A low-latency and high-reliability IP data transmission method for computing power network

Abstract

The application discloses a low-latency and high-reliability IP data transmission method for a computing power network, relates to the technical field of computing power network data transmission, and comprises the following steps: collecting network state panoramic information of a topology snapshot, link delay data, real-time packet loss rate statistics and node available computing power resources before data transmission session establishment; generating a transmission environment feature tensor through joint feature extraction; inputting the tensor into a pre-trained path quality evaluation model; outputting a plurality of candidate IP transmission path quantitative quality scores and reliability predictions; constructing a deterministic transmission scheme containing a main path, a backup path and a switching trigger condition; and generating flow table rules and strategy instructions and issuing the flow table rules and strategy instructions to corresponding path forwarding nodes. The method realizes accurate evaluation and deterministic transmission configuration of the IP transmission path of the computing power network, and guarantees low latency and high reliability of data transmission.

Description

Technical Field

[0001] This invention belongs to the field of computing power network data transmission technology, specifically a low-latency, high-reliability IP data transmission method for computing power networks. Background Technology

[0002] Traditional computing network IP data transmission often employs static routing or routing methods based on single network parameters. It relies solely on fragmented network data such as link latency and packet loss rate to determine the transmission path, failing to incorporate available computing resources at each node into the transmission planning considerations. Path quality assessment depends on simple threshold judgments or manually configured parameters. Existing transmission schemes only plan a single data transmission path, with path switching relying on passive triggering after a failure occurs, lacking any mechanism for pre-judging the path transmission status.

[0003] Traditional transmission methods cannot adapt to operating scenarios involving dynamic changes in computing network topology and real-time fluctuations in computing resources. Path performance can only be qualitatively assessed, lacking quantitative quality evaluation and reliability prediction methods, and the adaptability of transmission paths cannot be accurately distinguished. Passive path switching methods are prone to data transmission interruptions and increased transmission latency. Transmission schemes and data plane forwarding configurations are independent of each other, making it impossible to generate flow table rules specifically and distribute them to forwarding nodes, and transmission control and forwarding execution are difficult to coordinate.

[0004] It is necessary to construct feature tensors based on multi-dimensional network state information, realize the quantitative quality scoring and reliability prediction of transmission paths through pre-trained models, and build a transmission scheme with primary and backup paths and switching trigger conditions based on the scoring and prediction results, and complete the generation and distribution of corresponding flow table rules and policy instructions. Summary of the Invention

[0005] This invention aims to solve at least one of the technical problems existing in the prior art; To this end, the present invention proposes a low-latency, high-reliability IP data transmission method for computing power networks, comprising: Before a data transmission session is established, network status panoramic information associated with the transmission intention is collected. The network status panoramic information includes topology snapshots, link latency data, real-time packet loss rate statistics, and available computing power resources of nodes. Perform joint feature extraction on the network state panorama information to generate a transmission environment feature tensor; The transmission environment feature tensor is input into a pre-trained path quality assessment model, which outputs quantitative quality scores and reliability predictions for multiple candidate IP transmission paths. Based on the quantitative quality score and reliability prediction, a deterministic transmission scheme including the primary path, backup path and switching trigger conditions is constructed. Based on the deterministic transmission scheme, flow table rules and policy instructions for data plane execution are generated and sent to the path forwarding nodes involved in the deterministic transmission scheme.

[0006] Further, joint feature extraction is performed on the network state panorama information to generate a transmission environment feature tensor, including: Adjacency encoding is performed on the topology snapshot to convert the network topology into a spatial relationship matrix of nodes and links; Perform sliding window statistics on the link delay data to extract feature vectors characterizing the stability and trend of link delay; Time series analysis was performed on the real-time packet loss rate statistics to extract the occurrence patterns and intensity characteristics of packet loss events; The available computing resources of the nodes are normalized and combined to form a node load capacity vector; The spatial relationship matrix, the feature vector, the occurrence mode and intensity features, and the node load capacity vector are aligned and spliced ​​in a preset joint embedding space to construct a multidimensional transmission environment feature tensor.

[0007] Furthermore, the transmission environment feature tensor is input into a pre-trained path quality assessment model, which outputs quantitative quality scores and reliability predictions for multiple candidate IP transmission paths, including: In the coding layer of the path quality assessment model, a deep feature transformation is performed on the transmission environment feature tensor to extract high-dimensional path assessment features. In the path exploration layer of the path quality assessment model, based on the high-dimensional path assessment features, K candidate IP transmission paths that satisfy the transmission constraints are generated, and a path feature descriptor is constructed for each candidate IP transmission path. In the multi-task prediction layer of the path quality assessment model, the path feature descriptors corresponding to each candidate IP transmission path are processed in parallel, and the quantitative quality score and reliability prediction value of each path are output synchronously. The quantitative quality score comprehensively reflects the expected latency and jitter, and the reliability prediction value reflects the probability of successful transmission within a specified time.

[0008] Furthermore, the path quality assessment model is pre-built through the following steps: Acquire the historical transmission session dataset and the corresponding network status snapshot dataset collected by the computing power network control platform, label the historical transmission session dataset, and generate a real performance label for each historical transmission path. The real performance label includes latency score, jitter score and transmission success label, which constitute the sample pairs required for model training. A deep neural network comprising an encoding layer, a path exploration layer, and a multi-task prediction layer is constructed as the initial structure of the path quality assessment model; The network state snapshot dataset is input into the encoding layer of the initial structure to generate a historical transmission environment feature tensor. In the path exploration layer, multiple historical candidate paths and their path feature descriptors are simulated and generated based on the historical transmission environment feature tensor. In the multi-task prediction layer, the path feature descriptors of each historical candidate path are processed in parallel, and the corresponding prediction latency score, prediction jitter score and prediction transmission success probability are output. With the goal of minimizing the comprehensive loss function between the predicted latency score, predicted jitter score, predicted transmission success probability and the corresponding real performance label, the initial structure is trained end-to-end using the sample pairs. In an environment independent of data transmission services, the trained model is validated and optimized using a validation set to achieve the final model parameters that meet the preset prediction requirements, thus completing the construction of the path quality assessment model.

[0009] Furthermore, based on the quantified quality score and reliability prediction, a deterministic transmission scheme is constructed, including a primary path, a backup path, and switching trigger conditions, comprising: From the multiple candidate IP transmission paths output by the path quality assessment model, the path with the highest quantitative quality score is selected as the primary path; Among the remaining candidate IP transmission paths, the path with the highest comprehensive score and which conforms to the node and link separation constraint with the primary path is selected as the backup path based on the weighted comprehensive score of the quantitative quality score and the reliability prediction value. Based on the historical performance statistics and real-time predictions of the primary and backup paths, dynamic thresholds for corresponding switching trigger conditions are calculated, and these dynamic thresholds are set as the switching trigger conditions. The switching trigger conditions include a latency jitter upper limit threshold and a consecutive packet loss count threshold.

[0010] Furthermore, based on the deterministic transport scheme, flow table rules and policy instructions for data plane execution are generated, including: The primary and backup paths in the deterministic transmission scheme are analyzed, and each path is converted into a definite forwarding trajectory consisting of a string of nodes and outgoing port identifiers; Generate high-priority flow table matching entries and forwarding actions for the forwarding trajectory of the primary path; Generate flow table matching entries and forwarding actions with the second highest priority and path switching identifiers in the matching entries for the forwarding trajectory of the backup path; Generate an independent path switching control command, which is used to activate the path switching identifier when the conditions are met, and encapsulate it into a protocol message.

[0011] Furthermore, the method also includes: During the data transmission session, the end-to-end performance metrics of the primary path are continuously monitored, including latency jitter and packet loss sequence. When the monitored end-to-end performance indicators meet the preset switching trigger conditions in the deterministic transmission scheme, a lossless data stream switch from the primary path to the backup path is triggered. After the lossless data stream switching is completed, the status of the abandoned original primary path is asynchronously explored and its quality repair assessment is performed. Based on the results of asynchronous exploration and quality repair assessment, the transmission environment feature tensor is updated, and the assessment process of the path quality assessment model is restarted, forming a closed-loop optimization for transmission path selection. The continuous monitoring of end-to-end performance metrics of the primary path during the data transmission session includes: At the entry node of the primary path, a timestamp sequence is inserted into the data packets of each transmission batch; The data packets are received at the exit node of the primary path, the timestamp sequence is extracted, the one-way transmission delay is calculated, and a delay sample sequence is formed. Window statistics are performed on the time delay sample sequence to calculate the time delay jitter; By checking the sequence number continuity of the data packets, packet loss events are identified and recorded, forming the packet loss sequence.

[0012] Furthermore, when the monitored end-to-end performance indicators meet the preset switching trigger conditions in the deterministic transmission scheme, a lossless data stream switch from the primary path to the backup path is triggered, including: The latency jitter is compared with the latency jitter upper limit threshold in real time; Real-time statistics of the length of consecutively lost packets in the packet loss sequence; When the latency jitter continuously exceeds the upper limit threshold for a preset duration, or when the length of the continuously lost packets reaches the threshold for the number of consecutive packet losses, the handover condition is determined to be met. Within a preset response time, an alarm is sent to the network control plane and a pre-stored switching command is executed, wherein the switching command includes activating the path switching identifier; During the handover process, by sending data packets from the buffer queue and synchronizing the path status signaling, the first data packet transmitted on the backup path is seamlessly connected without data loss.

[0013] Furthermore, after completing the lossless data stream switch, the status of the deprecated primary path is asynchronously explored and its quality repair assessment is performed, including: Initiate an active detection task in the background for the abandoned original primary path, send detection data packets and record the response status; Analyze the response results of the active probing task to evaluate the current latency, packet loss rate, and available bandwidth of the abandoned primary path; The current latency, packet loss rate, and available bandwidth indicators obtained from the assessment are compared with the path performance baseline to determine whether the abandoned original primary path has been restored to the preset available state. Generate an assessment report that includes the repair status of the deprecated primary path.

[0014] Furthermore, the step of updating the transmission environment feature tensor based on the results of asynchronous exploration and quality repair assessment, and restarting the assessment process of the path quality assessment model to form a closed-loop optimization of transmission path selection includes: The repair status information of the abandoned original primary path in the evaluation report is integrated into the network status panoramic information collected from the computing power network control platform; Using the integrated network state panorama information, joint feature extraction is re-performed to generate an updated transport environment feature tensor. The updated transmission environment feature tensor is input into the path quality assessment model to start a new round of assessment process, and the updated quantitative quality score and reliability prediction of the candidate IP transmission path are output. Based on the updated evaluation results, determine whether the current deterministic transmission scheme needs to be adjusted. If adjustment is required, generate a new path switching instruction to optimize subsequent data transmission.

[0015] Compared with the prior art, the beneficial effects of the present invention are: By inputting the transmission environment feature tensor into a pre-trained path quality assessment model, the model can directly output quantitative quality scores and reliability prediction results for multiple candidate IP transmission paths. Through model inference, the model completes the numerical representation of path transmission performance, replacing the traditional qualitative judgment method of path performance. This enables accurate differentiation of the transmission quality of different candidate paths, early detection of potential abnormal states during path transmission, and a complete presentation of the service capabilities and stability of each transmission path. This provides objective numerical reference for path selection, reduces path adaptation bias caused by manual parameter configuration, and improves the accuracy and timeliness of path quality judgment.

[0016] Based on the quantitative quality scores and reliability prediction results of multiple candidate IP transmission paths, a deterministic transmission scheme is constructed, including a primary path, backup paths, and switching trigger conditions. This scheme enables standardized path scheduling logic, avoiding transmission interruptions caused by single-path dependence in traditional transmission methods. The switching trigger conditions set according to the prediction results allow for proactive path scheduling, changing the passive handling mode of switching only after a failure occurs. Based on this deterministic transmission scheme, flow table rules and policy instructions adapted to data plane execution are generated and issued to the corresponding path forwarding nodes. This allows the transmission planning logic to be directly translated into the execution actions of the forwarding nodes, achieving synchronous adaptation between the transmission scheme and data plane forwarding execution. This eliminates coordination deviations between transmission control and forwarding execution, shortens the configuration response time during path switching, adapts to the dynamically changing operating environment of the computing network, maintains latency stability during IP data transmission, preserves the continuity of data transmission, and enhances the reliability of the entire data transmission process. Attached Figure Description

[0017] Figure 1 This is a flowchart illustrating the steps of a low-latency, high-reliability IP data transmission method for computing power networks as described in this invention. Figure 2 A flowchart for generating transport environment feature tensors through joint feature extraction; Figure 3 A flowchart for outputting quantitative quality scores and reliability predictions for the path quality assessment model. Detailed Implementation

[0018] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] See Figure 1 This invention provides a low-latency, high-reliability IP data transmission method for computing power networks, the overall implementation scheme of which is as follows: Before establishing a data transmission session, a comprehensive network state information associated with the current transmission intent is collected from the computing power network control platform. This information includes network topology snapshots, latency data for each link, real-time packet loss statistics, and available computing resources for nodes along the transmission path. Subsequently, joint feature extraction is performed on this heterogeneous network state information, fusing topology, latency, packet loss, and computing power information to generate a multi-dimensional transmission environment feature tensor. This tensor is input into a pre-trained path quality assessment model, which performs in-depth analysis of the network environment and outputs quantitative quality scores and reliability predictions for multiple candidate IP transmission paths selected for this transmission session. Based on the model's evaluation results, a deterministic transmission scheme is constructed, explicitly specifying the primary path, backup paths, and the specific triggering conditions for switching from the primary to the backup path. Finally, based on this deterministic transmission scheme, a series of precise flow table matching rules and policy instructions that can be directly executed in the data plane are generated and distributed to all path forwarding nodes involved in the scheme, thus completing the pre-establishment and policy configuration of the transmission channel.

[0020] In one embodiment of the present invention, joint feature extraction is performed on the network state panorama information to generate a transmission environment feature tensor. This process includes multiple parallel and serial processing stages. See also... Figure 2 The process involves several steps: First, adjacency encoding is performed on the acquired topology snapshots, transforming the connections between nodes and between nodes and links into a computable spatial relation matrix. This matrix quantifies the relative positions and connection strengths of network elements. Second, sliding window statistics are performed on link latency data, calculating the mean, variance, skewness, and slope of the latency within the window to extract feature vectors characterizing link latency stability and short-term trends. Third, time series analysis is conducted on real-time packet loss statistics to identify periodic and bursty patterns of packet loss events, and to calculate the intensity and density of packet loss events per unit time, forming the occurrence patterns and intensity characteristics of packet loss events. Finally, the available computing resources of nodes are normalized to eliminate differences in units and magnitudes among different resource types, and indicators such as CPU, memory, storage, and accelerator resources are combined to form a unified node load capacity vector. Finally, the spatial relationship matrix, link delay feature vector, packet loss mode and intensity feature, and node load capacity vector obtained from the above processing are aligned and dimension-unified in a preset joint embedding space. Then, the aligned feature vectors are concatenated along the feature dimension to construct a high-dimensional, dense, multidimensional transmission environment feature tensor.

[0021] In the specific implementation, the process of performing joint feature extraction on the network state panorama information to generate a transmission environment feature tensor is implemented. In this implementation, the collection of network state panorama information is periodic or event-triggered. The collected topology snapshots exist in the form of graph-structured data, link latency data is a time series, real-time packet loss rate statistics are timestamped event logs, and available computing resources for nodes are a resource list in key-value pair format. After the joint feature extraction operation is initiated, the adjacency encoding step for the topology snapshot begins. A unique numerical identifier is assigned to each node and each link in the network, and a two-dimensional spatial relation matrix is ​​constructed. The row and column indices of the matrix correspond to the node identifiers, and the values ​​of the matrix elements represent the connection status and weight of the links between corresponding node pairs. For node pairs that are not directly connected, their matrix element values ​​are zero or infinity. The spatial relation matrix thus transforms the network topology into a mathematical representation that can be processed by numerical algorithms.

[0022] In some embodiments, the process of performing sliding window statistics on link delay data involves extracting the sequence of raw delay measurements within the most recent time window for each link. Within this time window, the mean, variance, skewness, and slope of the delay trend obtained through linear regression are calculated. The mean of the link delay reflects the average transmission speed, the variance reflects the degree of delay fluctuation, the skewness reflects the asymmetry of the delay distribution, and the slope of the delay trend reflects the recent trend of the delay. These statistics are combined into a multi-dimensional feature vector to characterize the stability and trend of the link delay. The dimension of this feature vector is fixed and independent of the number of links. Its calculation can be expressed as a window... Inner delay samples Statistics, such as variance The calculation follows the formula: in: It is a window Mean of internal delay, It refers to the window size.

[0023] In practical implementation, time series analysis of real-time packet loss rate statistics includes aligning discrete packet loss events on the time axis to form a continuous time series. This time series is then analyzed using Fourier transform or wavelet transform to extract the amplitude of periodic frequency components and identify periodic packet loss patterns. Sudden packet loss patterns are extracted by detecting packet loss rate abrupt changes exceeding a preset threshold. The number of packet loss events per unit time and the average number of lost packets are calculated to form the occurrence pattern and intensity characteristics of packet loss events. The available computing resources of each node are normalized and combined. For each node, indicators such as CPU utilization, memory idle rate, storage I / O throughput, and the number of idle cores on a specific accelerator are extracted. Each indicator is linearly scaled to the [0,1] interval using a min-max normalization method to eliminate dimensional differences. The normalized indicators are then sequentially concatenated to form a comprehensive node load capacity vector, which reflects the node's real-time computing and data carrying capacity.

[0024] In practical implementation, the spatial relationship matrix, feature vectors, occurrence mode and intensity features, and node load capacity vectors are aligned and concatenated in a predefined joint embedding space. The joint embedding space is a high-dimensional vector space with a predefined total dimension. The alignment operation ensures that feature vectors from different sources are projected or linearly transformed to representations with the same semantic subspace dimension. Essentially, the spatial relationship matrix is ​​first flattened into a one-dimensional vector. Feature vectors representing link latency stability and trends, and occurrence modes and intensity features representing packet loss events, are directly input. The node load capacity vector is also used as input. All vectors undergo dimension alignment and feature fusion through a fully connected layer, outputting intermediate features with a unified dimension. In practical implementation, the concatenation operation sequentially connects all the processed and aligned feature vectors along their feature dimensions to construct the final multi-dimensional transmission environment feature tensor. This transmission environment feature tensor serves as the input to the subsequent path quality assessment model, containing comprehensive information on network topology, dynamic performance, and node resources.

[0025] In one embodiment of the present invention, a transmission environment feature tensor is input into a pre-trained path quality assessment model, which outputs quantitative quality scores and reliability predictions for multiple candidate IP transmission paths. See also... Figure 3In the encoding layer of the path quality assessment model, multiple cascaded convolutional and fully connected layers perform nonlinear transformations and deep feature extraction on the input transmission environment feature tensor, mapping the original environment features to a high-dimensional path assessment feature space containing high-level semantic information. In the path exploration layer, based on the high-dimensional path assessment features output from the encoding layer, a graph search algorithm is used to explore the network topology, generating K candidate IP transmission paths that meet preset transmission constraints. A path feature descriptor is constructed for each generated candidate path, integrating the feature information of the nodes and links traversed by the path. In the multi-task prediction layer, a parallel processing network branch is used to synchronously and independently process the path feature descriptor of each candidate path. One branch outputs a quantitative quality score reflecting the expected path latency and jitter, while the other branch outputs a reliability prediction value reflecting the probability that the path can successfully complete the transmission task within a specified time window.

[0026] The path quality assessment model is pre-built through the following steps: A large dataset of historical transmission sessions and their corresponding network state snapshots is obtained from the historical records of the computing power network control platform. Performance data for each historical transmission path is labeled to generate true performance labels containing latency scores, jitter scores, and binary transmission success labels, thus forming the sample pairs required for model training. A deep neural network is constructed as the initial structure of the path quality assessment model, containing an encoding layer, a path exploration layer, and a multi-task prediction layer. The historical network state snapshot dataset is input into the encoding layer of the initial structure to generate a series of historical transmission environment feature tensors. In the path exploration layer, based on the historical transmission environment feature tensors, the path selection process under historical conditions is simulated, generating multiple historical candidate paths and their corresponding path feature descriptors. In the multi-task prediction layer, the descriptors of each historical candidate path are processed, outputting the corresponding predicted latency score, predicted jitter score, and predicted transmission success probability. With the goal of minimizing the combined loss function between the predicted performance metrics and the true performance labels, the entire initial network structure is trained end-to-end using the prepared sample pairs. In a simulation environment independent of actual data transmission services, the trained model is used to perform performance verification and hyperparameter tuning using a validation set until the model reaches the preset prediction accuracy requirements. At this point, the final model parameters are saved, and the construction of the path quality assessment model is completed.

[0027] In practice, the process of inputting the transmission environment feature tensor into a pre-trained path quality assessment model and outputting quantitative quality scores and reliability predictions for multiple candidate IP transmission paths is a hierarchical computational process. Specifically, the encoding layer of the path quality assessment model consists of multiple convolutional layers and fully connected layers stacked alternately. The transmission environment feature tensor is fed as input into the first convolutional layer of the encoding layer. The convolutional layer performs convolution operations on the space and feature dimensions of the transmission environment feature tensor through a set of learnable filters to extract local feature patterns. Subsequently, a nonlinear activation function is used to introduce the nonlinear expressive power of the model. The output of the convolutional layer is fed into subsequent fully connected layers. The fully connected layers globally integrate and nonlinearly transform the features abstracted by convolution, ultimately outputting a high-dimensional path assessment feature containing high-level semantic information. This high-dimensional path assessment feature is a dense vector with a dimension much lower than the original transmission environment feature tensor but contains key information for path assessment.

[0028] In some embodiments, the path exploration layer of the path quality assessment model operates based on the high-dimensional path evaluation features output by the encoding layer. Internally, the path exploration layer maintains a dynamic network topology representation, where the node and link weights are mapped from the high-dimensional path evaluation features. The path exploration layer employs an improved K-shortest path algorithm to explore the network topology, searching for K candidate IP transmission paths from the source node to the destination node that satisfy transmission constraints, including a maximum hop count limit and a minimum available bandwidth threshold. For each candidate IP transmission path generated during the exploration, the path exploration layer traverses all nodes and links along the path, indexes the corresponding feature slices from the high-dimensional path evaluation features, and aggregates these feature slices according to the path order. This constructs a unique path feature descriptor for each candidate IP transmission path, which comprehensively represents the complete state of a single path.

[0029] In its implementation, the multi-task prediction layer of the path quality assessment model comprises parallel fully connected network branches: one branch for quantizing quality score prediction and another for reliability prediction. The path feature descriptor of each candidate IP transmission path output by the path exploration layer is independently and synchronously fed into the multi-task prediction layer. The quantizing quality score prediction branch processes the input path feature descriptors and calculates a comprehensive score value through a series of fully connected layers. This comprehensive score value reflects the expected latency and jitter, and its calculation can be expressed as follows: in: Indicates a quantitative quality score. This represents the activation function. Represents the weight matrix. Represents a path feature descriptor vector. This represents the bias term. The reliability prediction branch processes the same path feature descriptor with similar but independent network parameters, outputting a probability value between 0 and 1. The reliability prediction value reflects the probability of successful transmission within a specified time.

[0030] The path quality assessment model is pre-built through the following steps: First, a historical transmission session dataset and a corresponding network state snapshot dataset collected by the computing power network control platform are acquired. The historical transmission session dataset contains detailed records of numerous completed sessions. For each historical transmission path in a record, offline annotation processing is performed based on its actual measured end-to-end latency, latency jitter data, and transmission success indicators. This generates a true latency score, a true jitter score, and a true transmission success label for each historical transmission path. The historical transmission path, network state snapshot, and true performance label together constitute the sample pairs required for model training. Second, a deep neural network containing an encoding layer, a path exploration layer, and a multi-task prediction layer is constructed as the initial structure of the path quality assessment model. The number of convolutional kernels in the encoding layer, the dimension of the fully connected layer, the search algorithm parameters of the path exploration layer, and the branch structure of the multi-task prediction layer are all configurable hyperparameters.

[0031] In the specific implementation, historical network state snapshot datasets are input into the encoding layer of the initial structure to generate a historical transmission environment feature tensor. In the path exploration layer, based on the historical transmission environment feature tensor, multiple historical candidate paths and their path feature descriptors are simulated and generated. In the multi-task prediction layer, the path feature descriptors of each historical candidate path are processed in parallel, outputting the corresponding predicted latency score, predicted jitter score, and predicted transmission success probability. With the objective of minimizing the comprehensive loss function between the predicted latency score, predicted jitter score, predicted transmission success probability, and the corresponding actual latency score, actual jitter score, and actual transmission success label, the initial structure is trained end-to-end using sample pairs. The comprehensive loss function is a weighted sum of mean squared error loss and binary cross-entropy loss. In an environment independent of data transmission services, the trained model is validated and optimized using a validation set. The validation process calculates the mean absolute error between the predicted score and the actual label, as well as the accuracy of the predicted success probability. Hyperparameters are adjusted and training is iterated repeatedly until the model's mean absolute error and accuracy on the validation set reach the preset prediction requirements. The final model parameters at the point of achieving the requirements are saved, completing the construction of the path quality assessment model. It is understandable that the model building process is offline, and the completed path quality assessment model is stored in the form of a parameter file for loading and use when an online transmission session is established.

[0032] In one embodiment of the present invention, a deterministic transmission scheme comprising a primary path, a backup path, and switching triggering conditions is constructed based on a quantitative quality score and reliability prediction. From multiple candidate IP transmission paths output by the path quality assessment model, the path with the highest quantitative quality score is directly selected as the primary path for this data transmission session. Among the remaining candidate IP transmission paths, a comprehensive score for each path is calculated using a preset weighted formula based on the quantitative quality score and reliability prediction value. The path with the highest comprehensive score, and which physically satisfies the node and link separation constraint with the primary path, is selected as the backup path. Based on the historical performance statistics of the primary and backup paths and the real-time prediction values ​​given by the model, a set of dynamic switching triggering thresholds is calculated. The upper limit threshold for latency jitter is set as a proportional coefficient to the predicted jitter value of the current primary path, and the threshold for consecutive packet loss is set as a dynamic value negatively correlated with the path reliability prediction value. These dynamic thresholds constitute the triggering conditions for path switching.

[0033] Based on a deterministic transmission scheme, flow table rules and policy instructions for data plane execution are generated. The primary and backup paths determined in the deterministic transmission scheme are parsed, converting the node sequence of each path into a definite forwarding trajectory consisting of a string of specific node identifiers and output port numbers. A set of high-priority flow table entries is generated for the forwarding trajectory of the primary path, where the matching item corresponds to the tuple information of the transmission session, and the forwarding action is to point to the output port of the next hop of this path. A set of second-highest priority flow table entries is generated for the forwarding trajectory of the backup path, where the matching item adds a path switching identifier to the session tuple, and the forwarding action is to point to the output port of the next hop of the backup path. An independent path switching control instruction is generated. This instruction logic is used to generate and activate the path switching identifier when the switching trigger condition is met. This instruction is encapsulated as a protocol message, which can be issued by the control plane or triggered by the data plane under specific conditions.

[0034] In practice, the path quality assessment model outputs a list containing multiple candidate IP transmission paths and their corresponding evaluation results. Each candidate IP transmission path in the list records a quantitative quality score and a reliability prediction value. The first step in constructing a deterministic transmission scheme is to select the primary path from this list. The selection logic is to directly traverse the list and select the candidate IP transmission path with the highest quantitative quality score as the primary path for this data transmission session. The quantitative quality score is a comprehensive value, and its value directly reflects the model's overall expectation of path latency and jitter performance.

[0035] In some embodiments, after determining the primary path, a backup path is selected from the remaining candidate IP transmission paths. This selection process requires satisfying node and link separation constraints and calculating a weighted composite score. The node and link separation constraint requires that the backup path and the primary path cannot share the same physical nodes and physical links to avoid single points of failure caused by shared physical equipment failures. For each remaining candidate IP transmission path that satisfies the separation constraint, a quantitative quality score is applied. Compared with reliability predictions Calculate the weighted composite score The calculation follows the formula: in: These are the weighting coefficients for quantifying quality scores. These are the weighting coefficients for the reliability prediction values. and The given positive real number is given and satisfies In practice, a weighted composite score is selected from all paths that satisfy the separation constraint. The path with the highest value is designated as the backup path. Evaluation data for the primary and backup paths are shown in Table 1. Table 1: Candidate Path Evaluation and Selection Table In practice, dynamic thresholds are calculated based on historical performance statistics and real-time predictions of the primary and backup paths to set switching trigger conditions and latency jitter upper limit thresholds. The calculation method is to use the path quality assessment model to predict the latency jitter of the primary path. Multiply by a coefficient greater than 1 ,Right now Threshold for consecutive packet loss counts Settings and reliability predictions The value is related and is obtained through a decreasing function mapping, for example: in: It is a basic constant for the number of packet losses. This indicates a rounding down operation; the reliability prediction value. The lower the threshold, the fewer consecutive packet loss times. The settings are relatively more lenient, in order to trigger a switch earlier when the path reliability is low.

[0036] In practical implementation, flow table rules and policy instructions for data plane execution are generated based on the deterministic transmission scheme. The primary and backup paths in the deterministic transmission scheme are parsed, and each path is converted from a node sequence into a defined forwarding trajectory. The format of the forwarding trajectory is a list of node IDs and output port numbers. For the primary path's forwarding trajectory, high-priority flow table matches and forwarding actions are generated. The matches precisely match the five-tuple of the transmission session (source IP address, destination IP address, protocol, source port, destination port), and the action is to forward hop-by-hop according to the primary path's forwarding trajectory. For the backup path's forwarding trajectory, second-highest priority flow table matches and forwarding actions are generated, with the matches carrying a path switching identifier. The matches add a path switching identifier field to the five-tuple, with an initial value of "inactive," and the action is to forward hop-by-hop according to the backup path's forwarding trajectory. Independent path switching control commands are generated. These commands are pieces of logic code whose function is to update the path switching identifier field to "active" when the switching trigger condition is detected, and to encapsulate this state change into a protocol message, which is then sent to the relevant path forwarding nodes. It can be understood that the generation of flow table rules and policy commands is a function of the control plane; the generated commands are compiled into configuration commands that the underlying network devices can recognize and then issued.

[0037] In one embodiment of the present invention, during the data transmission session, the end-to-end performance indicators of the primary path are continuously monitored, including latency jitter and packet loss sequence. At the entry node of the primary path, a precise timestamp sequence is inserted for each transmission batch or critical data packet. These data packets are received at the exit node of the primary path, the timestamp information is extracted, and the one-way transmission delay of each packet is calculated, forming a latency sample sequence. A sliding window statistical analysis is performed on this latency sample sequence to calculate the variance or standard deviation of the delay within the window, which is used as a measure of real-time latency jitter. By checking the continuity of data packet sequence numbers, the sequence numbers of lost packets are identified and recorded, forming a packet loss event sequence. When the monitored end-to-end performance indicators meet the preset switching trigger conditions in the deterministic transmission scheme, a lossless data stream switch from the primary path to the backup path is triggered. The calculated latency jitter is compared in real-time with a preset latency jitter upper limit threshold. The length of consecutively lost packets in the packet loss sequence is calculated in real-time. When the latency jitter metric continuously exceeds the upper limit threshold for a preset duration, or when the length of consecutively lost packets reaches the consecutive packet loss count threshold, the path switching condition is deemed met. Within a preset response time window, a performance alarm is sent to the network control plane, and a pre-defined switching command is automatically executed. The core operation of this command is to activate the path switching identifier associated with the backup path flow table entry. During the switching process, data packets in the transmission queue are temporarily buffered, and the status of the primary and backup paths is synchronized under the coordination of control signaling. This ensures that the first data packet transmitted on the backup path seamlessly connects to the last valid packet on the primary path, guaranteeing no data loss.

[0038] In the specific implementation, the continuous monitoring of end-to-end performance indicators of the primary path and the triggering of lossless data stream switching are implemented during the data transmission session. Specifically, the continuous monitoring function for the end-to-end performance indicators of the primary path is deployed at the ingress and egress nodes. The ingress node inserts a precise timestamp sequence into the data plane for each transmission batch or each critical data packet. The timestamp sequence contains the packet transmission time generated by a high-precision clock. The egress node receives these timestamp-carrying data packets, extracts the timestamp information from the packet header, and records the packet arrival time using a local high-precision clock. The one-way transmission delay is obtained by calculating the difference between the arrival time and the transmission time. The egress node arranges the calculated series of one-way transmission delay values ​​in chronological order to form a delay sample sequence. In the specific implementation, window statistics are performed on the delay sample sequence to calculate delay jitter. The system defines a fixed-length sliding time window, within which the variance of all delay samples is calculated, and the calculated variance value is used as the delay jitter at the current moment. Its calculation can be expressed as: in: It is the number of time-delay samples within the sliding window. It is the first one in the window Each time delay sample value, It is the arithmetic mean of all delay samples within the window. Monitoring of packet loss sequences is achieved by checking the continuity of data packet sequence numbers. The egress node maintains an expected received sequence number. When the received data packet sequence numbers are not continuous, the range of missing sequence numbers is recorded, forming a packet loss event record. A series of consecutive packet loss event records constitutes a packet loss sequence.

[0039] In some embodiments, the process of comparing the latency jitter with the upper limit threshold in real time is a continuous logical judgment operation. The monitoring module continuously calculates the latest latency jitter. Compared with the preset delay jitter upper limit threshold in deterministic transmission schemes The comparison is performed. Real-time statistics on the length of consecutively lost packets in a packet loss sequence are achieved by analyzing packet loss event records. The statistical logic involves finding the longest consecutively missing sequence number interval in the packet loss sequence; the length of this interval is the length of the consecutively lost packets. The monitoring logic determines whether the switching conditions are met, based on latency jitter. The latency jitter threshold is exceeded continuously. Has the duration reached the preset duration? Or the length of consecutively lost packets Has the threshold for consecutive packet loss been reached? These two conditions are logically related by "OR". See Table 2 for a monitoring and judgment scenario: Table 2: Example Table of Path Performance Monitoring and Switching Condition Judgment In practical implementation, when the monitoring logic determines that the switching conditions are met, a lossless data stream switch from the primary path to the backup path is triggered. Within a preset response time, the monitoring module sends a protocol message containing detailed performance alarm information to the network control plane and automatically executes a pre-stored local switching command. The core operation of the switching command is to activate the path switching identifier associated with the backup path flow table entry. During the switching process, to achieve seamless transition, at the moment the switching command is triggered, the data transmission buffer queue of the ingress node temporarily stores data packets that are being processed and those about to be sent. It can be understood that path status synchronization signaling is simultaneously sent on the control channel, notifying relevant nodes on the backup path to prepare to receive and forward data streams with specific sequence numbers. When the backup path is ready, the data packets in the buffer queue are re-marked and forwarded along the backup path. This ensures that the sequence number of the first data packet transmitted on the backup path is closely linked to the sequence number of the last successfully transmitted data packet on the original primary path, thus ensuring no data loss. Optionally, the command execution and data synchronization of the entire switching process are completed within milliseconds.

[0040] In one embodiment of the present invention, after the lossless data stream switch is completed, the status of the deprecated primary path is asynchronously probed and its quality repair assessed. A low-priority active probe task is initiated in the system background to periodically send probe data packets to the deprecated primary path and record its response time and whether it returns successfully. The response results of the active probe task are analyzed to assess key performance indicators such as the current latency, packet loss rate, and available bandwidth of the deprecated primary path. The assessed current performance indicators are compared with the historical performance baseline of the path to determine whether it has recovered to above a preset available state threshold. An assessment report is generated, which includes the assessment of the repair status of the deprecated primary path and its detailed performance indicator data.

[0041] Based on the results of asynchronous exploration and quality repair assessment, the transmission environment feature tensor is updated, and the evaluation process of the path quality assessment model is restarted, forming a closed-loop optimization for transmission path selection. The repair status and performance information of the deprecated primary paths in the assessment report are integrated as new inputs into the network status panorama information continuously collected from the computing power network control platform. Using the network status panorama information integrated with the latest path status information, the joint feature extraction process is re-executed to generate an updated transmission environment feature tensor. The updated transmission environment feature tensor is input into the path quality assessment model to start a new round of evaluation. The model will output a quantitative quality score and reliability prediction for the updated candidate IP transmission paths based on the latest network environment. Based on the updated evaluation results, it is determined whether the currently used deterministic transmission scheme needs adjustment. If the evaluation results show that a better path scheme exists, new path switching or optimization instructions are generated to optimize subsequent data transmission sessions.

[0042] In practice, the process of asynchronously probing and evaluating the status of the deprecated primary path after the lossless data stream switch is implemented. Specifically, after the system confirms that the data stream has been stably switched to the backup path, a low-priority background task is started. This background task is responsible for actively probing the deprecated primary path. The active probing task sends a series of encapsulated probe data packets to the entry node of the deprecated primary path at a low frequency. The destination address of the probe data packets is set to the exit node address of the primary path, and a special probe identifier is marked in the packet header. Upon receiving the probe data packets, the exit node immediately generates and returns a response packet. The background task records the sending time, receiving time, and whether the response arrived successfully, forming a response result log for the active probing task.

[0043] In some embodiments, the response results of the active probing mission are analyzed to assess the current performance metrics of the deprecated primary path, and the current latency is estimated by calculating half of the round-trip time of the response packets. The current packet loss rate is calculated by comparing the total number of probe data packets sent with the number of packets for which no response was received. Available bandwidth is estimated by measuring the total payload of probe data packets successfully transmitted within a fixed time window. The current delay obtained from the assessment. Current packet loss rate and currently available bandwidth Compared to the path performance baseline, which is a historical statistical analysis of the performance metrics of a deprecated primary path under normal operating conditions, including baseline latency. Baseline packet loss rate and baseline available bandwidth Determining whether a deprecated primary path has been restored to a preset usable state is achieved through a composite conditional function, for example, defining a path health score: in: It is a weighting coefficient, when The value exceeds the preset available state threshold. If necessary, the system determines that the previously deprecated primary path has been restored to a preset usable state. Optionally, an evaluation report containing the restoration status of the deprecated primary path is generated. The evaluation report records the path identifier and current performance metrics in a structured data format. , , Path health score And the Boolean result of whether to restore the available state.

[0044] In practice, the transmission environment characteristic tensor is updated based on the results of asynchronous exploration and quality repair assessment, and the assessment process of the path quality assessment model is restarted, forming a closed-loop optimization for transmission path selection. The assessment report includes information on the repair status of deprecated primary paths, particularly the current latency. Current packet loss rate and currently available bandwidth These metrics are integrated as new dynamic information into the network status overview information collected periodically from the computing power network control platform. This integration means using this new, measured performance data to cover or supplement the old statistical information about abandoned primary paths in the network status overview information.

[0045] In practice, the updated transmission environment characteristic tensor is input into the path quality assessment model to initiate a new round of assessment. The path quality assessment model recalculates and outputs updated quantitative quality scores and reliability predictions for candidate IP transmission paths based on the latest network environment characteristics. Based on the updated assessment results, it is determined whether the current deterministic transmission scheme needs adjustment. The judgment logic compares the score of the currently used backup path with the best score among all the latest evaluated candidate paths. If another path has a significantly higher score than the current backup path and satisfies the node and link separation constraint, then adjustment is deemed necessary. If adjustment is required, a new path switching instruction is generated to optimize subsequent data transmission. The new path switching instruction may specify a better path as the new backup path, or, if conditions permit, switch the data flow back to the restored primary path. Optionally, the entire closed-loop optimization process can be triggered periodically or when significant changes occur in the network state.

[0046] The above embodiments are only used to illustrate the technical methods of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical methods of the present invention without departing from the spirit and scope of the technical methods of the present invention.

Claims

1. A method for low-latency, high-reliability IP data transmission in computing power networks, characterized in that, include: Before a data transmission session is established, network status panoramic information associated with the transmission intention is collected. The network status panoramic information includes topology snapshots, link latency data, real-time packet loss rate statistics, and available computing power resources of nodes. Perform joint feature extraction on the network state panorama information to generate a transmission environment feature tensor; The transmission environment feature tensor is input into a pre-trained path quality assessment model, which outputs quantitative quality scores and reliability predictions for multiple candidate IP transmission paths. Based on the quantitative quality score and reliability prediction, a deterministic transmission scheme including the primary path, backup path and switching trigger conditions is constructed. Based on the deterministic transmission scheme, flow table rules and policy instructions for data plane execution are generated and sent to the path forwarding nodes involved in the deterministic transmission scheme.

2. The method for low-latency, high-reliability IP data transmission for computing power networks according to claim 1, characterized in that, Joint feature extraction is performed on the network state panorama information to generate a transmission environment feature tensor, including: Adjacency encoding is performed on the topology snapshot to convert the network topology into a spatial relationship matrix of nodes and links; Perform sliding window statistics on the link delay data to extract feature vectors characterizing the stability and trend of link delay; Time series analysis was performed on the real-time packet loss rate statistics to extract the occurrence patterns and intensity characteristics of packet loss events; The available computing resources of the nodes are normalized and combined to form a node load capacity vector; The spatial relationship matrix, the feature vector, the occurrence mode and intensity features, and the node load capacity vector are aligned and spliced ​​in a preset joint embedding space to construct a multidimensional transmission environment feature tensor.

3. The method for low-latency, high-reliability IP data transmission for computing power networks according to claim 1, characterized in that, The transmission environment feature tensor is input into a pre-trained path quality assessment model, which outputs quantitative quality scores and reliability predictions for multiple candidate IP transmission paths, including: In the coding layer of the path quality assessment model, a deep feature transformation is performed on the transmission environment feature tensor to extract high-dimensional path assessment features. In the path exploration layer of the path quality assessment model, based on the high-dimensional path assessment features, K candidate IP transmission paths that satisfy the transmission constraints are generated, and a path feature descriptor is constructed for each candidate IP transmission path. In the multi-task prediction layer of the path quality assessment model, the path feature descriptors corresponding to each candidate IP transmission path are processed in parallel, and the quantitative quality score and reliability prediction value of each path are output synchronously. The quantitative quality score comprehensively reflects the expected latency and jitter, and the reliability prediction value reflects the probability of successful transmission within a specified time.

4. The method for low-latency, high-reliability IP data transmission for computing power networks according to claim 1, characterized in that, The path quality assessment model is pre-built through the following steps: Acquire the historical transmission session dataset and the corresponding network status snapshot dataset collected by the computing power network control platform, label the historical transmission session dataset, and generate a real performance label for each historical transmission path. The real performance label includes latency score, jitter score and transmission success label, which constitute the sample pairs required for model training. A deep neural network comprising an encoding layer, a path exploration layer, and a multi-task prediction layer is constructed as the initial structure of the path quality assessment model; The network state snapshot dataset is input into the encoding layer of the initial structure to generate a historical transmission environment feature tensor. In the path exploration layer, multiple historical candidate paths and their path feature descriptors are simulated and generated based on the historical transmission environment feature tensor. In the multi-task prediction layer, the path feature descriptors of each historical candidate path are processed in parallel, and the corresponding prediction latency score, prediction jitter score and prediction transmission success probability are output. With the goal of minimizing the comprehensive loss function between the predicted latency score, predicted jitter score, predicted transmission success probability and the corresponding real performance label, the initial structure is trained end-to-end using the sample pairs. In an environment independent of data transmission services, the trained model is validated and optimized using a validation set to achieve the final model parameters that meet the preset prediction requirements, thus completing the construction of the path quality assessment model.

5. A low-latency, high-reliability IP data transmission method for computing power networks according to claim 1, characterized in that, Based on the quantified quality score and reliability prediction, a deterministic transmission scheme is constructed, including a primary path, a backup path, and switching trigger conditions, comprising: From the multiple candidate IP transmission paths output by the path quality assessment model, the path with the highest quantitative quality score is selected as the primary path; Among the remaining candidate IP transmission paths, the path with the highest comprehensive score and which conforms to the node and link separation constraint with the primary path is selected as the backup path based on the weighted comprehensive score of the quantitative quality score and the reliability prediction value. Based on the historical performance statistics and real-time predictions of the primary and backup paths, dynamic thresholds for corresponding switching trigger conditions are calculated, and these dynamic thresholds are set as the switching trigger conditions. The switching trigger conditions include a latency jitter upper limit threshold and a consecutive packet loss count threshold.

6. The method for low-latency, high-reliability IP data transmission for computing power networks according to claim 1, characterized in that, Based on the deterministic transport scheme, flow table rules and policy instructions for data plane execution are generated, including: The primary and backup paths in the deterministic transmission scheme are analyzed, and each path is converted into a definite forwarding trajectory consisting of a string of nodes and outgoing port identifiers; Generate high-priority flow table matching entries and forwarding actions for the forwarding trajectory of the primary path; Generate flow table matching entries and forwarding actions with the second highest priority and path switching identifiers in the matching entries for the forwarding trajectory of the backup path; Generate an independent path switching control command, which is used to activate the path switching identifier when the conditions are met, and encapsulate it into a protocol message.

7. A method for low-latency, high-reliability IP data transmission for computing power networks according to claim 6, characterized in that, The method further includes: During the data transmission session, the end-to-end performance metrics of the primary path are continuously monitored, including latency jitter and packet loss sequence. When the monitored end-to-end performance indicators meet the preset switching trigger conditions in the deterministic transmission scheme, a lossless data stream switch from the primary path to the backup path is triggered. After the lossless data stream switching is completed, the status of the abandoned original primary path is asynchronously explored and its quality repair assessment is performed. Based on the results of asynchronous exploration and quality repair assessment, the transmission environment feature tensor is updated, and the assessment process of the path quality assessment model is restarted, forming a closed-loop optimization for transmission path selection. The continuous monitoring of end-to-end performance metrics of the primary path during the data transmission session includes: At the entry node of the primary path, a timestamp sequence is inserted into the data packets of each transmission batch; The data packets are received at the exit node of the primary path, the timestamp sequence is extracted, the one-way transmission delay is calculated, and a delay sample sequence is formed. Window statistics are performed on the time delay sample sequence to calculate the time delay jitter; By checking the sequence number continuity of the data packets, packet loss events are identified and recorded, forming the packet loss sequence.

8. A method for low-latency, high-reliability IP data transmission for computing power networks according to claim 7, characterized in that, When the monitored end-to-end performance indicators meet the preset switching trigger conditions in the deterministic transmission scheme, a lossless data stream switch from the primary path to the backup path is triggered, including: The latency jitter is compared with the latency jitter upper limit threshold in real time; Real-time statistics of the length of consecutively lost packets in the packet loss sequence; When the latency jitter continuously exceeds the upper limit threshold for a preset duration, or when the length of the continuously lost packets reaches the threshold for the number of consecutive packet losses, the handover condition is determined to be met. Within a preset response time, an alarm is sent to the network control plane and a pre-stored switching command is executed, wherein the switching command includes activating the path switching identifier; During the handover process, by sending data packets from the buffer queue and synchronizing the path status signaling, the first data packet transmitted on the backup path is seamlessly connected without data loss.

9. A method for low-latency, high-reliability IP data transmission for computing power networks according to claim 8, characterized in that, After completing the lossless data stream switch, the status of the deprecated primary path is asynchronously explored and its quality repair is assessed, including: Initiate an active detection task in the background for the abandoned original primary path, send detection data packets and record the response status; Analyze the response results of the active probing task to evaluate the current latency, packet loss rate, and available bandwidth of the abandoned primary path; The current latency, packet loss rate, and available bandwidth indicators obtained from the assessment are compared with the path performance baseline to determine whether the abandoned original primary path has been restored to the preset available state. Generate an assessment report that includes the repair status of the deprecated primary path.

10. A low-latency, high-reliability IP data transmission method for computing power networks according to claim 9, characterized in that, The process of updating the transmission environment feature tensor based on the results of asynchronous exploration and quality repair assessment, and restarting the assessment process of the path quality assessment model to form a closed-loop optimization of transmission path selection includes: The repair status information of the abandoned original primary path in the evaluation report is integrated into the network status panoramic information collected from the computing power network control platform; Using the integrated network state panorama information, joint feature extraction is re-performed to generate an updated transport environment feature tensor. The updated transmission environment feature tensor is input into the path quality assessment model to start a new round of assessment process, and the updated quantitative quality score and reliability prediction of the candidate IP transmission path are output. Based on the updated evaluation results, determine whether the current deterministic transmission scheme needs to be adjusted. If adjustment is required, generate a new path switching instruction to optimize subsequent data transmission.

Images