Deterministic routing method, system and equipment based on deep learning and medium

Abstract

The invention relates to a deterministic routing method, system and device based on deep learning and a medium. The method comprises the following steps: collecting a plurality of flow requests arriving in a current time slot; performing path quality evaluation processing between a source subnet and a target subnet on a flow request needing to be transmitted across the subnets by combining the quotient graph to obtain a cross-subnet transmission path and a boundary node pair sequence; splitting an end-to-end service quality demand into service quality distribution values of each subnet; and generating a routing decision on the pre-calculated candidate path set for each piece of flow through the subnet agent, and generating a flow transmission instruction. By adopting the method, the static defect of a traditional routing protocol in a heterogeneous traffic scene and the decision-making efficiency bottleneck of centralized intelligent routing can be effectively solved, the processing efficiency of high-concurrency traffic is remarkably improved, local congestion is avoided through global collaborative optimization, the multi-dimensional QoS requirement is dynamically guaranteed, and the method is suitable for large-scale popularization and application. Therefore, extensible and self-adaptive deterministic service guarantee is provided for scenes such as the industrial internet.

Description

Technical Field

[0001] This invention belongs to the field of computer network communication technology, and in particular relates to deterministic routing methods, systems, devices and media based on deep learning. Background Technology

[0002] With the development and application of network technology, emerging services such as AR / VR, Industrial Internet, and telemedicine are constantly increasing their demands on network performance. Deterministic networking, as one solution, can provide stable and high-quality service guarantees for different service traffic. However, as the network scale expands, controlling and managing large-scale deterministic networks becomes increasingly challenging. Different service traffic has varying requirements in terms of QoS metrics such as latency, bandwidth, jitter, and packet loss rate. For example, control commands in the Industrial Internet require extremely low latency and packet loss rate; video conferencing traffic can tolerate moderate latency and slight packet loss, but its bandwidth requirements are relatively high, necessitating the reservation of more resources to fully guarantee deterministic transmission.

[0003] Traditional technologies, such as the Open Shortest Path First (OSPF) protocol, maintain a unified topology map across the network by flooding link-state advertisements. Based on this map, they use the static "cost" of each link as a single metric and employ Dijkstra's algorithm to calculate the shortest path to all destination networks. This cost is typically inversely proportional to link bandwidth. While mature and stable, and capable of quickly responding to network topology changes and rerouting, its reliance on a single, static cost metric for path calculation fails to comprehensively consider multi-dimensional QoS metrics crucial for industrial internet services, such as latency, jitter, and packet loss rate. This makes it difficult to meet differentiated needs in mixed service scenarios. In addition, there are centralized intelligent routing solutions based on deep learning networks. These deploy agents within a software-defined network architecture, using global network state (such as link utilization) as input and employing deep learning network models to serially calculate the routing path for each arriving traffic request, aiming to optimize overall network performance through long-term learning. Compared to traditional static protocols, this approach possesses the ability to dynamically perceive network state and make decisions, exhibiting a degree of environmental adaptability.

[0004] However, current traditional routing protocols and centralized intelligent routing methods still have significant limitations: static measurement mechanisms cannot flexibly adapt to the diverse QoS requirements of heterogeneous traffic in the Industrial Internet, especially failing to adequately guarantee latency-sensitive services; centralized serial decision-making models are inefficient when handling massive concurrent traffic, with decision-making time increasing linearly with the amount of traffic, becoming a bottleneck for large-scale deployments; existing intelligent solutions struggle to effectively model resource competition and dependencies between traffic flows, easily leading to local congestion and uneven resource utilization, failing to achieve global collaborative optimization. Therefore, a deterministic routing method is needed that can efficiently handle high-concurrency traffic, dynamically guarantee multi-dimensional QoS requirements, and possess global optimization capabilities. However, the centralized serial decision-making model it employs causes routing decision-making time to increase linearly with the amount of concurrent traffic, making decision-making efficiency a bottleneck when dealing with the massive concurrent traffic that may occur in the Industrial Internet. Furthermore, existing intelligent solutions struggle to effectively model resource competition and dependencies between traffic flows, easily leading to local congestion and uneven resource utilization, failing to achieve global collaborative optimization. Summary of the Invention

[0005] Therefore, it is necessary to provide a deterministic routing method that can efficiently handle high-concurrency traffic, dynamically guarantee multi-dimensional QoS requirements, and has global optimization capabilities to address the aforementioned technical problems.

[0006] Firstly, this application provides a deterministic routing method based on deep learning, including:

[0007] Collect multiple traffic requests arriving within the current time slot; traffic requests include source node, destination node, and end-to-end quality of service requirements; end-to-end quality of service requirements include latency requirements, jitter requirements, packet loss rate requirements, and bandwidth requirements;

[0008] Based on the business map constructed by subnet partitioning, for traffic requests that need to be transmitted across subnets, path quality assessment between the source subnet and the destination subnet is performed to obtain the sequence of cross-subnet transmission paths and boundary node pairs.

[0009] Based on the cross-subnet transmission path and boundary node pair sequence, the end-to-end quality of service requirements are broken down to obtain the quality of service allocation value for each subnet;

[0010] Based on the service quality allocation value of each subnet, the corresponding subnet agent generates routing decisions for each traffic on the pre-computed candidate path set; the subnet agent includes a policy network based on the Transformer-Encoder architecture;

[0011] Based on the routing decision, a traffic transmission command is generated. The traffic transmission command is used to instruct the underlying network devices to perform traffic transmission according to the routing decision and update the network link status. When the traffic transmission is completed, the occupied resources are released.

[0012] Furthermore, based on the commerce graph constructed from subnet partitioning, path quality assessment is performed between the source and destination subnets for traffic requests that require cross-subnet transmission, resulting in a sequence of cross-subnet transmission paths and boundary node pairs, including:

[0013] A commerce graph is constructed based on a set of subnets; in the commerce graph, vertices represent subnets and edges represent the set of connection links between subnets.

[0014] Obtain real-time status metrics for each connection link in the business map; real-time status metrics include real-time latency, real-time bandwidth, real-time jitter, and real-time packet loss rate;

[0015] For any two subnets in the business map, the node pair quality evaluation coefficient is calculated based on the real-time status indicators of the connection link; the node pair quality evaluation coefficient is used to evaluate the link performance.

[0016] Based on the node-pair quality assessment coefficient, the inter-subnet connectivity quality assessment coefficient is calculated using the following formula:

[0017]

[0018] in, This is a quality evaluation coefficient for inter-subnet connections. It is a set of connections between subnets. For inter-node links, and For subnet identification, The node pair quality evaluation coefficient;

[0019] Among all reachable paths from the source subnet to the destination subnet in the business map, the path with the best cumulative inter-subnet connection quality evaluation coefficient is determined as the cross-subnet transmission path;

[0020] For each segment of the cross-subnet transmission path, calculate the cross-segment compatibility index; and based on the cross-segment compatibility index and the node pair quality evaluation coefficient, determine the optimal boundary node pair.

[0021] Combine all the optimal boundary node pairs to obtain a sequence of boundary node pairs.

[0022] Furthermore, for any connection between two subnets in the business map, based on the real-time status indicators of the connection link, the node pair quality evaluation coefficient is calculated, including:

[0023] For connections between any subnets in the business map, the node pair quality evaluation coefficient is calculated based on the real-time status indicators of the connection links using the following formula:

[0024]

[0025] in, The node pair quality evaluation coefficient. This represents the connection link between nodes, where u and v are node identifiers. For the real-time latency of the connection link, The real-time bandwidth of the connection link. To address the real-time jitter of the connection link, For the real-time packet loss rate of the connection link, This is the maximum reference value for latency across the entire network. This is the maximum reference value for bandwidth across the entire network. This is the largest reference value for shaking across the entire network. This is the highest reference value for packet loss rate across the entire network. This is the time delay weighting coefficient. This is the bandwidth weighting coefficient. This is the jitter delay weighting coefficient. This is the weighting coefficient for packet loss rate.

[0026] Furthermore, for each segment of the inter-subnet connection in the cross-subnet transmission path, a cross-segment compatibility index is calculated; and based on the cross-segment compatibility index and the node pair quality evaluation coefficient, the optimal boundary node pair is determined, including:

[0027] Obtain the link performance metrics of the selected optimal node pair for the previous segment of the inter-subnet connection; the link performance metrics include latency, bandwidth, jitter, and packet loss rate;

[0028] The following formula is used to calculate the difference between the link performance index of each node pair in the candidate node pair set for inter-target subnet connections and the link performance index of the previously selected optimal node pair, representing a cross-segment compatibility index:

[0029]

[0030] in, For cross-segment compatibility indicators, This is a latency compatibility weighting coefficient. This is a bandwidth compatibility weighting coefficient. This is a jitter compatibility weighting coefficient. This is a compatibility weighting coefficient for packet loss rate. This is the delay of the previously selected optimal node pair. This represents the bandwidth of the previously selected optimal node pair. This is for the jitter of the previously selected optimal node pair. This represents the packet loss rate of the previously selected optimal node pair. The latency of node pairs connecting the target subnets. The bandwidth of the node pairs connecting the target subnets. Jitter of node pairs connecting target subnets. Let c be the packet loss rate of the node pairs connecting the target subnets. This serves as an identifier for inter-subnet connections.

[0031] Based on the node pair quality assessment coefficient and the differential cross-segment compatibility index, a comprehensive index value is calculated; and the node pair with the smallest comprehensive index value is determined as the optimal boundary node pair for the inter-target subnet connection.

[0032] Furthermore, based on the cross-subnet transmission path and boundary node pair sequence, the end-to-end quality of service (QoS) requirements are broken down to obtain the QoS allocation values ​​for each subnet, including:

[0033] Based on the cross-subnet transmission path and boundary node pair sequence, the subnet sequence through which the traffic will pass is determined; and historical performance data of each subnet in the subnet sequence is obtained; the historical transmission performance data includes average latency, average jitter, average packet loss rate and standard deviation;

[0034] Set the bandwidth allocation value within each subnet to the minimum bandwidth requirement of the traffic;

[0035] Based on the average latency in historical performance data, calculate the first proportional coefficient and the latency stability factor; and based on the first proportional coefficient and the latency stability factor, calculate the latency allocation value within the subnet.

[0036] Based on the average jitter in historical performance data, calculate the second proportional coefficient and jitter stability factor; and based on the second proportional coefficient and jitter stability factor, calculate the jitter allocation value within the subnet.

[0037] Calculate the third proportional coefficient based on the average packet loss rate in historical performance data; and calculate the packet loss rate allocation value within the subnet based on the third proportional coefficient.

[0038] Based on the bandwidth allocation values, latency allocation values, jitter allocation values, and packet loss rate allocation values, we verify whether the total latency, total jitter, and total packet loss rate after splitting meet the original traffic requirements and whether the minimum bandwidth meets the constraints, and obtain the verification results.

[0039] If the verification results show that the total latency, total jitter, and total packet loss rate after splitting meet the original traffic requirements and the minimum bandwidth meets the constraints, the bandwidth allocation value, latency allocation value, jitter allocation value, and packet loss rate allocation value in each subnet are summarized to obtain the service quality allocation value for each subnet.

[0040] Furthermore, based on the service quality allocation values ​​of each subnet, the corresponding subnet agent generates routing decisions for each traffic flow on the pre-computed candidate path set, including:

[0041] Based on the service quality allocation values ​​of each subnet, obtain the service quality requirements of all traffic to be decided within the current subnet; and based on the service quality requirements of the traffic to be decided and the real-time status indicators of the links within the current subnet, construct a traffic feature matrix and a link feature matrix.

[0042] Perform independent linear projection transformations on the traffic feature matrix and the link feature matrix to obtain the link feature tensor and the traffic feature tensor, respectively.

[0043] The link feature tensor is aggregated into a global network state context vector through an attention weighting mechanism; and the global network state context vector is fused with the traffic feature tensor to obtain a fused tensor.

[0044] The fused tensor is input into a multi-layer Transformer-Encoder, and the encoded feature representation is output.

[0045] Based on the encoded feature representation, the probability distribution of each traffic flow on the pre-computed candidate path set is generated through the classification head of the Actor network; and based on the probability distribution, a routing decision is generated through sampling.

[0046] The pre-computed candidate path set was obtained using the following method:

[0047] Obtain the link performance metrics for each link within the current subnet; and calculate the composite weight value of the link based on the link performance metrics and preset weight parameters using a composite weight formula:

[0048]

[0049] in, This represents the composite weight value of the link. , and The preset weight parameters, For time delay, For shaking, For packet loss rate, This is the bandwidth normalization adjustment factor. For broadband, It is a very small constant;

[0050] Based on the composite weight value, multiple short paths from the node to the destination node are calculated using the Dijkstra algorithm; and the short paths are summarized to obtain a pre-calculated candidate path set.

[0051] Secondly, this application also provides a deterministic routing system based on deep learning, including:

[0052] The request collection module is used to collect multiple traffic requests arriving within the current time slot; the traffic requests include source node, destination node, and end-to-end service quality requirements; the end-to-end service quality requirements include latency requirements, jitter requirements, packet loss rate requirements, and bandwidth requirements.

[0053] The path filtering module is used to construct a business map based on subnet partitioning. For traffic requests that need to be transmitted across subnets, it performs path quality assessment between the source subnet and the destination subnet to obtain a sequence of cross-subnet transmission paths and boundary node pairs.

[0054] The quality allocation module is used to break down the end-to-end quality of service requirements based on the cross-subnet transmission path and boundary node pair sequence, and obtain the quality of service allocation value for each subnet.

[0055] The decision generation module is used to generate routing decisions for each traffic flow on a pre-computed candidate path set based on the service quality allocation value of each subnet and through the subnet intelligent agent corresponding to each subnet; the subnet intelligent agent includes a policy network based on the Transformer-Encoder architecture;

[0056] The instruction generation module is used to generate traffic transmission instructions based on routing decisions. The traffic transmission instructions are used to instruct the underlying network devices to perform traffic transmission according to the routing decisions, update the network link status, and release occupied resources when the traffic transmission is completed.

[0057] Thirdly, this application also provides a computer device, including a memory and a processor, wherein the memory stores at least one instruction, at least one program, code set, or instruction set, and the at least one instruction, the at least one program, the code set, or the instruction set is loaded and executed by the processor to implement any of the deep learning-based deterministic routing methods described in the embodiments of this application.

[0058] Fourthly, this application also provides a computer-readable storage medium storing at least one piece of program code, which is loaded and executed by a processor to implement any of the deep learning-based deterministic routing methods described in the embodiments of this application.

[0059] The aforementioned deep learning-based deterministic routing method, system, device, and medium collect multiple traffic requests arriving within the current time slot; combine commerce maps to perform path quality assessment between the source and destination subnets for traffic requests requiring cross-subnet transmission, obtaining a sequence of cross-subnet transmission paths and boundary node pairs; decompose end-to-end service quality requirements into service quality allocation values ​​for each subnet; and generate routing decisions for each traffic flow on a pre-computed candidate path set through the subnet-specific intelligent agent, generating traffic transmission instructions. This effectively addresses the static limitations of traditional routing protocols in heterogeneous traffic scenarios and the decision-making efficiency bottleneck of centralized intelligent routing, significantly improving the processing efficiency of high-concurrency traffic. It avoids local congestion through global collaborative optimization and dynamically guarantees multi-dimensional QoS requirements, thus providing scalable and adaptive deterministic service guarantees for scenarios such as the Industrial Internet. Attached Figure Description

[0060] To more clearly illustrate the technical solutions in the embodiments or related technologies of this application, the accompanying drawings used in the description of the embodiments or related technologies will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0061] Figure 1 This is a flowchart illustrating a deterministic routing method based on deep learning in one embodiment.

[0062] Figure 2 This is a schematic diagram of the structure of a deterministic routing system based on deep learning in one embodiment.

[0063] Figure 3 This is a schematic diagram of the structure of a computer device using a deterministic routing method based on deep learning in one embodiment. Detailed Implementation

[0064] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0065] In one embodiment, a deterministic routing method based on deep learning is provided. This embodiment illustrates the application of this method to a terminal. It is understood that this method can also be applied to a server, and further to a system including both a terminal and a server, and implemented through the interaction between the terminal and the server. Figure 1 As shown, in this embodiment, the method includes the following steps:

[0066] Step S101: Collect multiple traffic requests arriving in the current time slot; the traffic requests include source node, destination node, and end-to-end quality of service requirements; the end-to-end quality of service requirements include latency requirements, jitter requirements, packet loss rate requirements, and bandwidth requirements.

[0067] In this context, a time slot refers to a fixed time window used for batch processing of network traffic to improve decision-making efficiency; end-to-end (Quality of Service, QoS) requirements reflect key performance indicators of business traffic. For example, control commands in the Industrial Internet may require extremely low latency and packet loss rates, while video streaming may focus more on bandwidth and jitter; latency requirements refer to the maximum allowable transmission time of data packets from source to destination; jitter requirements refer to the maximum allowable range of latency variation; packet loss rate requirements refer to the allowable percentage of lost data packets; and bandwidth requirements refer to the minimum required transmission rate.

[0068] For example, the terminal, acting as the top-level controller, is responsible for collecting multiple traffic requests arriving within the current time slot and organizing them into a request set. Each traffic request includes a source node, a destination node, and end-to-end QoS requirements, specifically including latency requirements, jitter requirements, packet loss rate requirements, and bandwidth requirements. During the collection process, each request undergoes preliminary verification, such as checking the validity of the source and destination nodes and the rationality of the QoS parameters, to avoid invalid data affecting subsequent processing.

[0069] Step S102: Based on the business map constructed by subnet partitioning, perform path quality assessment processing between the source subnet and the destination subnet for traffic requests that need to be transmitted across subnets, and obtain the sequence of cross-subnet transmission paths and boundary node pairs.

[0070] Among them, path quality assessment refers to judging whether the data transmission path between different subnets is stable and efficient. It needs to be carried out around three core objectives: whether it can transmit normally, how fast it can transmit, and how stable it can transmit. In essence, it is to test indicators such as path connectivity, latency, packet loss rate, and bandwidth utilization. A commerce graph is an abstract network representation, where vertices represent subnets, that is, smaller areas of the network that are logically divided, and edges represent the set of physical connections between subnets.

[0071] For example, a commerce map is initialized based on the network topology configuration (such as the connection relationship of subnet boundary routers), and the status indicators of each connection link are acquired in real time, including latency, bandwidth, jitter, and packet loss rate. Based on these indicators, a cross-subnet transmission path is selected by calculating the inter-subnet connection quality assessment coefficient. A cross-segment compatibility index is calculated for each segment of the inter-subnet connection along the path. This index measures the performance difference between adjacent links and, combined with the node pair quality assessment coefficient, selects the optimal boundary node pair. The optimal boundary node pairs for each segment of the inter-subnet connection are combined to form a boundary node pair sequence, ensuring a smooth transition of QoS characteristics throughout the entire path.

[0072] Step S103: Based on the cross-subnet transmission path and boundary node pair sequence, the end-to-end quality of service requirements are split to obtain the quality of service allocation value for each subnet.

[0073] For example, based on the determined cross-subnet transmission path and boundary node pair sequence, the subnet sequence through which the traffic will pass is determined, and historical performance data such as average latency, average jitter, average packet loss rate and standard deviation of each subnet are queried. Based on the historical performance data, the end-to-end quality of service requirements are decomposed to obtain the quality of service allocation value for each subnet.

[0074] Step S104: Based on the service quality allocation value of each subnet, the subnet agent corresponding to each subnet generates a routing decision on the pre-computed candidate path set for each traffic. The subnet agent includes a policy network based on the Transformer-Encoder architecture.

[0075] The subnet agents employ a policy network based on the Transformer-Encoder architecture, which utilizes a self-attention mechanism to process all pending traffic requests in parallel.

[0076] For example, the QoS allocation values ​​and real-time link states received by the subnet agent are used to construct a feature matrix: the traffic feature matrix contains the QoS requirements (such as latency and bandwidth) for each traffic flow, and the link feature matrix contains the performance metrics (such as available bandwidth and latency) of each link within the subnet. The traffic feature matrix and the link feature matrix are fused into a unified representation, which is input into a multi-layer Transformer-Encoder and outputs the encoded feature representation. Based on these feature representations, a probability distribution for each traffic flow on a pre-computed candidate path set is generated, and the final path, i.e., routing decision, is determined by sampling (during training) or taking the maximum probability (at runtime).

[0077] Step S105: Generate a traffic transmission instruction based on the routing decision; the traffic transmission instruction is used to instruct the underlying network device to perform traffic transmission according to the routing decision and update the network link status, and release the occupied resources when the traffic transmission is completed.

[0078] For example, routing decisions are sent to the underlying network devices, specifying the transmission path for each traffic item. The instructions include source and destination nodes, matching rules (such as IP addresses), and actions (such as forwarding ports). Traffic begins transmission and occupies the resources of the corresponding link. The network environment updates accordingly, such as reducing available bandwidth on the link. When traffic transmission is complete and its end time is reached, a resource release mechanism is triggered to reclaim bandwidth and other resources, ensuring dynamic management of network resources.

[0079] In this embodiment, cross-subnet path evaluation is performed by collecting traffic requests and abstracting quotient maps, QoS requirements are finely segmented, and routing decisions are generated in parallel using an intelligent agent based on the Transformer-Encoder architecture. Based on the routing decisions, traffic transmission instructions are generated, and transmission is completed through instruction issuance and state updates. This effectively solves the static defects of traditional routing protocols in heterogeneous traffic scenarios and the decision efficiency bottleneck of centralized intelligent routing, significantly improving the processing efficiency of high-concurrency traffic. It avoids local congestion through global collaborative optimization and dynamically guarantees multi-dimensional QoS requirements, thereby providing scalable and adaptive deterministic service guarantees for scenarios such as the Industrial Internet.

[0080] In one embodiment, based on the commerce graph constructed by subnet partitioning, path quality assessment is performed between the source subnet and the destination subnet for traffic requests that need to be transmitted across subnets, resulting in a sequence of cross-subnet transmission paths and boundary node pairs, including:

[0081] Step S201: Construct a commerce graph based on the subnet set; in the commerce graph, vertices represent subnets and edges represent the set of connection links between subnets.

[0082] For example, loading network configuration information, including a set of subnets. Each subnet represents a logical partition, such as an autonomous system or a management domain. A quotient graph is constructed based on the set of subnets. The vertices of this quotient graph correspond to each subnet in the set, and the edges represent the set of physical connections between subnets; that is, if there are one or more physical links between two subnets, an edge is added to the quotient graph. During the construction process, graph theory algorithms can be used to store and represent the quotient graph structure, ensuring accurate mapping between vertices and edges. The construction of the quotient graph is based on the static configuration of the network topology but supports dynamic updates to cope with topology changes. The graph theory algorithms are computer algorithms designed based on graph theory (a branch of mathematics that studies graph data structures) and are used to efficiently solve various problems in graph structures, such as pathfinding, node association analysis, and network optimization.

[0083] Step S202: Obtain the real-time status indicators of each connection link in the business map; the real-time status indicators include real-time latency, real-time bandwidth, real-time jitter, and real-time packet loss rate.

[0084] Among them, real-time latency refers to the one-way or round-trip delay of data packets in transmission on the link, which is usually measured by active probing or passive monitoring; real-time bandwidth refers to the current available capacity of the link, which is used to reflect the load situation; real-time jitter refers to the variance of latency, which is used to evaluate the stability of the link; and real-time packet loss rate refers to the proportion of data packets lost, which can be calculated by a statistical counter.

[0085] For example, real-time status indicators such as latency, bandwidth, jitter, and packet loss rate of each connection link in the business map are obtained.

[0086] Step S203: For any two subnets in the business map, calculate the node pair quality evaluation coefficient based on the real-time status index of the connection link; the node pair quality evaluation coefficient is used to evaluate the link performance.

[0087] Among them, the node pair quality evaluation coefficient is a comprehensive indicator that quantifies link performance and is used to eliminate differences in units.

[0088] For example, for the connection between any two subnets in the business map, the multi-dimensional real-time status indicators are transformed into a single score, namely the node pair quality estimation coefficient, by linear weighting. The lower the score, the better the link performance.

[0089] Step S204: Based on the node pair quality evaluation coefficient, calculate the inter-subnet connection quality evaluation coefficient using the following formula:

[0090]

[0091] in, This is a quality evaluation coefficient for inter-subnet connections. It is a set of connections between subnets. For inter-node links, and For subnet identification, is the quality evaluation coefficient for node pairs.

[0092] in, It is the set of connection links between subnets Sp and Sq. The value indicates the size of the set; the inter-subnet connection quality evaluation coefficient is used to characterize the average performance of all links between two subnets, and the lower the value, the better the connection quality.

[0093] For example, the set of connection links corresponding to each edge in the business graph is traversed, and the quality evaluation coefficients between subnets are calculated based on the quality evaluation coefficients of each node and the results are cached.

[0094] Step S205: Among all reachable paths from the source subnet to the destination subnet in the business map, the path with the best cumulative inter-subnet connection quality evaluation coefficient is determined as the cross-subnet transmission path.

[0095] Among them, the reachable path refers to the continuous edge sequence enumerated in the business graph by the breadth-first search algorithm; the breadth-first search algorithm is a graph and tree traversal algorithm. The core idea is to "first visit all the neighbors of the current node, and then visit the neighbors of the neighbors layer by layer", which is similar to the ripple diffusion process, ensuring that all reachable nodes are traversed in order from the nearest to the farthest from the current starting point.

[0096] For example, for all reachable paths from the source subnet to the destination subnet in the business map, the cumulative quality assessment coefficient value of the inter-subnet connection for each reachable path is calculated by summation or weighted average, and the path with the smallest cumulative value is selected as the cross-subnet transmission path.

[0097] Step S206: For each segment of the cross-subnet transmission path, calculate the cross-segment compatibility index; and based on the cross-segment compatibility index and the node pair quality evaluation coefficient, determine the optimal boundary node pair.

[0098] For example, for each segment of the selected cross-subnet transmission path, a cross-segment compatibility index is calculated. Combined with the node pair quality evaluation coefficients of candidate node pairs, a comprehensive index value is calculated, and the node pair with the minimum value is selected as the optimal boundary node pair for the corresponding segment's inter-subnet connection. This ensures a smooth performance transition between adjacent links, avoiding the impact of sudden delays or bandwidth changes on traffic stability. During implementation, each connection segment is processed iteratively according to the path sequence, dynamically updating the predecessor node data.

[0099] Step S207: Combine all optimal boundary node pairs to obtain a sequence of boundary node pairs.

[0100] For example, the optimal boundary node pairs are stored as a list or vector in path order, and the continuity of the sequence (such as IP address reachability) is verified, ultimately resulting in a sequence of boundary node pairs. This sequence is a concrete implementation of the cross-subnet transmission path, with each node pair representing a physical connection point between subnets (such as a border router).

[0101] In this embodiment, a business map abstraction network topology is constructed; link status indicators are acquired in real time, and the quality coefficients between node pairs and subnets are dynamically calculated based on the link status indicators; cross-subnet paths are selected based on the cumulative optimal principle, and the boundary node sequence is determined using compatibility indicators. This approach combines macro-path planning with micro-link optimization, ensuring the globality and adaptability of path selection, effectively improving routing decision efficiency, enhancing network awareness, optimizing path performance, and supporting the deterministic transmission requirements of heterogeneous traffic.

[0102] In one embodiment, for a connection between any two subnets in the commerce map, a node pair quality assessment coefficient is calculated based on the real-time status indicators of the connection link, including:

[0103] Step S301: For any connection between subnets in the business map, based on the real-time status indicators of the connection links, calculate the node pair quality evaluation coefficient using the following formula:

[0104]

[0105] in, The node pair quality evaluation coefficient. This represents the connection link between nodes, where u and v are node identifiers. For the real-time latency of the connection link, The real-time bandwidth of the connection link. To address the real-time jitter of the connection link, For the real-time packet loss rate of the connection link, This is the maximum reference value for latency across the entire network. This is the maximum reference value for bandwidth across the entire network. This is the largest reference value for shaking across the entire network. This is the highest reference value for packet loss rate across the entire network. This is the time delay weighting coefficient. This is the bandwidth weighting coefficient. This is the jitter delay weighting coefficient. This is the weighting coefficient for packet loss rate.

[0106] in, , , and These represent the relative importance of latency, bandwidth, jitter, and packet loss rate, respectively, and can be dynamically configured according to business needs. For example, assigning certain parameters to low-latency traffic can be done by... Higher values; , , and It is a network-wide normalized reference value used to eliminate differences in indicator dimensions, for example... This can be obtained through statistics from historical monitoring data.

[0107] For example, by loading weights and reference values, for connections between any subnets in the commerce graph, the node pair quality assessment coefficient for each node pair is calculated in real time based on the real-time status indicators of the connection links using a coefficient calculation formula. This formula employs weighted summation and normalization to transform multi-dimensional QoS indicators into a single score; a lower score indicates better link performance, thus supporting fair comparison.

[0108] In this embodiment, real-time link status indicators are acquired, and weighted calculations are performed based on predefined weights and normalized reference values ​​to output the node pair quality evaluation coefficient for each node pair. This method integrates multi-dimensional indicators into a single evaluation value, eliminates differences in units, and ensures the comparability and objectivity of link performance.

[0109] In one embodiment, for each segment of the inter-subnet connection across the subnet transmission path, a cross-segment compatibility index is calculated; and based on the cross-segment compatibility index and the node pair quality evaluation coefficient, the optimal boundary node pair is determined, including:

[0110] Step S401: Obtain the link performance metrics of the previously selected optimal node pair for the inter-target subnet connection; the link performance metrics include latency, bandwidth, jitter, and packet loss rate.

[0111] Among them, the target subnet connection refers to the inter-subnet connection currently being processed in each segment of the cross-subnet transmission path.

[0112] For example, the optimal boundary node pair of the previous segment of the inter-subnet connection of the target inter-subnet connection is retrieved from the path decision results, and the performance index data of the corresponding link of the node pair is obtained, including latency, bandwidth, jitter and packet loss rate.

[0113] Step S402: Calculate the difference between the link performance index of each node pair in the target subnet inter-connection candidate node pair set and the link performance index of the previously selected optimal node pair using the following formula:

[0114]

[0115] in, For cross-segment compatibility indicators, This is a latency compatibility weighting coefficient. This is a bandwidth compatibility weighting coefficient. This is a jitter compatibility weighting coefficient. This is a compatibility weighting coefficient for packet loss rate. This is the delay of the previously selected optimal node pair. This represents the bandwidth of the previously selected optimal node pair. This is for the jitter of the previously selected optimal node pair. This represents the packet loss rate of the previously selected optimal node pair. The latency of node pairs connecting the target subnets. The bandwidth of the node pairs connecting the target subnets. Jitter of node pairs connecting target subnets. Let c be the packet loss rate of the node pairs connecting the target subnets. This serves as an identifier for inter-subnet connections.

[0116] Among them, the compatibility weight coefficient , , and These are preset parameters used to reflect the sensitivity to differences in various indicators, such as high. These coefficients indicate that latency changes have a significant impact, and they can be dynamically adjusted according to business strategies.

[0117] For example, all candidate node pairs in the candidate node pair set for inter-target subnet connections are traversed, the link performance index of each node pair is extracted, and based on the link performance index of the previously selected optimal node pair, the difference cross-segment compatibility index is calculated in turn, and the results are temporarily stored.

[0118] Step S403: Calculate the comprehensive index value based on the node pair quality evaluation coefficient and the differential cross-segment compatibility index; and determine the node pair with the smallest comprehensive index value as the optimal boundary node pair for the inter-target subnet connection.

[0119] For example, based on the node pair quality evaluation coefficient values ​​and the obtained cross-segment compatibility index values ​​of the candidate node pairs, a comprehensive score for each candidate pair is calculated using a weighted or summative method. All candidate node pairs are iterated over, the comprehensive index values ​​are compared, and the node pair with the smallest value is selected as the optimal boundary node pair for the current segment, and the path status record is updated.

[0120]

[0121] in, For optimal boundary node pairs, The node pair quality evaluation coefficient. This is a cross-segment compatibility indicator.

[0122] In this embodiment, the real-time performance index of the previous optimal node pair is used as a benchmark to calculate the difference compatibility index of the current candidate node pair to quantify performance fluctuations. This is combined with the node pair quality coefficient for comprehensive evaluation and selection of the optimal solution. This significantly improves the continuity and stability of the path, reduces traffic jitter and packet loss risks by avoiding performance abrupt changes, and enhances the adaptive capability of routing by combining historical and real-time data, providing reliable inter-subnet transmission guarantees for deterministic services.

[0123] In one embodiment, the end-to-end quality of service (QoS) requirements are broken down based on the cross-subnet transmission path and boundary node pair sequence to obtain the QoS allocation value for each subnet, including:

[0124] Step S501: Based on the cross-subnet transmission path and boundary node pair sequence, determine the subnet sequence through which the traffic will pass; and obtain the historical performance data of each subnet in the subnet sequence; the historical transmission performance data includes average latency, average jitter, average packet loss rate and standard deviation.

[0125] For example, based on the cross-subnet transmission path and boundary node pair sequence, a subnet list is extracted according to the path order, and historical performance data for each subnet is obtained, including average latency, average jitter, average packet loss rate, and their standard deviation. The cross-subnet transmission path consists of a series of subnet identifiers, such as from the source subnet. via intermediate subnet To the destination subnet sequence.

[0126] Step S502: Set the bandwidth allocation value in each subnet to the minimum bandwidth requirement of the traffic.

[0127] For example, the end-to-end bandwidth requirement is extracted from the traffic request. This requirement is the minimum guaranteed bandwidth for the service traffic. For each subnet in the subnet sequence, the bandwidth allocation value is directly assigned to this minimum bandwidth requirement without adjustment based on historical data, because bandwidth resources need to remain consistent throughout the entire path.

[0128] Step S503: Calculate the first proportional coefficient and the delay stability factor based on the average delay in historical performance data; and calculate the delay allocation value within the subnet based on the first proportional coefficient and the delay stability factor.

[0129] For example, based on the historical average latency data of each subnet in the subnet sequence, the first proportional coefficient and the latency stability factor are calculated using the following formula:

[0130]

[0131]

[0132] in, This is the first proportionality coefficient for each subnet. Let N be the average latency of each subnet, and N be the total number of subnets. This is the sum of the average latency across all subnets. For each subnet, the delay stability factor is... Let be the standard deviation of latency for each subnet. For example, based on a first proportionality coefficient and a latency stability factor, the latency allocation value within the subnet is calculated using the following formula:

[0133]

[0134] in, Assign latency values ​​to each subnet. The end-to-end latency requirements for each subnet are defined as the maximum reference value for latency across the entire network.

[0135] Step S504: Calculate the second proportional coefficient and jitter stability factor based on the average jitter in historical performance data; and calculate the jitter allocation value within the subnet based on the second proportional coefficient and jitter stability factor.

[0136] For example, based on the average jitter data of each subnet, the second proportional coefficient and jitter stability factor are calculated using the following formula:

[0137]

[0138]

[0139] in, This is the second proportionality coefficient for each subnet. The average jitter of each subnet. Let N be the sum of the average jitter across all subnets, where N is the total number of subnets. The jitter stability factor for each subnet. Let be the jitter standard deviation for each subnet. For example, based on the obtained second proportionality coefficient and jitter stability factor, the jitter assignment value is calculated using the following formula:

[0140]

[0141] in, Assign jitter values ​​to each subnet. This represents the end-to-end jitter requirement for each subnet, i.e., the maximum reference value for jitter across the entire network.

[0142] Step S505: Calculate the third proportional coefficient based on the average packet loss rate in the historical performance data; and calculate the packet loss rate allocation value within the subnet based on the third proportional coefficient.

[0143] For example, based on the average packet loss rate data of each subnet, the third proportional coefficient is calculated using the following formula:

[0144]

[0145] in, This is the third proportionality coefficient for each subnet. The average packet loss rate for each subnet. Let N be the sum of the average packet loss rates of all subnets, where N is the total number of subnets. For example, based on the obtained third proportionality coefficient, the packet loss rate allocation for each subnet is calculated using the following formula:

[0146]

[0147] in, Assign values ​​to the packet loss rate of each subnet. This refers to the end-to-end packet loss rate requirement for each subnet, i.e., the maximum reference value for packet loss rate across the entire network.

[0148] Step S506: Based on the bandwidth allocation value, latency allocation value, jitter allocation value, and packet loss rate allocation value, verify whether the total latency, total jitter, and total packet loss rate after splitting meet the original traffic requirements and whether the minimum bandwidth meets the constraints, and obtain the verification results.

[0149] For example, based on the assigned values ​​of each indicator in the calculated subnet sequence and preset constraints, the verification results are obtained to confirm whether the total latency, total jitter, and total packet loss rate after splitting meet the original traffic requirements and whether the minimum bandwidth meets the constraints. The preset constraints are as follows:

[0150]

[0151]

[0152]

[0153]

[0154] This is the sum of the delay allocation values ​​within the subnet. Let N be the delay allocation value for the i-th subnet, and N be the total number of subnets traversed by the traffic transmission path. The sum of cross-subnet link delays For a given cross-subnet transmission path, it is the set of all inter-subnet connection links; Assign a jitter value to the i-th subnet. This represents the probability that transmission is successful across all subnets. Assign a value to the packet loss rate of the i-th subnet. The minimum guaranteed end-to-end bandwidth required for traffic requests; It represents the minimum available bandwidth across all subnets along the traffic path; Let be the bandwidth allocated to the current traffic in the i-th subnet. The maximum end-to-end packet loss rate allowed for traffic requests. The maximum permissible end-to-end jitter required for traffic requests. The maximum permissible end-to-end latency required for a traffic request.

[0155] Step S507: If the verification results show that the total latency, total jitter, and total packet loss rate after splitting meet the original traffic requirements and the minimum bandwidth meets the constraints, the bandwidth allocation value, latency allocation value, jitter allocation value, and packet loss rate allocation value in each subnet are summarized to obtain the service quality allocation value for each subnet.

[0156] For example, if the constraints are met, the bandwidth allocation value, latency allocation value, jitter allocation value, and packet loss rate allocation value of each subnet are organized into structured data, including the subnet ID and the corresponding QoS indicator value; then the set of allocation values ​​is stored or distributed as input to the subnet agent.

[0157] In this embodiment, by determining the subnet sequence and acquiring historical indicator data, fixing the bandwidth allocation value, and calculating the allocation values ​​for latency, jitter, and packet loss rate based on the proportional coefficient and stability factor, the rationality is ensured through mathematical verification. Finally, the service quality allocation values ​​for each subnet are obtained by summarizing. This enables fine-grained and adaptive splitting of QoS requirements, considers the performance differences of subnets, improves reliability, and thus provides a deterministic guarantee of high satisfaction rate for heterogeneous traffic, while optimizing network resource utilization and avoiding local congestion or resource waste.

[0158] In one embodiment, based on the service quality allocation value of each subnet, the subnet agent corresponding to each subnet generates a routing decision on a pre-computed candidate path set for each traffic flow, including:

[0159] Step S601: Based on the service quality allocation value of each subnet, obtain the service quality requirements of all traffic to be decided in the current subnet; and based on the service quality requirements of the traffic to be decided and the real-time status indicators of the links in the current subnet, construct a traffic feature matrix and a link feature matrix.

[0160] The traffic feature matrix is ​​a two-dimensional numerical array, where the row index corresponds to the traffic identifier and the column index corresponds to the QoS requirement dimension, such as latency requirement, bandwidth requirement, etc. The link feature matrix has a similar structure to the traffic feature matrix, with rows representing link identifiers and columns representing link status indicators.

[0161] For example, the service quality allocation values ​​for each subnet are parsed. These values ​​include the specific allocation amounts for QoS indicators such as latency, bandwidth, jitter, and packet loss rate. Then, the pending traffic queue within the current subnet is queried, and detailed information for each traffic item is extracted, such as the source node address, destination node address, and end-to-end QoS requirement parameters. Based on the specific allocation amounts for QoS indicators, a traffic feature matrix is ​​constructed. Simultaneously, performance data for all links within the subnet is collected in real time, including available bandwidth capacity, transmission latency measurements, jitter statistics, and packet loss rate counts. Based on the performance data for all links within the subnet, a link feature matrix is ​​constructed.

[0162] Step S602: Perform independent linear projection transformation on the traffic feature matrix and the link feature matrix to obtain the link feature tensor and the traffic feature tensor.

[0163] Among them, linear projection transformation processing refers to a linear transformation technique that aims to solve the problem of linear inseparability of original features and introduce more feature interaction possibilities through dimensionality increase operations.

[0164] For example, the traffic feature matrix and the link feature matrix are input into two independent linear projection layers for linear projection transformation, resulting in link feature tensors and traffic feature tensors. These projection layers are essentially fully connected neural network layers, each containing a weight matrix and a bias vector. The projection process performs matrix multiplication based on the weight matrix and bias vector, mapping the input matrix to a higher-dimensional space and hiding the input features in a higher-dimensional space, thereby enhancing the expressive power and discriminative power of the features. Matrix multiplication must strictly follow the rules of dimension matching and element-wise computation; the core is to perform an inner product operation between the rows of the previous matrix and the columns of the next matrix. The weight parameters of the linear projection layers are learned during the model training phase through backpropagation and are used to capture key information in the features.

[0165] Step S603: Through the attention weighting mechanism, the link feature tensor is aggregated into a global network state context vector; and the global network state context vector is fused with the traffic feature tensor to obtain a fused tensor.

[0166] Among them, the attention weighting mechanism is an algorithmic logic that simulates the characteristics of human attention. When processing information, it assigns different importance weights to different parts of the information, thereby prioritizing and utilizing key information and weakening the interference of irrelevant or secondary information.

[0167] For example, an attention mechanism is applied to the link feature tensor to calculate the attention weight for each link feature. These weights reflect the importance of the link in the current state. The link feature tensor is then aggregated into a context vector through weighted summation, which represents the overall network state. The context vector is then fused with the traffic feature tensor, for example, by adding the context vector to each traffic feature through a broadcast mechanism. The function for calculating the attention weights can be predefined, such as using a query-key-value model. The broadcast mechanism is a rule for simplifying operations on mismatched dimensional data in numerical computation (especially matrix / array operations). By automatically expanding smaller-dimensional data to be compatible with larger-dimensional data, element-wise operations can be performed directly without manually copying data, simplifying the code and saving memory.

[0168] Step S604: Input the fused tensor into the multilayer Transformer-Encoder and output the encoded feature representation.

[0169] The multi-layer Transformer-Encoder is one of the core components of the Transformer model. It is used to perform deep semantic encoding on input data such as text and image features, extracting feature representations rich in contextual relationships. The Transformer model is divided into two main modules: Encoder and Decoder. The Encoder focuses only on the input data itself, capturing the contextual dependencies within the input (such as "he" referring to "Xiaoming" in the text) and transforming the original input (such as word vectors) into structured feature vectors. Its core is understanding the input content. The Decoder focuses more on generating the output (such as generating a sentence in the target language during translation), and combines the encoding results of the Encoder with the generated output. The output content is handled by the encoder, which focuses solely on input modeling without needing to consider the output. The multi-layered structure refers to the encoder not being a single layer, but rather composed of N identical encoder layers stacked together. Each layer performs a local semantic processing step, and the stacking of multiple layers achieves a progression of semantics from shallow to deep. Each encoder layer contains a multi-head self-attention feedforward neural network, and the training process is stabilized between layers through residual connections and layer normalization. The self-attention mechanism allows each traffic feature to interact with all other traffic features, learning dependencies and resource competition patterns between traffic. The feedforward network provides non-linear transformations, enhancing the model's expressive power. The architectural parameters of the Transformer model, such as the number of attention heads, hidden layer dimensions, and number of layers, are set during model initialization.

[0170] For example, the fusion tensor, as the input sequence, is fed into a network structure consisting of multiple stacked Transformer encoder layers. Through a self-attention mechanism, it learns the complex dependencies and resource competition patterns between flows, outputting encoded feature representations. In implementation, pre-trained Transformer model weights can be loaded, or online fine-tuning can be performed to adapt to specific subnet environments.

[0171] Step S605: Based on the encoded feature representation, generate the probability distribution of each traffic flow on the pre-computed candidate path set through the classification head of the Actor network; and generate a routing decision by sampling based on the probability distribution.

[0172] The Actor network is a neural network consisting of multiple fully connected layers. The final output layer is a classification head that uses the Softmax function to convert the original output into a probability distribution, where each probability value corresponds to a path in the pre-computed candidate path set.

[0173] For example, the encoded feature representation is input into the Actor network to generate the probability distribution of candidate paths for each traffic flow. Then, a random sampling method can be used to obtain the specific path allocation decision, i.e., the routing decision. Here, the random sampling method refers to the method of selecting samples from the research population (the set of all individuals or objects to be studied). Its principle is that each individual in the population has an equal and independent chance of being selected into the sample, thereby reducing subjective bias and allowing the sample to reflect the true characteristics of the population as much as possible.

[0174] The pre-computed candidate path set was obtained using the following method:

[0175] Step S6051: Obtain the link performance index of each link in the current subnet; and calculate the composite weight value of the link based on the link performance index and the preset weight parameters using the composite weight formula:

[0176]

[0177] in, This represents the composite weight value of the link. , and The preset weight parameters, For time delay, For shaking, For packet loss rate, This is the bandwidth normalization adjustment factor. For broadband, It is a very small constant.

[0178] Among them, the preset weight parameters , and and bandwidth normalization adjustment factor It is dynamically adjusted based on the stringency of the service quality requirements for the traffic. For example, for latency-sensitive traffic, Set a small value to amplify the effect of time delay; minimal constant. This is used to prevent a denominator of 0 when the link bandwidth is fully occupied.

[0179] For example, real-time performance data of all links within the subnet are collected, including metrics such as latency, bandwidth, jitter, and packet loss rate. The composite weight value of the links is calculated using a composite weight formula.

[0180] Step S6052: Based on the composite weight value, calculate multiple short paths from the node to the destination node using the Dijkstra algorithm; and summarize the short paths to obtain a pre-calculated candidate path set.

[0181] Dijkstra's algorithm is a single-source shortest path algorithm used to solve the problem of finding the shortest path from a fixed starting point in a graph to all other vertices. Essentially, it uses a "greedy strategy" to gradually select the vertex that is closest to the starting point and uses that vertex to optimize the shortest path estimate of other vertices, ultimately obtaining the shortest path for all vertices.

[0182] For example, the calculated composite weight values ​​are used as the cost values ​​of links in the graph to construct a network topology graph within the subnet, where nodes represent network devices (such as switches), edges represent links, and costs are weight values. Then, Dijkstra's algorithm is applied to enumerate the top K shortest paths from the current node to the destination node. These top K shortest paths are then summarized to obtain a pre-calculated candidate path set. Here, K is the preset number of candidate paths. Dijkstra's algorithm traverses the graph by maintaining a priority queue, gradually expanding the paths and updating the costs.

[0183] In this embodiment, a feature matrix is ​​constructed to numerically represent the network environment by collecting traffic demand and link status within the subnet. A linear projection layer is used to enhance the feature dimension, and an attention mechanism aggregates the global context and fuses it with traffic features to form a rich environmental representation. Deep features are further extracted through a multi-layer Transformer-Encoder to obtain the encoded feature representation. This encoded feature representation is then input into an Actor network to generate a path probability distribution, and a random sampling method is used to output the final routing decision. Simultaneously, a pre-computed candidate path set is used to dynamically evaluate link costs using a composite weight formula, and Dijkstra's algorithm is employed to generate a diverse set of shortest paths. This significantly improves the intelligence and real-time performance of routing decisions, effectively handling high-concurrency traffic scenarios through parallel processing and high-dimensional feature learning.

[0184] To further illustrate the solutions of the embodiments of this application, a specific example is provided below.

[0185] Input N deterministic flows of different types

[0186] Output routing strategy for N traffic streams

[0187] (1) The top-level controller will first collect n traffic F that need to be routed in the current time slot.

[0188] (2) According to , Calculate the cross-subnet transmission path P and its optimal node pair sequence.

[0189] (3) Based on the transmission performance within each subnet [ , , ] and its stability[ , Split traffic.

[0190] (4) The split traffic is grouped and sent to the corresponding subnet intelligent agent controller.

[0191] (5) After receiving the traffic information, the intelligent agent controller generates a traffic-routing probability distribution relationship based on the current network state.

[0192] (6) Perform probability sampling to obtain the final routing strategy and update the network environment, add traffic to the network and occupy the corresponding resources. When the system time slot reaches the end time of the traffic, remove the traffic and release the resources.

[0193] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0194] Based on the same inventive concept, this application also provides a deep learning-based deterministic routing system for implementing the aforementioned deep learning-based deterministic routing method. The solution provided by this system is similar to the implementation described in the above method; therefore, the specific limitations in one or more deep learning-based deterministic routing system embodiments provided below can be found in the limitations of the deep learning-based deterministic routing method described above, and will not be repeated here.

[0195] In one exemplary embodiment, such as Figure 2 As shown, a deterministic routing system 200 based on deep learning is provided, including:

[0196] The request collection module 201 is used to collect multiple traffic requests arriving in the current time slot; the traffic requests include source node, destination node and end-to-end service quality requirements; the end-to-end service quality requirements include latency requirements, jitter requirements, packet loss rate requirements and bandwidth requirements.

[0197] The path filtering module 202 is used to construct a business map based on subnet partitioning. For traffic requests that need to be transmitted across subnets, it performs path quality assessment between the source subnet and the destination subnet to obtain a sequence of cross-subnet transmission paths and boundary node pairs.

[0198] The quality allocation module 203 is used to split the end-to-end quality of service requirements according to the cross-subnet transmission path and boundary node pair sequence to obtain the quality of service allocation value for each subnet.

[0199] The decision generation module 204 is used to generate routing decisions for each traffic on a pre-computed candidate path set based on the service quality allocation value of each subnet and through the subnet intelligent agent corresponding to each subnet; the subnet intelligent agent includes a policy network based on the Transformer-Encoder architecture;

[0200] The instruction generation module 205 is used to generate traffic transmission instructions based on routing decisions. The traffic transmission instructions are used to instruct the underlying network devices to perform traffic transmission according to the routing decisions and update the network link status. When the traffic transmission is completed, the occupied resources are released.

[0201] In one embodiment, the path filtering module 202 is further configured to:

[0202] A commerce graph is constructed based on a set of subnets; in the commerce graph, vertices represent subnets and edges represent the set of connection links between subnets.

[0203] Obtain real-time status metrics for each connection link in the business map; real-time status metrics include real-time latency, real-time bandwidth, real-time jitter, and real-time packet loss rate;

[0204] For any two subnets in the business map, the node pair quality evaluation coefficient is calculated based on the real-time status indicators of the connection link; the node pair quality evaluation coefficient is used to evaluate the link performance.

[0205] Based on the node-pair quality assessment coefficient, the inter-subnet connectivity quality assessment coefficient is calculated using the following formula:

[0206]

[0207] in, This is a quality evaluation coefficient for inter-subnet connections. It is a set of connections between subnets. For inter-node links, and For subnet identification, The node pair quality evaluation coefficient;

[0208] Among all reachable paths from the source subnet to the destination subnet in the business map, the path with the best cumulative inter-subnet connection quality evaluation coefficient is determined as the cross-subnet transmission path;

[0209] For each segment of the cross-subnet transmission path, calculate the cross-segment compatibility index; and based on the cross-segment compatibility index and the node pair quality evaluation coefficient, determine the optimal boundary node pair.

[0210] Combine all the optimal boundary node pairs to obtain a sequence of boundary node pairs.

[0211] In one embodiment, the path filtering module 202 is further configured to:

[0212] For connections between any subnets in the business map, the node pair quality evaluation coefficient is calculated based on the real-time status indicators of the connection links using the following formula:

[0213]

[0214] in, The node pair quality evaluation coefficient. This represents the connection link between nodes, where u and v are node identifiers. For the real-time latency of the connection link, The real-time bandwidth of the connection link. To address the real-time jitter of the connection link, For the real-time packet loss rate of the connection link, This is the maximum reference value for latency across the entire network. This is the maximum reference value for bandwidth across the entire network. This is the largest reference value for shaking across the entire network. This is the highest reference value for packet loss rate across the entire network. This is the time delay weighting coefficient. This is the bandwidth weighting coefficient. This is the jitter delay weighting coefficient. This is the weighting coefficient for packet loss rate.

[0215] In one embodiment, the path filtering module 202 is further configured to:

[0216] Obtain the link performance metrics of the selected optimal node pair for the previous segment of the inter-subnet connection; the link performance metrics include latency, bandwidth, jitter, and packet loss rate;

[0217] The following formula is used to calculate the difference between the link performance index of each node pair in the candidate node pair set for inter-target subnet connections and the link performance index of the previously selected optimal node pair, representing a cross-segment compatibility index:

[0218]

[0219] in, For cross-segment compatibility indicators, This is a latency compatibility weighting coefficient. This is a bandwidth compatibility weighting coefficient. This is a jitter compatibility weighting coefficient. This is a compatibility weighting coefficient for packet loss rate. This is the delay of the previously selected optimal node pair. This represents the bandwidth of the previously selected optimal node pair. This is for the jitter of the previously selected optimal node pair. This represents the packet loss rate of the previously selected optimal node pair. The latency of node pairs connecting the target subnets. The bandwidth of the node pairs connecting the target subnets. Jitter of node pairs connecting target subnets. Let c be the packet loss rate of the node pairs connecting the target subnets. This serves as an identifier for inter-subnet connections.

[0220] Based on the node pair quality assessment coefficient and the differential cross-segment compatibility index, a comprehensive index value is calculated; and the node pair with the smallest comprehensive index value is determined as the optimal boundary node pair for the inter-target subnet connection.

[0221] In one embodiment, the mass allocation module 203 is further configured to:

[0222] Based on the cross-subnet transmission path and boundary node pair sequence, the subnet sequence through which the traffic will pass is determined; and historical performance data of each subnet in the subnet sequence is obtained; the historical transmission performance data includes average latency, average jitter, average packet loss rate and standard deviation;

[0223] Set the bandwidth allocation value within each subnet to the minimum bandwidth requirement of the traffic;

[0224] Based on the average latency in historical performance data, calculate the first proportional coefficient and the latency stability factor; and based on the first proportional coefficient and the latency stability factor, calculate the latency allocation value within the subnet.

[0225] Based on the average jitter in historical performance data, calculate the second proportional coefficient and jitter stability factor; and based on the second proportional coefficient and jitter stability factor, calculate the jitter allocation value within the subnet.

[0226] Calculate the third proportional coefficient based on the average packet loss rate in historical performance data; and calculate the packet loss rate allocation value within the subnet based on the third proportional coefficient.

[0227] Based on the bandwidth allocation values, latency allocation values, jitter allocation values, and packet loss rate allocation values, we verify whether the total latency, total jitter, and total packet loss rate after splitting meet the original traffic requirements and whether the minimum bandwidth meets the constraints, and obtain the verification results.

[0228] If the verification results show that the total latency, total jitter, and total packet loss rate after splitting meet the original traffic requirements and the minimum bandwidth meets the constraints, the bandwidth allocation value, latency allocation value, jitter allocation value, and packet loss rate allocation value in each subnet are summarized to obtain the service quality allocation value for each subnet.

[0229] In one embodiment, the decision generation module 204 is further configured to:

[0230] Based on the service quality allocation values ​​of each subnet, obtain the service quality requirements of all traffic to be decided within the current subnet; and based on the service quality requirements of the traffic to be decided and the real-time status indicators of the links within the current subnet, construct a traffic feature matrix and a link feature matrix.

[0231] Perform independent linear projection transformations on the traffic feature matrix and the link feature matrix to obtain the link feature tensor and the traffic feature tensor, respectively.

[0232] The link feature tensor is aggregated into a global network state context vector through an attention weighting mechanism; and the global network state context vector is fused with the traffic feature tensor to obtain a fused tensor.

[0233] The fused tensor is input into a multi-layer Transformer-Encoder, and the encoded feature representation is output.

[0234] Based on the encoded feature representation, the probability distribution of each traffic flow on the pre-computed candidate path set is generated through the classification head of the Actor network; and based on the probability distribution, a routing decision is generated through sampling.

[0235] The pre-computed candidate path set was obtained using the following method:

[0236] Obtain the link performance metrics for each link within the current subnet; and calculate the composite weight value of the link based on the link performance metrics and preset weight parameters using a composite weight formula:

[0237]

[0238] in, This represents the composite weight value of the link. , and The preset weight parameters, For time delay, For shaking, For packet loss rate, This is the bandwidth normalization adjustment factor. For broadband, It is a very small constant;

[0239] Based on the composite weight value, multiple short paths from the node to the destination node are calculated using the Dijkstra algorithm; and the short paths are summarized to obtain a pre-calculated candidate path set.

[0240] In one embodiment, such as Figure 3 As shown, a computer device is provided, including:

[0241] At least one processor 301, and a memory 302 communicatively connected to at least one of the processors 301: the memory stores application code executable by at least one of the processors, the application code being executed by at least one of the processors to enable at least one of the processors to perform the software-based load computing resource virtualization allocation method as described above.

[0242] The computer device may also include: sensor 303.

[0243] The processor 301, memory 302 and sensor 303 can be connected via a bus or other means, with the bus being an example in the figure.

[0244] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the steps in the above method embodiments.

[0245] For the device embodiments, since they basically correspond to the method embodiments, the relevant parts can be referred to in the description of the method embodiments. The device embodiments described above are merely illustrative. The components described as separate parts may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this disclosure according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0246] The above-described embodiments are merely illustrative of several implementation methods of the embodiments of this application, and their descriptions are relatively specific and detailed. However, they should not be construed as limiting the scope of the patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the embodiments of this application, and these modifications and improvements all fall within the protection scope of the embodiments of this application.

Claims

1. A deterministic routing method based on deep learning, characterized in that, The method includes: Collect multiple traffic requests arriving within the current time slot; the traffic requests include source node, destination node, and end-to-end quality of service requirements; the end-to-end quality of service requirements include latency requirements, jitter requirements, packet loss rate requirements, and bandwidth requirements; Based on the business map constructed by subnet partitioning, the path quality assessment between the source subnet and the destination subnet is performed on the traffic requests that need to be transmitted across subnets, and the sequence of cross-subnet transmission paths and boundary node pairs is obtained. Based on the cross-subnet transmission path and the boundary node pair sequence, the end-to-end quality of service requirements are broken down to obtain the quality of service allocation value for each subnet; Based on the service quality allocation values ​​of each subnet, a routing decision is generated for each traffic flow on a pre-computed candidate path set through the subnet agent corresponding to each subnet; the subnet agent includes a policy network based on the Transformer-Encoder architecture; Based on the routing decision, a traffic transmission instruction is generated; the traffic transmission instruction is used to instruct the underlying network device to perform traffic transmission according to the routing decision, update the network link status, and release the occupied resources when the traffic transmission is completed.

2. The method according to claim 1, characterized in that, The commerce graph constructed based on subnet partitioning performs path quality assessment between the source and destination subnets for traffic requests that require cross-subnet transmission, resulting in a sequence of cross-subnet transmission paths and boundary node pairs, including: A commerce graph is constructed based on a set of subnets; in the commerce graph, vertices represent subnets and edges represent connections between subnets. Obtain real-time status indicators for each connection link in the business map; the real-time status indicators include real-time latency, real-time bandwidth, real-time jitter, and real-time packet loss rate; For any two subnets in the business map, a node pair quality evaluation coefficient is calculated based on the real-time status index of the connection link; the node pair quality evaluation coefficient is used to evaluate link performance. Based on the node pair quality evaluation coefficient, the inter-subnet connection quality evaluation coefficient is calculated using the following formula: in, This is a quality evaluation coefficient for inter-subnet connections. It is a set of links connecting subnets. For inter-node links, and For subnet identification, The node pair quality evaluation coefficient; Among all reachable paths from the source subnet to the destination subnet in the business map, the path with the best cumulative inter-subnet connection quality evaluation coefficient is determined as the cross-subnet transmission path; For each segment of the cross-subnet transmission path, calculate the cross-segment compatibility index; and based on the cross-segment compatibility index and the node pair quality evaluation coefficient, determine the optimal boundary node pair. Combine all the optimal boundary node pairs to obtain a sequence of boundary node pairs.

3. The method according to claim 2, characterized in that, For any two subnets in the business map, based on the real-time status indicators of the connection link, the node pair quality evaluation coefficient is calculated, including: For any connection between subnets in the business map, based on the real-time status indicators of the connection links, the node pair quality evaluation coefficient is calculated using the following formula: in, The node pair quality evaluation coefficient. This represents the connection link between nodes, where u and v are node identifiers. For the real-time latency of the connection link, The real-time bandwidth of the connection link. To address the real-time jitter of the connection link, For the real-time packet loss rate of the connection link, This is the maximum reference value for latency across the entire network. This is the maximum reference value for bandwidth across the entire network. This is the largest reference value for shaking across the entire network. This is the highest reference value for packet loss rate across the entire network. This is the time delay weighting coefficient. This is the bandwidth weighting coefficient. This is the jitter delay weighting coefficient. This is the weighting coefficient for packet loss rate.

4. The method according to claim 2, characterized in that, For each segment of the cross-subnet transmission path, an inter-subnet compatibility index is calculated; and based on the cross-subnet compatibility index and the node pair quality evaluation coefficient, the optimal boundary node pair is determined, including: Obtain the link performance metrics of the previously selected optimal node pair for the inter-target subnet connection; the link performance metrics include latency, bandwidth, jitter, and packet loss rate; The following formula is used to calculate the difference between the link performance index of each node pair in the target subnet inter-connection candidate node pair set and the link performance index of the previously selected optimal node pair, representing a cross-segment compatibility index: in, For cross-segment compatibility indicators, This is a latency compatibility weighting coefficient. This is a bandwidth compatibility weighting coefficient. This is a jitter compatibility weighting coefficient. This is a compatibility weighting coefficient for packet loss rate. This is the delay of the previously selected optimal node pair. This represents the bandwidth of the previously selected optimal node pair. This is for the jitter of the previously selected optimal node pair. This represents the packet loss rate of the previously selected optimal node pair. The latency of node pairs connecting the target subnets. The bandwidth of the node pairs connecting the target subnets. Jitter of node pairs connecting target subnets. Let c be the packet loss rate of the node pairs connecting the target subnets. This serves as an identifier for inter-subnet connections. Based on the node pair quality assessment coefficient and the differential cross-segment compatibility index, a comprehensive index value is calculated; and the node pair with the smallest comprehensive index value is determined as the optimal boundary node pair for the inter-target subnet connection.

5. The method according to claim 1, characterized in that, The step of decomposing the end-to-end quality of service (QoS) requirement based on the cross-subnet transmission path and the boundary node pair sequence to obtain the QoS allocation value for each subnet includes: Based on the cross-subnet transmission path and the boundary node pair sequence, the subnet sequence through which the traffic will pass is determined; and historical performance data of each subnet in the subnet sequence is obtained; the historical transmission performance data includes average latency, average jitter, average packet loss rate, and standard deviation; Set the bandwidth allocation value within each subnet to the minimum bandwidth requirement of the traffic. Based on the average latency in the historical performance data, a first proportional coefficient and a latency stability factor are calculated; and based on the first proportional coefficient and the latency stability factor, a latency allocation value within the subnet is calculated. Based on the average jitter in the historical performance data, a second proportional coefficient and a jitter stability factor are calculated; and based on the second proportional coefficient and the jitter stability factor, a jitter allocation value within the subnet is calculated. Based on the average packet loss rate in the historical performance data, a third proportional coefficient is calculated; and based on the third proportional coefficient, a packet loss rate allocation value within the subnet is calculated. Based on the bandwidth allocation value, latency allocation value, jitter allocation value, and packet loss rate allocation value, verify whether the total latency, total jitter, and total packet loss rate after splitting meet the original traffic requirements, and whether the minimum bandwidth meets the constraints, and obtain the verification results. If the verification results show that the total latency, total jitter, and total packet loss rate after the split meet the original traffic requirements, and the minimum bandwidth meets the constraints, the bandwidth allocation value, latency allocation value, jitter allocation value, and packet loss rate allocation value in each subnet are summarized to obtain the service quality allocation value for each subnet.

6. The method according to claim 1, characterized in that, The step of generating routing decisions for each traffic flow on a pre-computed candidate path set based on the service quality allocation values ​​of each subnet, through the subnet agents corresponding to each subnet, includes: Based on the service quality allocation values ​​of each subnet, obtain the service quality requirements of all traffic to be decided within the current subnet; and based on the service quality requirements of the traffic to be decided and the real-time status indicators of the links within the current subnet, construct a traffic feature matrix and a link feature matrix. Perform independent linear projection transformations on the traffic feature matrix and the link feature matrix respectively to obtain the link feature tensor and the traffic feature tensor. The link feature tensor is aggregated into a global network state context vector through an attention weighting mechanism; and the global network state context vector is fused with the traffic feature tensor to obtain a fused tensor. The fused tensor is input into a multi-layer Transformer-Encoder, which outputs the encoded feature representation. Based on the encoded feature representation, the probability distribution of each traffic flow on the pre-computed candidate path set is generated through the classification head of the Actor network; and based on the probability distribution, a routing decision is generated through sampling. The pre-computed candidate path set is obtained through the following method: Obtain the link performance index of each link within the current subnet; and based on the link performance index and preset weight parameters, calculate the composite weight value of the link using a composite weight formula: in, This represents the composite weight value of the link. , and The preset weight parameters, For time delay, For shaking, For packet loss rate, This is the bandwidth normalization adjustment factor. For broadband, It is a very small constant; Based on the composite weight value, multiple short paths from the node to the destination node are calculated using the Dijkstra algorithm; and the short paths are summarized to obtain a pre-calculated candidate path set.

7. A deterministic routing system based on deep learning, characterized in that, The system includes: The request collection module is used to collect multiple traffic requests arriving within the current time slot; the traffic requests include source nodes, destination nodes, and end-to-end quality of service requirements; the end-to-end quality of service requirements include latency requirements, jitter requirements, packet loss rate requirements, and bandwidth requirements. The path filtering module is used to construct a business map based on subnet partitioning. For the traffic requests that need to be transmitted across subnets, it performs path quality assessment between the source subnet and the destination subnet to obtain a sequence of cross-subnet transmission paths and boundary node pairs. The quality allocation module is used to split the end-to-end quality of service requirements according to the cross-subnet transmission path and the boundary node pair sequence, and obtain the quality of service allocation value for each subnet. The decision generation module is used to generate routing decisions for each traffic flow on a pre-computed candidate path set based on the service quality allocation value of each subnet and through the subnet intelligent agent corresponding to each subnet; the subnet intelligent agent includes a policy network based on the Transformer-Encoder architecture; The instruction generation module is used to generate traffic transmission instructions based on the routing decision; the traffic transmission instructions are used to instruct the underlying network devices to perform traffic transmission according to the routing decision, update the network link status, and release occupied resources when the traffic transmission is completed.

8. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 6.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 6.

Images