NoC low-delay data transmission method based on dynamic routing algorithm

Abstract

The invention provides an NoC low-delay data transmission method based on a dynamic routing algorithm, relates to the technical field of data transmission, and aims to solve the technical problems that an existing NoC routing algorithm cannot adapt to a link dynamic load, is lack of congestion trend prejudgment, is slow in path search convergence and is insufficient in QoS differential scheduling. The method comprises the following steps: constructing a dynamic sensing module containing a double-branch LSTM time sequence prediction model, collecting states such as link bandwidth and queue length in real time, and predicting a congestion trend in 50-100ms in the future; constructing a multi-objective evaluation function taking delay as a core, and dynamically adjusting the weight according to the QoS level; solving an optimal path by adopting an improved ant colony algorithm which introduces a congestion penalty and pruning strategy; during transmission, dynamic reselection is triggered through a pre-calculation + fast matching mechanism; and iteratively optimizing the prediction model based on incremental learning. According to the method, congestion is avoided in advance, the path search efficiency and the transmission reliability are improved, the delay is remarkably reduced compared with a traditional algorithm, and the method is adaptive to delay sensitive scenes such as a high-performance processor and an AI chip.

Description

Technical Field

[0001] This invention belongs to the field of data transmission technology, and more specifically, relates to a NoC low-latency data transmission method based on a dynamic routing algorithm. Background Technology

[0002] With the continuous improvement of chip integration, NoC (No-Chip) architecture, as an alternative to traditional buses, has become a core technology for solving data interaction in multi-core chips. Existing NoC routing algorithms have three major drawbacks:

[0003] First, although static routing algorithms (such as XY routing) are simple to implement, they are completely unable to adapt to the dynamic fluctuations of link load. In multi-core concurrent communication scenarios, they are very likely to form continuous congestion on hot links, resulting in a 3-5 fold increase in transmission latency.

[0004] Second, traditional dynamic routing algorithms rely solely on real-time state decisions and lack the ability to predict congestion trends. They often only make passive adjustments after congestion has formed. Furthermore, during path search, they are prone to excessive detours or slow iterative convergence due to an overemphasis on finding the optimal solution, resulting in a reduction in convergence efficiency of more than 40%.

[0005] Third, most algorithms lack a differentiated scheduling mechanism, making it impossible to allocate transmission resources based on the QoS level of data packets. This leads to high real-time data packets such as control commands competing for the link with ordinary data, resulting in uncontrollable transmission delays.

[0006] The aforementioned defects directly cause a bottleneck in NoC transmission performance, making it difficult to meet the stringent requirements of microsecond-level latency and high-reliability transmission in scenarios such as high-performance computing, AI inference, and real-time embedded control. To address this, this invention proposes a low-latency transmission method that integrates dynamic perception, trend prediction, and intelligent optimization. Summary of the Invention

[0007] To address the aforementioned technical problems, this invention provides a NoC low-latency data transmission method based on a dynamic routing algorithm. This method solves the technical problems of existing NoC routing algorithms, such as inability to adapt to dynamic link load, lack of congestion trend prediction, slow path search convergence, and failure to fully meet the QoS differentiation requirements of data packets, resulting in high transmission latency and insufficient reliability.

[0008] A NoC low-latency data transmission method based on a dynamic routing algorithm includes the following steps:

[0009] S1. Construct a dynamic perception model for the NoC topology. The dynamic perception model collects data on the transmission bandwidth, data packet queue length, link error rate, and node computing load of each link in real time. At the same time, the time series prediction model trained with historical transmission data outputs the predicted value of link congestion trend within the next 50-100ms.

[0010] S2. Based on the output of the dynamic perception model, a multi-objective path evaluation function is constructed. The evaluation function takes the minimum transmission delay as the core objective, while taking into account link reliability and node energy consumption. The delay weight coefficient is dynamically adjusted according to the QoS level of the data packet. The delay weight of data packets with real-time requirements higher than a preset threshold shall account for no less than 70%.

[0011] S3. An improved ant colony algorithm is used to solve the multi-objective path evaluation function. A link congestion trend penalty factor is introduced to correct the pheromone update rule, which improves the pheromone evaporation rate of predicted congested links by 30-50%. At the same time, a path hop count upper limit constraint is set to avoid excessive path detours.

[0012] S4. The routing execution module establishes a data transmission channel based on the optimal path obtained by solving. During the transmission process, it detects the deviation between the actual link status and the predicted value in real time. When the absolute value of the deviation exceeds the preset deviation threshold, it triggers the path dynamic reselection mechanism. During the reselection process, it uses a pre-established backup path segment to shorten the switching delay.

[0013] S5. After the data transmission is completed, the actual transmitted data, delay data and reselection records of the link are fed back to the training set of the time series prediction model of the dynamic perception model to realize online iterative optimization of the prediction model.

[0014] Preferably, the time-series prediction model in step S1 adopts a dual-branch LSTM network structure. The first branch takes in the real-time link status data of the past 10-20 sampling periods and outputs the short-term congestion trend prediction value; the second branch takes in the historical transmission data of the past 100-200 sampling periods and outputs the long-term congestion trend prediction value. The outputs of the two branches are weighted and fused through an attention mechanism to obtain the final congestion trend prediction value.

[0015] Preferably, the specific expression of the multi-objective path evaluation function in step S2 is:

[0016] F(P)=α×D(P)+β×R(P)+γ×E(P), where F(P) is the comprehensive evaluation value of path P, D(P) is the predicted transmission delay of path P, R(P) is the reliability coefficient of path P, and E(P) is the average energy consumption coefficient of path P.

[0017] α, β, and γ are weighting coefficients, and α + β + γ = 1. The value of α ranges from 0.7 to 0.9, while the values ​​of β and γ are dynamically adjusted according to the data packet type. The β value of control data packets is 20-30% higher than that of ordinary data packets.

[0018] Preferably, the improved ant colony algorithm in step S3 also includes a path pruning strategy: when there are three consecutive links in the path explored by the ants whose predicted congestion values ​​all exceed the congestion threshold, the exploration of the path is terminated and marked as an invalid path. At the same time, a temporary taboo mark is added to the link combination corresponding to the path, and the taboo duration is positively correlated with the predicted duration of link congestion.

[0019] Preferably, the path dynamic reselection mechanism in step S4 adopts a pre-calculation + fast matching strategy:

[0020] When establishing the initial path, simultaneously calculate 2-3 alternative paths with an overlap rate of no more than 50% with the optimal path and store the path information;

[0021] When a reselection is triggered, the optimal path is matched from the candidate paths based on the current link status. If none of the candidate paths meet the requirements, a simplified improved ant colony algorithm is started to perform a fast path search, reducing the number of search iterations to 40-60% of the initial search.

[0022] Preferably, step S1 also includes a link status classification mechanism: the link is divided into four levels: unobstructed, slightly congested, moderately congested, and heavily congested, based on the transmission bandwidth utilization, queue length, and bit error rate. The sampling frequency of links of different levels is different, and the sampling frequency of links with slight congestion or above is 2-3 times higher than that of unobstructed links.

[0023] Preferably, step S4 also includes a data packet priority scheduling mechanism: a priority queue is set up in the routing execution module, and different scheduling weights are assigned according to the QoS level of the data packets. High-priority data packets adopt a preemptive scheduling strategy. When a high-priority data packet arrives, the transmission of the current low-priority data packet can be paused. After the high-priority data packet is transmitted, the transmission of the low-priority data packet is resumed. The pause position is recorded by frame marking technology to avoid data loss.

[0024] Preferably, the pheromone update rule of the improved ant colony algorithm in step S3 is as follows:

[0025] τ_ij(t+1)=(1-ρ)×τ_ij(t)+ρ×Δτ_ij, where τ_ij(t) is the pheromone concentration of link (i,j) at time t, ρ is the pheromone evaporation coefficient, and Δτ_ij is the pheromone increment;

[0026] For predicted congested links, the ρ value increases to 0.4-0.6, while for unobstructed links, the ρ value decreases to 0.1-0.2. Δτ_ij is positively correlated with the reciprocal of the path evaluation function F(P).

[0027] Preferably, step S2 also includes a node computing power load balancing strategy: during the path evaluation process, if the computing power load of a node exceeds the load threshold, a load penalty term is added to the path passing through that node. The value of the penalty term is positively correlated with the proportion of the node load exceeding the threshold. When the node load exceeds the overload threshold, new data packets are prohibited from selecting the path passing through that node.

[0028] Preferably, the online iterative optimization in step S5 adopts an incremental learning strategy: the model is only trained and updated on new transmitted data whose distribution differs from the historical data by more than a difference threshold. At the same time, the training set is periodically pruned, outdated historical data is deleted, and transmitted data from the last 3-5 working days is retained as training samples to ensure the adaptability and computational efficiency of the prediction model.

[0029] Compared with the prior art, the present invention has the following beneficial effects:

[0030] 1. Significantly improved congestion prediction accuracy and response speed: The dual-branch LSTM model combines short-term 10-20 sampling periods and long-term 100-200 sampling periods of data. After fusion through the attention mechanism, the congestion prediction accuracy reaches over 92%. Compared with traditional non-prediction algorithms, it avoids congestion 50-100ms earlier. Transmission latency is reduced by 37% in low-load scenarios and by 65% ​​in high-load scenarios.

[0031] 2. Improved path search efficiency and optimization accuracy: The improved ant colony algorithm improves convergence speed by 30-50% and reduces path hops by 15-25% compared to the traditional ant colony algorithm by using a congestion trend penalty factor (predicting pheromone evaporation rate of congested links by 30-50%) and a continuous congestion pruning strategy, thus avoiding excessive detours.

[0032] 3. Significantly enhanced QoS differentiation guarantee capability: Based on the 3-bit QoS level field, dynamic weight allocation is achieved. High-priority data packets are preemptively scheduled and frame marking technology is used to avoid loss. The transmission delay fluctuation of control data packets is controlled within ±5%, and the retransmission rate is less than 2%.

[0033] 4. Continuous optimization of system robustness and adaptability: The incremental learning strategy only updates new data parameters with distribution differences exceeding KL divergence of 0.3. Combined with the training set pruning cycle of 4 working days, it reduces training computing power consumption by 80% while ensuring model adaptability, and can stably adapt to NoC scenarios with different load intensities and topologies.

[0034] 5. High hardware feasibility: It adopts a mature FPGA+GPU+ARM architecture, and the processing latency of each module is controlled in the millisecond level (prediction latency ≤2ms, solution latency ≤5ms, switching latency ≤1ms), which is easy to integrate with existing NoC routers. Attached Figure Description

[0035] Figure 1 This is a schematic diagram of the method flow of the present invention. Detailed Implementation

[0036] The embodiments of the present invention will be described in further detail below with reference to the accompanying drawings and examples. The following examples are for illustrative purposes only and should not be construed as limiting the scope of the invention.

[0037] Please see Figure 1 The NoC low-latency data transmission system based on dynamic routing algorithms of this invention mainly comprises five functional modules: a dynamic perception module, a path evaluation module, an improved ant colony solving module, a routing execution module, and a model optimization module. Each module interacts with the others via a high-speed bus. The dynamic perception module collects link and node status data in real time and performs trend prediction; the path evaluation module constructs a multi-objective evaluation function to quantify path performance; the improved ant colony solving module outputs the optimal transmission path; the routing execution module is responsible for path establishment, data transmission, and dynamic reselection; and the model optimization module iteratively optimizes the prediction model using historical data. The overall system architecture adopts a modular design and is adaptable to various NoC topologies such as Mesh, Torus, and Fat-Tree.

[0038] The core parameters are as follows:

[0039] Sampling period: 10ms for unobstructed links, 5ms for lightly / moderately congested links, and 3ms for heavily congested links;

[0040] Prediction duration: 80ms from now (configurable within the range of 50-100ms);

[0041] The weight coefficients are in the range of α∈[0.7,0.9], β∈[0.05,0.2], γ∈[0.05,0.2], and α+β+γ=1;

[0042] Pheromone evaporation coefficient ρ: 0.1-0.2 for unobstructed links, 0.4-0.6 for predicted congested links;

[0043] Path hop limit: Configured according to the NoC topology size; for Mesh topology, it is √(2N) (N is the total number of nodes);

[0044] Deviation threshold: 20% (the relative deviation between the actual link status and the predicted value);

[0045] Number of alternative paths: 2 (expandable to 3), overlap rate ≤ 50%;

[0046] Load threshold: Node computing power utilization 70%, overload threshold: 90%;

[0047] Difference threshold: KL divergence 0.3 (measures the difference in distribution between new and historical data);

[0048] Training set retention period: 4 business days (can be adjusted within 3-5 business days).

[0049] Specific implementation steps:

[0050] Step S1: Dynamic Perception Model Construction and State Prediction

[0051] Data acquisition unit implementation:

[0052] The dynamic sensing module adopts a distributed acquisition + centralized processing architecture. Each NoC node is configured with one local acquisition unit. The core components include:

[0053] Bandwidth monitoring unit: Employs a counter-based traffic statistics chip (such as MAX31855) to count the number of data packets transmitted per second in real time and calculate bandwidth utilization = actual transmission rate / theoretical link bandwidth;

[0054] Queue length monitoring unit: Obtains the data packet queue length by reading the current data volume of the link buffer FIFO register, with the unit being "number of data packets";

[0055] Bit error rate monitoring unit: 16-bit CRC check code is added to each transmitted data packet. After verification by the receiving end, the number of erroneous data packets is counted. Bit error rate = number of erroneous data packets / total number of transmitted data packets;

[0056] Computing load monitoring unit: Through the CPU's built-in performance monitoring counter (such as ARMPMU), it collects the instruction execution rate and cache hit rate of the node processor, and calculates the computing load = number of currently executed instructions / maximum number of instructions executed by the node.

[0057] The local acquisition unit adjusts the sampling frequency according to the link status classification mechanism: by comparing bandwidth utilization (smooth operation ≤30%, mild operation 30%-50%, moderate operation 50%-80%, severe operation >80%), queue length (smooth operation ≤5 queues, mild operation 5-10 queues, moderate operation 10-20 queues, severe operation >20 queues), and bit error rate (smooth operation ≤10 ... -6 Mild 10 -6 -10 -4 Moderate 10 -4 -10 -2 Severe >10 -2 The system categorizes links into four levels, with the sampling frequency for lightly congested and more congested links increased by 2-3 times compared to uncongested links. Collected data is transmitted to the centralized processing unit via an on-chip high-speed serial bus (1Gbps speed).

[0058] Implementation of a two-branch LSTM prediction model:

[0059] The centralized processing unit uses an FPGA (Xilinx Kintex-7) to implement a dual-branch LSTM network. The network structure parameters are as follows:

[0060] Input layer: The first branch (short-term prediction) has an input dimension of 20×4 (20 sampling periods, 4 features per period: bandwidth utilization, queue length, bit error rate, and computing load); the second branch (long-term prediction) has an input dimension of 150×4 (150 sampling periods, with the same features as above);

[0061] Hidden layers: Both branches use a 2-layer LSTM structure. The first branch has 64 hidden units per layer, and the second branch has 128 hidden units per layer. The activation function is ReLU.

[0062] Attention fusion layer: Calculate the attention weights of the two branch outputs. The short-term prediction weight ω1 = 0.6 and the long-term prediction weight ω2 = 0.4. The weights can be optimized online using the gradient descent algorithm. The fusion formula is: Predicted value = ω1 × short-term output + ω2 × long-term output.

[0063] Output layer: Outputs the predicted link congestion trend value within the next 80ms, normalized to [0,1] using the Sigmoid activation function. The larger the value, the higher the congestion risk.

[0064] The model training adopts an offline pre-training + online iteration approach. In the offline stage, it is trained with 100,000 sets of historical data under different load scenarios until the loss function (MSE) is lower than 0.01. In the online stage, the parameters are updated in real time through the model optimization module.

[0065] Step S2: Construction of Multi-Objective Path Evaluation Function

[0066] Evaluation index calculation:

[0067] The path evaluation module runs on the ARM Cortex-A9 processor of the NoC control node. The multi-objective evaluation function F(P) = α×D(P) + β×R(P) + γ×E(P) is implemented in C language. The calculation methods for each index are as follows:

[0068] Predicted transmission delay D(P): For each link contained in path P, calculate the single link delay based on the predicted bandwidth utilization rate = link length / (link bandwidth × (1 - predicted bandwidth utilization rate)) + queue delay (queue length × single packet transmission time), where D(P) is the sum of the delays of all links.

[0069] Reliability coefficient R(P): Single link reliability = 1 - predicted bit error rate, R(P) is the product of the reliability of all links in the path, and the value ranges from [0,1]. The larger the value, the higher the reliability.

[0070] Average energy consumption coefficient E(P): Static energy consumption (idle power consumption) and dynamic energy consumption (data processing power consumption) of each node in the acquisition path. E(P) = (Σ node energy consumption) / number of nodes, in mW.

[0071] Dynamic adjustment mechanism for weighting coefficients:

[0072] The weighting coefficients are dynamically allocated based on the QoS level of the data packet. The QoS level is identified by a 3-bit priority field in the packet header (level 0-7, with level 7 being the highest):

[0073] High real-time data packets (QoS≥5, such as control commands, real-time sensor data): α=0.9, β=0.08, γ=0.02, latency weighting accounts for 90%;

[0074] For medium-real-time data packets (3≤QoS<5, such as video stream data): α=0.8, β=0.15, γ=0.05;

[0075] Normal data packets (QoS<3, such as storage read and write data): α=0.7, β=0.15, γ=0.15.

[0076] At the same time, a node computing power load balancing strategy is introduced: if the computing power load of the nodes traversed by path P exceeds 70%, a load penalty term ΔF = k × (load rate - 70%) is added to F(P), where k is the penalty coefficient (value 0.1); if the node load exceeds 90%, the path is directly marked as an invalid path and is prohibited from being selected.

[0077] Step S3: Solve using the improved ant colony algorithm

[0078] Algorithm initialization configuration:

[0079] The improved ant colony solving module uses GPU (NVIDIA Jetson TX2) for accelerated computation. The algorithm initialization parameters are as follows: number of ants = 2 × N (N is the total number of NoC nodes), maximum number of iterations = 100, initial pheromone concentration τ0 = 0.5, and maximum path hop count = √(2N) (Mesh topology). Each ant carries the QoS level information of the data packet, which is used to adapt the evaluation function weights.

[0080] Pathfinding and pruning strategies:

[0081] Ants start from the source node and select the next hop node based on pheromone concentration and a heuristic function, η_ij = 1 / D_ij (where D_ij is the predicted delay of link (i,j)). A pruning strategy is implemented during path search: the predicted congestion trend of links in the ant's explored path is monitored in real time. If the predicted values ​​of three consecutive links are all >0.8 (the congestion threshold), the ant's path exploration is terminated, the path is marked as invalid, and a temporary tabu is added to the link combination. The tabu duration is equal to the predicted congestion duration × 0.8 (e.g., if the predicted congestion duration is 100ms, the tabu duration is 80ms).

[0082] Implementation of pheromone update rules:

[0083] After each iteration, a pheromone update is performed, with the update formula being τ_ij(t+1)=(1-ρ)×τ_ij(t)+ρ×Δτ_ij, where:

[0084] Pheromone evaporation coefficient ρ: dynamically adjusted according to the predicted link status. For smooth links (predicted value ≤ 0.3), ρ = 0.1; for lightly congested links (0.3 < predicted value ≤ 0.5), ρ = 0.2; for moderately congested links (0.5 < predicted value ≤ 0.8), ρ = 0.4; and for heavily congested links (predicted value > 0.8), ρ = 0.6.

[0085] Pheromones increment Δτ_ij: Only links on the optimal path (with the minimum F(P)) receive an increment, Δτ_ij=K / F(P) (K is a constant, with a value of 10), and links on non-optimal paths do not receive an increase in pheromones.

[0086] This update rule enables the algorithm to converge quickly to the optimal path while suppressing the selection of predicted congested links.

[0087] Step S4: Route Execution and Dynamic Reselection

[0088] Path establishment and packet scheduling:

[0089] The routing execution module is integrated into the NoC router and uses ASIC circuitry for high-speed processing. Upon receiving the optimal path output by the improved ant colony module, the routing execution module establishes data transmission channels by sending configuration commands to the routers at each node along the path (configuring the routers' routing tables and specifying the next-hop node). Simultaneously, it initiates a packet priority scheduling mechanism.

[0090] Set up 3 priority queues (high, medium, and low) to correspond to data packets with QoS≥5, 3≤QoS<5, and QoS<3, respectively;

[0091] The high-priority queue uses a "preemptive" scheduling method: when a high-priority data packet arrives, if a low-priority data packet is currently being transmitted, the router records the pause position of the low-priority data packet (marking the byte offset of the current transmission frame) using frame marking technology, pauses the low-priority transmission, and switches to high-priority transmission; after the high-priority transmission is completed, the transmission of the low-priority data packet is resumed according to the frame marking to avoid data loss.

[0092] Implementation of dynamic reselection mechanism:

[0093] During data transmission, the routing execution module compares the deviation between the actual link status and the predicted value in real time (deviation rate = |actual value - predicted value| / predicted value). When the deviation rate > 20% (deviation threshold), dynamic reselection is triggered.

[0094] Prioritize matching of alternative paths: When the initial path is established, two alternative paths are calculated simultaneously (overlap rate ≤ 50%, calculated by the number of intersection nodes: overlap rate = number of intersection nodes / (number of nodes in path 1 + number of nodes in path 2 - number of intersection nodes)). The path with the smallest current F(P) is selected as the new path from the alternative paths.

[0095] When alternative paths are invalid, a fast search is initiated: if the F(P) of both alternative paths is more than 1.5 times that of the optimal path, a simplified improved ant colony algorithm is initiated, reducing the number of iterations to 50 (50% of the initial number), and only searching the local region between the source node and the destination node (hop count ≤ 3), thus shortening the search delay.

[0096] Path switching: After a new path is established, uncompleted data packets are forwarded to the new path using "seamless switching" technology, with a switching delay of ≤1ms.

[0097] Step S5: Online Iterative Optimization of the Model

[0098] The model optimization module runs on the embedded Linux system of the control node, and uses Python to implement the incremental learning algorithm. The specific process is as follows:

[0099] Data filtering: After data transmission is completed, collect the actual transmission data of the link (bandwidth, queue length, etc.), latency data and reselection records, calculate the KL divergence between the new data and the historical training set, and if the KL divergence is >0.3 (difference threshold), then the data is included in the incremental training set.

[0100] Incremental training: The SGD optimizer is used with a learning rate of 0.001. Only the output layer and attention layer parameters of the dual-branch LSTM network are updated, and the training epochs are 10 to avoid the computational consumption caused by full training.

[0101] Training set pruning: The training set is pruned every day at 2:00 AM, deleting historical data older than 4 working days and retaining valid data from the last 4 days (approximately 500,000 sets) to ensure the timeliness and storage efficiency of the training set.

[0102] Model Deployment: After training is complete, the optimized model parameters are downloaded to the FPGA's storage unit via the PCIe bus to replace the original model parameters, thus realizing online upgrades of the prediction model.

[0103] Hardware implementation example:

[0104] This invention uses a Mesh topology NoC architecture (4×4 nodes, 16 nodes in total) for hardware verification. The core hardware configuration is as follows:

[0105]

[0106] Test verification results:

[0107] On the aforementioned hardware platform, tests were conducted under different load scenarios (low load: 30% node activity, medium load: 60% node activity, high load: 90% node activity) to compare the performance metrics of traditional XY routing, traditional ant colony routing, and the method of this invention:

[0108]

[0109] Test results show that the method of the present invention can significantly reduce transmission latency under different load scenarios, reducing it by 37%-65% compared with traditional XY routing and 19%-45% compared with traditional ant colony routing, and the retransmission rate is less than 2%, which verifies the effectiveness and reliability of the method.

[0110] It is important to note that:

[0111] The prediction duration, weighting coefficients, and other parameters of this invention need to be debugged offline according to the specific NoC topology scale and application scenario to ensure adaptability.

[0112] The maximum number of hops in a path needs to be configured reasonably. Too high a number will increase the delay of path detours, while too low a number will reduce the flexibility of path search.

[0113] The difference threshold and training set retention time for incremental learning need to be adjusted according to the data update frequency to avoid overfitting or underfitting the model.

[0114] Preemptive scheduling of high-priority data packets requires strict control of the preemption frequency to avoid starvation of low-priority data packets (a minimum transmission bandwidth guarantee for low-priority data packets can be set).

[0115] The embodiments of the present invention are given for illustrative and descriptive purposes only, and are not intended to be exhaustive or to limit the invention to the forms disclosed. Many modifications and variations will be apparent to those skilled in the art. The embodiments were chosen and described in order to better illustrate the principles and practical application of the invention, and to enable those skilled in the art to understand the invention and to design various embodiments with various modifications suitable for a particular purpose.

Claims

1. A NoC low-latency data transmission method based on a dynamic routing algorithm, characterized in that, Includes the following steps: S1. Construct a dynamic perception model for the NoC topology. The dynamic perception model collects data on the transmission bandwidth, data packet queue length, link error rate, and node computing load of each link in real time. At the same time, the time series prediction model trained with historical transmission data outputs the predicted value of link congestion trend within the next 50-100ms. S2. Based on the output of the dynamic perception model, a multi-objective path evaluation function is constructed. The evaluation function takes the minimum transmission delay as the core objective, while taking into account link reliability and node energy consumption. The delay weight coefficient is dynamically adjusted according to the QoS level of the data packet. The delay weight of data packets with real-time requirements higher than a preset threshold shall account for no less than 70%. S3. An improved ant colony algorithm is used to solve the multi-objective path evaluation function. A link congestion trend penalty factor is introduced to correct the pheromone update rule, which improves the pheromone evaporation rate of predicted congested links by 30-50%. At the same time, a path hop count upper limit constraint is set to avoid excessive path detours. S4. The routing execution module establishes a data transmission channel based on the optimal path obtained by solving. During the transmission process, it detects the deviation between the actual link status and the predicted value in real time. When the absolute value of the deviation exceeds the preset deviation threshold, it triggers the path dynamic reselection mechanism. During the reselection process, it uses a pre-established backup path segment to shorten the switching delay. S5. After the data transmission is completed, the actual transmitted data, delay data and reselection records of the link are fed back to the training set of the time series prediction model of the dynamic perception model to realize online iterative optimization of the prediction model.

2. The method according to claim 1, characterized in that, The time-series prediction model in step S1 adopts a dual-branch LSTM network structure. The first branch takes in the real-time link status data of the last 10-20 sampling periods and outputs the short-term congestion trend prediction value. The second branch takes in historical transmission data from the past 100-200 sampling periods and outputs a long-term congestion trend prediction. The outputs of the two branches are weighted and fused through an attention mechanism to obtain the final congestion trend prediction.

3. The method according to claim 1, characterized in that, The specific expression of the multi-objective path evaluation function in step S2 is: F(P)=α×D(P)+β×R(P)+γ×E(P), where F(P) is the comprehensive evaluation value of path P, D(P) is the predicted transmission delay of path P, R(P) is the reliability coefficient of path P, and E(P) is the average energy consumption coefficient of path P. α, β, and γ are weighting coefficients, and α + β + γ = 1. The value of α ranges from 0.7 to 0.9, while the values ​​of β and γ are dynamically adjusted according to the data packet type. The β value of control data packets is 20-30% higher than that of ordinary data packets.

4. The method according to claim 1, characterized in that, The improved ant colony algorithm in step S3 also includes a path pruning strategy: when there are three consecutive links in the path explored by the ants whose predicted congestion values ​​all exceed the congestion threshold, the exploration of the path is terminated and marked as an invalid path. At the same time, a temporary taboo mark is added to the link combination corresponding to the path, and the taboo duration is positively correlated with the predicted duration of link congestion.

5. The method according to claim 1, characterized in that, The path dynamic reselection mechanism in step S4 adopts a pre-calculation + fast matching strategy: When establishing the initial path, simultaneously calculate 2-3 alternative paths with an overlap rate of no more than 50% with the optimal path and store the path information; When a reselection is triggered, the optimal path is matched from the candidate paths based on the current link status. If none of the candidate paths meet the requirements, a simplified improved ant colony algorithm is started to perform a fast path search, reducing the number of search iterations to 40-60% of the initial search.

6. The method according to claim 1, characterized in that, Step S1 also includes a link status classification mechanism: the link is divided into four levels: smooth, lightly congested, moderately congested, and heavily congested, based on the transmission bandwidth utilization, queue length, and bit error rate. The sampling frequency of links of different levels is different, and the sampling frequency of links with light or more congestion is 2-3 times higher than that of smooth links.

7. The method according to claim 1, characterized in that, Step S4 also includes a packet priority scheduling mechanism: a priority queue is set up in the routing execution module, and different scheduling weights are assigned according to the QoS level of the packets. High-priority packets adopt a preemptive scheduling strategy. When a high-priority packet arrives, the transmission of the current low-priority packet can be paused. After the high-priority packet is transmitted, the transmission of the low-priority packet is resumed. The pause position is recorded by frame marking technology to avoid data loss.

8. The method according to claim 1, characterized in that, In step S3, the pheromone update rule of the improved ant colony algorithm is: τ_ij(t+1)=(1-ρ)×τ_ij(t)+ρ×Δτ_ij, where τ_ij(t) is the pheromone concentration of link (i,j) at time t, ρ is the pheromone evaporation coefficient, and Δτ_ij is the pheromone increment; For predicted congested links, the ρ value increases to 0.4-0.6, while for unobstructed links, the ρ value decreases to 0.1-0.

2. Δτ_ij is positively correlated with the reciprocal of the path evaluation function F(P).

9. The method according to claim 1, characterized in that, Step S2 also includes a node computing power load balancing strategy: during the path evaluation process, if the computing power load of a node exceeds the load threshold, a load penalty term is added to the path passing through that node. The value of the penalty term is positively correlated with the proportion of the node load exceeding the threshold. When the node load exceeds the overload threshold, new data packets are prohibited from choosing the path passing through that node.

10. The method according to claim 1, characterized in that, The online iterative optimization in step S5 adopts an incremental learning strategy: the model is only trained and updated on new transmitted data whose distribution differs from the historical data distribution by more than a difference threshold. At the same time, the training set is periodically pruned, outdated historical data is deleted, and transmitted data from the last 3-5 working days is retained as training samples to ensure the adaptability and computational efficiency of the prediction model.

Images