A CXL switching structure and adaptive data optimization method based on dynamic link prediction

Abstract

The application discloses a CXL switching structure based on dynamic link prediction, comprising: a traffic monitoring unit configured to monitor real-time traffic data at the input port and the output port of each node; a link prediction unit comprising a data cache module and a link prediction engine, wherein the data cache module is configured to record historical traffic data within a unit time, and the link prediction engine is configured to predict the congestion probability of each path according to the historical traffic data by using a decision tree algorithm model; an adaptive path control unit configured to assign corresponding data transmission weights to each path based on a probability threshold mapping strategy according to the congestion probability of each path, and to adjust a path mapping table in real time; and a feedback optimization unit configured to update the decision tree algorithm model according to the historical traffic data and actual congestion results.

Description

Technical Field

[0001] This invention relates to the field of computer technology, and in particular to a CXL switching structure and adaptive data optimization method based on dynamic link prediction. Background Technology

[0002] With the ever-increasing demand for low-latency, high-bandwidth, and high-reliability interconnects in scenarios such as high-performance computing (HPC), artificial intelligence (AI), and data centers, the traditional PCIe (Peripheral Component Interconnect Express) architecture is gradually becoming insufficient to meet the needs of elastic expansion and dynamic communication. Therefore, Compute Express Link (CXL), as an emerging high-speed interconnect standard, has become one of the key directions for the evolution of future data center interconnect architectures due to its low latency, high bandwidth, and inherent support for cache coherency. The CXL Switch Fabric is a core component of the CXL architecture. Its function is similar to a PCIe switch, but it supports more device types, more complex interconnect topologies, and higher levels of resource sharing and isolation.

[0003] Current CXL switching architectures mostly employ static or semi-static path selection mechanisms, such as fixed routing policies based on the target device's BDF (Bus / Device / Function) number or Global ID. CXL routing uses static routing tables, also known as policy-based routing tables; each switch maintains an address mapping table to route host requests (by address or device ID) to the next-level port within the switch, i.e., from upstream to downstream. Routing between switches is managed by the Fabric Manager.

[0004] These methods often lack sufficient dynamic adaptability when faced with sudden traffic surges, high-concurrency data transmission, or link degradation, easily leading to link congestion, path imbalance, or even a decrease in overall throughput. Furthermore, in multi-path reachable scenarios, these methods cannot fully utilize bandwidth redundancy resources, lack intelligent prediction mechanisms for path switching, and suffer from high path switching latency and error rates.

[0005] In a CXL switching fabric, there may be multiple reachable paths between a host and a device. Intelligently selecting the optimal path for data forwarding while ensuring Quality of Service (QoS) is one of the key issues in the design of a CXL switching fabric.

[0006] The above background information is provided only to assist in understanding the inventive concept and technical solution of this invention. It does not necessarily belong to the prior art of this application, nor does it necessarily provide technical teaching. In the absence of clear evidence that the above information was disclosed before the filing date of this application, the above background information should not be used to evaluate the novelty and inventiveness of this application. Summary of the Invention

[0007] The purpose of this invention is to provide a CXL switching structure based on dynamic link prediction and an adaptive data optimization method.

[0008] To achieve the above objectives, the first aspect of the technical solution adopted by the present invention is as follows:

[0009] A CXL switching architecture based on dynamic link prediction includes:

[0010] A traffic monitoring unit is configured to monitor real-time traffic data at the input and output ports of each node;

[0011] The link prediction unit includes a data caching module and a link prediction engine. The data caching module is configured to record historical traffic data within a unit of time, and the link prediction engine is configured to use a decision tree algorithm model to predict the congestion probability of each path based on the historical traffic data.

[0012] An adaptive path control unit is configured to assign corresponding data transmission weights to each path based on a probability threshold mapping strategy according to the congestion probability of each path, and adjust the path mapping table in real time.

[0013] The feedback optimization unit is configured to update the decision tree algorithm model based on the historical traffic data and the actual congestion results.

[0014] In one embodiment, the traffic data includes packet queue count, throughput, and latency.

[0015] In one embodiment, the link prediction engine has preset path judgment conditions, and the parameters in the path judgment conditions include the number of data packets queued on the path, which is the average number of data packets queued on each link in the path.

[0016] In one embodiment, the link prediction engine has preset path judgment conditions, and the parameters in the path judgment conditions include path delay time, which is the average delay time of each link in the path.

[0017] In one embodiment, the link prediction engine has preset path determination conditions, the parameters of which include path bandwidth utilization, wherein path bandwidth utilization = 0.4*AVG + 0.4*MAX + 0.2*S 2 Where AVG is the average bandwidth utilization of each link in the path, MAX is the maximum bandwidth utilization of each link in the path, and S 2 Let V be the variance of bandwidth utilization for each link in the path. The bandwidth utilization of a link = link throughput / link bandwidth.

[0018] In one embodiment, the congestion probability of each path is defined as P, and the congestion probability threshold satisfies 0 < P1 < P2 < ... < PN-1 < PN < 1, where PN represents the Nth threshold. The probability threshold mapping strategy is as follows: if PN < P < 1, the data transmission weight of the path is set to 0; if PN-1 < P ≤ PN, the data transmission weight of the path is set to AN; if P1 < P ≤ P2, the data transmission weight of the path is set to A2, and 1 > A2 > AN > 0; if 0 < P ≤ P1, the path is recorded as a path with data transmission weight to be increased, and the total scheduling weight released by all paths with reduced data transmission weight is proportionally allocated to the path with data transmission weight to be increased.

[0019] In one embodiment, the feedback optimization unit includes a timed training module, which is configured to periodically update the decision tree algorithm model based on the historical traffic data and the actual congestion results.

[0020] In one embodiment, the timed training module is configured to perform the following steps:

[0021] If the current time is greater than or equal to T since the last training time of the decision tree algorithm model, then the decision tree algorithm model is retrained.

[0022] The historical traffic data is matched with the actual congestion results to form multiple sets of sample data. It is determined whether the number of sample data is greater than the threshold H. If so, the decision tree algorithm model is retrained.

[0023] Replace the current decision tree algorithm model with the retrained decision tree algorithm model.

[0024] In one embodiment, the decision tree algorithm model obtained after retraining directly takes over the decision tree algorithm model of the link prediction unit.

[0025] In one embodiment, the feedback optimization unit further includes an error feedback module, which is configured to compare the actual congestion result with the congestion probability and calculate the congestion probability error, and use the congestion probability error to increase or decrease the complexity of the decision tree algorithm model.

[0026] In one embodiment, the error feedback module is configured to perform the following steps:

[0027] Read the complexity of the current decision tree algorithm model;

[0028] Set up an error sliding window queue to record the congestion probability error for each path;

[0029] If the congestion probability error is greater than the threshold I for N consecutive times, and the number of trees in the current decision tree algorithm model is less than the threshold J, then the number of trees in the decision tree algorithm model is increased by one step, where N≥2.

[0030] In one embodiment, the error feedback module is configured to perform the following steps:

[0031] Read the complexity of the current decision tree algorithm model;

[0032] Set up an error sliding window queue to record the congestion probability error for each path;

[0033] If the congestion probability error is less than the threshold L for N consecutive times, and the number of trees in the current decision tree algorithm model is greater than the threshold M, then the number of trees in the decision tree algorithm model is reduced by one step, where N≥2.

[0034] In one embodiment, changes to the number of trees in the decision tree algorithm model do not take effect immediately on the current decision tree algorithm model, but rather when the timing training module retrains the decision tree algorithm model the next time.

[0035] To achieve the above objectives, the second aspect of the technical solution adopted by the present invention is as follows:

[0036] An adaptive data optimization method, applicable to the CXL switching structure based on dynamic link prediction as described above, specifically includes the following steps:

[0037] Use a traffic monitoring unit to monitor real-time traffic data at the input and output ports of each node;

[0038] The data caching module in the link prediction unit records historical traffic data per unit time.

[0039] Using the link prediction engine in the link prediction unit, the congestion probability of each path is predicted based on the historical traffic data using the decision tree algorithm model.

[0040] The adaptive path control unit assigns corresponding data transmission weights to each path based on the congestion probability of each path and a probability threshold mapping strategy, and adjusts the path mapping table in real time.

[0041] The switch forwards data according to the adjusted path mapping table;

[0042] The feedback optimization unit updates the decision tree algorithm model based on the historical traffic data and the actual congestion results.

[0043] The updated decision number algorithm model is input into the link prediction unit.

[0044] The beneficial effects of the technical solution provided by this invention are as follows:

[0045] a. Improve the intelligence and real-time performance of path selection: This invention uses a link prediction unit to estimate the future load of each path in real time, breaking through the traditional routing method based on static BDF or Global ID, and realizing "path selection based on prediction results", which significantly improves the adaptability and intelligence of the switching structure;

[0046] b. Alleviating link congestion and reducing latency fluctuations: This invention can identify link nodes that are about to become congested in advance, and can proactively avoid high-load paths before data is sent, thus preventing data packets from queuing and waiting in congested links, thereby reducing overall latency and jitter;

[0047] c. Improve bandwidth utilization and overall throughput: In a multi-path reachable CXL switching structure, this invention can dynamically distribute traffic according to the current link availability, achieve path-level load balancing, effectively eliminate bandwidth bottlenecks, and improve the overall utilization and throughput performance of the system data channel.

[0048] d. Enhanced system scalability and fault tolerance: This invention supports automatic adaptation to link changes or failures through the path reconstruction mechanism inside the switching structure without requiring device reconfiguration, achieving system self-healing and hot-swappability, and is suitable for future large-scale CXL topologies.

[0049] e. It has versatility and compatibility, and is suitable for various traffic types: The optimization mechanism is well adapted to different communication modes such as elephant stream and mouse stream, and can be widely used in various CXL application scenarios such as AI training clusters, storage interconnection, and heterogeneous computing;

[0050] f. This invention introduces a feedback optimization adjustment unit, which greatly improves the accuracy and real-time performance of the system. Attached Figure Description

[0051] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0052] Figure 1 A schematic diagram of a CXL switching structure is provided for an exemplary embodiment of the present invention;

[0053] Figure 2 A flowchart of an adaptive data optimization method provided as an exemplary embodiment of the present invention;

[0054] Figure 3 The process executed by the error feedback module provided in an exemplary embodiment of the present invention Figure 1 ;

[0055] Figure 4 The process executed by the error feedback module provided in an exemplary embodiment of the present invention Figure 2 . Detailed Implementation

[0056] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. 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 should fall within the scope of protection of the present invention.

[0057] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, apparatus, product, or device that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or devices.

[0058] In one embodiment of the present invention, a CXL switching structure is provided, such as Figure 1As shown, the CXL switching structure includes a host, a structure manager, upstream switch A, downstream switches B, C, and D, and terminals E, F, and G. In other embodiments, multiple upstream switches or different numbers of downstream switches and terminals can be configured. The structure manager manages the paths from the upstream switches to the downstream switches and then to the terminals. In this embodiment, for a single terminal, there are three paths from upstream switch A to the terminal. That is, in this embodiment, for each terminal, the CXL switching structure of the present invention is configured to dynamically adjust the transmission weights of the three paths from upstream switch A to the terminal based on the predicted congestion probability to prevent congestion.

[0059] In one embodiment, the structure manager includes a traffic monitoring unit, a link prediction unit, an adaptive path control unit, and a feedback optimization unit.

[0060] The Traffic Monitoring Unit (TM) is configured to monitor real-time traffic data at the input and output ports of each switch and the input port of an endpoint. Specifically, in one embodiment, the traffic data includes packet queuing count, throughput, and latency.

[0061] The number of data packets queued can be measured using a data traffic counter, specifically referring to the number of data packets queued at each input port of a switch at a given time point. For example... Figure 1 As shown, in one embodiment, a data traffic counter is used to measure the number of data packets queued at upstream switch A and downstream switches B, C, and D.

[0062] Throughput can be measured using flow sensors, specifically referring to the amount of data successfully transmitted by each switch node per unit time. For example... Figure 1 As shown, in one embodiment, the flow sensor is used to measure the packet throughput of upstream switch A and downstream switches B, C, and D, in bits per second (bps).

[0063] The latency can be measured using flow sensors, specifically referring to the time it takes for data to travel from the source to the destination. For example... Figure 1 In one embodiment, the data flow counter is used to measure the delay time of data from the input port of upstream switch A to the input ports of downstream switches B, C, and D, and the delay time from the input ports of downstream switches B, C, and D to the input ports of terminals E, F, and G, respectively, in milliseconds.

[0064] The Link Prediction Unit (LPU) consists of a data caching module and a link prediction engine. The data caching module is configured to record historical traffic data per unit time, and the link prediction engine is configured to use a decision tree algorithm model to predict the congestion probability of each path based on the historical traffic data.

[0065] The link prediction unit processes the obtained historical traffic data to obtain the packet count, path delay time (avg_delay), and path bandwidth utilization for each path.

[0066] In one embodiment, the path packet queuing number is the average packet queuing number of each link in the path. For example... Figure 1 As shown, in one embodiment, the path from upstream switch A to downstream switch B and then to terminal E is called path ABE, which consists of link AB and link BE. The number of data packets queued on path ABE = (number of data packets queued on link AB + number of data packets queued on link BE) / 2. In other embodiments, different weights can be assigned to each link according to the actual situation.

[0067] In one embodiment, the path delay time is the average delay time of each link in the path. In one embodiment, the path ABE delay time = (link AB delay time + link BE delay time) / 2. In other embodiments, different weights can be assigned to each link according to the actual situation.

[0068] In one embodiment, the path bandwidth utilization rate = 0.4 * AVG + 0.4 * MAX + 0.2 * S 2 Where AVG is the average bandwidth utilization of each link in the path, MAX is the maximum bandwidth utilization of each link in the path, and S 2 Let V be the variance of bandwidth utilization for each link in the path. Link bandwidth utilization = link throughput / link bandwidth. In one embodiment, path ABE bandwidth utilization = 0.4 * AVG + 0.4 * MAX + 0.2 * S 2 Where AVG = (Link AB bandwidth utilization + Link BE bandwidth utilization) / 2, MAX is the larger of the link AB bandwidth utilization and the link BE bandwidth utilization, and S 2 = (Average value - Link AB bandwidth utilization) 2 + (Average value - Link BE bandwidth utilization) 2 ) / 2, where the bandwidth utilization of link AB = the throughput of link AB / the bandwidth of link AB.

[0069] As shown in Table 1, in one embodiment, the link prediction engine is provided with a correspondence between path judgment conditions and probability distributions, which is obtained according to the current decision tree algorithm model. When the decision tree algorithm model changes, it may lead to a change in the probability distribution result.

[0070]

[0071] The Adaptive Path Control Unit (APCU) is configured to assign corresponding data transmission weights to each path based on the congestion probability of each path and the probability threshold mapping strategy.

[0072] In one embodiment, the congestion probability of each path is defined as P, and the congestion probability thresholds satisfy 0 < P1 < P2 < … < PN-1 < PN < 1, where PN represents the Nth threshold. The probability threshold mapping strategy is as follows: If PN < P < 1, set the data transmission weight of the path to 0; if PN-1 < P ≤ PN, set the data transmission weight of the path to AN; if P1 < P ≤ P2, set the data transmission weight of the path to A2, and 1 > A2 > AN > 0; if 0 < P ≤ P1, mark the path as a path with data transmission weight to be increased, and equally distribute the total scheduling weight released by all paths with reduced data transmission weights to the paths with data transmission weight to be increased.

[0073] Specifically, in one embodiment, after receiving the congestion probability of each path, the Adaptive Path Control Unit adjusts the path weights according to the following probability threshold mapping strategy:

[0074] Divide the congestion probability returned by the link prediction engine into multiple interval levels, corresponding to different scheduling weight adjustment strategies. For example:

[0075] P > 0.9: It is considered that the path is very likely to be congested, and immediately set its path weight to 0 (temporarily disabled);

[0076] 0.7 < P ≤ 0.9: Lower the path weight, multiply the current weight by 0.2 or 0.3;

[0077] 0.4 < P ≤ 0.7: Slightly lower, keep it in an available state, multiply the current weight by 0.4 or 0.5;

[0078] P ≤ 0.4: Mark this path as a path with data transmission weight to be increased;

[0079] The total scheduling weight released by all paths with reduced weights will be equally distributed to the paths with data transmission weight to be increased to ensure that the overall utilization rate of scheduling resources does not decrease.

[0080] After the path weight update is completed, the adaptive path control unit will quickly write the weight results into the forwarding table of the switch through the internal control channel. The switch forwards data according to the adjusted path mapping table, without affecting the existing global routing table structure, thereby avoiding the latency caused by reconstruction and global synchronization.

[0081] The Feedback Optimization Unit (FOU) is configured to update the decision tree algorithm model based on historical traffic data and actual congestion results.

[0082] In one embodiment, the feedback optimization unit includes a timed training module configured to periodically update the decision tree algorithm model based on historical traffic data and actual congestion results.

[0083] In one embodiment, the timed training module is configured to perform the following steps:

[0084] If the current time is greater than or equal to T since the last training of the decision tree algorithm model, then the decision tree algorithm model will be retrained.

[0085] By correlating historical traffic data with actual congestion results, multiple sets of sample data are formed. It is determined whether the number of sample data exceeds the threshold H. If so, the decision tree algorithm model is retrained.

[0086] The retrained decision tree algorithm model replaces the current decision tree algorithm model. In one embodiment, the retrained decision tree algorithm model directly takes over the decision tree algorithm model of the link prediction unit.

[0087] In some embodiments, T can be set to a value greater than zero, such as 15 seconds, 20 seconds, 30 seconds, etc., and the threshold H can be set to a value greater than zero, such as 60, 70, 80, etc. Specifically, in one embodiment, T can be set to 10 seconds and the threshold H can be set to 50, resulting in the extraction of 86 sample data. Table 2 below lists the sample data corresponding to four sets of historical traffic data and actual congestion results:

[0088]

[0089] The decision tree algorithm model is retrained using 86 sets of sample data and then replaced with the current decision tree algorithm model. This drives the link prediction unit to change the congestion probability in real time, adapting to the changes in data traffic patterns over time, and realizing an end-to-end adaptive path scheduling closed loop.

[0090] In one embodiment, the feedback optimization unit further includes an error feedback module, which is configured to compare the actual congestion result with the congestion probability, calculate the congestion probability error, and update the decision tree algorithm model complexity using the congestion probability error, that is, adjust the decision tree algorithm model complexity according to the error rate.

[0091] Specifically, such as Figure 3 As shown, in one embodiment, the error feedback module is configured to perform the following steps:

[0092] Read the complexity of the current decision tree algorithm model;

[0093] Set up an error sliding window queue to record the congestion probability error of each path;

[0094] Compare the congestion probability error of each path with the threshold;

[0095] If the congestion probability error is greater than the threshold I for N consecutive times, and the number of trees in the current decision tree algorithm model is less than the threshold J, then the number of trees in the decision tree algorithm model is increased by one step, where N≥2;

[0096] The number of trees in the decision tree algorithm model changes;

[0097] Changes to the number of trees in a decision tree algorithm model do not take effect immediately on the current decision tree algorithm model, but only when the decision tree algorithm model is retrained in the next scheduled training module.

[0098] The above error feedback module is used to dynamically determine whether the performance of the decision tree algorithm model has deteriorated. If it has, the number of trees in the decision tree algorithm model is increased to enhance the expressive power of the decision tree algorithm model, better fit the changes in link state, improve prediction accuracy, and thus reduce the future error rate, thereby enhancing the adaptability of the decision tree algorithm model to changes in data distribution.

[0099] Specifically, such as Figure 4 As shown, in one embodiment, the error feedback module can also be configured to perform the following steps:

[0100] Read the complexity of the current decision tree algorithm model;

[0101] Set up an error sliding window queue to record the congestion probability error of each path;

[0102] Compare the congestion probability error of each path with the threshold;

[0103] If the congestion probability error is less than the threshold L for N consecutive times, and the number of trees in the current decision tree algorithm model is greater than the threshold M, then reduce the number of trees in the decision tree algorithm model by one step, where N≥2;

[0104] The number of trees in the decision tree algorithm model changes;

[0105] Changes to the number of trees in a decision tree algorithm model do not take effect immediately on the current decision tree algorithm model, but only when the decision tree algorithm model is retrained in the next scheduled training module.

[0106] The error feedback module described above is used to gradually restore the complexity of the decision tree algorithm model to a better range, avoiding overfitting and wasting computational resources. Figure 3 and Figure 4 The error feedback module can be applied within the same feedback optimization unit, allowing the unit to dynamically increase or decrease the complexity of the decision tree algorithm model to better fit the actual situation. Thresholds I and L must be greater than zero and less than 1, and thresholds J and M must be greater than zero. A step size can also be set to a value greater than zero. In one embodiment, the error feedback module can simultaneously execute... Figure 3 He Ru Figure 4 When performing the steps shown, 0 < L < I < 1 must be satisfied.

[0107] like Figure 2 As shown, in one embodiment, the adaptive data optimization method of the present invention includes the following steps:

[0108] Use a traffic monitoring unit to monitor real-time traffic data at the input and output ports of each node;

[0109] The data caching module in the link prediction unit records historical traffic data per unit time.

[0110] The link prediction engine in the link prediction unit is used to predict the congestion probability of each path based on historical traffic data using a decision tree algorithm model.

[0111] The adaptive path control unit assigns corresponding data transmission weights to each path based on the congestion probability of each path and a probability threshold mapping strategy.

[0112] The switch forwards data based on the adjusted path mapping table and adjusts the path mapping table in real time.

[0113] The feedback optimization unit updates the decision tree algorithm model based on historical traffic data and actual congestion results;

[0114] The updated decision number algorithm model is input into the link prediction unit.

[0115] The above description is only a specific embodiment of this application. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of this application, and these improvements and modifications should also be considered within the scope of protection of this application.

Claims

1. A CXL switching structure based on dynamic link prediction, characterized in that, include: A traffic monitoring unit is configured to monitor real-time traffic data at the input and output ports of each node; The link prediction unit includes a data caching module and a link prediction engine. The data caching module is configured to record historical traffic data within a unit of time, and the link prediction engine is configured to use a decision tree algorithm model to predict the congestion probability of each path based on the historical traffic data. An adaptive path control unit is configured to assign corresponding data transmission weights to each path based on a probability threshold mapping strategy according to the congestion probability of each path, and adjust the path mapping table in real time. A feedback optimization unit is configured to update the decision tree algorithm model based on the historical traffic data and the actual congestion results; The traffic data includes packet queuing count, throughput, and latency. The link prediction engine has preset path determination conditions, the parameters of which include path bandwidth utilization, which is calculated as 0.4*AVG + 0.4*MAX + 0.2*S. 2 Where AVG is the average bandwidth utilization of each link in the path, MAX is the maximum bandwidth utilization of each link in the path, and S 2 Let V be the variance of bandwidth utilization for each link in the path. Link bandwidth utilization = link throughput / link bandwidth.

2. The CXL switching structure based on dynamic link prediction according to claim 1, characterized in that: The link prediction engine has preset path judgment conditions. The parameters in the path judgment conditions include the number of data packets queued on the path. The number of data packets queued on the path is the average number of data packets queued on each link in the path.

3. The CXL switching structure based on dynamic link prediction according to claim 1, characterized in that: The link prediction engine has preset path judgment conditions, and the parameters in the path judgment conditions include path delay time, which is the average delay time of each link in the path.

4. The CXL switching structure based on dynamic link prediction according to claim 1, characterized in that: The congestion probability of each path is defined as P. The congestion probability threshold satisfies 0 < P1 < P2 < ... < PN-1 < PN < 1, where PN represents the Nth threshold. The probability threshold mapping strategy is as follows: if PN < P < 1, the data transmission weight of the path is set to 0; if PN-1 < P ≤ PN, the data transmission weight of the path is set to AN; if P1 < P ≤ P2, the data transmission weight of the path is set to A2, and 1 > A2 > AN > 0; if 0 < P ≤ P1, the path is recorded as a path with data transmission weight to be increased, and the total scheduling weight released by all paths with reduced data transmission weight is proportionally allocated to the path with data transmission weight to be increased.

5. The CXL switching structure based on dynamic link prediction according to claim 1, characterized in that: The feedback optimization unit includes a timed training module, which is configured to update the decision tree algorithm model at regular intervals based on the historical traffic data and the actual congestion results.

6. The CXL switching structure based on dynamic link prediction according to claim 5, characterized in that: The timed training module is configured to perform the following steps: If the current time is greater than or equal to T since the last training time of the decision tree algorithm model, then the decision tree algorithm model is retrained. The historical traffic data is matched with the actual congestion results to form multiple sets of sample data. It is determined whether the number of sample data is greater than the threshold H. If so, the decision tree algorithm model is retrained. Replace the current decision tree algorithm model with the retrained decision tree algorithm model.

7. The CXL switching structure based on dynamic link prediction according to claim 6, characterized in that: The decision tree algorithm model obtained after retraining directly takes over the decision tree algorithm model of the link prediction unit.

8. The CXL switching structure based on dynamic link prediction according to claim 5, characterized in that: The feedback optimization unit further includes an error feedback module, which is configured to compare the actual congestion result with the congestion probability and calculate the congestion probability error, and use the congestion probability error to increase or decrease the complexity of the decision tree algorithm model.

9. The CXL switching structure based on dynamic link prediction according to claim 8, characterized in that: The error feedback module is configured to perform the following steps: Read the complexity of the current decision tree algorithm model; Set up an error sliding window queue to record the congestion probability error for each path; If the congestion probability error is greater than the threshold I for N consecutive times, and the number of trees in the current decision tree algorithm model is less than the threshold J, then the number of trees in the decision tree algorithm model is increased by one step, where N≥2.

10. The CXL switching structure based on dynamic link prediction according to claim 8, characterized in that: The error feedback module is configured to perform the following steps: Read the complexity of the current decision tree algorithm model; Set up an error sliding window queue to record the congestion probability error for each path; If the congestion probability error is less than the threshold L for N consecutive times, and the number of trees in the current decision tree algorithm model is greater than the threshold M, then the number of trees in the decision tree algorithm model is reduced by one step, where N≥2.

11. The CXL switching structure based on dynamic link prediction according to claim 9 or 10, characterized in that: Changes to the number of trees in a decision tree algorithm model do not take effect immediately on the current decision tree algorithm model, but only when the timed training module retrains the decision tree algorithm model next time.

12. An adaptive data optimization method, applicable to the CXL switching structure based on dynamic link prediction as described in any one of claims 1-11, specifically comprising the following steps: Use a traffic monitoring unit to monitor real-time traffic data at the input and output ports of each node; The data caching module in the link prediction unit records historical traffic data per unit time. Using the link prediction engine in the link prediction unit, the congestion probability of each path is predicted based on the historical traffic data using the decision tree algorithm model. The adaptive path control unit assigns corresponding data transmission weights to each path based on the congestion probability of each path and a probability threshold mapping strategy, and adjusts the path mapping table in real time. The switch forwards data according to the adjusted path mapping table; The feedback optimization unit updates the decision tree algorithm model based on the historical traffic data and the actual congestion results. The updated decision number algorithm model is input into the link prediction unit.

Images