AI Data Center Switch Traffic Intelligent Scheduling Method, Device, Equipment and Media

Abstract

This invention relates to the field of traffic scheduling technology, specifically to a method, apparatus, device, and medium for intelligent traffic scheduling of AI data center switches. The method includes the following steps: collecting port rate changes, identifying hopping ports, filtering schedulable paths, locating release paths, filtering target paths, and updating the traffic scheduling scheme. This invention dynamically identifies periodic changes in switch port rates, filters scheduling entry paths based on real-time hopping behavior, and accurately locates release paths by combining tail path operation trends. Through unified updates of path states and synchronous issuance of traffic migration commands, it achieves rapid response and accurate allocation of multi-path traffic scheduling, effectively avoiding resource waste, response lag, and path congestion problems existing in traditional static path configuration and time-slice scheduling. This improves the link stability and resource regulation efficiency of AI data centers under high-frequency interaction during training tasks.

Description

Technical Field

[0001] This invention relates to the field of traffic scheduling technology, and in particular to intelligent traffic scheduling methods, devices, equipment and media for AI data center switches. Background Technology

[0002] Traffic scheduling technology encompasses various methods for dynamically distributing and managing data flows in computer network environments. These methods include determining data transmission paths, allocating bandwidth resources, identifying and mitigating traffic congestion, and formulating and executing scheduling strategies. By constructing scheduling algorithms, formulating scheduling strategies, and combining them with network device status information, dynamic control of data flows is achieved to ensure the rational utilization of network resources and the timeliness and reliability of data transmission. Among these methods, traditional intelligent traffic scheduling for AI data center switches addresses issues such as traffic congestion, scheduling delays, and path conflicts that occur at the switch level during large-scale data interaction within AI data centers. This typically employs priority scheduling based on static configuration or time-slice scheduling based on port polling to manage traffic.

[0003] Existing technologies mainly rely on static priority settings and fixed time slice polling during traffic scheduling. In the context of sudden and high concurrency AI training tasks, it is difficult to effectively identify real-time changes in path status, resulting in some ports being continuously congested while some paths are idle. Especially in scenarios where data flow changes frequently, there is a lack of scheduling feedback mechanisms based on path load trends and terminal operating status. The scheduling behavior lacks real-time and target-oriented characteristics, which can easily lead to scheduling delays and path conflicts. For example, when multiple high-priority tasks occupy a certain path resource, the system has difficulty adjusting the flow direction quickly, which can easily form a link bottleneck and affect the overall training efficiency. Summary of the Invention

[0004] The purpose of this invention is to address the shortcomings of existing technologies by proposing an AI data center switch traffic intelligent scheduling method, device, equipment, and medium.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: an AI data center switch traffic intelligent scheduling method, comprising the following steps:

[0006] S1: Obtain the real-time rate sampling values ​​of all uplink and downlink ports of the switch in the active node of the AI ​​training task within the period, calculate the absolute difference of each port, and mark it as a hopping port if it exceeds the dynamic traffic threshold, and generate a hopping port identification list.

[0007] S2: Based on the jump port identification list, compare the scheduling task status identifier item by item, construct the set of schedulable path entry points in the current period, and generate a schedulable port task list;

[0008] S3: Based on the confirmed path of the port in the schedulable port task list, obtain the rate sampling value of the switch port at the end of the corresponding path for three consecutive cycles, mark the valid path of release status, and generate the tail release path location result;

[0009] S4: Based on the tail release path location result, count the current path switching node hop count, the average load rate of the entire path and the end-to-end propagation delay of the path, filter and schedule paths, and generate a path filtering and scheduling combination list.

[0010] S5: Based on the path filtering and scheduling combination list, update the path information to the switch accordingly, send traffic migration signals to all affected paths, and generate an AI data center switch traffic intelligent scheduling scheme.

[0011] As a further embodiment of the present invention, the hopping port identification list includes the hopping port number, hopping time, and hopping amplitude; the schedulable port task list includes the scheduling status, path start identifier, and path number; the tail-end release path location result includes the path tail-end port number, load change trend, and bandwidth-delay product evaluation record; the path screening and scheduling combination list includes the path hop count, path average load rate, and path propagation delay; and the AI ​​data center switch traffic intelligent scheduling scheme includes path update information, traffic migration signal, and target scheduling path.

[0012] As a further aspect of the present invention, the step of obtaining the switching port identification list is as follows:

[0013] S111: Based on all uplink and downlink ports of the switch in the active node of the AI ​​training task, obtain the real-time rate sampling value of each port within the period, collect the rate sampling value sequence of the current period and the previous period, and combine the port rate sampling data with the period division interval to generate the port rate change difference sequence.

[0014] S112: Based on the port rate change difference sequence, read the rate sampling data of the previous period, organize the change trend of each port, and obtain the port rate jump evaluation record.

[0015] S113: Based on the port rate jump evaluation record, a judgment is made according to the dynamic traffic threshold, and the ports that meet the conditions are marked. The marked information is integrated to establish a jump port index set and generate a jump port identification list.

[0016] As a further aspect of the present invention, the step of obtaining the schedulable port task list is as follows:

[0017] S211: Based on all port numbers in the jump port identification list, extract the task status identifier of the corresponding port in the scheduling system one by one, associate and match the task status identifier with the scheduling status reference code, filter the set of port numbers with the identifier as ready, and establish a ready port number list.

[0018] S212: Based on the list of ready port numbers and the path position mapping relationship between ports and nodes in the network topology table, determine whether each port in the list corresponds to the first node of the path, determine whether it is in the starting node position in the scheduling path structure, and generate a set of first-end ports of the path.

[0019] S213: Based on the intersection of the set of port at the beginning of the path and the list of ready port numbers, perform a filtering operation on all port sets that meet the dual conditions of scheduling status and path structure in the current period, calculate and obtain the scheduling priority ranking coefficient, sort according to the coefficient, complete the construction of the task scheduling entry queue, and establish a list of schedulable port tasks.

[0020] As a further aspect of the present invention, the step of obtaining the tail release path positioning result is as follows:

[0021] S311: Based on the confirmed path of the port in the schedulable port task list, extract the outgoing rate sampling value of three consecutive cycles from the end switch port connected to each path, record the specific rate value and corresponding time tag in each cycle, and bind the collected rate data to the port and cycle to obtain the tail continuous rate data group.

[0022] S312: Based on the continuous rate data set at the tail end, analyze the rate change trend between two adjacent cycles, determine whether the rate shows a continuous periodic decrease, and compare it with the decrease range of the bandwidth delay product. If the path tail end port continuously decreases and the change falls within the decrease range of the bandwidth delay product, then record the path as a path that meets the condition of the decreasing state, and obtain the set of bandwidth decrease matching interval identifiers.

[0023] S313: Based on the bandwidth decrease conformity interval identifier set, the average receiving rate, scheduling waiting time and switching hops of the path tail port in three cycles are statistically analyzed to identify and confirm the path tail release trend, determine whether the current path has a release tendency, and perform a marking operation on the path that meets the release conditions to obtain the tail release path location result.

[0024] As a further aspect of the present invention, the step of obtaining the path filtering and scheduling combination list is as follows:

[0025] S411: Based on the tail release path location result, record the exchange node numbers that are passed sequentially from the source end to the tail end in the path structure for each path, organize all nodes in the path according to the connection order, filter unreachable nodes in combination with path connectivity information, confirm the complete node hop count path, and perform statistics and archiving according to the path number to obtain the path exchange hop count list.

[0026] S412: Based on the path switching hop count list, perform load status reading operations on the switching node ports involved in each path, obtain the load occupancy of each egress port in the current period, organize them into a load rate dataset, perform segmentation merging and classification operations on the organized dataset, determine whether the overall load performance of the path is within the PFC trigger threshold range, filter the path numbers that meet the interval conditions, and obtain the path average load rate matching result set.

[0027] S413: Based on the path average load rate matching result set, retrieve the propagation delay value of each link in the corresponding path in the topology table according to the path number, read the path link transmission delay information segment by segment, and determine whether it meets the RDMA propagation delay upper limit standard. If the path propagation delay index meets the threshold condition, combine the path hop count condition and the load rate matching condition to filter the path information that meets all scheduling requirements and obtain the path filtering scheduling combination list.

[0028] As a further aspect of the present invention, the steps for obtaining the AI ​​data center switch traffic intelligent scheduling scheme are as follows:

[0029] S511: Based on the path filtering and scheduling combination list, sequentially read the forwarding table records and MAC address mapping data of each hop switching node in the path, replace the original forwarding entries with the new forwarding entries of the target path, and establish a path switching configuration update state set.

[0030] S512: According to the path switching configuration update status set, identify all updated path numbers, construct corresponding path migration signaling data structures, inject the constructed signaling structures into the control channel of the source switching node, send synchronization signaling to all relevant paths, if the path signaling response status is valid and the feedback success flag is written, record the corresponding path number and response flag, and summarize to obtain the path migration signaling confirmation flag table.

[0031] S513: Based on the path migration signaling confirmation identifier table, perform a strategy configuration summary operation on all path numbers with valid signaling status identifiers, integrate them into a structured parameter set, construct a scheduling control instruction group based on path reconstruction behavior, and generate an AI data center switch traffic intelligent scheduling scheme.

[0032] The AI ​​data center switch traffic intelligent scheduling device is used to execute the AI ​​data center switch traffic intelligent scheduling method, and the AI ​​data center switch traffic intelligent scheduling device includes:

[0033] The rate jump identification module is used to perform S1: obtain the real-time rate sampling values ​​of all uplink and downlink ports of the switch in the active node of the AI ​​training task within the period, calculate the absolute difference of each port, and mark it as a jump port if it exceeds the dynamic traffic threshold, and generate a jump port identification list.

[0034] The scheduling port filtering module is used to perform S2: based on the jump port identification list, compare the scheduling task status identifier item by item, construct the set of schedulable path entry points in the current period, and generate a list of schedulable port tasks;

[0035] The tail trend recognition module is used to execute S3: based on the path to which the port has been confirmed in the schedulable port task list, obtain the rate sampling value of the switch port at the end of the corresponding path for three consecutive cycles, mark the valid path of release status, and generate the tail release path location result;

[0036] The path optimization calculation module is used to execute S4: based on the tail release path positioning result, to count the current path exchange node hop count, the average load rate of the entire path and the end-to-end propagation delay of the path, to filter and schedule paths, and to generate a path filtering and scheduling combination list;

[0037] The traffic scheduling execution module is used to execute S5: based on the path filtering and scheduling combination list, update the path information to the switch accordingly, send traffic migration signals to all affected paths, and generate an AI data center switch traffic intelligent scheduling scheme.

[0038] Compared with the prior art, the advantages and positive effects of the present invention are as follows:

[0039] In this invention, the periodic changes in switch port rates are dynamically identified, and entry paths are selected based on real-time hopping behavior. The release path is accurately located by combining the running trend of the tail path. A comprehensive evaluation mechanism for hop count, load rate, and propagation latency is introduced to ensure that the selected path achieves the optimal balance between latency control and load balancing. By uniformly updating the path status and synchronously issuing traffic migration instructions, rapid response and accurate allocation of multi-path traffic scheduling are achieved. This effectively avoids the resource waste, response lag, and path congestion problems existing in traditional static path configuration and time-slice scheduling, and improves the link stability and resource regulation efficiency of AI data centers under high-frequency interaction of training tasks. Attached Figure Description

[0040] Figure 1 This is a schematic diagram of the main steps of the present invention;

[0041] Figure 2 This is a flowchart of the process for obtaining the jump port identification list in this invention;

[0042] Figure 3 This is a flowchart of the process for obtaining the schedulable port task list in this invention;

[0043] Figure 4 This is a flowchart of the process for obtaining the tail release path positioning result of the present invention;

[0044] Figure 5 This is a flowchart of the path filtering and scheduling combination list acquisition process of the present invention;

[0045] Figure 6 The flowchart for the AI ​​data center switch traffic intelligent scheduling scheme of this invention is as follows. Detailed Implementation

[0046] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0047] In the description of this invention, it should be understood that the terms "length," "width," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," and "outer," etc., indicating orientation or positional relationships, are based on the orientation or positional relationships shown in the accompanying drawings and are only for the convenience of describing the invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of the invention. Furthermore, in the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0048] Please see Figure 1 The AI-powered intelligent traffic scheduling method for data center switches includes the following steps:

[0049] S1: Obtain the real-time rate sampling values ​​of all uplink and downlink ports of the switch in the active node of the AI ​​training task within the period, calculate the absolute difference between the current period rate and the previous period rate of each port in turn, and determine whether it exceeds the dynamic traffic threshold (meeting the standard setting of 10% rate change rate). If it exceeds the threshold, mark the corresponding port as a jump port and generate a jump port identification list.

[0050] S2: Based on all port numbers in the jump port identification list, compare the scheduling task status identifier item by item, retain only the ports with the status of ready, and confirm whether the port is the first node of the path, construct the set of schedulable path entry points in the current period, and generate a list of schedulable port tasks.

[0051] S3: Based on the path to which the port has been confirmed in the schedulable port task list, obtain the rate sampling value of the switch port at the end of the corresponding path for three consecutive cycles, calculate the difference between the previous and next cycles and determine whether it is continuously negative, and whether the difference is within the bandwidth-delay product decrease range during idle operation (according to the congestion control algorithm). If the condition is met, mark the tail end of the corresponding path as a valid path in the release state and generate the tail end release path location result.

[0052] S4: Based on the tail release path location results, count the number of hops through the switching nodes traversed by the current path, the average load rate of the entire path, and the end-to-end propagation delay of the path for each path, and simultaneously filter and schedule paths (filter paths whose number of hops does not exceed the maximum number of hops of ECMP equivalent multipath, whose average load rate is not higher than the PFC trigger threshold: the standard setting is 60-70%, and whose propagation delay does not exceed the upper limit of RDMA delay), and generate a list of path screening and scheduling combinations.

[0053] S5: Based on the path filtering and scheduling combination list, update the path information to the switch accordingly, send traffic migration signals to all affected paths, and generate an AI data center switch traffic intelligent scheduling solution.

[0054] The hop port identification list includes the hop port number, hop time, and hop magnitude; the schedulable port task list includes the scheduling status, path start identifier, and path number; the tail release path location result includes the path tail port number, load change trend, and bandwidth-delay product evaluation record; the path screening and scheduling combination list includes the path hop count, path average load rate, and path propagation delay; and the AI ​​data center switch traffic intelligent scheduling solution includes path update information, traffic migration signals, and target scheduling path.

[0055] Please see Figure 2 The specific steps of S1 are as follows:

[0056] S111: Based on all uplink and downlink ports of the switch in the active node of the AI ​​training task, obtain the real-time rate sampling value of each port within the period, collect the rate sampling value sequence of the current period and the previous period, and combine the port rate sampling data with the period division interval to generate the port rate change difference sequence.

[0057] Based on all uplink and downlink ports of the switch in the active node of the AI ​​training task, it is necessary to first clarify the port numbers and functional status of all ports on the switch so as to accurately define the sampling targets in actual deployment. Assuming that the active ports configured in the current deployment node are P01 to P04, the system uses 5 minutes as the time boundary in each analysis cycle, and the sampling frequency is fixed at 1Hz within the cycle, that is, sampling rate data once per second. Each port forms 300 sampling points within the cycle. The sampled data is the instantaneous value of the port transmission rate in Mbps. The current cycle and the previous cycle sampling generate two sets of 300-point data sequences respectively. After the data collection is completed, the two sets of sequences are aligned according to the port number and the difference is processed. That is, the numerical difference of the rate value between the two cycles of data at each time point is calculated. For example, under port P01, the first sampling point in the current cycle is 820Mbps and the previous cycle is 750Mbps, then the difference is 70Mbps. The system takes the absolute value of the difference and adds it to the difference sequence. This process is repeated to complete the processing of 300 sampling points. The generated difference data sequence is stored as the port rate change difference sequence. The specific data structure is shown in the table below:

[0058] Table 1. Examples of Port Periodic Rate Sampling and Difference

[0059]

[0060] As shown in Table 1, the rate difference of port P01 is the largest and that of port P04 is the smallest. Both are recorded in Mbps and used for subsequent difference judgment logic. This sequence reflects the actual amplitude of the rate fluctuation of the port in a continuous period and finally generates the port rate change difference sequence.

[0061] S112: Based on the port rate change difference sequence, read the rate sampling data of the previous period, organize the change trend of each port, and obtain the port rate jump evaluation record.

[0062] Based on the port rate change difference sequence, the original rate value within the current period is extracted, and a historical reference rate is introduced as a comparison basis. The historical reference rate can be calculated by statistically analyzing the average rate of each port over the previous m periods (m=3 here). For example, the average rate of P01 over the previous 3 periods is 780Mbps, and that of P02 is 890Mbps. The system performs a linked analysis of the current period rate, the difference, and the historical reference rate in port number order. First, the difference is matched with the reference rate. If the current period rate is higher than the reference average and the difference is significant, it is determined that the port has a tendency to jump. If the current cycle rate is close to or lower than the historical average and the difference is small, it is considered to be in a stable state. For example, the current cycle rate of P01 is 820Mbps, the historical reference is 780Mbps, and the difference is 70Mbps. This is a case where the rate is higher than the reference average and the difference is large. Therefore, its offset rate can be marked as 70 / 780≈8.97%. Further, each port is sorted according to the degree of offset, and the fluctuation intensity label of each port in this cycle is output. The label value is the fluctuation evaluation result of the port in this cycle, which constitutes the port rate jump evaluation record. This type of data is output in vector form and associated with the port number.

[0063] S113: Based on the port rate jump evaluation record, perform judgment according to the dynamic traffic threshold, mark the ports that meet the conditions at the time, integrate the marking information to establish a jump port index set, and generate a jump port identification list;

[0064] Based on the aforementioned generated port rate hopping evaluation records, a judgment operation is performed according to the set hopping judgment threshold. Here, the hopping judgment threshold is set to 10%, that is, if the rate change rate of a port within a certain period is greater than 10%, it is considered that a hopping behavior has occurred. For example, the hopping evaluation value of P02 is 50Mbps, and its rate in the previous period is 860Mbps, so the change rate is 50 / 860≈5.81%, which does not reach the hopping threshold, so no hopping event is recorded. The change rate of P01 is 70 / 750≈9.33%, which is still less than 10%, so it also does not constitute a hopping event. If the evaluation value of a port is 95Mbps, and the reference rate is 860Mbps, the change rate is about 11.05%, so it is recorded as a hopping event. The system marks all port numbers that meet the hopping conditions and their corresponding sampling time points, writes the results into the hopping event list, and finally summarizes and constructs a unified hopping port index record structure to output a hopping port identification list.

[0065] Please see Figure 3 The specific steps of S2 are as follows:

[0066] S211: Based on all port numbers in the jump port identification list, extract the task status identifier of the corresponding port in the scheduling system one by one, associate and match the task status identifier with the scheduling status reference code, filter the set of port numbers with the identifier as ready, and establish a list of ready port numbers.

[0067] Based on all port numbers in the jump port identification list, the task status identifier field corresponding to each port in the current period scheduling system needs to be extracted item by item. This field is then used to determine whether the port is in a "ready" state. The task status identifier field is generally composed of status codes, with common identifiers including R (ready), B (occupied), I (idle), and F (faulty). The system scans the port task table to read the port status field one by one and performs a status code comparison operation. The comparison method is equal value matching, excluding non-scheduled states such as B, I, and F, and only retaining ports with the status identifier R. For example, if the system obtains that port P01 is R, P02 is B, P03 is R, and P04 is I during the sampling period, the filtering result will only retain ports P01 and P03. The results are summarized in the table below:

[0068] Table 2. Switching Port Scheduling Status Judgment Table

[0069]

[0070] As shown in Table 2, the system marks P01 and P03, which are in state R, as schedulable ports and writes them into the next stage of the processing flow to establish a list of ready port numbers.

[0071] S212: Based on the list of ready port numbers and the path position mapping relationship between ports and nodes in the network topology table, determine whether each port in the list corresponds to the first node of the path, determine whether it is in the starting node position in the scheduling path structure, and generate a set of first-end ports of the path.

[0072] After obtaining the list of ready port numbers, to determine whether the above ports are the starting entry ports of the scheduling path, it is necessary to reference the path topology table, which lists the structural relationship between the ports and their corresponding transmission paths. The field information includes path number, port number, and whether it is the starting point (is_entry), etc. The system performs a mapping lookup operation based on the port number and the path topology, reads the is_entry field value of each port, and determines whether it is the starting point of the path based on the value. For example, if the is_entry field value of port P01 is 1, it means that the port is the starting point of the path, and the corresponding field value of P03 is 0, it means that the port is a relay or end node. The system determines whether all ready ports meet the path starting point condition in this way. The filtering results are shown in the table below:

[0073] Table 3. Path Structure Start-End Identification Table

[0074]

[0075] As shown in Table 3, port P01 is the first entry port of path Path01. P03 does not meet the scheduling entry conditions. Therefore, the system uses P01 as the valid scheduling starting point of the current period and writes it into the scheduling path entry table to construct the set of first entry ports of the path.

[0076] S213: Based on the intersection of the set of ports at the beginning of the path and the list of ready port numbers, perform a filtering operation on all port sets that meet both the scheduling status and path structure conditions within the current period, using the following formula:

[0077] ;

[0078] The scheduling priority coefficients are calculated, and tasks are sorted according to these coefficients to construct the task scheduling entry queue and establish a list of schedulable port tasks. Indicates port Scheduling priority coefficient (unit: Mbps). Indicates port Total expected scheduling rate for the current week (in Mbps). Indicates port Remaining bandwidth capacity for the current period (in Mbps). Indicates port forward Total resources used by each scheduled path segment (in Mbps). Indicates port The basic rate reference value (in Mbps). Indicates port Number of state transitions within the current period Indicates the duration (in seconds) of a scheduling cycle. Indicates port Current path entry point actual load rate (in Mbps). Indicates port Ideal scheduling rate benchmark for adjacent paths (in Mbps). For port The number of adjacent path entries;

[0079] Based on the intersection of the set of ports at the beginning of the path and the list of ready port numbers, only port P01 simultaneously satisfies both the conditions of being in a ready state and being at the beginning of the structure. The system collects the scheduling-related parameters of this port for the current period for priority ranking calculation. The collected fields include: the rate of the task to be scheduled (…). ), remaining bandwidth capacity ( ), path segment resource consumption ( ), base rate reference value ( ), number of state transitions ( ), period duration ( ), actual load at the path inlet ( Ideal scheduling rate ( The parameters are summarized below:

[0080] Table 4. Parameters for Scheduling Priority Calculation

[0081]

[0082] Based on the above parameters, substitute them into the formula to perform the calculation:

[0083] First item:

[0084] ;

[0085] Second item:

[0086] ;

[0087] Third item:

[0088] ;

[0089] Comprehensive calculations yield:

[0090] ;

[0091] The final scheduling priority ranking coefficient of port P01 in the current period is 35.327. This value will be used as a reference for the scheduler's task scheduling priority, to generate the sorted task enqueue instructions, and to establish a list of schedulable port tasks.

[0092] The scheduling priority ranking coefficient is a quantitative indicator that measures the priority allocation of task resources to each port in the switching node within the current scheduling cycle. This coefficient comprehensively reflects the port's matching degree between task load intensity, available bandwidth capacity, state stability, and scheduling behavior and system expectations. The larger the value, the better the scheduling conditions of the port in the current cycle, including being able to handle more tasks, having smaller state fluctuations, and having a lower deviation from the ideal scheduling path of the system. Therefore, the ranking coefficient is used as the core basis for the scheduler's decision-making, constructing the scheduling enqueue sequence, determining the port order in which tasks are issued, and thus realizing the scheduling arrangement and path guidance among multiple ports within the cycle.

[0093] The formula calculation is based on a comprehensive consideration of three aspects: scheduling task intensity, link stability, and scheduling deviation. The first term... The current scheduling load ratio of the port is represented by the rate of the tasks to be scheduled. Remaining bandwidth With occupied path resources This is combined to reflect the ratio between the current port scheduling resource availability and the total load, and this ratio is divided by the base rate reference value. The results can be normalized to a general scheduling dimension, facilitating comparisons between different ports; intermediate terms This is used to measure the correlation between the frequency of port state changes and the cycle length, and the number of state transitions. More cycles indicate poorer scheduling stability, multiplied by the period duration. Then, time-weighted adjustments are made to the long-term fluctuation trend, followed by square root calculations to reduce the impact of extreme jumps and balance the amplification effect caused by drastic changes; the last term This represents the average scheduling deviation per unit time between a port and its adjacent path's ideal scheduling baseline. It is used to evaluate the port's deviation in terms of scheduling consistency. The larger this value is, the greater the difference between the port and the ideal scheduling rate set by the system, and the lower the priority should be. Therefore, this value is negatively included in the total value to weight and suppress the scheduling ranking index of jumping ports. Overall, the three parts represent task scheduling intensity, scheduling stability, and scheduling matching, respectively. Operations such as addition, subtraction, square roots, and absolute values ​​are used to unify dimensions and transform logical relationships to form an effective ranking basis.

[0094] Please see Figure 4 The specific steps of S3 are as follows:

[0095] S311: Based on the confirmed path of the port in the schedulable port task list, extract the outgoing rate sampling value of three consecutive cycles from the end switch port connected to each path, record the specific rate value and corresponding time tag in each cycle, and bind the collected rate data to the port and cycle to obtain the tail continuous rate data group.

[0096] Based on the confirmed paths to which ports belong in the schedulable port task list, the outbound rate data of the switch ports at the end of the path are collected within the most recent three consecutive scheduling cycles. The collection steps are executed according to the cycle number, and the cycle number is called from the scheduling control module. , and The scheduling record table is used to retrieve the corresponding path's end port number field. After locating the port using the port index, the real-time outgoing rate field value for each period is extracted, in Mbps. The data is then combined sequentially according to the period to construct a rate time series structure. The structure includes fields such as port number, period label, period timestamp, and sampling rate value. For example, taking the end port E45 of path number Path07 as an example, the recorded rate values ​​in scheduling periods 2401, 2402, and 2403 are 190.2 Mbps, 184.7 Mbps, and 179.9 Mbps, respectively. These are stored sequentially in the data structure, as shown in the following data example:

[0097] Table 5 Continuous Rate Sampling Table at the Tail Port of the Path

[0098]

[0099] As shown in Table 5, the data structure uses a uniform period span of 300 seconds for time segmentation, which ensures equal-interval sampling in the time dimension and finally generates a continuous rate data set at the tail end.

[0100] S312: Based on the continuous rate data set at the tail end, analyze the rate change trend between two adjacent cycles. By judging whether the rate shows a continuous periodic decrease, and comparing it with the decrease range of the bandwidth delay product, if the path tail end port continuously decreases and the change falls within the decrease range of the bandwidth delay product, then record the path as a path that meets the condition of the decreasing state, and obtain the set of bandwidth decrease matching interval identifiers.

[0101] Based on the continuous rate data set at the tail end, a trend recognition operation is performed on the three consecutive periodic rate sequences recorded at the end ports of each path. The rate change of each port between two adjacent periods is extracted, and it is determined whether the current period rate is less than the previous period rate. If the two consecutive changes are both decreasing, the port is determined to be in a downward trend. Then, the two change amplitudes are further matched with the idle operating rate decrease interval set by the system. The interval parameters are set by the bandwidth-delay product model, with a lower limit of 4.0 Mbps and an upper limit of 15.0 Mbps. If the rate of a port decreases to 5.5 Mbps and 6.2 Mbps in two consecutive periods, then the two change values ​​of 5.5 and 6.2 are both within the given interval, thus meeting the requirement that the decrease amplitude is in the idle interval. The system records this path as a path that meets the decrease condition and generates a set of bandwidth decrease compliance interval identifiers.

[0102] S313: Based on the bandwidth decrease conforms to the interval identifier set, the average receiving rate, scheduling waiting time and switching hops of the trailing port in three cycles are statistically analyzed to identify and confirm the trailing release trend of the path, determine whether the current path has a release tendency, and perform a marking operation on the path that meets the release conditions to obtain the trailing release path location result.

[0103] Based on the bandwidth decrease matching interval identifier set, the system further calls the average receiving rate field of the port at the end of the path, the total scheduling waiting time within three periods, and the switching hop number field of the corresponding path in the path topology table. These fields are processed together. The average receiving rate field is the arithmetic mean of the receiving rate of the port in all unit time periods during the sampling period, in Mbps. For example, the average receiving rate of port E45 in periods 2401 to 2403 is 210.4 Mbps. The scheduling waiting time field is the cumulative value of the time the port waits for scheduling in the scheduling system, in seconds, such as 25 seconds. The switching hop number field comes from the number of switching path segments recorded for Path07 in the path topology table, which is 5 hops. The system categorizes the above three items and constructs a path release trend data segment. By comparing it with the release status identification strategy, it determines whether the current path meets the judgment conditions of the release tendency status, and finally obtains the end release path location result.

[0104] Please see Figure 5 The specific steps of S4 are as follows:

[0105] S411: Based on the tail release path location result, record the number of the exchange nodes passed sequentially from the source end to the tail end in the path structure for each path, organize all nodes in the path according to the connection order, filter unreachable nodes in combination with path connectivity information, confirm the complete node hop count path, and perform statistics and archiving according to the path number to obtain the path exchange hop count list.

[0106] Based on the tail-end release path location results, the names of the exchange nodes from the source to the tail end of each path structure are traversed. The path number and corresponding exchange node index need to be extracted from the path topology to confirm the connection order between the source and tail nodes. Node numbers are extracted sequentially through hop-by-hop connections, and a path relay structure list is established. During this process, node pairs with broken links or loops are excluded. Simultaneously, the path integrity verification module checks the path connectivity. After confirming that there are no discontinuous hop segments within the path, the node numbers are entered into a statistics table in sequence. In the example, path number Path-A has exchange nodes SW01, SW07, SW12, SW18, SW21, and SW29, a total of 6 hops. These are structured and recorded in the path structure hop count statistics table, as follows:

[0107] Table 6. Path Switching Hop Count Statistics

[0108]

[0109] As shown in Table 6, the hop count is obtained by subtracting 1 from the node sequence length. After the statistics are completed, the system summarizes the hop count for each path and generates a path exchange hop count list.

[0110] S412: Based on the path switching hop count list, perform load status reading operations on the switching node ports involved in each path, obtain the load occupancy of each egress port in the current period, organize them into a load rate dataset, perform segmentation, merging and classification operations on the organized dataset, determine whether the overall load performance of the path is within the PFC trigger threshold range, filter the path numbers that meet the interval conditions, and obtain the path average load rate matching result set.

[0111] Based on the path switching hop count list, the load rate of the egress ports of each switching node within the path is collected. This requires extracting the egress bandwidth utilization rate of each port from the network interface statistics module corresponding to the scheduling cycle, expressed as a percentage. During the collection process, the egress port of each path segment needs to be matched with the physical location field in the path structure table to confirm that the port belongs to the egress channel on the path. Then, the load rate values ​​of the egress ports of all path segments are integrated, and a path load rate array is constructed using the path number as the key. In the example, the load rates of the six egress ports in Path-A are 58%, 62%, 66%, 64%, 61%, and 63%, respectively. The calculated average path load rate is 62.3%. According to the PFC trigger threshold, the average path load rate should be between 60% and 70%. Here, 62.3% meets the requirement, therefore path A meets the filtering criteria. The path numbers that meet the filtering criteria are summarized and recorded to obtain the path average load rate matching result set.

[0112] S413: Based on the path average load rate matching result set, retrieve the propagation delay value of each link in the corresponding path in the topology table according to the path number, read the path link transmission delay information segment by segment, and determine whether it meets the RDMA propagation delay upper limit standard. If the path propagation delay index meets the threshold condition, combine the path hop count condition and the load rate matching condition to filter the path information that meets all scheduling requirements and obtain the path filtering scheduling combination list.

[0113] Based on the path average load rate matching result set, the propagation delay information of the links involved in the path is extracted and processed. For each path, the start and end port numbers of the path segments and the corresponding propagation delay values ​​in microseconds are read from the topology delay field. A path propagation delay structure list is then established. Subsequently, the propagation delays of the links corresponding to each path are added together to form the total propagation delay index of the path. Taking Path-A as an example, the propagation delays of its six links are 2.5, 2.6, 2.3, 2.4, 2.7, and 2.3 microseconds, respectively, with a total propagation delay of 14.8 microseconds. According to the standard that the upper limit of RDMA transmission delay is set at 15 microseconds, the path is judged to be within the acceptable range. Then, the path information that meets the three conditions of hop count, load rate, and propagation delay is merged and recorded to form a path filtering structure combination dataset. The system marks the filtering results as valid scheduling path combination information and finally generates a path filtering scheduling combination list.

[0114] Please see Figure 6 The specific steps of S5 are as follows:

[0115] S511: Based on the path filtering and scheduling combination list, read the forwarding table records and MAC address mapping data of each hop switching node in the path in sequence, replace the original forwarding entries with the new forwarding entries of the target path, and establish a path switching configuration update state set.

[0116] Based on the path filtering and scheduling combination list, the node number, switch identifier, and forwarding table address of each path are extracted item by item. For each path, the node numbers involved are read sequentially, and the corresponding path forwarding table entries of the switch are queried through the path structure table to obtain the target address, priority field, and next-hop information of the current forwarding entry. Then, the update template is called to replace the old path record with the forwarding structure of the new path. The replacement items include the destination MAC address field, forwarding interface index, and matching priority. After the operation is completed, a switch instruction receipt is generated. In the example, path PA passes through switches SW01, SW07, SW12, and SW19. Its forwarding table entry is updated as follows: the next-hop address is updated from 00:1A:4B:33:22:88 to 00:1B:5C:55:44:11, the forwarding interface is changed from eth2 to eth3, and the priority is changed from 60 to 30. After the instruction is injected, the switch returns the status code "200-ACK". After the record is completed, the status is written to the status registration table, forming a structured data set according to the path number, and generating a path switching configuration update status set.

[0117] S512: Based on the path switching configuration update status set, identify all updated path numbers, construct the corresponding path migration signaling data structure, inject the constructed signaling structure into the control channel of the source switching node, send synchronization signaling to all relevant paths, if the path signaling response status is valid and the feedback success flag is written, record the corresponding path number and response flag, and summarize to obtain the path migration signaling confirmation flag table.

[0118] Based on the path exchange configuration update status set, identify the path number for which the forwarding table update has been performed, construct the signaling data body structure for each path, and set the signaling structure field content to path number, source node switch ID, target path number, migration trigger flag, signaling timestamp, and path version number. Write the signaling body to the starting switch of the corresponding path in the control channel. The signaling writing action is executed by the controller according to the path number mapping. After the distribution is completed, read the signaling response structure and determine whether it is valid if the "write flag" is 1 and the "migration response flag" is ACK. In the example, the target path number in the path PB signaling is PC, the source switch number is SW05, the trigger flag is "1", and the version number is v4. The controller reads the status code "ACK-1" in the receipt content, records the path number and the return status, and the signaling injection time is 1.42 milliseconds. Aggregate the recorded structure content by path number to generate a path migration signaling confirmation flag table.

[0119] S513: Based on the path migration signaling confirmation identifier table, perform policy configuration summary operations on all path numbers with valid signaling status identifiers, integrate them into a structured parameter set, construct a scheduling control instruction group based on path reconstruction behavior, and generate an AI data center switch traffic intelligent scheduling solution.

[0120] Based on the path migration signaling confirmation identifier table, all path numbers with valid confirmation status are aggregated and processed. For each path, the configuration update record field, migration flag status, path load statistical average, path hop count field, and migration response time value are extracted and archived according to the scheduling cycle label. The information of multiple paths in the same cycle is integrated to form a path behavior indicator set. The field structure in the indicator set is converted into scheduling policy table entries. Each policy table entry contains the source path number, update frequency, switching node distribution, forwarding interface rewrite count, scheduling cycle bandwidth index, and migration latency field, and is uniformly encoded and written into the policy cluster control system. In the example, path PB has an update frequency of 3 times, an average load rate of 62%, a hop count of 5 hops, and an average response latency of 1.35 milliseconds. Its identifier number "PB-R4" is written into the control plan table, and finally, an AI data center switch traffic intelligent scheduling solution is generated.

[0121] AI data center switch traffic intelligent scheduling device, including:

[0122] The rate jump identification module is used to perform S1: obtain the real-time rate sampling values ​​of all uplink and downlink ports of the switch in the active node of the AI ​​training task within the period, calculate the absolute difference of each port, and mark it as a jump port if it exceeds the dynamic traffic threshold, and generate a jump port identification list.

[0123] The scheduling port filtering module is used to execute S2: based on the jump port identification list, it compares the status identifier of the scheduling task item by item, constructs the set of schedulable path entry points in the current period, and generates a list of schedulable port tasks;

[0124] The tail trend recognition module is used to execute S3: based on the path to which the confirmed port belongs in the schedulable port task list, obtain the rate sampling value of the switch port at the end of the corresponding path for three consecutive cycles, mark the valid path of release status, and generate the tail release path location result;

[0125] The path optimization calculation module is used to execute S4: based on the tail release path location results, it calculates the current path switching node hop count, the average load rate of the entire path and the end-to-end propagation delay of the path, filters and schedules paths, and generates a path filtering and scheduling combination list.

[0126] The traffic scheduling execution module is used to execute S5: based on the path filtering and scheduling combination list, update the path information to the switch accordingly, send traffic migration signals to all affected paths, and generate an AI data center switch traffic intelligent scheduling solution.

[0127] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments that can be applied to other fields. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. An AI-powered intelligent traffic scheduling method for data center switches, characterized in that: Includes the following steps: S1: Obtain the real-time rate sampling values ​​of all uplink and downlink ports of the switch in the active node of the AI ​​training task within the period, calculate the absolute difference of each port, and mark it as a hopping port if it exceeds the dynamic traffic threshold, and generate a hopping port identification list. S2: Based on the jump port identification list, compare the scheduling task status identifier item by item, construct the set of schedulable path entry points in the current period, and generate a schedulable port task list; S3: Based on the confirmed path of the port in the schedulable port task list, obtain the rate sampling value of the switch port at the end of the corresponding path for three consecutive cycles, mark the valid path of release status, and generate the tail release path location result; S4: Based on the tail release path location result, count the current path switching node hop count, the average load rate of the entire path and the end-to-end propagation delay of the path, filter and schedule paths, and generate a path filtering and scheduling combination list. S5: Based on the path filtering and scheduling combination list, update the path information to the switch accordingly, send traffic migration signals to all affected paths, and generate an AI data center switch traffic intelligent scheduling scheme.

2. The AI ​​data center switch traffic intelligent scheduling method according to claim 1, characterized in that, The hop port identification list includes the hop port number, hop time, and hop magnitude; the schedulable port task list includes the scheduling status, path start identifier, and path number; the tail-end release path location result includes the path tail-end port number, load change trend, and bandwidth-delay product evaluation record; the path filtering and scheduling combination list includes the path hop count, path average load rate, and path propagation delay; and the AI ​​data center switch traffic intelligent scheduling scheme includes path update information, traffic migration signals, and target scheduling path.

3. The AI ​​data center switch traffic intelligent scheduling method according to claim 1, characterized in that, The steps for obtaining the switching port identification list are as follows: S111: Based on all uplink and downlink ports of the switch in the active node of the AI ​​training task, obtain the real-time rate sampling value of each port within the period, collect the rate sampling value sequence of the current period and the previous period, and combine the port rate sampling data with the period division interval to generate the port rate change difference sequence. S112: Based on the port rate change difference sequence, read the rate sampling data of the previous period, organize the change trend of each port, and obtain the port rate jump evaluation record. S113: Based on the port rate jump evaluation record, a judgment is made according to the dynamic traffic threshold, and the ports that meet the conditions are marked. The marked information is integrated to establish a jump port index set and generate a jump port identification list.

4. The AI ​​data center switch traffic intelligent scheduling method according to claim 1, characterized in that, The steps for obtaining the list of schedulable port tasks are as follows: S211: Based on all port numbers in the jump port identification list, extract the task status identifier of the corresponding port in the scheduling system one by one, associate and match the task status identifier with the scheduling status reference code, filter the set of port numbers with the identifier as ready, and establish a ready port number list. S212: Based on the list of ready port numbers and the path position mapping relationship between ports and nodes in the network topology table, determine whether each port in the list corresponds to the first node of the path, determine whether it is in the starting node position in the scheduling path structure, and generate a set of first-end ports of the path. S213: Based on the intersection of the set of port at the beginning of the path and the list of ready port numbers, perform a filtering operation on all port sets that meet the dual conditions of scheduling status and path structure in the current period, calculate and obtain the scheduling priority ranking coefficient, sort according to the coefficient, complete the construction of the task scheduling entry queue, and establish a list of schedulable port tasks.

5. The AI ​​data center switch traffic intelligent scheduling method according to claim 1, characterized in that, The steps for obtaining the tail release path location result are as follows: S311: Based on the confirmed path of the port in the schedulable port task list, extract the outgoing rate sampling value of three consecutive cycles from the end switch port connected to each path, record the specific rate value and corresponding time tag in each cycle, and bind the collected rate data to the port and cycle to obtain the tail continuous rate data group. S312: Based on the continuous rate data set at the tail end, analyze the rate change trend between two adjacent cycles, determine whether the rate shows a continuous periodic decrease, and compare it with the decrease range of the bandwidth delay product. If the path tail end port continuously decreases and the change falls within the decrease range of the bandwidth delay product, then record the path as a path that meets the condition of the decreasing state, and obtain the set of bandwidth decrease matching interval identifiers. S313: Based on the bandwidth decrease conformity interval identifier set, the average receiving rate, scheduling waiting time and switching hops of the path tail port in three cycles are statistically analyzed to identify and confirm the path tail release trend, determine whether the current path has a release tendency, and perform a marking operation on the path that meets the release conditions to obtain the tail release path location result.

6. The AI ​​data center switch traffic intelligent scheduling method according to claim 1, characterized in that, The steps for obtaining the path filtering and scheduling combination list are as follows: S411: Based on the tail release path location result, record the exchange node numbers that are passed sequentially from the source end to the tail end in the path structure for each path, organize all nodes in the path according to the connection order, filter unreachable nodes in combination with path connectivity information, confirm the complete node hop count path, and perform statistics and archiving according to the path number to obtain the path exchange hop count list. S412: Based on the path switching hop count list, perform load status reading operations on the switching node ports involved in each path, obtain the load occupancy of each egress port in the current period, organize them into a load rate dataset, perform segmentation merging and classification operations on the organized dataset, determine whether the overall load performance of the path is within the PFC trigger threshold range, filter the path numbers that meet the interval conditions, and obtain the path average load rate matching result set. S413: Based on the path average load rate matching result set, retrieve the propagation delay value of each link in the corresponding path in the topology table according to the path number, read the path link transmission delay information segment by segment, and determine whether it meets the RDMA propagation delay upper limit standard. If the path propagation delay index meets the threshold condition, combine the path hop count condition and the load rate matching condition to filter the path information that meets all scheduling requirements and obtain the path filtering scheduling combination list.

7. The AI ​​data center switch traffic intelligent scheduling method according to claim 1, characterized in that, The steps for obtaining the AI ​​data center switch traffic intelligent scheduling solution are as follows: S511: Based on the path filtering and scheduling combination list, sequentially read the forwarding table records and MAC address mapping data of each hop switching node in the path, replace the original forwarding entries with the new forwarding entries of the target path, and establish a path switching configuration update state set. S512: According to the path switching configuration update status set, identify all updated path numbers, construct corresponding path migration signaling data structures, inject the constructed signaling structures into the control channel of the source switching node, send synchronization signaling to all relevant paths, if the path signaling response status is valid and the feedback success flag is written, record the corresponding path number and response flag, and summarize to obtain the path migration signaling confirmation flag table. S513: Based on the path migration signaling confirmation identifier table, perform a strategy configuration summary operation on all path numbers with valid signaling status identifiers, integrate them into a structured parameter set, construct a scheduling control instruction group based on path reconstruction behavior, and generate an AI data center switch traffic intelligent scheduling scheme.

8. An AI data center switch traffic intelligent scheduling device, characterized in that, The AI ​​data center switch traffic intelligent scheduling device is used to execute the AI ​​data center switch traffic intelligent scheduling method according to any one of claims 1 to 7, and the AI ​​data center switch traffic intelligent scheduling device includes: The rate jump identification module is used to perform S1: obtain the real-time rate sampling values ​​of all uplink and downlink ports of the switch in the active node of the AI ​​training task within the period, calculate the absolute difference of each port, and mark it as a jump port if it exceeds the dynamic traffic threshold, and generate a jump port identification list. The scheduling port filtering module is used to perform S2: based on the jump port identification list, compare the scheduling task status identifier item by item, construct the set of schedulable path entry points in the current period, and generate a list of schedulable port tasks; The tail trend recognition module is used to execute S3: based on the path to which the port has been confirmed in the schedulable port task list, obtain the rate sampling value of the switch port at the end of the corresponding path for three consecutive cycles, mark the valid path of release status, and generate the tail release path location result; The path optimization calculation module is used to execute S4: based on the tail release path positioning result, to count the current path exchange node hop count, the average load rate of the entire path and the end-to-end propagation delay of the path, to filter and schedule paths, and to generate a path filtering and scheduling combination list; The traffic scheduling execution module is used to execute S5: based on the path filtering and scheduling combination list, update the path information to the switch accordingly, send traffic migration signals to all affected paths, and generate an AI data center switch traffic intelligent scheduling scheme.

9. An AI data center switch traffic intelligent scheduling device, comprising a memory and a processor, characterized in that, The memory stores a computer program, and when the processor executes the computer program, it implements the steps of the AI ​​data center switch traffic intelligent scheduling method according to any one of claims 1 to 7.

10. A computer-readable 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 AI ​​data center switch traffic intelligent scheduling method according to any one of claims 1 to 7.

Images