Load balancing optimization method, related server and system
By using the generative diffusion model in the SDN network to generate priority weights and update the flow table entries, the load imbalance problem of traditional load balancing methods under network congestion and asymmetric topology is solved, and the dynamic and real-time load balancing of the network is realized. Efficient transmission.
Patent Information
- Application Number
- CN202410090534.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-01-23
- Publication Date
- 2024-11-01
- Estimated Expiration
- 2044-01-23
AI Technical Summary
Traditional load balancing methods such as ECMP cannot effectively achieve load balancing under network congestion and asymmetric topology, resulting in inefficient network transmission.
In a software-defined network (SDN), by obtaining bandwidth utilization data for each link, a generative diffusion model (GDM) is used to generate priority weights, and the flow table entries are updated to achieve dynamic and real-time load balancing.
Dynamic and real-time load balancing of various network links is realized, network transmission efficiency is improved, and load imbalance problem of traditional methods under congestion and asymmetric topology is solved.
Smart Images

Figure CN117880205B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of network technology, and in particular to a load balancing optimization method, a related server and a system. Background Art
[0002] Link load balancing is an important issue in the field of network routing optimization. It refers to coordinating the routing of traffic within the network to make the traffic load distribution on each link relatively even, thereby avoiding network congestion and improving the overall transmission efficiency of the network.
[0003] Equal Cost Multi-path (ECMP) is a typical traditional load balancing method. Due to its ease of deployment, it is widely used in data center networks and large-scale enterprise networks. Figure 1 As shown in the figure, when a data packet arrives at switch s1 and is to be sent to the target host, there are two forwarding paths with equal hop counts, namely s1-s2-s4 and s1-s3-s4. Switch s1 will select a forwarding path based on the ECMP routing strategy in operation. Common ECMP routing strategies include polling and hashing algorithms. Its purpose is to ensure that the traffic passing through switch s1 can be evenly distributed to the two paths s1-s2-s4 and s1-s3-s4.
[0004] Although traditional load balancing solutions such as ECMP are easy to deploy, they lack a congestion awareness mechanism. Therefore, if the average forwarding logic in ECMP is still used when congestion has already occurred in the network, the network congestion and load imbalance may be aggravated. In addition, ECMP performs very poorly in asymmetric topology networks because the asymmetry of the network topology will cause the average forwarded traffic to cause load imbalance on the devices in the network. Based on the dynamic and real-time nature of the load of each link in a complex network, it cannot meet the optimization requirements of load balancing. Summary of the invention
[0005] In view of the above problems, the present invention is proposed to provide a load balancing optimization method, related server and system that overcome the above problems or at least partially solve the above problems.
[0006] In a first aspect, an embodiment of the present invention provides a method for optimizing load balancing. In an SDN network, the following steps are iteratively performed until convergence:
[0007] Obtaining bandwidth utilization data of each link in the network collected by the data plane of the software defined network SDN, and generating network status information reflecting the load status of each link at the current moment;
[0008] In the control plane of the SDN network, the network state information is used as input to train a generative diffusion model GDM, and the priority weight of each link in the network is output through the GDM model;
[0009] According to the priority weights of the links, flow entries are regenerated and sent to the switches in the network, so that the switches forward data according to the forwarding priorities defined by the updated flow entries.
[0010] In one embodiment, the bandwidth utilization data of each link in the network collected by the data plane of the software defined network SDN is obtained, and the network status information reflecting the load status of each link at the current moment is generated, including:
[0011] The metadata of the bandwidth utilization corresponding to each link recorded by the switch that is the last hop on the routing path of each link and serves as the in-band telemetry INT terminal is obtained;
[0012] The switch on the link, as an INT source, as an INT intermediate transmission node or as an INT terminal, obtains the bandwidth utilization data of all links before the node through the metadata of the data packet received by itself, calculates the bandwidth utilization data of its own link, records the bandwidth utilization data of all links before the node and the bandwidth utilization data of its own link as metadata in the data packet and forwards it to the next hop switch;
[0013] The acquired data set of bandwidth utilization of each link in the network is used as the network status information at the current moment.
[0014] In one embodiment, the data of the bandwidth utilization of the link itself is calculated in the following way:
[0015] The switch calculates the bandwidth utilization of its own link by the following formula:
[0016] ;
[0017] in, is the link bandwidth utilization, Indicates the bandwidth occupied by the link. For time interval The data packets passing through the switch, Indicates the size of the data packet. is the inherent bandwidth of the link.
[0018] In one embodiment, the generative diffusion model GDM is trained with the network state information as input, including:
[0019] A GDM is defined as a solution generating network that maps network status information to a set of priority weights of each link, and the load balancing strategy outputted by the GDM is a set of priority weights of each link in the network;
[0020] Defining Neural Networks As the solution evaluation network, the neural network Used to output the evaluation value of the expected objective function for the current network status information-each link priority weight set ;
[0021] Through the neural network The training of GDM is guided by the training of GDM to generate the optimal solution generation network; the optimal solution generation network generates the highest expected evaluation value The training target of the solution evaluation network is the evaluation value predicted by the current network. The actual evaluation value Minimize the difference between them.
[0022] In one embodiment, the neural network The training of GDM is guided by the training of GDM to generate the optimal solution generation network, including:
[0023] Input hyper parameters, including: denoising step size N and exploration noise ε;
[0024] Initializing the solution generation network and the solution evaluation network respectively;
[0025] If it does not converge, repeat the iterations as follows until it converges:
[0026] At the i-th moment, obtain the network status information at the current moment;
[0027] set up is Gaussian noise, and the network generated by the solution is De-noising, generating a set of current priority weights for each link , adding exploration noise to middle;
[0028] Regenerate a flow table entry according to a set of priority weights of each link generated by the solution generation network, send it to a switch of each link in the network, and calculate a value of an objective function of the solution generation network;
[0029] The solution generation network is updated according to the following formula:
[0030] argmin e i ℒ e ( i ) = − W 0 ~ e i [ U v ( or , W 0 ) ] ;
[0031] in, is the set of priority weights of each link output by the reverse diffusion process of the solution generation network, and n is the set of bandwidth utilization of each link in the network, which is input into the solution generation network as network status information;
[0032] The solution evaluation network is updated according to the following formula :
[0033] argmin U v ℒ U ( v ) = W 0 ~ π i [ ‖ r ( or , W 0 ) − U u ( or , W 0 ) ‖ 2 ] ;
[0034] in, To implement load balancing strategy in environment η When , the solution generates the value of the objective function of the network;
[0035] Returns the trained solution generating network.
[0036] In one embodiment, the objective function of the solution generation network is:
[0037] ;
[0038] is the value of the objective function, is the bandwidth utilization of each link, It is the collection of links in the network.
[0039] In a second aspect, an embodiment of the present invention provides a control plane server, including:
[0040] The network status information generating module is used to iteratively execute the following steps until convergence: obtaining the bandwidth utilization data of each link in the network collected by the data plane of the software defined network SDN, and generating network status information reflecting the load status of each link at the current moment;
[0041] A priority weight calculation module, used to take the network status information obtained by the network status information acquisition module as input, train a generative diffusion model GDM, and output the priority weight of each link in the network through the GDM model;
[0042] The flow entry sending module is used to regenerate the flow entry according to the priority weight of each link and send it to each switch in the network, so that the switch forwards data according to the forwarding priority defined by the updated flow entry.
[0043] In a third aspect, an embodiment of the present invention provides a load balancing optimization system, the load balancing optimization system comprising a control plane server and a plurality of switches as described above; the control plane server is connected to the plurality of switches;
[0044] The control plane server is used to iteratively execute the following steps until convergence: obtain the bandwidth utilization data of each link in the network collected by the data plane of the software defined network SDN, and generate network status information reflecting the load status of each link at the current moment; in the control plane of the SDN network, use the network status information as input to train the generative diffusion model GDM, and output the priority weight of each link in the network through the GDM model; according to the priority weight of each link, regenerate the flow table entry and send it to each switch in the network;
[0045] The switch is used to forward data according to the forwarding priority defined by the updated flow table entry.
[0046] In a fourth aspect, an embodiment of the present invention provides a computer-readable storage medium, wherein the computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the load balancing optimization method as described above is implemented.
[0047] The beneficial effects of the above technical solution provided by the embodiment of the present invention include at least:
[0048] The embodiment of the present invention adopts the SDN network architecture and designs a load balancing solution based on programmable data plane technology and a generative diffusion model (GDM). By adopting the SDN network architecture, the bandwidth utilization of each link is collected on the data plane, so that the current network status information can be perceived in real time and in a fine-grained manner, thereby establishing a network global information perception mechanism; at the same time, a generative diffusion model (GDM) is deployed in the control plane to output the load balancing strategy, i.e., the priority weight of each link, and is sent to the switches in the network in the form of a flow table, so as to realize the dynamic and real-time load balancing optimization process of each link in the network. In addition, since the generative diffusion model is simple, direct and robust, the embodiment of the present invention can effectively solve the problems of slow training and difficult optimal convergence.
[0049] Other features and advantages of the present invention will be described in the following description, and partly become apparent from the description, or understood by practicing the present invention. The purpose and other advantages of the present invention can be realized and obtained by the structures particularly pointed out in the written description, claims, and drawings.
[0050] The technical solution of the present invention is further described in detail below through the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS
[0051] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings:
[0052] Figure 1 This is a schematic diagram of the network architecture of the traditional load balancing method;
[0053] Figure 2 A flow chart of a load balancing optimization method provided by an embodiment of the present invention;
[0054] Figure 3 A schematic diagram of obtaining network status information through an in-band telemetry mechanism in an embodiment of the present invention;
[0055] Figure 4 In the embodiment of the present invention, through the neural network training, a flowchart to guide the training process of GDM;
[0056] Figure 5 4 is a structural block diagram of a control plane server in an embodiment of the present invention. DETAILED DESCRIPTION
[0057] The exemplary embodiments of the present disclosure will be described in more detail below with reference to the accompanying drawings. Although the exemplary embodiments of the present disclosure are shown in the accompanying drawings, it should be understood that the present disclosure can be implemented in various forms and should not be limited by the embodiments set forth herein. On the contrary, these embodiments are provided to enable a more thorough understanding of the present disclosure and to fully convey the scope of the present disclosure to those skilled in the art.
[0058] The inventors of the present invention have discovered that in order to solve the problem of network link load balancing in the prior art, in addition to the traditional load balancing method, the prior art has also proposed the following load balancing methods, such as a load balancing method based on a heuristic algorithm, a load balancing method based on deep reinforcement learning, etc.
[0059] Common heuristic algorithms include simulated annealing algorithm, ant colony algorithm and greedy algorithm. Taking the ant colony algorithm as an example, in network load balancing, the target host can be regarded as a food source, ants represent traffic, and pheromones represent path selection information. Ants select paths based on pheromone concentrations and update pheromone concentrations after passing through the paths. Through the iterative process, the pheromone concentration is gradually adjusted, making ants more inclined to choose paths with relatively balanced loads. The ant colony algorithm can consider global information and local information, and has the characteristics of adaptability and distributed optimization.
[0060] The load balancing method based on heuristic algorithms needs to establish a model based on specific network conditions. When faced with a dynamically changing complex network, heuristic algorithms are often powerless, and their scalability is limited, and they cannot guarantee the quality of network services. At the same time, due to the complex iterative solution process of the heuristic algorithm, this will cause the problem of slow convergence of the load balancing strategy.
[0061] Based on the problems existing in various load balancing optimization methods in the above-mentioned prior art, for complex networks such as SDN, the inventors of this application proposed a network-aware load balancing optimization method based on a generative diffusion model, which collects the bandwidth utilization of each link in the network as network status information reflecting the link load situation. At the same time, a generative diffusion model (GDM) is run in the control plane, with the network status information collected by telemetry as input. After GDM training, the priority weights of each link in the network are output, and the controller generates new flow table entries based on the priority weights and sends them to each switch in the network.
[0062] The specific implementation process of the load balancing optimization method provided by the embodiment of the present invention is described in detail below with reference to the accompanying drawings.
[0063] The load balancing optimization method provided by the embodiment of the present invention refers to Figure 2 As shown, in the SDN network, the following steps are iteratively executed until convergence:
[0064] S21, obtaining bandwidth utilization data of each link in the network collected by a data plane of a software defined network (SDN), and generating network status information reflecting the load status of each link at the current moment;
[0065] S22, in the control plane of the SDN network, taking the network status information as input, training a generative diffusion model GDM, and outputting the priority weight of each link in the network through the GDM model;
[0066] S23. Regenerate a flow entry according to the priority weight of each link and send it to each switch in the network, so that each switch forwards data according to the forwarding priority defined by the updated flow entry.
[0067] The embodiment of the present invention utilizes the architecture of a software defined network (SDN). In the SDN network, a data plane for implementing data forwarding and a control plane for implementing forwarding control are included.
[0068] In the above process, steps S21-S23 will be iteratively executed until the convergence state. In the convergence state, each link in the SDN network reaches or is close to the load balancing state. At this time, the parameters of the generative diffusion model will tend to be stable and will not change or will change slightly. Each time the network status information is used as input, the priority weights of each link output will also tend to be stable.
[0069] The embodiment of the present invention adopts the SDN network architecture and designs a load balancing solution based on programmable data plane technology and a generative diffusion model (GDM). By adopting the SDN network architecture, the bandwidth utilization of each link is collected on the data plane, so that the current network status information can be perceived in real time and in a fine-grained manner, thereby establishing a network global information perception mechanism; at the same time, a generative diffusion model (GDM) is deployed in the control plane to output the load balancing strategy, i.e., the priority weight of each link, and is sent to the switches in the network in the form of a flow table, so as to realize the dynamic and real-time load balancing optimization process of each link in the network. In addition, since the generative diffusion model is simple, direct and robust, the embodiment of the present invention can effectively solve the problems of slow training and difficult optimal convergence.
[0070] In one embodiment, the above step S21 obtains the bandwidth utilization data of each link in the network collected by the data plane of the software defined network SDN, and generates network status information reflecting the load status of each link at the current moment, which can be achieved, for example, by the following method:
[0071] The metadata of the bandwidth utilization of each link recorded by the switch that is the last hop on the routing path of each link and serves as the in-band network telemetry (INT) terminal is obtained;
[0072] The switch on the link, as an INT source, as an INT intermediate transmission node or as an INT terminal, can obtain the bandwidth utilization data of all links before the node through the metadata of the data packet received by itself, and calculate the bandwidth utilization data of its own link, and record the bandwidth utilization data of all links before the node and the bandwidth utilization data of its own link as metadata in the data packet and forward it to the next hop switch;
[0073] The acquired data set of bandwidth utilization of each link in the network is used as the network status information at the current moment.
[0074] The above method uses in-band telemetry technology. INT technology is a network measurement technology that is convenient for collecting information in complex networks. Its architecture can be referred to Figure 3As shown in the figure, in this network architecture, there are three links, including the link from the client-S1 node (the corresponding bandwidth utilization is INT metadata1), the link from the INT source to the INT intermediate transmission node (the corresponding bandwidth utilization is INT metadata2), and the link from the INT intermediate transmission node to the INT terminal (the corresponding bandwidth utilization is INT metadata3). The network status information finally obtained by the control plane server is a data set consisting of three values: INTmetadata1, INT metadata2, and INT metadata3.
[0075] Reference Figure 3 As shown, there are some business flow data in the network, and the in-band telemetry technology uses these business flows as carriers of network status information. In the embodiment of the present invention, when a data packet passes through a switch node, the corresponding link bandwidth utilization can be calculated according to the data packet counter set in the switch. , as shown in the following formula (1):
[0076] ; (1)
[0077] in Indicates the bandwidth occupied by the link. For time interval The data packets passing through the switch can be counted by the counters set on the switch. Indicates the size of the data packet in bytes. is the inherent bandwidth of the link. The calculated link bandwidth utilization will be embedded into the INT header of the data packet by the switch in the form of metadata. When the data packet is forwarded to the last hop on the routing path, i.e., the INT terminal in the figure, the switch at the last hop will send the previously collected information of each link to the control plane server. In this way, the control plane server can obtain real-time, fine-grained network status information.
[0078] In one embodiment, in the above step S22, the network state information is used as input to train the generative diffusion model GDM, which can be specifically implemented in the following manner:
[0079] 1. Define a GDM that maps network status information to a set of priority weights of each link as a solution generation network, and the load balancing strategy output by the GDM is a set of priority weights of each link in the network;
[0080] 2. Define Neural Network As the solution evaluation network, the neural network Used to output the evaluation value of the expected objective function for the current network status information-each link priority weight set ;
[0081] 3. Through the neural network The training of GDM is guided by the training of GDM to generate the optimal solution generation network; the optimal solution generation network generates the highest expected evaluation value The training target of the solution evaluation network is the evaluation value predicted by the current network. The actual evaluation value Minimize the difference between them.
[0082] First, the GDM model is briefly explained.
[0083] The generative diffusion model (GDM model) is a type of probabilistic model for generating data. It simulates the diffusion process in thermodynamics, constructs forward and reverse diffusion processes, and can accurately capture complex data distributions and has high-quality data generation capabilities. And because GDM is built based on the theory of stochastic differential equations, it can be applied to a wide range of types of data and applications, with high flexibility and powerful data generation capabilities. In addition, compared with other data generation models, GDM has a simple architecture and training process, which makes it fast to train and easy to converge. The inventors of this application found that these advantages of GDM enable it to solve complex problems in the field of network optimization. Generally speaking, the network state is complex and dynamic. To solve this type of optimization problem, it is necessary to establish a high-dimensional abstract network management model. Due to its diffusion characteristics, GDM can well capture the characteristics of this high-dimensional complex network structure, and can effectively handle the decision-making process and optimization problems in network optimization.
[0084] The process of the GDM model includes forward diffusion process and reverse diffusion process. Taking image data as an example, the forward diffusion process and reverse diffusion process are briefly described as follows:
[0085] (1) Forward diffusion process:
[0086] The forward diffusion process can be modeled as a Step Markov chain, let represents the original image, in the Markov chain Step 2: The variance is Add the Gaussian noise to the The image of the step is and produces a distribution of Image , as shown in formula (2):
[0087] ; (2)
[0088] in So is the mean, is a normal distribution with a variance of To the final image data , the posterior probability can be expressed as formula (3):
[0089] ; (3)
[0090] According to formula (2), Sampling requires Calculate the time When it is large, the amount of calculation will become very large, so we define ,and then can be expressed as:
[0091] ; (4)
[0092] in So we can use the following distribution to get :
[0093] ; (5)
[0094] is a preset hyperparameter, which can be calculated , which allows you to Noise Sampling to obtain .
[0095] (2) Reverse diffusion process:
[0096] when When it is very big, Approximately an isotropic Gaussian distribution. If the inverse distribution can be learned , then from Sampling in And carry out the reverse diffusion process, we can obtain the distribution and generate sample data from it. However The statistical estimation of requires the calculation of data distribution, which is a difficult problem to handle, so a parametric model is used. To estimate , as shown in formula (6):
[0097] ; (6)
[0098] Further, we can obtain arrive The trajectory:
[0099] ; (7)
[0100] By adjusting the model at time t, we can learn to predict the Gaussian parameter, which is the mean at time t. and variance The training of GDM involves optimizing the negative log-likelihood of the training data, adding conditional information such as , you can Modeled as a noise prediction model with a fixed covariance matrix, the covariance can be calculated by formula (8):
[0101] ; (8)
[0102] The mean can be calculated by formula (9):
[0103] ; (9)
[0104] You can first start from Sampling in , and then calculate the parameters for the reverse diffusion chain as Equation (10):
[0105] ; (10)
[0106] in And by ignoring certain weight terms, the original calculated loss function can be simplified to obtain the loss function shown in formula (11):
[0107] ℒ t = X 0 , t , e [ ‖ e − e i ( a ¯ t X 0 + 1 − a ¯ t e , t ) ‖ 2 ] ; (11)
[0108] The reverse diffusion process of GDM indicates that at each time The noise is predicted within.
[0109] The above is the GDM training process illustrated by using image data as an example. The basic principle of GDM training is briefly described as follows: In the initial input, GDM gradually introduces Gaussian noise through a series of steps (i.e., the forward diffusion process), which generates targets for the denoising neural network. Subsequently, the neural network is trained to reverse the noise process and restore the data and content. The reverse diffusion process allows the generation of new data.
[0110] The inventors of the present application have found that, for example, network optimization problems with load balancing usually do not have access to large datasets suitable for offline training. The lack of expert datasets poses a challenge when applying GDM to load balancing network optimization. The reverse diffusion process of GDM (involving a denoising network) can be effectively utilized.
[0111] Specifically, rather than relying on a standard loss function, a denoising network can be trained to maximize the value of an objective function output by the resulting solution, which is designed to maximize or minimize a specific outcome depending on the given application. In network optimization, this value can be a network performance metric. This training process can be achieved by executing the generated solution in the network environment, and then adjusting the network parameters based on the feedback received.
[0112] In scenarios where no expert database exists, GDM can learn to generate the best solution for given environmental conditions by actively exploring unknown environments.
[0113] With respect to the problem of link load balancing, combined with the introduction of the aforementioned GDM model, in an embodiment of the present invention, the network status information is used as input, and a set of priority weights of each link is output through the GDM model. According to the performance target to be achieved by network load balancing, the target of load balancing can be expressed by formula (12):
[0114] ; (12)
[0115] in Represents the priority weight set of links in the network. It is the output of the GDM model and can be used as the basis for the controller to generate flow table entries.
[0116] The above formula (12) shows that the goal of load balancing is to make the bandwidth utilization of the link with the highest bandwidth utilization among all links in the network as low as possible.
[0117] In order to effectively judge the reliability of the GDM model output strategy, the embodiment of the present invention introduces a neural network Output The value is used to evaluate the effectiveness of the output load balancing strategy (i.e., the solution output by the solution generation network) to guide the training process of the GDM model and the optimal noise prediction model. The final load balancing strategy should be the priority weight set of the link. When Therefore, the optimal solution generation network can be expressed as:
[0118] argmin e i ℒ e ( i ) = − W 0 ~ e i [ U v ( or , W 0 ) ] ; (13)
[0119] In the above formula (13): It is the load balancing strategy output by the GDM reverse diffusion process. It is the set of network link bandwidth utilization, which is input into GDM as the network environment status value. Neural Network The training goal is to minimize the current network prediction Value and truth The difference between the values, therefore, the optimal It can be expressed as:
[0120] argmin U v ℒ U ( v ) = W 0 ~ π i [ ‖ r ( or , W 0 ) − U u ( or , W 0 ) ‖ 2 ] ; (14)
[0121] in To implement load balancing strategy under environment η (i.e. current network status information) The value of the objective function when , according to formula (12), can be designed as:
[0122] ; (15)
[0123] is the value of the objective function, is the bandwidth utilization of each link, It is the collection of links in the network.
[0124] In one embodiment, the above-mentioned neural network The training of GDM is guided by the training of GDM to generate the optimal solution generation network. Figure 4 As shown, this can be achieved through the following process:
[0125] Input hyperparameters, where the hyperparameters include: denoising step size N and exploration noise ε;
[0126] Initialize the solution generation network and the solution evaluation network respectively; initialize the parameters of the solution generation network and the solution evaluation network.
[0127] If it does not converge, repeat the iterations as follows until it converges:
[0128] At the i-th moment, obtain the network status information of the current moment; i is an integer greater than or equal to 0;
[0129] set up is Gaussian noise, and the network generated by the solution is De-noising, generating a set of current priority weights for each link , adding exploration noise to middle.
[0130] Regenerate the flow table entry according to the set of priority weights of each link generated by the solution generation network, send it to each switch in the network, and calculate the value r of the objective function of the solution generation network;
[0131] The solution generation network is updated according to the following formula:
[0132] argmin e i ℒ e ( i ) = − W 0 ~ e i [ U v ( or , W 0 ) ] ;
[0133] in, is the set of priority weights of each link output by the reverse diffusion process of the solution generation network, and n is the set of bandwidth utilization of each link in the network, which is input into the solution generation network as network status information;
[0134] The solution evaluation network is updated according to the following formula :
[0135] argmin U v ℒ U ( v ) = W 0 ~ π i [ ‖ r ( or , W 0 ) − U u ( or , W 0 ) ‖ 2 ] ;
[0136] in, For the environment Execute load balancing strategy When , the solution generates the value of the objective function of the network;
[0137] Returns the trained solution generating network.
[0138] The above-mentioned convergence state can be judged by solving the objective function of the generated network, or by the above-mentioned loss function, which is not limited in the embodiment of the present invention.
[0139] The inventors of the present application have discovered that, compared with the prior art method of using deep reinforcement learning to solve the network load balancing problem, the embodiment of the present invention introduces the GDM model as part of the overall learning model framework in the learning framework for seeking the optimal solution to the load balancing strategy (the optimal solution for the priority weights of each link), and adds a denoising process to the overall framework. This method not only makes the model easier to converge, but also, according to experimental data, it can be measured that the optimality gap between the obtained optimal solution and the true value is smaller, thereby ensuring a better network load balancing effect.
[0140] Based on the same inventive concept, the embodiment of the present invention further provides a control plane server, referring to Figure 5 As shown, including:
[0141] The network status information generating module 51 is used to iteratively execute the following steps until convergence: obtaining the bandwidth utilization data of each link in the network collected by the data plane of the software defined network SDN, and generating network status information reflecting the load status of each link at the current moment;
[0142] A priority weight calculation module 52, configured to use the network status information obtained by the network status information acquisition module as input, train a generative diffusion model GDM, and output the priority weight of each link in the network through the GDM model;
[0143] The flow entry sending module 53 is used to regenerate the flow entry according to the priority weight of each link and send it to each switch in the network, so that the switch forwards data according to the forwarding priority defined by the updated flow entry.
[0144] The embodiment of the present invention further provides a load balancing optimization system, the load balancing optimization system comprising the control plane server and multiple switches as described above; wherein the control plane server is connected to the multiple switches;
[0145] The control plane server is used to iteratively execute the following steps until convergence: obtain the bandwidth utilization data of each link in the network collected by the data plane of the software defined network SDN, and generate network status information reflecting the load status of each link at the current moment; in the control plane of the SDN network, use the network status information as input to train the generative diffusion model GDM, and output the priority weight of each link in the network through the GDM model; according to the priority weight of each link, regenerate the flow table entry and send it to each switch in the network;
[0146] The switch is used to forward data according to the forwarding priority defined by the updated flow table entry.
[0147] An embodiment of the present invention further provides a control plane server, comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the program, the load balancing optimization method as described in the previous embodiment is implemented.
[0148] An embodiment of the present invention further provides a computer-readable storage medium, wherein the computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the processor implements the load balancing optimization method as described in the previous embodiment.
[0149] It should be understood by those skilled in the art that the embodiments of the present invention may be provided as methods, systems, or computer program products. Therefore, the present invention may take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Moreover, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage and optical storage, etc.) containing computer-usable program codes.
[0150] The present invention is described with reference to flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowchart and / or block diagram, as well as the combination of processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure one A process or multiple processes and / or boxes Figure one A device that provides the functions specified in a block or multiple blocks.
[0151] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure one A process or multiple processes and / or boxes Figure one A function specified in one or more boxes.
[0152] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process in the computer or other programmable device. Figure one A process or multiple processes and / or boxes Figure one A step that specifies a function in one or more boxes.
[0153] Obviously, those skilled in the art can make various changes and modifications to the present invention without departing from the spirit and scope of the present invention. Thus, if these modifications and variations of the present invention fall within the scope of the claims of the present invention and their equivalents, the present invention is also intended to include these modifications and variations.
Claims
1. A load balancing optimization method, characterized in that: In a software defined network (SDN), the following steps are iteratively executed until convergence: Obtaining bandwidth utilization data of each link in the network collected by the data plane of the SDN, and generating network status information reflecting the load status of each link at the current moment; In the control plane of the SDN network, the network state information is used as input to train a generative diffusion model GDM, and the priority weight of each link in the network is output through the GDM model; The method uses the network state information as input to train the generative diffusion model GDM, including: defining the GDM that maps the network state information to a set of priority weights of each link as a solution generation network, and the load balancing strategy output by the GDM is a set of priority weights of each link in the network; defining a neural network U v As the solution evaluation network, the U v Used to output an evaluation value U of the value of the expected objective function for the current network state information-each link priority weight set; through the neural network U v The training of GDM is guided by the training of GDM to generate an optimal solution generation network; the optimal solution generation network is a network that generates a priority weight allocation scheme with the highest expected evaluation value U; the training goal of the solution evaluation network is to minimize the difference between the evaluation value U predicted by the current network and the actual evaluation value U; According to the priority weights of the links, flow entries are regenerated and sent to the switches in the network, so that the switches forward data according to the forwarding priorities defined in the updated flow entries.
2. The load balancing optimization method according to claim 1, characterized in that: Obtaining bandwidth utilization data of each link in the network collected by the data plane of the software defined network SDN, and generating network status information reflecting the load status of each link at the current moment, including: The metadata of the bandwidth utilization corresponding to each link recorded by the switch that is the last hop on the routing path of each link and serves as the in-band telemetry INT terminal is obtained; The switch on the link, as an INT source, as an INT intermediate transmission node or as an INT terminal, obtains the bandwidth utilization data of all links before the node through the metadata of the data packet received by itself, calculates the bandwidth utilization data of its own link, records the bandwidth utilization data of all links before the node and the bandwidth utilization data of its own link as metadata in the data packet and forwards it to the next hop switch; The acquired data set of bandwidth utilization of each link in the network is used as the network status information at the current moment.
3. The load balancing optimization method according to claim 2, characterized in that: The data of the bandwidth utilization of the link itself is calculated in the following way: The switch calculates the bandwidth utilization of its own link by the following formula: Among them, η l is the link bandwidth utilization, C represents the bandwidth occupied by the link, p is the data packet passing through the switch within the time interval t, S(p) represents the size of the data packet, and B is the inherent bandwidth of the link.
4. The load balancing optimization method according to claim 1, characterized in that: Through the neural network U v The training of GDM is guided by the training of GDM to generate the optimal solution generation network, including: Input hyper parameters, including: denoising step size N and exploration noise ε; Initializing the solution generation network and the solution evaluation network respectively; If it does not converge, repeat the iterations as follows until it converges: At the i-th moment, obtain the network status information at the current moment; Setting W N is Gaussian noise, and the solution generation network is W N De-noising, generating a set of priority weights W0 for each link (i) , add exploration noise to W0 (i) middle; Regenerate a flow table entry according to a set of priority weights of each link generated by the solution generation network, send it to a switch of each link in the network, and calculate a value of an objective function of the solution generation network; The solution generation network is updated according to the following formula: Wherein, W0 is the set of priority weights of each link output by the reverse diffusion process of the solution generation network, and n is the set of bandwidth utilization of each link in the network, which is input into the solution generation network as network status information; Update the solution evaluation network U according to the following formula v : Where r is the value of the objective function of the solution generating network when the load balancing strategy W0 is executed in the environment η; Returns the trained solution generating network.
5. The load balancing optimization method according to claim 4, characterized in that: The objective function of the solution generation network is: r is the value of the objective function, η l (l∈L) is the bandwidth utilization of each link, and L is the set of links in the network.
6. A control plane server, characterized in that: include: The network status information generating module is used to iteratively execute the following steps until convergence: obtaining the bandwidth utilization data of each link in the network collected by the data plane of the software defined network SDN, and generating network status information reflecting the load status of each link at the current moment; A priority weight calculation module, used to take the network status information obtained by the network status information acquisition module as input, train a generative diffusion model GDM, and output the priority weight of each link in the network through the GDM model; The method uses the network state information as input to train the generative diffusion model GDM, including: defining the GDM that maps the network state information to a set of priority weights of each link as a solution generation network, and the load balancing strategy output by the GDM is a set of priority weights of each link in the network; defining a neural network U v As the solution evaluation network, the U v Used to output an evaluation value U of the value of the expected objective function for the current network state information-each link priority weight set; through the neural network U v The training of GDM is guided by the training of GDM to generate an optimal solution generation network; the optimal solution generation network is a network that generates a priority weight allocation scheme with the highest expected evaluation value U; the training goal of the solution evaluation network is to minimize the difference between the evaluation value U predicted by the current network and the actual evaluation value U; The flow entry sending module is used to regenerate the flow entry according to the priority weight of each link and send it to each switch in the network, so that the switch forwards data according to the forwarding priority defined by the updated flow entry.
7. A control plane server, characterized in that: include: A memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the program, the load balancing optimization method according to any one of claims 1 to 5 is implemented.
8. A load balancing optimization system, characterized in that: The load balancing optimization system comprises the control plane server as claimed in claim 7 and a plurality of switches; the control plane server is connected to the plurality of switches; The control plane server is used to iteratively execute the following steps until convergence: obtain the bandwidth utilization data of each link in the network collected by the data plane of the software defined network SDN, and generate network status information reflecting the load status of each link at the current moment; in the control plane of the SDN network, use the network status information as input to train the generative diffusion model GDM, and output the priority weight of each link in the network through the GDM model; according to the priority weight of each link, regenerate the flow table entry and send it to each switch in the network; The switch is used to forward data according to the forwarding priority defined by the updated flow table entry.
9. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the load balancing optimization method according to any one of claims 1 to 5 is implemented.
