minimum total load: Server-010 → Leaf3 → Spine1 → Leaf8 → Server-150. It then converts this path into an SRv6 segment routing list [SID-Leaf3, SID-Spine1, SID-Leaf8]. Each SID corresponds to a one-hop logical path in the network, implementing initial traffic scheduling.

[0072] At 9:20 a.m., the packet loss rate on the Leaf3-Spine1 link suddenly increased to 6% (exceeding the 5% threshold) and persisted for four seconds, indicating short-term congestion and triggering a local path adjustment mechanism. The controller identified the backup path, Server-010 → Leaf3 → Spine3 → Leaf8 → Server-150, with 1.2 Gbps of remaining bandwidth, capable of carrying a peak burst of 900 Mbps. It then generated a new SRv6 segment routing table [SID-Leaf3, SID-Spine3, SID-Leaf8] and switched 50% of traffic from Server-010 to Server-150 to the backup path. Using the DQN algorithm to train an intelligent agent, after a week of optimization, the agent could identify potential congestion two seconds in advance, reducing average latency by 12%, and enabling autonomous optimization of traffic scheduling strategies.

[0073] The SDN controller directly distributes the adjusted SRv6 traffic scheduling policy through a graphical user interface or northbound interface, sending instructions to the Leaf3 and Spine3 devices to update the flow table to identify the new SID sequence. This completes the path switch within 5 seconds, ensuring uninterrupted transaction traffic. This process requires no code, automating traffic redistribution and monitoring the effects of adjustments in real time, ensuring data center network stability and transaction continuity during peak hours.

[0074] In summary, the present invention collects node and link status in real time (5-second cycle) through the SDN controller, combines the multi-dimensional weight model to dynamically and accurately identify overloaded nodes (such as Server-037 load 0.89) and congested links (such as Leaf1-Spine2 load 0.86), and generates the initial optimized path based on SRv6 segment routing and Dijkstra algorithm; in the face of sudden congestion (such as Leaf3-Spine1 packet loss rate 6%), it can determine short-term congestion within 4 seconds and trigger local path adjustment, using the backup path (remaining bandwidth 1.2Gbps to carry 90 0Mbps peak traffic) achieved 50% traffic switching, and at the same time, trained the intelligent agent through the DQN algorithm, enabling potential congestion identification to be advanced by 2 seconds and reducing average latency by 12%. Finally, through automated policy delivery (flow table updates completed within 5 seconds), without business interruption, node load balancing was improved by more than 20%, link bandwidth utilization was optimized to 72%, transaction latency was stabilized within 8ms, and packet loss rate was controlled below 1%. This enhanced the stability, reliability, and resource utilization efficiency of the data center network during high-frequency trading periods, and ensured the continuity and real-time performance of core trading and clearing businesses.

[0075] Next, an SRv6-based SDN data center load balancing system proposed according to an embodiment of the present application is described with reference to the accompanying drawings.

[0076] Specifically, Figure 8 A schematic diagram of the composition of an SRv6-based SDN data center load balancing system provided in an embodiment of the present application.

[0077] like Figure 8 As shown, the SRv6-based SDN data center load balancing system 10 includes: an acquisition module 100, a construction module 200, a generation module 300, a monitoring module 400, and an update module 500.

[0078] Among them, the acquisition module 100 is used to obtain the real-time load status of each node in the data center and the available resource information of each route segment in the SRv6 network through the SDN controller; the construction module 200 is used to build a load balancing model based on the real-time load status and available resource information, and identify overloaded nodes and congested links in the network; the generation module 300 is used to generate a traffic scheduling path based on the overloaded nodes and congested links, combined with the segmented routing characteristics of SRv6. The path is an ordered list containing multiple SRv6 segment identifiers, and each SID corresponds to a logical or physical path segment in the data center network; the monitoring module 400 is used to generate a traffic scheduling path based on the overloaded nodes and congested links. Based on the traffic scheduling path, combined with network status information, the load situation of the data center network is monitored in real time. When a node load imbalance or a link packet loss rate greater than 5% is detected, the load balancing adjustment mechanism is triggered, the imbalance type is distinguished, and the relevant traffic set is determined. The load balancing adjustment mechanism is optimized using the DQN algorithm to maximize network throughput and minimize load imbalance, and the traffic scheduling strategy is generated and dynamically adjusted. The update module 500 is used to send the adjusted SRv6 traffic scheduling strategy to the data center network equipment, update the device's flow table information through the Netconf protocol, and redistribute the traffic.

[0079] It should be noted that the above explanation of an embodiment of an SRv6-based SDN data center load balancing method is also applicable to an SRv6-based SDN data center load balancing system of this embodiment, and will not be repeated here.

[0080] According to the embodiment of the present application, an SRv6-based SDN data center load balancing system is proposed. The SDN controller collects the data center node load status and SRv6 routing resource information in real time, accurately identifies overloaded nodes and congested links based on a dynamic load balancing model, breaks through the blindness of traditional ECMP hash distribution and the limitations of the SDN controller's response lag, generates an ordered scheduling path containing multiple segment identifiers (SIDs) based on the SRv6 segment routing characteristics, realizes fine-grained path orchestration of traffic, and can provide differentiated path selection for diverse scenarios such as real-time services and high-bandwidth demand services. The DQN algorithm is introduced to construct an intelligent optimization mechanism. Dynamic policy adjustment is triggered by real-time monitoring of node load imbalance or link packet loss rate (threshold > 5%). With the goal of maximizing network throughput and minimizing load imbalance, the traffic distribution strategy is dynamically optimized, and the global optimization capability for complex scenarios such as multi-path transmission and service chain orchestration is improved. By automatically issuing SRv6 scheduling policies and updating device flow tables through the Netconf protocol, traffic can be quickly redirected, reducing manual intervention costs, improving network response efficiency, and dynamically adapting to changes in data center traffic tidal patterns. This provides efficient and stable network support for emerging services such as cloud computing and AI computing power scheduling, enhancing the intelligent and flexible operational capabilities of data center networks. This solves the problems of low load balancing efficiency, insufficient dynamic adaptability, and low resource utilization in existing technologies.

[0081] In the description of this specification, the reference terms "one embodiment", "some embodiments", "example", "specific example", or "some examples" mean that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present application. In this specification, the schematic expressions of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described can be combined in any one or N embodiments or examples in a suitable manner. In addition, those skilled in the art can combine and combine different embodiments or examples described in this specification and the features of different embodiments or examples without contradiction.

[0082] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features being referred to. Thus, a feature specified as "first" or "second" may explicitly or implicitly include at least one such feature. Throughout the description of this application, "N" means at least two, for example, two, three, etc., unless otherwise specifically defined.

[0083] Any process or method description in a flowchart or otherwise described herein may be understood to represent a module, fragment or portion of code comprising one or more executable instructions for implementing the steps of a custom logical function or process, and the scope of the preferred embodiments of the present application includes alternative implementations in which functions may be performed out of the order shown or discussed, including performing functions in a substantially simultaneous manner or in reverse order depending on the functions involved, which should be understood by those skilled in the art to which the embodiments of the present application belong.

[0084] It should be understood that various parts of this application can be implemented using hardware, software, firmware, or a combination thereof. In the above embodiment, the N steps or methods can be implemented using software or firmware stored in a memory and executed by a suitable instruction execution system. For example, if implemented using hardware, as in another embodiment, any one of the following technologies known in the art or a combination thereof can be used: a discrete logic circuit having logic gate circuits for implementing logical functions on data signals, an application-specific integrated circuit having suitable combinational logic gate circuits, a programmable gate array (PGA), a field-programmable gate array (FPGA), etc.

[0085] Those skilled in the art will appreciate that all or part of the steps in the method for implementing the above-mentioned embodiment can be completed by instructing related hardware through a program, and the program can be stored in a computer-readable storage medium. When the program is executed, it includes one or a combination of the steps of the method embodiment.< / commit> < / lock>

Claims

1. A method for load balancing in an SDN data center based on SRv6, characterized in that: The following steps are involved: Step A1: Obtain the real-time load status of each node in the data center and the available resource information of each route segment in the SRv6 network through the SDN controller; Step A2: Building a load balancing model based on the real-time load status and the available resource information to identify overloaded nodes and congested links in the network; Step A3: Based on the overloaded nodes and congested links, and in combination with the segment routing feature of SRv6, a traffic scheduling path is generated. The path is an ordered list of multiple SRv6 segment identifiers, where each SID corresponds to a logical or physical path segment in the data center network. Step A4: Based on the traffic scheduling path and combined with network status information, the data center network load is monitored in real time. When a node load imbalance or a link packet loss rate greater than 5% is detected, a load balancing adjustment mechanism is triggered, the imbalance type is distinguished, the relevant traffic set is determined, and the load balancing adjustment mechanism is optimized using the DQN algorithm to maximize network throughput and minimize load imbalance, thereby generating and dynamically adjusting the traffic scheduling strategy. Step A5: Send the adjusted SRv6 traffic scheduling policy to the data center network device, update the device's flow table information through the Netconf protocol, and redistribute the traffic.

2. The SRv6-based SDN data center load balancing method according to claim 1, characterized in that: Building a load balancing model based on the real-time load status and the available resource information to identify overloaded nodes and congested links in the network, the specific steps are as follows: Step B1: Extract the real-time load status parameters of the node, including CPU utilization , memory usage , network throughput , and the link's available resource information, including bandwidth utilization , delay , packet loss rate ; Step B2: Constructing a multi-dimensional dynamic load weight function , ;in, ; ; is the weight coefficient of each node parameter; It is the multi-dimensional dynamic load weight function of the link, and the weight coefficient is optimized through historical data training; Step B3: If the node satisfies ( =0.8) or the link satisfies ( =0.75), To set a threshold, it is marked as an overloaded node or congested link.

3. The SRv6-based SDN data center load balancing method according to claim 2, characterized in that: The training method of the weight coefficient in step B2 includes: Step B201: Collect historical load data to build a training set, and define the loss function as mean square error (MSE): ,in, is the normalization factor; ; is the predicted value of node load; is the true value of the node load; is the square of the prediction error of a single sample; Step B202: Optimize by gradient descent method .

4. The SRv6-based SDN data center load balancing method according to claim 1, characterized in that: Based on the load balancing model and the segment routing feature of SRv6, a traffic scheduling path is generated. The specific steps are as follows: Step C1: Abstract the data center network into a graph structure G(N, L, W), where N is the node set, L is the link set, and W is the network load matrix; Step C2: For the source node S and the destination node D, use the Dijkstra algorithm to find the initial path in the graph structure G. The selection of the initial path aims to minimize the sum of the comprehensive load coefficients on the path, that is, ,in, To find the minimum value; Summation symbol, path Perform sum operation on each element i in; is the weight of the link or node related to node i in the path; Step C3: Convert the initial path into an SRv6 segment route list, and configure corresponding SRv6 header information according to the endpoint behavior (EndPointBehavior) of the segment route.

5. According to the SRv6-based SDN data center load balancing method of claim 1, the traffic scheduling strategy is dynamically adjusted, and the specific steps are as follows: Step D1: Constructing an environment model for reinforcement learning, wherein the environment model includes nodes and links of the data center network and their load status and resource information; Step D2: Initialize the reinforcement learning agent based on the environment model and set the agent's state space, action space, and reward function. The state space is a combination of information such as load balancing, congestion, and remaining bandwidth for all nodes and links in the network. The action space includes operations such as adding, deleting, or modifying SRv6 segment routes. Step D3: The agent selects an action to adjust the traffic scheduling path based on the current network status. After executing the action, it observes the new network status and the reward value obtained, and uses the Q-learning algorithm to update the agent's strategy. The formula is: , where s is the current state; a is the current action; is the next state; R is the reward value; is the learning rate; is the discount factor; is the expected long-term reward of performing action a in state s; is the maximum Q value of all possible actions a′ after transitioning to the new state s′; Step D4: After multiple trainings, the agent is able to autonomously select the target SRv6 segment routing adjustment strategy according to changes in network status and dynamically optimize the traffic scheduling path.

6. The SRv6-based SDN data center load balancing method according to claim 1, characterized in that: When a node load imbalance is detected or the link packet loss rate is greater than 5%, the load balancing adjustment mechanism is triggered to distinguish the imbalance type and determine the relevant traffic set. The specific steps are as follows: Step E1: When a sudden increase in traffic on a node or link is detected, exceeding a preset burst traffic threshold, a load balancing adjustment mechanism is triggered; Step E2: The SDN controller immediately queries whether there is a backup segment routing path in the SRv6 network. The backup path must satisfy the requirement that the remaining bandwidth is greater than the peak bandwidth of the burst traffic. Step E3: If a backup path exists, direct the burst traffic to the backup path for transmission; if no backup path exists, temporarily create a new SRv6 segment routing path and allocate the required network resources to the burst traffic through the load balancing adjustment mechanism.

7. The SRv6-based SDN data center load balancing method according to claim 6, characterized in that: In step E1, when a sudden increase in traffic on a node or link is detected, exceeding a preset burst traffic threshold, a load balancing adjustment mechanism is triggered. The specific rules for triggering the load balancing adjustment mechanism are as follows: Step E101: If the node load imbalance continues to exceed =5s or link packet loss rate If the traffic congestion rate is >5%, it is considered short-term congestion and triggers local path adjustment. Step E102: If the node load is unbalanced, >90% and If the traffic volume exceeds 85%, it is considered as a long-term overload and triggers global traffic redistribution.

8. According to the SRv6-based SDN data center load balancing method of claim 1, the load balancing adjustment mechanism is optimized using the DQN algorithm, and the specific design of the DQN algorithm is as follows: Step F1: Define the state space S including the node load weights , link weight ,SRv6 segment list availability; Step F2: define action space A as insert, delete, or replace operations of the SRv6 segment list; Step F3: Design reward function ,in, + + =1; is the total weight of the path; ; is the load imbalance.

9. The SRv6-based SDN data center load balancing method according to claim 8, characterized in that: The training process of the DQN algorithm includes: Step G1: Use a dual network architecture (target network + online network) and synchronize parameters every K = 1000 steps; Step G2: Store the transferred samples (s, a, r, s') in the experience replay pool and randomly sample to break the data correlation; Step G3: Use Huber loss function to optimize Q value update: in, is the true value; is the predicted value of the model, where s is the state and a is the action; hour, When , a form similar to L2 loss is adopted, the gradient changes smoothly and is more sensitive to smaller errors; , adopts a form similar to L1 loss, is insensitive to large errors (outliers), and avoids gradient explosion.

10. An SDN data center load balancing system based on SRv6, characterized in that: include: The acquisition module is used to obtain the real-time load status of each node in the data center and the available resource information of each route segment in the SRv6 network through the SDN controller; A construction module, configured to construct a load balancing model based on the real-time load status and the available resource information, and identify overloaded nodes and congested links in the network; A generation module is configured to generate a traffic scheduling path based on the overloaded node and the congested link in combination with the segment routing feature of SRv6, wherein the path is an ordered list of multiple SRv6 segment identifiers, where each SID corresponds to a logical or physical path segment in the data center network; A monitoring module is configured to monitor the load of the data center network in real time based on the traffic scheduling path and in combination with network status information. When a node load imbalance or a link packet loss rate greater than 5% is detected, a load balancing adjustment mechanism is triggered, the imbalance type is distinguished, the relevant traffic set is determined, and the load balancing adjustment mechanism is optimized using a DQN algorithm to maximize network throughput and minimize load imbalance, thereby generating and dynamically adjusting the traffic scheduling strategy. The update module is used to send the adjusted SRv6 traffic scheduling policy to the data center network device, update the flow table information of the device through the Netconf protocol, and redistribute the traffic.

Images