AOC active optical cable transmission performance optimization method for supercomputer high-speed interconnection

Abstract

The application discloses an AOC active optical cable transmission performance optimization method for high-speed interconnection of a wisdom calculation center, relates to the technical field of high-speed interconnection communication optimization, and comprises the following steps: collecting training operation data of a GPU cluster of the wisdom calculation center, identifying a current GPU training stage according to the training operation data, determining communication requirements of a next scheduling period based on the GPU training stage, and predicting a micro-burst traffic state of an AOC link according to the communication requirements; determining an AOC link to be evaluated according to the micro-burst traffic state, obtaining link state data of the AOC link to be evaluated, calculating a sub-health score of the AOC link to be evaluated based on the link state data, and generating a link risk topology in combination with the micro-burst traffic state. Through the fusion of GPU training stage sensing, AOC link risk topology construction and improved sea squirt swarm intelligent optimization mechanism, the transmission stability and dynamic response capability of the AOC link in the high-speed interconnection scene of the wisdom calculation center are improved.

Description

Technical Field

[0001] This invention relates to the field of high-speed interconnect communication optimization technology, and in particular to a method for optimizing the transmission performance of AOC active optical cables used for high-speed interconnection in intelligent computing centers. Background Technology

[0002] As large-scale AI training tasks continue to evolve towards ultra-large parameter models and multi-node parallel computing, intelligent computing centers are gradually forming a high-density, high-speed interconnect architecture centered on GPU clusters. To meet the requirements of low latency, high bandwidth, and high stability communication links for gradient synchronization, parameter exchange, and model partitioning computation during distributed training, AOC (Alternating Current Components) has been widely used in high-speed interconnect scenarios between intelligent computing center racks, switching nodes, and GPU servers due to its high transmission rate, low electromagnetic interference sensitivity, and good deployment flexibility. In existing technologies, optimizing AOC link transmission performance typically employs a link monitoring mechanism based on fixed thresholds, combined with forward error correction level adjustment, buffer parameter configuration, transmission window control, and digital signal processing equalization parameter adjustment to optimize and control link bit error rate, transmission latency, and throughput performance.

[0003] However, in current real-world scenarios for large-scale GPU collaborative training, the training tasks exhibit significant phased characteristics, with marked differences in data exchange patterns across different training phases. This can easily trigger sudden bursts of high-concurrency communication, resulting in micro-burst traffic characterized by temporality, dynamism, and non-stationarity. Existing technologies typically focus on parameter adjustment based on the static operating state of the link or historical average load, lacking the ability to jointly model the correlation between the evolutionary patterns of the GPU training phase and the micro-burst traffic of the AOC link. Consequently, it is difficult to promptly identify the sub-health evolution trend of the link and achieve targeted predictive optimization. In this situation, when localized link performance degradation overlaps with burst traffic, it can easily lead to a lag in link transmission parameter adjustments, affecting the dynamic adaptability of link load allocation. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides a method for optimizing the transmission performance of AOC active optical cables for high-speed interconnection in intelligent computing centers, solving the problems of lag in micro-burst traffic perception and insufficient multi-parameter collaborative adaptive optimization capabilities of AOC active optical cable links in high-speed interconnection scenarios of intelligent computing centers.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution: In a first aspect, the present invention provides a method for optimizing the transmission performance of active optical cables (AOCs) used for high-speed interconnection in intelligent computing centers, comprising, Collect training and running data of the GPU cluster in the intelligent computing center, identify the current GPU training stage based on the training and running data, determine the communication requirements for the next scheduling cycle based on the GPU training stage, and predict the micro-burst traffic status of the AOC link based on the communication requirements. The AOC link to be evaluated is determined based on the micro-burst traffic status, and the link status data of the AOC link to be evaluated is obtained. The sub-health score of the AOC link to be evaluated is calculated based on the link status data, and the link risk topology is generated by combining the micro-burst traffic status. Based on the link risk topology, an optimization objective function is constructed with FEC level, buffer depth, sending window, DSP equalization parameters and link traffic carrying ratio as optimization variables. An improved tunic group algorithm with fused sine and cosine update mechanism is used to iteratively optimize the optimization objective function and generate the optimal transmission strategy. According to the optimal transmission strategy, a gradual traffic migration is performed on the sub-healthy AOC link, and transmission parameters are sent to the corresponding AOC link.

[0007] As a preferred embodiment of the method for optimizing the transmission performance of AOC active optical cables for high-speed interconnection in intelligent computing centers according to the present invention, the step of collecting training and running data of the GPU cluster in the intelligent computing center and identifying the current GPU training stage based on the training and running data includes the following steps: Collect running status data, communication behavior data, and task scheduling data of each GPU node in the GPU cluster of the intelligent computing center, and preprocess the collected data to generate a training running dataset; Based on the training dataset, we extract GPU computing load variation features, inter-node communication frequency features, gradient synchronization cycle features, and data transmission scale features to generate a training behavior feature parameter set. Based on the training behavior feature parameter set and combined with the preset training stage identification rules, the running status of the current training task is classified into stages to obtain the current GPU training stage.

[0008] As a preferred embodiment of the method for optimizing the transmission performance of AOC active optical cables for high-speed interconnection of intelligent computing centers as described in this invention, the step of determining the communication requirements for the next scheduling cycle based on the GPU training phase and predicting the micro-burst traffic status of the AOC link according to the communication requirements includes the following steps: Based on the training behavior characteristics corresponding to the current GPU training phase, determine the data interaction method and communication load change trend of the GPU cluster in the next scheduling cycle; Based on the data interaction method and communication load change trend, the communication frequency, data transmission scale, synchronization cycle and bandwidth usage of nodes in the GPU cluster are statistically analyzed to generate the communication requirements for the next scheduling cycle. Based on communication requirements, and combined with historical AOC link traffic data, switch port queue length, and RDMA message arrival interval, the traffic change trend of the AOC link is predicted. Based on the predicted traffic change trend, the peak value, duration and probability of occurrence of micro-burst traffic on the AOC link are determined, and the corresponding micro-burst traffic status is generated.

[0009] As a preferred embodiment of the method for optimizing the transmission performance of active optical cables (AOCs) for high-speed interconnection in intelligent computing centers according to the present invention, the step of determining the AOC link to be evaluated based on the micro-burst traffic status and obtaining the link status data of the AOC link to be evaluated includes the following steps: Based on the traffic peak, duration and probability of occurrence in the micro-burst traffic state, AOC links whose load change exceeds the preset threshold in the next scheduling cycle are identified as AOC links to be evaluated. Obtain the link operation information of the AOC link to be evaluated, including the link's real-time throughput, port utilization, and queue buffer usage; Based on the link operation information, physical layer status detection is performed on the AOC link to be evaluated to obtain the corresponding link status data.

[0010] Based on the link operation information of the AOC link to be evaluated, determine the high-load operation period and abnormal traffic fluctuation range of the corresponding AOC active optical cable, and trigger the link physical layer status detection within the corresponding time window. The physical layer operating parameters of the AOC link to be evaluated are collected, including bit error rate, received optical power, transmitted optical power, optical power margin, link temperature and signal jitter information, and link status data is generated.

[0011] As a preferred embodiment of the method for optimizing the transmission performance of active optical cables (AOCs) for high-speed interconnection in intelligent computing centers as described in this invention, the step of calculating the sub-health score of the AOC link to be evaluated based on link status data and generating a link risk topology by combining micro-burst traffic status includes the following steps: The link status data of the AOC link to be evaluated is normalized, and the bit error rate characteristics, optical power margin characteristics, temperature change characteristics and signal jitter characteristics reflecting the link operation stability are extracted. Based on the extracted features reflecting the stability of link operation, the link health status index of each AOC link to be evaluated is calculated, and the corresponding sub-health score is generated according to the link health status index.

[0012] Based on the preset link operation impact weights corresponding to each feature, the bit error rate feature, optical power margin feature, temperature change feature and signal jitter feature are weighted and fused to generate a link health status index. Based on historical AOC link normal operation status data, establish health threshold ranges and abnormal threshold ranges for link health status indicators. The link health status indicators corresponding to each AOC link to be evaluated are compared with the health threshold range and the abnormal threshold range to determine the link health status level of the AOC link. A sub-health score is generated based on the link health status level and the degree of deviation between the link health status indicators and the abnormal threshold range.

[0013] Based on the peak traffic volume, duration, and probability of occurrence in the micro-burst traffic state, calculate the link load risk level of each AOC link to be evaluated in the next scheduling cycle.

[0014] Based on the relationship between the peak micro-burst traffic and the current bandwidth utilization of the corresponding AOC link, the instantaneous traffic impact of each AOC link to be evaluated is calculated. Based on the duration of micro-bursts and changes in queue buffer usage, calculate the degree of sustained load growth for each AOC link to be evaluated; Based on the probability of micro-bursts, the risk trend of each AOC link to be evaluated becoming high-load in the next scheduling cycle is determined. By combining the instantaneous traffic surge, the continuous load growth, and the risk trend of high load conditions, the link load risk level corresponding to each AOC link to be evaluated is generated.

[0015] By combining the sub-health score and link load risk level of each AOC link to be evaluated, a corresponding comprehensive link risk value is generated; Using server and switching devices in the GPU cluster as network nodes, AOC links as inter-node connections, and the comprehensive risk value of each AOC link as the edge weight parameter, a corresponding link risk topology is generated.

[0016] As a preferred embodiment of the method for optimizing the transmission performance of AOC active optical cables for high-speed interconnection of intelligent computing centers as described in this invention, the step of constructing an optimization objective function based on link risk topology with FEC level, buffer depth, transmission window, DSP equalization parameters, and link traffic carrying ratio as optimization variables includes the following steps: Based on the comprehensive risk value and link connection relationship of each AOC link in the link risk topology, combined with the current bandwidth usage, queue buffer usage and link status data, the available carrying capacity of each AOC link is determined. Based on the comprehensive risk value and available carrying capacity of each AOC link, the variables to be optimized for the corresponding AOC link are determined, including FEC level, buffer depth, sending window, DSP equalization parameters and link traffic carrying ratio. Based on the variables to be optimized, corresponding constraints are established for bit error rate, transmission delay, link load, and traffic balancing.

[0017] Based on FEC level, DSP equalization parameters and link status data, establish bit error rate constraints to limit the range of bit error rate variation in AOC links; Based on the buffer depth, transmission window, and current queue buffer occupancy of the link, establish transmission delay constraints to limit the AOC link transmission waiting time and data forwarding delay; Based on the link traffic carrying ratio, the current bandwidth usage of the link, and the available carrying capacity of each AOC link, establish a link load constraint to limit the load growth of a single AOC link. Based on the differences in link traffic carrying ratios and overall link risk values ​​among the AOC links, a traffic balancing constraint is established to balance the traffic allocation status of each AOC link within the GPU cluster.

[0018] Based on bit error rate constraints, transmission delay constraints, link load constraints, and traffic balancing constraints, an optimization objective function is constructed to evaluate the transmission performance of the AOC link.

[0019] Obtain the constraint parameters corresponding to bit error rate constraint, transmission delay constraint, link load constraint and traffic balancing constraint, and determine the degree of impact of each constraint parameter on AOC link transmission performance; Based on the degree of impact of each constraint parameter on the transmission performance of the AOC link, evaluation items for bit error rate, latency, link load, and traffic balance are established respectively. Based on the bit error rate evaluation item, latency evaluation item, link load evaluation item, and traffic balance evaluation item, and combined with the communication requirements of the GPU cluster in the next scheduling cycle and the available carrying capacity of each AOC link, the weights of each evaluation item are assigned to generate an optimization objective function.

[0020] As a preferred embodiment of the method for optimizing the transmission performance of AOC active optical cables for high-speed interconnection of intelligent computing centers as described in this invention, the improved tunic swarm algorithm using a fused sine and cosine update mechanism to iteratively optimize the objective function and generate the optimal transmission strategy includes the following steps: An initial set of candidate solutions is generated based on each variable to be optimized. Each candidate solution corresponds to a search individual in the tunic group algorithm and a set of AOC link transmission parameter configuration schemes. Based on the optimization objective function, the objective function value corresponding to each candidate solution is calculated, and the current optimal candidate solution is determined according to the objective function value. The search individual corresponding to the optimal candidate solution is then determined as the leader individual. Based on the comprehensive risk value, micro-burst traffic status, and available carrying capacity of each AOC link, a risk weight factor is constructed, and the search step size, search range, and position update speed of each candidate solution in the tunic swarm algorithm are dynamically adjusted using the risk weight factor.

[0021] The comprehensive risk value, peak micro-burst traffic, duration of micro-burst, probability of micro-burst occurrence, and available carrying capacity of each AOC link are obtained as link risk status parameters. Based on the link risk status parameters, the risk weight factor corresponding to each AOC link is calculated. The higher the comprehensive risk value of the link, the more severe the micro-burst traffic status, and the lower the available carrying capacity of the link, the larger the corresponding risk weight factor. Based on the risk weight factor corresponding to each AOC link, the search step size of each candidate solution in the tunic group algorithm is dynamically adjusted so that the search step size and search range of candidate solutions with risk weight factors greater than the preset threshold are expanded, thereby increasing the search priority of high-risk AOC link areas. Based on the risk weight factor, the position update speed of each candidate solution is adaptively adjusted so that the candidate solutions corresponding to high-risk AOC links converge in the direction of reducing link load and reducing bit error rate, while the candidate solutions corresponding to low-risk AOC links converge in the direction of increasing link throughput and traffic balance. During the candidate solution search and update process, the risk weight factor is dynamically adjusted based on the changes in the risk status parameters of each AOC link.

[0022] The leader individual in the tunic group algorithm after risk weight adjustment is used to perform a global search for candidate solutions, and the search direction of each candidate solution is updated according to the current best candidate solution.

[0023] The current position of the navigator is used as the global guiding position in the current iteration round; Based on the global guidance position of the navigator, the positional deviation between the search individual and the navigator corresponding to each candidate solution is calculated, and the corresponding search direction adjustment parameters are generated. The parameters are adjusted according to the search direction, and the position update direction of each candidate solution corresponding to the search individual is adjusted so that each search individual gradually moves closer to the search region corresponding to the current best candidate solution.

[0024] During the candidate solution update process, a sine-cosine update mechanism is introduced, which uses sine and cosine functions to dynamically perturb and update the position change direction and magnitude of the candidate solution.

[0025] Determine the current position, search direction, and position update step size of each candidate solution's corresponding search individual in the current iteration round, and determine the position update parameters for each search individual; Based on the positional deviation between the current position and the current optimal candidate solution, corresponding sine perturbation terms and cosine perturbation terms are constructed respectively. The sine perturbation term is used to adjust the direction of positional change of the search individual, and the cosine perturbation term is used to adjust the magnitude of positional change of the search individual. Based on the sine and cosine perturbation terms, the position update parameters of each search individual are dynamically perturbed and adjusted so that each search individual forms a nonlinear position change trajectory near its current position. For search individuals located in high-risk AOC link search areas, increase the positional change amplitude corresponding to the sine and cosine perturbation terms to improve the jump search capability in high-risk areas; For search individuals located in the low-risk AOC link search region, reduce the positional change amplitude of the sine and cosine perturbation terms to improve the convergence stability in the local stable region. Based on the position update parameters adjusted by dynamic perturbation, update the position of each candidate solution corresponding to the search individual, and generate the corresponding candidate solution update results; After the candidate solution position is updated, the updated candidate solution is subjected to boundary range detection. When the candidate solution exceeds the preset parameter boundary range, the candidate solution that exceeds the boundary range is subjected to boundary correction processing.

[0026] After each iteration, the objective function value corresponding to the updated candidate solution is recalculated, and the current best candidate solution is updated based on the objective function value, and the current leader individual is determined. When the preset number of iterations is reached, the optimal candidate solution is obtained, and the AOC link transmission parameter configuration scheme corresponding to the optimal candidate solution is determined as the optimal transmission strategy.

[0027] As a preferred embodiment of the method for optimizing the transmission performance of active optical cables for high-speed interconnection of intelligent computing centers according to the present invention, the step of performing gradual traffic migration on sub-healthy AOC links according to the optimal transmission strategy and sending transmission parameters to the corresponding AOC links includes the following steps: Based on the link traffic carrying ratio in the optimal transmission strategy, gradually reduce the traffic carrying ratio of the sub-healthy AOC links and gradually increase the traffic carrying ratio of the corresponding target AOC links. Based on the changes in the link traffic carrying ratio, the business traffic in the sub-healthy AOC link is migrated in stages, and the link traffic distribution after migration is generated.

[0028] Obtain the current service traffic information carried by the sub-health AOC link, and determine the migration ratio and migration stage of the corresponding service traffic based on the changes in the link traffic carrying ratio in the optimal transmission strategy. Based on the service priority, data transmission latency requirements, and bandwidth usage of the service traffic, the service traffic in the sub-healthy AOC link is classified to generate a set of service traffic to be migrated. Based on the link traffic carrying ratio and available carrying capacity of each target AOC link, the traffic set of services to be migrated will be gradually allocated to the corresponding target AOC links according to the preset migration stages. In each migration phase, the proportion of business traffic in the sub-healthy AOC link is gradually reduced, while the proportion of business traffic in the target AOC link is increased simultaneously. After completing all migration phases, the traffic usage of each AOC link is statistically analyzed, and the post-migration link traffic distribution is generated.

[0029] During the traffic migration process, the bit error rate, transmission latency, queue buffer usage, and link throughput of the AOC link are monitored in real time, and the traffic migration rate is dynamically adjusted based on the monitoring results.

[0030] Acquire real-time operating status data of each AOC link during traffic migration, and extract the corresponding bit error rate, transmission latency, queue buffer usage and link throughput parameters as link transmission status evaluation parameters; Each parameter in the link transmission status assessment parameters is compared with its corresponding preset threshold to generate the corresponding link status judgment result. When the link status determination result indicates that the corresponding AOC link is in a stable transmission state, the traffic migration rate of the current stage is increased, and the proportion of service traffic migrated to the target AOC link is increased. When the link status determination result indicates that the corresponding AOC link has an increased bit error rate, increased transmission latency, increased queue buffer occupancy rate, or decreased link throughput, the traffic migration rate at the current stage is reduced, and the proportion of service traffic migrated to the target AOC link is reduced. Based on the adjusted traffic migration rate, the service traffic of each AOC link is redistributed, and data collection and migration rate adjustment are performed for the next monitoring cycle.

[0031] When the traffic migration meets the stable transmission conditions, the transmission parameters are sent to the corresponding AOC link according to the FEC level, buffer depth, sending window and DSP equalization parameters in the optimal transmission strategy. After the transmission parameters are issued, the transmission performance feedback data of the corresponding AOC link is obtained and uploaded to the database for storage.

[0032] In a second aspect, the present invention provides a computer device, including a memory and a processor, wherein the memory stores a computer program, wherein: when the computer program is executed by the processor, it implements any step of the method for optimizing the transmission performance of AOC active optical cable for high-speed interconnection of intelligent computing centers as described in the first aspect of the present invention.

[0033] Thirdly, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein: when the computer program is executed by a processor, it implements any step of the method for optimizing the transmission performance of AOC active optical cable for high-speed interconnection of intelligent computing centers as described in the first aspect of the present invention.

[0034] The beneficial effects of this invention are as follows: By collecting and analyzing the training and operation data of the GPU cluster in the intelligent computing center and identifying the current training stage, and combining the communication behavior characteristics corresponding to different training stages to predict the micro-burst traffic status of the AOC link, a forward-looking perception of the load change trend of the high-speed interconnect link is achieved. This transforms the link regulation process from a traditional post-event response mode to a proactive prediction mode oriented towards future scheduling cycles, providing more timely and accurate dynamic basis for link status assessment and resource scheduling. By combining link status data to calculate sub-health scores and constructing link risk topology, a visualized correlation expression between link health status and traffic risk distribution is achieved, expanding link risk identification from single-point static judgment to global correlation dynamic analysis. By constructing a multivariate optimization model that includes FEC level, buffer depth, sending window, DSP equalization parameters, and link traffic carrying ratio, the invention achieves the following benefits. The objective function is evaluated using an improved tunic group algorithm that integrates sine and cosine update mechanisms for iterative optimization. This algorithm enables global search and dynamic perturbation-enhanced search of complex coupled parameter spaces, giving the parameter optimization process both global exploration and local convergence capabilities. This effectively improves the adaptability and optimization efficiency of multi-dimensional transmission parameter collaborative adjustment, generating optimal transmission strategies that better suit the current link operating state. By performing gradual traffic migration on sub-healthy AOC links based on the optimal transmission strategy and simultaneously distributing optimized transmission parameters, the algorithm achieves smooth reconfiguration of link load and online dynamic adjustment of transmission performance. This avoids sudden shocks and service jitter during link switching, enabling the high-speed interconnection network of the intelligent computing center to maintain stable data transmission capabilities even in sub-healthy link states. This improves AOC link transmission stability, enhances communication continuity, and increases overall resource utilization efficiency. Attached Figure Description

[0035] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0036] Figure 1 A flowchart illustrating the method for optimizing the transmission performance of AOC active optical cables for high-speed interconnection of intelligent computing centers.

[0037] Figure 2 A flowchart for generating micro-burst flow states.

[0038] Figure 3 A flowchart for generating link risk topology.

[0039] Figure 4 A flowchart for generating the optimal transmission strategy. Detailed Implementation

[0040] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0041] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0042] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0043] Reference Figures 1-4 This is one embodiment of the present invention, which provides a method for optimizing the transmission performance of active optical cables (AOCs) for high-speed interconnection in intelligent computing centers, comprising the following steps: The system collects training and operation data from the GPU cluster in the intelligent computing center, identifies the current GPU training stage based on the training and operation data, determines the communication requirements for the next scheduling cycle based on the GPU training stage, and predicts the micro-burst traffic status of the AOC link based on the communication requirements.

[0044] Specifically, the system collects the running status data, communication behavior data, and task scheduling data of each GPU node in the GPU cluster of the intelligent computing center, and preprocesses the collected data to generate a training dataset. Based on the training dataset, we extract GPU computing load variation features, inter-node communication frequency features, gradient synchronization cycle features, and data transmission scale features to generate a training behavior feature parameter set. Based on the training behavior feature parameter set and combined with the preset training stage identification rules, the running status of the current training task is classified into stages to obtain the current GPU training stage.

[0045] By constructing a training dataset and extracting features such as GPU computational load variation, inter-node communication frequency, gradient synchronization cycle, and data transmission scale from the dataset, a characteristic representation of the changing patterns of communication and computational behavior in GPU training tasks is achieved. This enables effective differentiation of resource occupancy states and communication modes in different training stages, providing support for communication demand prediction and link load analysis. By combining preset training stage identification rules to classify the current training task's running state into stages, accurate identification of the dynamic stage evolution process of the GPU training workflow is achieved. This facilitates the prediction of micro-burst traffic on AOC links and the early optimization of link resource scheduling strategies, improving the high-speed interconnection network of the intelligent computing center's ability to perceive changes in training tasks and enhancing the adaptability of communication resource scheduling.

[0046] Furthermore, based on the training behavior characteristics corresponding to the current GPU training phase, the data interaction method and communication load change trend of the GPU cluster in the next scheduling cycle are determined; Based on the data interaction method and communication load change trend, the communication frequency, data transmission scale, synchronization cycle and bandwidth usage of nodes in the GPU cluster are statistically analyzed to generate the communication requirements for the next scheduling cycle. Based on communication requirements, and combined with historical AOC link traffic data, switch port queue length, and RDMA message arrival interval, the traffic change trend of the AOC link is predicted. Based on the predicted traffic change trend, the peak value, duration and probability of occurrence of micro-burst traffic on the AOC link are determined, and the corresponding micro-burst traffic status is generated.

[0047] By determining the data interaction methods and communication load change trends of the GPU cluster in the next scheduling cycle based on the training behavior characteristics corresponding to the current GPU training phase, dynamic perception of the evolution law of training task communication behavior is achieved, providing a forward-looking basis for communication resource scheduling and link load prediction. By statistically analyzing node communication frequency, data transmission scale, synchronization cycle, and bandwidth usage based on data interaction methods and communication load change trends, refined quantitative analysis of GPU cluster communication needs is achieved, enabling accurate characterization of link resource occupancy status for different training tasks, improving the pertinence of communication demand prediction and the rationality of resource allocation. By combining historical AOC link traffic data, switch port queue length, and RDMA packet arrival interval to predict AOC link traffic change trends, early identification of instantaneous load fluctuations and potential congestion trends in links is achieved. By determining the peak value, duration, and probability of occurrence of micro-burst traffic in AOC links and generating corresponding micro-burst traffic states, dynamic characterization and risk quantification analysis of micro-burst traffic characteristics are achieved, enabling the link control process to more accurately adapt to sudden high-concurrency communication scenarios, improving the dynamic response capability and link load prediction capability of the high-speed interconnection network of the intelligent computing center to complex training communication scenarios.

[0048] The AOC link to be evaluated is determined based on the micro-burst traffic status, and the link status data of the AOC link to be evaluated is obtained. The sub-health score of the AOC link to be evaluated is calculated based on the link status data, and the link risk topology is generated by combining the micro-burst traffic status.

[0049] Specifically, based on the traffic peak, duration, and probability of occurrence in the micro-burst traffic state, AOC links whose load change exceeds a preset threshold in the next scheduling cycle are identified as AOC links to be evaluated. Obtain the link operation information of the AOC link to be evaluated, including the link's real-time throughput, port utilization, and queue buffer usage; Based on the link operation information, physical layer status detection is performed on the AOC link to be evaluated to obtain the corresponding link status data.

[0050] Based on the link operation information of the AOC link to be evaluated, determine the high-load operation period and abnormal traffic fluctuation range of the corresponding AOC active optical cable, and trigger the link physical layer status detection within the corresponding time window. The physical layer operating parameters of the AOC link to be evaluated are collected, including bit error rate, received optical power, transmitted optical power, optical power margin, link temperature and signal jitter information, and link status data is generated.

[0051] By identifying AOC links whose load changes exceed a preset threshold in the next scheduling cycle based on the peak traffic, duration, and probability of occurrence in micro-burst traffic states, the system achieves early screening and key location of potentially high-risk links, improving the pertinence of link risk analysis and the efficiency of control. Through dynamic perception of link operating status and resource occupancy status, combined with link operating information, the system determines the high-load operating periods and abnormal traffic fluctuation ranges of AOC active optical cables, and triggers physical layer status detection within the corresponding time window. This achieves a dynamic correlation between the link status detection process and the service load change process, making physical layer detection more closely reflect the actual link operating status and reducing the risk of misjudgment caused by detection during non-critical periods. By collecting physical layer operating parameters such as bit error rate, received optical power, transmitted optical power, optical power margin, link temperature, and signal jitter information, the system achieves multi-dimensional status characterization of the physical layer operating quality of AOC links. This allows for early identification and quantification of potential performance degradation trends, improving the AOC link's ability to identify sub-health states and perceive link anomalies.

[0052] Furthermore, the link status data of the AOC link to be evaluated is normalized, and features reflecting the link's operational stability, such as bit error rate, optical power margin, temperature change, and signal jitter, are extracted. Based on the extracted features reflecting the stability of link operation, the link health status index of each AOC link to be evaluated is calculated, and the corresponding sub-health score is generated according to the link health status index.

[0053] Based on the preset link operation impact weights corresponding to each feature, the bit error rate feature, optical power margin feature, temperature change feature and signal jitter feature are weighted and fused to generate a link health status index. Based on historical AOC link normal operation status data, establish health threshold ranges and abnormal threshold ranges for link health status indicators. The link health status indicators corresponding to each AOC link to be evaluated are compared with the health threshold range and the abnormal threshold range to determine the link health status level of the AOC link. A sub-health score is generated based on the link health status level and the degree of deviation between the link health status indicators and the abnormal threshold range.

[0054] Based on the peak traffic volume, duration, and probability of occurrence in the micro-burst traffic state, calculate the link load risk level of each AOC link to be evaluated in the next scheduling cycle.

[0055] Based on the relationship between the peak micro-burst traffic and the current bandwidth utilization of the corresponding AOC link, the instantaneous traffic impact of each AOC link to be evaluated is calculated. Based on the duration of micro-bursts and changes in queue buffer usage, calculate the degree of sustained load growth for each AOC link to be evaluated; Based on the probability of micro-bursts, the risk trend of each AOC link to be evaluated becoming high-load in the next scheduling cycle is determined. By combining the instantaneous traffic surge, the continuous load growth, and the risk trend of high load conditions, the link load risk level corresponding to each AOC link to be evaluated is generated.

[0056] By combining the sub-health score and link load risk level of each AOC link to be evaluated, a corresponding comprehensive link risk value is generated; Using server and switching devices in the GPU cluster as network nodes, AOC links as inter-node connections, and the comprehensive risk value of each AOC link as the edge weight parameter, a corresponding link risk topology is generated.

[0057] By normalizing the link status data of the AOC links to be evaluated and extracting stability features such as bit error rate, optical power margin, temperature change, and signal jitter, a unified characterization of physical layer status parameters of different types is achieved, making link health analysis more consistent and comparable. By weighting and fusing the link operation impact weights of each feature, a link health status index is generated. Combined with historical normal operation data, health threshold ranges and abnormal threshold ranges are established, enabling a graded judgment of AOC link health levels. This allows for the quantitative identification of sub-healthy links, facilitating the early detection of potentially degraded links. By generating a sub-health score based on the link health status level and its deviation from the abnormal threshold range, a refined characterization of the degree of link physical performance degradation is achieved. By combining the peak, duration, and probability of micro-burst traffic, the instantaneous traffic impact level, the degree of continuous load growth, and the high load risk trend are calculated, enabling dynamic assessment of link load risk for the next scheduling cycle. This allows risk judgment to not only rely on the current link status but also reflect the impact of future burst traffic. By integrating sub-health scores with link load risk levels to generate a comprehensive link risk value, and using this risk value as an edge weight parameter to construct a link risk topology, a unified mapping of AOC link physical health status, traffic impact risk, and GPU cluster network connectivity is achieved. This provides a global and quantifiable risk basis for subsequent transmission parameter optimization, traffic migration, and link scheduling, improving the intelligent computing center's high-speed interconnect network's ability to identify sub-healthy links, predict risks, and enhance overall transmission stability.

[0058] Based on the link risk topology, an optimization objective function is constructed with FEC level, buffer depth, sending window, DSP equalization parameters and link traffic carrying ratio as optimization variables. An improved tunic group algorithm with fused sine and cosine update mechanism is used to iteratively optimize the optimization objective function and generate the optimal transmission strategy.

[0059] Specifically, based on the comprehensive risk value and link connectivity of each AOC link in the link risk topology, combined with the current bandwidth usage, queue buffer usage, and link status data, the available carrying capacity of each AOC link is determined using the following formula: ; in, For AOC link index, For the first The available carrying capacity of each AOC link, For the first The maximum carrying capacity of each AOC link For the first The current bandwidth utilization of the AOC links. For the first Queue cache utilization rate of each AOC link For the first The overall risk value of each AOC link.

[0060] Based on the comprehensive risk value and available carrying capacity of each AOC link, the variables to be optimized for the corresponding AOC link are determined, including FEC level, buffer depth, sending window, DSP equalization parameters and link traffic carrying ratio. Based on the variables to be optimized, corresponding constraints are established for bit error rate, transmission delay, link load, and traffic balancing.

[0061] Based on FEC level, DSP equalization parameters and link status data, establish bit error rate constraints to limit the range of bit error rate variation in AOC links; Based on the buffer depth, transmission window, and current queue buffer occupancy of the link, establish transmission delay constraints to limit the AOC link transmission waiting time and data forwarding delay; Based on the link traffic carrying ratio, the current bandwidth usage of the link, and the available carrying capacity of each AOC link, establish a link load constraint to limit the load growth of a single AOC link. Based on the differences in link traffic carrying ratios and overall link risk values ​​among the AOC links, a traffic balancing constraint is established to balance the traffic allocation status of each AOC link within the GPU cluster.

[0062] Based on bit error rate constraints, transmission delay constraints, link load constraints, and traffic balancing constraints, an optimization objective function is constructed to evaluate the transmission performance of the AOC link.

[0063] Obtain the constraint parameters corresponding to bit error rate constraint, transmission delay constraint, link load constraint and traffic balancing constraint, and determine the degree of impact of each constraint parameter on AOC link transmission performance; Based on the degree of impact of each constraint parameter on the transmission performance of the AOC link, evaluation items for bit error rate, latency, link load, and traffic balance are established respectively. Based on the bit error rate evaluation item, latency evaluation item, link load evaluation item, and traffic balancing evaluation item, and combined with the communication requirements of the GPU cluster in the next scheduling cycle and the available carrying capacity of each AOC link, the weights of each evaluation item are assigned to generate the optimization objective function, the formula of which is: ; in, To optimize the objective function, For the bit error rate evaluation item, For time delay evaluation items, For link load evaluation items, For flow balance evaluation items, , , , The weights are assigned to the bit error rate evaluation item, latency evaluation item, link load evaluation item, and traffic balance evaluation item based on communication requirements and the available carrying capacity of the AOC link.

[0064] By determining the available carrying capacity of each AOC link based on the comprehensive risk value of the link, link connectivity, current bandwidth usage, queue buffer usage, and link status data, a joint assessment of the actual carrying capacity and risk status of the link is achieved. This allows subsequent parameter optimization to move beyond relying solely on static bandwidth information and enable dynamic decision-making based on link health and operational load. By defining FEC level, buffer depth, transmission window, DSP equalization parameters, and link traffic carrying ratio as variables to be optimized, collaborative modeling of AOC link error correction capabilities, buffer adjustment capabilities, transmission control capabilities, signal compensation capabilities, and traffic allocation capabilities is achieved, providing a unified optimization object for multi-parameter joint tuning. By establishing constraints on bit error rate, transmission delay, link load, and traffic balancing, comprehensive limitations on link reliability, real-time performance, bearer security, and traffic distribution balance are achieved. This ensures that the optimization process avoids degrading other transmission performance aspects due to improvements in a single performance indicator. By constructing evaluation items for bit error rate, delay, link load, and traffic balancing, and weighting them based on communication requirements and available link capacity in the next scheduling cycle, a dynamic objective function oriented towards business needs and link status is constructed. This allows the optimization objective to adapt to changes in GPU cluster communication load, improving the accuracy of AOC link transmission parameter optimization and the rationality of link resource allocation.

[0065] Furthermore, an initial set of candidate solutions is generated based on each variable to be optimized. Each candidate solution corresponds to a search individual in the tunic group algorithm and a set of AOC link transmission parameter configuration schemes. Based on the optimization objective function, the objective function value corresponding to each candidate solution is calculated, and the current optimal candidate solution is determined according to the objective function value. The search individual corresponding to the optimal candidate solution is then determined as the leader individual. Based on the comprehensive risk value, micro-burst traffic status, and available carrying capacity of each AOC link, a risk weight factor is constructed, and the search step size, search range, and position update speed of each candidate solution in the tunic swarm algorithm are dynamically adjusted using the risk weight factor.

[0066] The comprehensive risk value, peak micro-burst traffic, duration of micro-burst, probability of micro-burst occurrence, and available carrying capacity of each AOC link are obtained as link risk status parameters. Based on the link risk status parameters, the risk weight factor corresponding to each AOC link is calculated. The higher the overall risk value of the link, the more severe the micro-burst traffic status, and the lower the available carrying capacity of the link, the larger the corresponding risk weight factor. The formula is as follows: ; in, For the first Risk weighting factor for each AOC link For the first The microburst traffic state intensity value of each AOC link is obtained by linearly weighting the normalized microburst traffic peak value, microburst duration, and microburst occurrence probability. For the first Normalized available carrying capacity values ​​for each AOC link. , , These are the weighting coefficients for the overall link risk value, the micro-burst traffic state intensity value, and the available carrying capacity value, respectively.

[0067] Based on the risk weight factor corresponding to each AOC link, the search step size of each candidate solution in the *Scalypsus swarm* algorithm is dynamically adjusted. This allows candidate solutions with risk weight factors greater than a preset threshold to have their search step size and search range expanded, thereby increasing the search priority for high-risk AOC link regions. The formula is as follows: ; in, Index for candidate solutions For the first The current search step size of each candidate solution. For the first Adjusted search step size for each candidate solution.

[0068] Based on the risk weight factor, the position update speed of each candidate solution is adaptively adjusted so that the candidate solutions corresponding to high-risk AOC links converge in the direction of reducing link load and reducing bit error rate, while the candidate solutions corresponding to low-risk AOC links converge in the direction of increasing link throughput and traffic balance. During the candidate solution search and update process, the risk weight factor is dynamically adjusted based on the changes in the risk status parameters of each AOC link.

[0069] The leader individual in the tunic group algorithm after risk weight adjustment is used to perform a global search for candidate solutions, and the search direction of each candidate solution is updated according to the current best candidate solution.

[0070] The current position of the navigator is used as the global guiding position in the current iteration round; Based on the global guidance position of the navigator, the positional deviation between the search individual and the navigator corresponding to each candidate solution is calculated, and the corresponding search direction adjustment parameters are generated. Based on the search direction adjustment parameters, the position update direction of each candidate solution's corresponding search individual is adjusted, so that each search individual gradually moves closer to the search region corresponding to the current optimal candidate solution. The formula is: ; in, For the iteration round index, For the first The candidate solution at the th... Position in the round of iteration, For the updated number The candidate solution at the th... Position in the round of iteration, For the first The position update speed of each candidate solution For the leading individual in the first Position in the round of iteration, Adjust parameters for the search direction.

[0071] During the candidate solution update process, a sine-cosine update mechanism is introduced, which uses sine and cosine functions to dynamically perturb and update the position change direction and magnitude of the candidate solution.

[0072] Determine the current position, search direction, and position update step size of each candidate solution's corresponding search individual in the current iteration round, and determine the position update parameters for each search individual; Based on the positional deviation between the current position and the current optimal candidate solution, corresponding sine perturbation terms and cosine perturbation terms are constructed respectively. The sine perturbation term is used to adjust the direction of positional change of the search individual, and the cosine perturbation term is used to adjust the magnitude of positional change of the search individual. Based on the sine and cosine perturbation terms, the position update parameters of each search individual are dynamically adjusted to form a nonlinear position change trajectory around its current position. The formula is as follows: ; in, For disturbance amplitude control parameters, The angle parameter of the sine and cosine functions. The influence coefficient of the navigation position. These are the adjustment coefficients for sinusoidal and cosine disturbances.

[0073] For search individuals located in high-risk AOC link search areas, increase the positional change amplitude corresponding to the sine and cosine perturbation terms to improve the jump search capability in high-risk areas; For search individuals located in the low-risk AOC link search region, reduce the positional change amplitude of the sine and cosine perturbation terms to improve the convergence stability in the local stable region. Based on the position update parameters adjusted by dynamic perturbation, update the position of each candidate solution corresponding to the search individual, and generate the corresponding candidate solution update results; After the candidate solution position is updated, the updated candidate solution is subjected to boundary range detection. When the candidate solution exceeds the preset parameter boundary range, the candidate solution that exceeds the boundary range is subjected to boundary correction processing.

[0074] After each iteration, the objective function value corresponding to the updated candidate solution is recalculated, and the current best candidate solution is updated based on the objective function value, and the current leader individual is determined. When the preset number of iterations is reached, the optimal candidate solution is obtained, and the AOC link transmission parameter configuration scheme corresponding to the optimal candidate solution is determined as the optimal transmission strategy.

[0075] By mapping each variable to be optimized to a search individual in the tunic swarm optimization algorithm, and determining the current optimal candidate solution and the leading individual based on the objective function value, a unified expression of the AOC link transmission parameter configuration scheme under the swarm intelligence optimization framework is achieved, providing a clear search foundation for subsequent iterative optimization. By constructing a risk weight factor by combining the link's comprehensive risk value, micro-burst traffic status, and available carrying capacity, and using the risk weight factor to dynamically adjust the search step size, search range, and position update speed, adaptive matching between the search process and the link risk status is achieved. This ensures that candidate solutions corresponding to high-risk links can converge preferentially towards reducing load and bit error rate, while candidate solutions corresponding to low-risk links can converge towards reducing load and bit error rate. The optimization process prioritizes convergence towards improving throughput and flow balance. A global guiding position is determined by a lead individual, and the search direction is adjusted based on the positional deviation between candidate solutions and the lead individual, achieving global guided search of the optimization space and improving the directionality and effectiveness of multi-parameter combination optimization. By introducing a sine and cosine update mechanism, the direction and magnitude of positional changes in candidate solutions are dynamically perturbed nonlinearly, achieving coordination between global exploration and local convergence and reducing the possibility of the optimization process getting trapped in local optima. Iterative updates to the objective function value, re-determining the optimal candidate solution and the lead individual, and boundary correction for out-of-bounds candidate solutions ensure the feasibility and stability of the optimization results.

[0076] According to the optimal transmission strategy, a gradual traffic migration is performed on the sub-healthy AOC link, and transmission parameters are sent to the corresponding AOC link.

[0077] Specifically, based on the link traffic carrying ratio in the optimal transmission strategy, the traffic carrying ratio of the sub-healthy AOC links is gradually reduced, while the traffic carrying ratio of the corresponding target AOC links is gradually increased. Based on the changes in the link traffic carrying ratio, the business traffic in the sub-healthy AOC link is migrated in stages, and the link traffic distribution after migration is generated.

[0078] Obtain the current service traffic information carried by the sub-health AOC link, and determine the migration ratio and migration stage of the corresponding service traffic based on the changes in the link traffic carrying ratio in the optimal transmission strategy. Based on the service priority, data transmission latency requirements, and bandwidth usage of the service traffic, the service traffic in the sub-healthy AOC link is classified to generate a set of service traffic to be migrated. Based on the link traffic carrying ratio and available carrying capacity of each target AOC link, the traffic set of services to be migrated will be gradually allocated to the corresponding target AOC links according to the preset migration stages. In each migration phase, the proportion of business traffic in the sub-healthy AOC link is gradually reduced, while the proportion of business traffic in the target AOC link is increased simultaneously. After completing all migration phases, the traffic usage of each AOC link is statistically analyzed, and the post-migration link traffic distribution is generated.

[0079] During the traffic migration process, the bit error rate, transmission latency, queue buffer usage, and link throughput of the AOC link are monitored in real time, and the traffic migration rate is dynamically adjusted based on the monitoring results.

[0080] Acquire real-time operating status data of each AOC link during traffic migration, and extract the corresponding bit error rate, transmission latency, queue buffer usage and link throughput parameters as link transmission status evaluation parameters; Each parameter in the link transmission status assessment parameters is compared with its corresponding preset threshold to generate the corresponding link status judgment result. When the link status determination result indicates that the corresponding AOC link is in a stable transmission state, the traffic migration rate of the current stage is increased, and the proportion of service traffic migrated to the target AOC link is increased. When the link status determination result indicates that the corresponding AOC link has an increased bit error rate, increased transmission latency, increased queue buffer occupancy rate, or decreased link throughput, the traffic migration rate at the current stage is reduced, and the proportion of service traffic migrated to the target AOC link is reduced. Based on the adjusted traffic migration rate, the service traffic of each AOC link is redistributed, and data collection and migration rate adjustment are performed for the next monitoring cycle.

[0081] When the traffic migration meets the stable transmission conditions, the transmission parameters are sent to the corresponding AOC link according to the FEC level, buffer depth, sending window and DSP equalization parameters in the optimal transmission strategy. After the transmission parameters are issued, the transmission performance feedback data of the corresponding AOC link is obtained and uploaded to the database for storage.

[0082] By progressively adjusting the traffic share of sub-healthy AOC links and target AOC links according to the link traffic carrying ratio in the optimal transmission strategy, a smooth transfer of link load from high-risk links to capable links is achieved, enabling the traffic scheduling process to avoid the impact of sudden switching on the continuity of service transmission. By acquiring the service traffic information currently carried by sub-healthy AOC links and generating a set of service traffic to be migrated based on service priority, transmission latency requirements, and bandwidth occupancy, a fine-grained classification of migrated services is achieved, allowing critical and ordinary services to be migrated in an orderly manner according to different transmission needs. By allocating the service traffic to be migrated in stages based on the traffic carrying ratio and available carrying capacity of the target AOC links, dynamic reconstruction of link traffic distribution is achieved, which helps reduce the operating pressure of sub-healthy links and improve the resource utilization efficiency of target links. By monitoring the bit error rate, transmission latency, queue buffer occupancy, and link throughput in real time during the migration process, and dynamically adjusting the migration rate according to the link status judgment results, closed-loop adaptive control of the traffic migration process is achieved, enabling the migration rate to automatically increase or decrease with changes in link operating status, improving the stability and security of the migration process.

[0083] This embodiment also provides a computer device applicable to the method for optimizing the transmission performance of AOC active optical cables for high-speed interconnection of intelligent computing centers, including: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to implement the method for optimizing the transmission performance of AOC active optical cables for high-speed interconnection of intelligent computing centers as proposed in the above embodiment.

[0084] The computer device can be a terminal, comprising a processor, memory, communication interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, carrier networks, NFC (Near Field Communication), or other technologies. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the computer device's casing, or an external keyboard, touchpad, or mouse.

[0085] This embodiment also provides a storage medium storing a computer program. When executed by a processor, the program implements the method for optimizing the transmission performance of AOC active optical cables for high-speed interconnection of intelligent computing centers as proposed in the above embodiments. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0086] In summary, this invention achieves forward-looking perception of high-speed interconnect link load change trends by: collecting and analyzing training and operation data of GPU clusters in intelligent computing centers and identifying the current training stage; combining communication behavior characteristics corresponding to different training stages to predict the micro-burst traffic status of AOC links; transforming the link regulation process from a traditional post-event response mode to a proactive prediction mode oriented towards future scheduling cycles; providing more timely and accurate dynamic basis for link status assessment and resource scheduling; calculating sub-health scores by combining link status data and constructing link risk topology; realizing a visual correlation expression between link health status and traffic risk distribution; expanding link risk identification from single-point static judgment to global correlation dynamic analysis; and constructing a multi-variable optimization target that includes FEC level, buffer depth, sending window, DSP equalization parameters, and link traffic carrying ratio. The algorithm uses a standardized function and an improved tunic group algorithm with a fusion of sine and cosine update mechanisms for iterative optimization. This achieves global search and dynamic perturbation-enhanced search in the complex coupled parameter space, enabling the parameter optimization process to have both global exploration and local convergence capabilities. This effectively improves the adaptability and optimization efficiency of multi-dimensional transmission parameter collaborative adjustment, and can generate optimal transmission strategies that better suit the current link operating state. By performing gradual traffic migration on the sub-healthy AOC link according to the optimal transmission strategy and simultaneously distributing the optimized transmission parameters, the smooth reconstruction of link load and online dynamic adjustment of transmission performance are achieved. This avoids sudden shocks and service jitter during link switching, enabling the high-speed interconnection network of the intelligent computing center to maintain stable data transmission capabilities in the sub-healthy state of the link, improving the transmission stability of the AOC link, enhancing communication continuity, and improving the overall resource utilization efficiency.

[0087] It should be noted that the above embodiments are only used to illustrate the technical solutions 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 solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for optimizing the transmission performance of active optical cables (AOCs) for high-speed interconnection in intelligent computing centers, characterized in that: include, Collect training and running data of the GPU cluster in the intelligent computing center, identify the current GPU training stage based on the training and running data, determine the communication requirements for the next scheduling cycle based on the GPU training stage, and predict the micro-burst traffic status of the AOC link based on the communication requirements. The AOC link to be evaluated is determined based on the micro-burst traffic status, and the link status data of the AOC link to be evaluated is obtained. The sub-health score of the AOC link to be evaluated is calculated based on the link status data, and the link risk topology is generated by combining the micro-burst traffic status. Based on the link risk topology, an optimization objective function is constructed with FEC level, buffer depth, sending window, DSP equalization parameters and link traffic carrying ratio as optimization variables. An improved tunic group algorithm with fused sine and cosine update mechanism is used to iteratively optimize the optimization objective function and generate the optimal transmission strategy. According to the optimal transmission strategy, a gradual traffic migration is performed on the sub-healthy AOC link, and transmission parameters are sent to the corresponding AOC link.

2. The method for optimizing the transmission performance of AOC active optical cable for high-speed interconnection of intelligent computing centers as described in claim 1, characterized in that: The process of collecting training and execution data from the GPU cluster in the intelligent computing center, and identifying the current GPU training stage based on the training and execution data, includes the following steps: Collect running status data, communication behavior data, and task scheduling data of each GPU node in the GPU cluster of the intelligent computing center, and preprocess the collected data to generate a training running dataset; Based on the training dataset, we extract GPU computing load variation features, inter-node communication frequency features, gradient synchronization cycle features, and data transmission scale features to generate a training behavior feature parameter set. Based on the training behavior feature parameter set and combined with the preset training stage identification rules, the running status of the current training task is classified into stages to obtain the current GPU training stage.

3. The method for optimizing the transmission performance of AOC active optical cable for high-speed interconnection of intelligent computing centers as described in claim 2, characterized in that: The process of determining the communication requirements for the next scheduling cycle based on the GPU training phase and predicting the micro-burst traffic status of the AOC link based on the communication requirements includes the following steps: Based on the training behavior characteristics corresponding to the current GPU training phase, determine the data interaction method and communication load change trend of the GPU cluster in the next scheduling cycle; Based on the data interaction method and communication load change trend, the communication frequency, data transmission scale, synchronization cycle and bandwidth usage of nodes in the GPU cluster are statistically analyzed to generate the communication requirements for the next scheduling cycle. Based on communication requirements, and combined with historical AOC link traffic data, switch port queue length, and RDMA message arrival interval, the traffic change trend of the AOC link is predicted. Based on the predicted traffic change trend, the peak value, duration and probability of occurrence of micro-burst traffic on the AOC link are determined, and the corresponding micro-burst traffic status is generated.

4. The method for optimizing the transmission performance of AOC active optical cable for high-speed interconnection of intelligent computing centers as described in claim 3, characterized in that: The process of determining the AOC link to be evaluated based on the micro-burst traffic status and obtaining the link status data of the AOC link to be evaluated includes the following steps: Based on the traffic peak, duration and probability of occurrence in the micro-burst traffic state, AOC links whose load change exceeds the preset threshold in the next scheduling cycle are identified as AOC links to be evaluated. Obtain the link operation information of the AOC link to be evaluated, including the link's real-time throughput, port utilization, and queue buffer usage. Based on the link operation information, physical layer status detection is performed on the AOC link to be evaluated to obtain the corresponding link status data.

5. The method for optimizing the transmission performance of AOC active optical cable for high-speed interconnection of intelligent computing centers as described in claim 4, characterized in that: The process of calculating the sub-health score of the AOC link to be evaluated based on link status data and generating a link risk topology by combining micro-burst traffic status includes the following steps: The link status data of the AOC link to be evaluated is normalized, and the bit error rate characteristics, optical power margin characteristics, temperature change characteristics and signal jitter characteristics reflecting the link operation stability are extracted. Based on the extracted features reflecting the stability of link operation, the link health status index of each AOC link to be evaluated is calculated, and the corresponding sub-health score is generated according to the link health status index. Based on the peak traffic, duration and probability of occurrence in the micro-burst traffic state, calculate the link load risk level of each AOC link to be evaluated in the next scheduling cycle. By combining the sub-health score and link load risk level of each AOC link to be evaluated, a corresponding comprehensive link risk value is generated; Using server and switching devices in the GPU cluster as network nodes, AOC links as inter-node connections, and the comprehensive risk value of each AOC link as the edge weight parameter, a corresponding link risk topology is generated.

6. The method for optimizing the transmission performance of AOC active optical cable for high-speed interconnection of intelligent computing centers as described in claim 5, characterized in that: The method for constructing an optimization objective function based on link risk topology, with FEC level, buffer depth, sending window, DSP equalization parameters, and link traffic carrying ratio as optimization variables, includes the following steps: Based on the comprehensive risk value and link connection relationship of each AOC link in the link risk topology, combined with the current bandwidth usage, queue buffer usage and link status data, the available carrying capacity of each AOC link is determined. Based on the comprehensive risk value and available carrying capacity of each AOC link, the variables to be optimized for the corresponding AOC link are determined, including FEC level, buffer depth, sending window, DSP equalization parameters and link traffic carrying ratio. Based on the variables to be optimized, corresponding bit error rate constraints, transmission delay constraints, link load constraints, and traffic balancing constraints are established respectively. Based on bit error rate constraints, transmission delay constraints, link load constraints, and traffic balancing constraints, an optimization objective function is constructed to evaluate the transmission performance of the AOC link.

7. The method for optimizing the transmission performance of AOC active optical cable for high-speed interconnection of intelligent computing centers as described in claim 6, characterized in that: The improved tunic group algorithm using a fused sine and cosine update mechanism iteratively optimizes the objective function and generates the optimal transmission strategy, including the following steps: An initial set of candidate solutions is generated based on each variable to be optimized. Each candidate solution corresponds to a search individual in the tunic group algorithm and a set of AOC link transmission parameter configuration schemes. Based on the optimization objective function, the objective function value corresponding to each candidate solution is calculated, and the current optimal candidate solution is determined according to the objective function value. The search individual corresponding to the optimal candidate solution is then determined as the leader individual. Based on the comprehensive risk value, micro-burst traffic status and available carrying capacity of each AOC link, a risk weight factor is constructed, and the risk weight factor is used to dynamically adjust the search step size, search range and position update speed of each candidate solution in the tunic group algorithm. The leader individuals in the tunic group algorithm after risk weight adjustment are used to perform a global search for candidate solutions, and the search direction of each candidate solution is updated according to the current best candidate solution. In the process of updating candidate solutions, a sine and cosine update mechanism is introduced, which uses sine and cosine functions to dynamically perturb and update the position change direction and change amplitude of candidate solutions. After each iteration, the objective function value corresponding to the updated candidate solution is recalculated, and the current optimal candidate solution is updated based on the objective function value. When the preset number of iterations is reached, the optimal candidate solution is obtained, and the AOC link transmission parameter configuration scheme corresponding to the optimal candidate solution is determined as the optimal transmission strategy.

8. The method for optimizing the transmission performance of AOC active optical cable for high-speed interconnection of intelligent computing centers as described in claim 7, characterized in that: The process of performing gradual traffic migration on sub-healthy AOC links according to the optimal transmission strategy and sending transmission parameters to the corresponding AOC links includes the following steps: Based on the link traffic carrying ratio in the optimal transmission strategy, gradually reduce the traffic carrying ratio of the sub-healthy AOC links and gradually increase the traffic carrying ratio of the corresponding target AOC links. Based on the changes in the link traffic carrying ratio, the business traffic in the sub-health AOC link is migrated in stages, and the link traffic distribution after migration is generated. During the traffic migration process, the bit error rate, transmission latency, queue buffer usage, and link throughput of the AOC link are monitored in real time, and the traffic migration rate is dynamically adjusted based on the monitoring results. When the traffic migration meets the stable transmission conditions, the transmission parameters are sent to the corresponding AOC link according to the FEC level, buffer depth, sending window and DSP equalization parameters in the optimal transmission strategy. After the transmission parameters are issued, the transmission performance feedback data of the corresponding AOC link is obtained and uploaded to the database for storage.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that: When the processor executes the computer program, it implements the steps of the method for optimizing the transmission performance of AOC active optical cable for high-speed interconnection of intelligent computing centers as described in any one of claims 1 to 8.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by the processor, it implements the steps of the method for optimizing the transmission performance of AOC active optical cable for high-speed interconnection of intelligent computing centers as described in any one of claims 1 to 8.

Images