Computation power network-oriented burst flow deterministic transmission method and medium

Abstract

This invention relates to a deterministic transmission method and medium for bursty flows in computing power networks. The method involves clustering analysis of the real-time traffic set in the computing power network to obtain periodic traffic and aperiodic bursty flows, processing the period of each flow, establishing a dual-agent adaptive time slot allocation strategy learning mechanism, and deploying this mechanism to a centralized network controller. The centralized network controller generates a scheduling configuration table based on the real-time traffic status in the computing power network and distributes this scheduling configuration table to each switch. Each switch then executes time slot scheduling according to the scheduling configuration table distributed by the centralized network controller, achieving adaptive scheduling and deterministic transmission of bursty flows in mixed-flow scenarios. Thus, while ensuring high bandwidth throughput performance, the deterministic transmission capability of aperiodic bursty flows is improved.

Description

Technical Field

[0001] This invention relates to the field of computing power networks, and more particularly to a method and medium for deterministic transmission of burst streams in computing power networks. Background Technology

[0002] In recent years, with the widespread deployment of artificial intelligence and large-scale model training tasks in various computing centers, the traffic patterns within computing networks have become highly heterogeneous. Tasks such as uploading training datasets and distributing model files in computing networks place extremely high demands on network bandwidth and throughput, constituting high-bandwidth traffic. Meanwhile, control-related traffic, such as task scheduling, heartbeat detection, and progress feedback, is primarily latency-sensitive. Although the data volume of this type of control traffic is relatively small, it is extremely sensitive to end-to-end transmission latency and jitter, directly impacting task scheduling efficiency and user experience.

[0003] Within the internal network of a computing center, high-bandwidth-demand traffic and latency-sensitive control traffic typically share the same physical link. Large-scale training tasks generate continuous high bandwidth usage, easily leading to link queuing and cache contention. This significantly increases the transmission latency of control traffic (such as scheduling instructions, status reporting, and heartbeat feedback), and can even result in feedback loss. Furthermore, this significantly increased transmission latency and even feedback loss is particularly pronounced in multi-task concurrency or large-scale model distribution scenarios, severely restricting the availability and service reliability of the computing platform. Time-Sensitive Networking (TSN) provides deterministic transmission guarantees within the local area network to address these issues. Through time synchronization, time slot scheduling, and priority shaping mechanisms, it can achieve latency control for high-priority traffic.

[0004] However, existing time-sensitive networking (TSN) mechanisms are primarily designed for deterministic scheduling of periodic traffic. For bursty control flows (or bursts) with non-strict periodicity, such as training scheduling traffic, heartbeat traffic, and progress feedback traffic, effective bandwidth reservation and deterministic transmission remain difficult. These bursty control flows exhibit uneven temporal distribution, and their frequency and load dynamically change with the task. Pre-allocating bursty control flows strictly according to periodic TSN mechanisms leads to decreased bandwidth utilization and resource waste; conversely, leaving them uncontrolled results in contention and exclusion by bandwidth-sensitive traffic, causing scheduling delays and state lags. Furthermore, the multi-layered structure of computing networks (inter-node interconnection, exchange aggregation, and central management) further amplifies this contradiction, leading to significant differences in transmission characteristics and priority conflicts between bursty control flows and high-bandwidth-demand traffic in cross-node training or multi-task parallel scenarios.

[0005] Therefore, how to ensure high bandwidth throughput while improving the deterministic transmission of bursty control flows in the time-sensitive network architecture of computing power networks has become an urgent technical problem to be solved in the field of time-sensitive scheduling of computing power networks. Summary of the Invention

[0006] The first technical problem to be solved by the present invention is to provide a deterministic transmission method for burst streams in computing networks, which is in contrast to the above-mentioned prior art.

[0007] The second technical problem to be solved by the present invention is to provide a readable storage medium that can realize the above-mentioned burst stream deterministic transmission method for computing power networks.

[0008] The technical solution adopted by this invention to solve the first technical problem is: a deterministic transmission method for burst streams in computing power networks, characterized by comprising the following steps: Step 1: Obtain the real-time traffic set in the computing power network formed by the computing power center server cluster, and perform cluster analysis on all traffic in the real-time traffic set to classify periodic traffic and non-periodic burst traffic. Step 2: Analyze and process each periodic flow and non-periodic burst flow obtained from the classification to obtain the period of each flow; Step 3: Establish a dual-agent adaptive time slot allocation strategy learning mechanism for the scheduled flow, based on a time slot quantity decision agent and a time slot location decision agent; wherein; The time slot quantity decision agent is configured to: determine the number of time slots allocated to the flow to be scheduled in each hop based on the network status, link load, and characteristics of the flow to be scheduled in the computing power network; wherein the flow to be scheduled includes periodic traffic and aperiodic burst traffic; The time slot location decision agent is configured to: determine the time slot location of each time slot within the period of each scheduled flow, based on the time slot quantity decision agent having determined the number of time slots allocated to each scheduled flow; Step 4: Deploy the established dual-agent adaptive time slot allocation strategy learning mechanism to the centralized network controller. The centralized network controller generates a scheduling configuration table based on the real-time traffic status in the computing network. This table sets the forwarding priority of non-periodic burst traffic relative to periodic traffic for all types of traffic in the real-time traffic set. The scheduling configuration table is then distributed to each switch. Step 5: Each switch performs time slot scheduling according to the scheduling configuration table issued by the centralized network controller, and completes the joint scheduling and priority forwarding of periodic traffic and non-periodic burst traffic, so as to realize adaptive scheduling and deterministic transmission of burst traffic in mixed flow scenarios.

[0009] Improved, in the burst flow deterministic transmission method for computing power networks, the real-time traffic set includes task scheduling flow, training data flow, heartbeat feedback flow, and model distribution flow.

[0010] Furthermore, in the aforementioned deterministic transmission method for burst flows in computing power networks, step 1, which involves performing cluster analysis on all traffic data within the real-time traffic set to obtain periodic traffic and aperiodic burst flows, includes steps a1 to a6: Step a1: The multidimensional features of each traffic stream in the real-time traffic set are modeled using time-series feature extraction and cluster analysis methods, and the main statistical features of each traffic stream are calculated. Step a2: Based on the main statistical characteristics of each traffic stream, calculate the burst index of each traffic stream and the bandwidth requirement of that traffic stream. Step a3: Take the number of packets in each traffic flow within a unit time as the time series of that traffic flow, and perform autocorrelation and spectral analysis on each time series to obtain the autocorrelation function of each time series; Step a4: Perform fast Fourier transform on the autocorrelation function of the time series corresponding to each flow to extract the main spectral energy peak of each flow. Step a5: Based on the autocorrelation function and main spectral energy peaks of the time series corresponding to each flow, calculate the periodicity score index characterizing the periodicity of each flow. Step a6: Based on the obtained periodic scoring index representing the periodicity of each flow and the preset periodic scoring index threshold, determine whether each flow in the real-time flow set is a periodic flow or an aperiodic burst flow.

[0011] Furthermore, in the aforementioned deterministic transmission method for burst flows oriented towards computing power networks, step 2, which uses a weighted quantile analysis method to analyze and process the aperiodic burst flow to obtain the period of the aperiodic burst flow, includes the following steps: Step b1: Predefine the arrival interval sequence of the aperiodic burst flow within the time window; Step b2, define the set of quantiles that statistically describe the temporal distribution characteristics of aperiodic burst flows; Step b3: Calculate the dynamic weight of each quantile in the set of quantiles for the non-periodic burst flow. Step b4: Based on the obtained dynamic weights and using a weighted average method, calculate the representative interval for non-periodic bursts. Step b5: Based on the representative interval of the obtained aperiodic burst flow, calculate the real-time update estimate period of the aperiodic burst flow within the sliding window, and use the real-time update estimate period as the approximate period of the aperiodic burst flow, and use the approximate period as the period of the aperiodic burst flow. Alternatively, in the aforementioned burst flow deterministic transmission method for computing power networks, after step b5 is completed, the method further includes: Step b61: Calculate the variance of the arrival interval sequence of the non-periodic burst flow within the time window; Step b62: Determine whether the calculated variance exceeds a preset variance threshold. When the variance exceeds the preset variance threshold, a recalculation of the approximate period of the aperiodic burst is performed, and the recalculated approximate period of the aperiodic burst is used as the approximate period of the aperiodic burst, and the process proceeds to step b63; otherwise, the currently calculated approximate period of the aperiodic burst is maintained as the approximate period of the aperiodic burst, and the process proceeds to step b63. Step b63: Calculate a stability index that quantifies the confidence level of the periodicity of non-periodic bursts.

[0012] Furthermore, in the aforementioned burst flow deterministic transmission method for computing power networks, in step 3, the dual-agent adaptive time slot allocation strategy learning mechanism is established as follows: Step c1: Obtain the burst flow information and network status of each current flow to be scheduled; Step c2: Obtain the matrix feature representation of the current flow to be scheduled, and concatenate the matrix feature representation with the statistical feature vector of the flow to be scheduled to obtain the flow feature embedding of the time slot number decision agent corresponding to the flow to be scheduled. Step c3: After the time slot quantity decision agent allocates the number of time slots to the flow to be scheduled, the time slot location decision agent outputs the time slot location once according to the flow feature embedding of the corresponding flow to be scheduled. After updating the link state of the flow to be scheduled, the flow feature embedding of the flow to be scheduled is extracted again to output the time slot location. Step c4: Based on the number of time slots allocated to the corresponding scheduled flow by the time slot position decision agent and the time position of each time slot allocated to the scheduled flow within one cycle of the scheduled flow, a scheduling strategy for the scheduled flow is formed.

[0013] Further improvements include, in step 3, the method for deterministic transmission of burst flows for computing power networks further includes: adopting a global reward convergence strategy to optimize the scheduling strategy for each flow to be scheduled, thereby forming a stable scheduling strategy.

[0014] Improved, in the aforementioned burst flow deterministic transmission method for computing power networks, step 4, the process by which the centralized network controller generates a scheduling configuration table based on the real-time traffic status in the computing power network, includes the following steps: Step d1: The centralized network controller obtains the scheduling policies of all flows to be scheduled, and forms a scheduling policy set from all the obtained scheduling policies; Step d2: The centralized network controller performs unified integration processing on the formed scheduling policy set to obtain a link-level scalable time slot-stream mapping structure. Step d3: The centralized network controller generates a time-triggered gating list based on the obtained link-level scalable time slot-stream mapping structure. Step d4: The centralized network controller performs queue binding and priority configuration for all scheduled flows, so as to map each scheduled flow to different priority queues according to the policy, and bind non-periodic burst flows to the time-triggered gating list of high-priority queues.

[0015] Furthermore, in the aforementioned burst flow deterministic transmission method for computing power networks, step 5, in which each switch performs time slot scheduling according to the scheduling configuration table issued by the centralized network controller, includes the following steps: Step e1: The switch refreshes its own configured time-triggered gating list according to the newly received scheduling configuration table to update the on / off status of the corresponding time window. Step e2: The switch rebuilds the queue scheduler according to the received scheduling configuration table to ensure that high-priority periodic time-triggered gating can be opened on time. Step e3: The switch, according to the received scheduling configuration table and flow table, correctly classifies, queues and schedules different aperiodic burst flows based on the path. In step e4, the switch executes synchronization mode to synchronize the clock and period, and switches to the new scheduling plan.

[0016] Further improvements include, in this invention, the burst stream deterministic transmission method for computing power networks further comprising: The centralized network controller periodically collects the network status of the computing network and generates feedback information based on this network status. The collected network status includes time slot occupancy, actual flow latency, and burst density changes. The feedback information includes the current network link occupancy rate, time slot collision rate, end-to-end latency, and burst flow update statistics. The centralized network controller inputs the feedback information generated to the dual agent in real time to guide the dual agent to update the scheduling strategy in the execution of a new round of dual agent adaptive time slot allocation strategy learning mechanism.

[0017] The second technical problem to be solved by the present invention is: a readable storage medium having a computer program stored thereon, characterized in that, when the computer program is executed by a processor, it implements any of the burst stream deterministic transmission methods for computing power networks described in the present invention.

[0018] Compared with existing technologies, the advantages of this invention are as follows: The deterministic transmission method for burst flows in computing power networks uses cluster analysis on the real-time traffic set in the acquired computing power network to obtain periodic traffic and aperiodic burst flows. Then, it analyzes and processes the data to obtain the period of each flow (including periodic and aperiodic burst flows). Then, it establishes a dual-agent adaptive time slot allocation strategy learning mechanism for the flow to be scheduled, based on the decision agent for the number of time slots and the decision agent for the time slot location. After deploying this adaptive time slot allocation strategy learning mechanism to a centralized network controller, the centralized network controller generates a scheduling configuration table based on the real-time traffic status in the computing power network. This table sets a forwarding priority for aperiodic burst flows relative to periodic traffic, and distributes the scheduling configuration table to each switch. Each switch then performs time slot scheduling according to the scheduling configuration table distributed by the centralized network controller, completing the joint scheduling and priority forwarding of periodic traffic and aperiodic burst flows, and realizing adaptive scheduling and deterministic transmission of burst flows in mixed flow scenarios. Thus, by using multi-agent reinforcement learning technology, combined with the transmission characteristics of mixed traffic within the computing center, and by sensing the dynamic behavior characteristics of different types of traffic, the system achieves near-periodic identification, adaptive time slot allocation, and deterministic enhanced scheduling of burst flows. This improves the deterministic transmission capability of non-periodic burst flows while ensuring high bandwidth throughput performance. Attached Figure Description

[0019] Figure 1 This is a flowchart illustrating the deterministic transmission method for burst streams in computing power networks according to an embodiment of the present invention. Detailed Implementation

[0020] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments.

[0021] This embodiment provides a method for deterministic transmission of bursty streams in computing power networks. See also... Figure 1 As shown, the burst stream deterministic transmission method for computing power networks in this embodiment includes the following steps: Step 1: Obtain the real-time traffic set in the computing power network formed by the computing power center server cluster, and perform cluster analysis on all traffic in the real-time traffic set to classify periodic traffic and non-periodic burst traffic; wherein, in this embodiment, the real-time traffic set includes task scheduling flow, training data flow, heartbeat feedback flow and model distribution flow; Step 2: Analyze and process each periodic flow and non-periodic burst flow obtained from the classification to obtain the period of each flow; wherein, each flow with the period obtained here is used as the flow to be scheduled in the subsequent step 3. Step 3: Establish a dual-agent adaptive time slot allocation strategy learning mechanism for the scheduled flow, based on a time slot quantity decision agent and a time slot location decision agent; wherein; The time slot number decision agent is configured to: determine the number of time slots allocated to the flow to be scheduled in each hop based on the network status, link load, and characteristics of the flow to be scheduled in the computing power network; the flow to be scheduled includes periodic traffic and aperiodic burst traffic; The time slot location decision agent is configured to: determine the time slot location of each time slot within the period of each scheduled flow, based on the time slot quantity decision agent having determined the number of time slots allocated to each scheduled flow; Step 4: Deploy the established dual-agent adaptive time slot allocation strategy learning mechanism to the centralized network configuration controller (CNC). The centralized network controller generates a scheduling configuration table based on the real-time traffic status in the computing network, which sets forwarding priority for all types of traffic within the real-time traffic set and prioritizes non-periodic burst traffic over periodic traffic. This scheduling configuration table is then distributed to each switch. For example, in this embodiment, the centralized network controller distributes the scheduling configuration table to each switch through the SDN southbound API or the TSN management interface (such as NETCONF / YANG). Switches generally comply with IEEE 802.1Qbv (time-based scheduling) or related configuration interfaces. Step 5: Each switch performs time slot scheduling according to the scheduling configuration table issued by the centralized network controller, and completes the joint scheduling and priority forwarding of periodic traffic and non-periodic burst traffic, so as to realize adaptive scheduling and deterministic transmission of burst traffic in mixed flow scenarios.

[0022] Specifically, in step 1 of this embodiment, the process of performing cluster analysis on all traffic data within the aforementioned real-time traffic set to obtain periodic traffic and non-periodic burst flow includes steps a1 to a6: Step a1 involves using time-series feature extraction and cluster analysis methods to model the multidimensional features of each traffic stream within the real-time traffic set, and calculating the main statistical features of each traffic stream; wherein, for each traffic stream within the real-time traffic set... i During the time window W Inside, traffic flow is collected separately. i arrival time series and package size sequence The main statistical characteristics of each flow include the mean and variance of its arrival interval; flow i The mean arrival interval is denoted as ,flow i The variance of the arrival interval is denoted as : , ;1≤ i ≤ I ; in, I This represents the total number of traffic entries in the real-time traffic dataset. N This represents the total number of arrival times within the arrival time series. Indicates flow rate i The k Arrival time; t k Indicates the first k Arrival time, t k-1 Indicates the first k -1 arrival time; Indicates flow rate i The k Data packet size; It should be noted that by performing time-series feature extraction and cluster analysis methods, identifiers for traffic types can be established, which can provide input features for subsequent burst flow modeling and scheduling; in this embodiment, "non-periodic burst flow" can be simply referred to as "burst flow"; for each traffic flow within the real-time traffic set, "traffic" can be simply referred to as "flow". Step a2: Based on the main statistical characteristics of each traffic stream, calculate the burst index and bandwidth requirement for each traffic stream; where, traffic... i The suddenness index is marked as B i This traffic i The required bandwidth is marked as r i : , ;in, Indicates flow rate i It is highly unpredictable; W Indicates a time window; Step a3: The number of packets per unit time for each traffic flow is taken as the time series of that traffic flow, and autocorrelation and spectral analysis are performed on each time series to obtain the autocorrelation function of each time series; wherein, the time series corresponding to any traffic flow is labeled as... N t Time series N tThe autocorrelation function is denoted as R ( t ): ; ;in, T To collect traffic i The length of the time series when the packet size sequence is given. t For delay amount, Representing time series N t The average number of packets per unit time; t represents the discrete-time index; N t+τ This represents the number of packets per unit time after a delay of τ at time t; Step a4: Perform Fast Fourier Transform on the autocorrelation function of the time series corresponding to each flow rate to extract the main spectral energy peak of each flow rate; whereby, the flow rate... i The main spectral energy peak is labeled as E( f max,i ), f max,i This indicates the frequency corresponding to the peak energy level in the main spectrum; Step a5: Based on the autocorrelation function and main spectral energy peaks of the time series corresponding to each flow, calculate the periodicity score index characterizing the periodicity of each flow; where, the index characterizing the flow... i Periodic scoring is marked as P i : ; ; in, α As weight, t * The lag behind the autocorrelation peak. R ( t * ) represents the autocorrelation function R ( t In lag t * The corresponding function value at the location; max | R ( t | indicates a lag. t The corresponding autocorrelation function within the search range R ( t The maximum absolute value, | R ( t * | indicates a lag. t * Within the search range for autocorrelation function R ( t *Take the absolute value; Step a6: Based on the obtained periodicity score index representing the periodicity of each flow and the preset periodicity score index threshold, determine whether each flow in the real-time flow set is a periodic flow or an aperiodic burst flow; wherein, when the flow... i Periodic scoring indicators P i Greater than the preset periodic scoring indicator threshold i P If the flow is periodic, it is determined to be a periodic flow; otherwise, it is determined to be a non-periodic burst flow. Flows determined to be non-periodic burst flows... i The statistical feature vector is labeled as : ; P 95 ( r This indicates the 95th percentile of bandwidth demand sorted in ascending order. J delay This indicates the limit value for traffic jitter.

[0023] Additionally, it should be noted that in step 2 of this embodiment, the process of using weighted quantile analysis to analyze the aperiodic burst flow and obtain its period includes the following steps: Step b1: Predefine the arrival interval sequence of aperiodic bursts within a time window; wherein, the aperiodic burst is denoted as... i ' The time window is marked as W Corresponding to sudden flow The arrival interval sequence is labeled as : ; This represents the total number of arrival times within the arrival interval sequence; Indicates sudden flow The k Arrival time, Indicates sudden flow The k +1 arrival time; Step b2, define the set of quantiles that statistically describe the temporal distribution characteristics of aperiodic bursts; where, aperiodic bursts... The set of quantiles is labeled as : ; q 0.25 This indicates the 25th percentile of the interval sequence. q 0.5 This represents the median of the arrival interval sequence. q0.75 This indicates the 75th percentile of the interval sequence. q 0.9 This represents the 90th percentile of the arrival interval sequence. Of course, introducing quantile statistics to describe the temporal distribution characteristics of burst flows can effectively avoid the bias of the mean caused by extremely short intervals. Step b3: Based on the quantile set of the aperiodic burst flow, calculate the dynamic weights of each quantile within that quantile set; where the quantile set... The first j The quantiles are marked as Corresponding quantiles The dynamic weights are labeled as w j ; 1≤j≤4; ; λ is the smoothing coefficient. ; Indicates non-periodic burst flow quantile set The mean of all quantiles in the middle; Step b4: Based on the obtained dynamic weights and using a weighted average method, calculate the representative interval for aperiodic bursts; whereby, for aperiodic bursts... The representative interval is marked as : Compared to traditional methods, the method used here to calculate the representative interval of non-periodic bursts is better able to reflect the rhythmic pattern of bursts in high quantile regions, making the results less sensitive to the occurrence of concentrated bursts. Step b5: Based on the representative interval of the obtained aperiodic burst flow, calculate the real-time update estimate period of the aperiodic burst flow within the sliding window, and use this real-time update estimate period as the approximate period of the aperiodic burst flow, and use this approximate period as the period of the aperiodic burst flow; wherein, the aperiodic burst flow... The approximate period is marked as : ; ; β This is the time smoothing coefficient; Indicates non-periodic burst flow i ' The approximate period within the previous sliding window; where the approximate period is... Reflecting non-periodic bursts The triggering pattern in the time domain. The real-time update estimation cycle calculation here can adapt to the situation where traffic characteristics change over time.

[0024] It should be noted that in the hybrid streaming scenario of a computing center, burst flows typically consist of small data packets such as scheduling instructions, heartbeats, and progress feedback, and their arrival time series exhibits a non-uniform distribution. Although these flows superficially lack periodicity, statistically, their arrival intervals still show a concentrated trend. Therefore, this embodiment, by performing a calculation process for the approximate period of the aforementioned aperiodic burst flows, can provide a temporal constraint reference for the reinforcement learning agent while ensuring the fidelity of burst flow characteristics, enabling it to learn stable scheduling strategies in aperiodic scenarios. In other words, by determining the approximate period of the burst flows, a time constraint basis can be provided for subsequent time slot allocation strategies.

[0025] As an improvement, in the burst flow deterministic transmission method for computing power networks in this embodiment, after step b5 is completed, the method further includes: Step b61: Calculate the variance of the arrival interval sequence of the aperiodic burst flow within the time window; where, the aperiodic burst flow... In the time window W Inner arrival interval sequence The variance is labeled as ; Step b62: Determine whether the calculated variance exceeds a preset variance threshold. When the variance exceeds the preset variance threshold, a recalculation of the approximate period of the aperiodic burst is performed, and the recalculated approximate period of the aperiodic burst is used as the approximate period of the aperiodic burst, and the process proceeds to step b63; otherwise, the currently calculated approximate period of the aperiodic burst is maintained as the approximate period of the aperiodic burst, and the process proceeds to step b63. Step b63: Calculate a stability index that quantifies the periodicity confidence of aperiodic bursts; whereby aperiodic bursts... The stability index is marked as Stability indicators The calculation method is as follows , This indicates the arrival interval sequence. The mean.

[0026] It should be noted that by executing steps b61 to b63 here, the periodic evaluation is automatically triggered, which realizes the adaptive characterization of the "dynamic periodicity" of the burst flow, so that the periodic estimation can be adjusted in real time according to the task stage and load fluctuation.

[0027] Due to the approximate period and stability indicators The setting limits the agent's time slot search space, enabling scheduling decisions to maintain flexibility while also having statistical constraints, thereby effectively improving the deterministic transmission of burst flows.

[0028] Specifically, in step 3 of this embodiment, the aforementioned dual-agent adaptive time slot allocation strategy learning mechanism is established as follows: steps c1 to c4: Step c1: Obtain burst flow information and network status for each currently scheduled flow; whereby, the currently scheduled flow... u The burst flow information is marked as , ; R For the current flow to be scheduled u The route, length For the current flow to be scheduled u Size, period For the current flow to be scheduled u The cycle, delay For the current flow to be scheduled u The latency requirement; the network state is the current flow to be scheduled. u The time slot occupancy matrix of the links traversed by the route; Step c2: Obtain the matrix feature representation of the current flow to be scheduled, and concatenate this matrix feature representation with the statistical feature vector of the flow to be scheduled to obtain the flow feature embedding corresponding to the time slot number decision agent of the flow to be scheduled; wherein, the current flow to be scheduled... u The matrix feature representation is labeled as Current flow to be scheduled u The statistical feature vector is labeled as , corresponding to the flow to be scheduled u The flow feature embedding label of the time slot number decision agent is as follows: : ; ; This represents the characteristic representation of the matrix. With statistical eigenvectors To splice or connect; Representation matrix The number of rows, ( l + d ) represents a matrix The number of columns; l d represents the dimension of the load feature vector for each integrated time slot, and d represents the statistical feature vector. Dimensions h 1 represents the characteristic of a matrix. The first integrated time slot load feature vector in the data. h 2 represents the characteristic of a matrix. The second integrated time slot load feature vector in the data. Matrix characteristic representation The first in A comprehensive time-slot load feature vector; Step c3: After the time slot quantity decision agent allocates the number of time slots to the flow to be scheduled by the time slot location decision agent, the time slot location decision agent outputs the time slot location once based on the flow feature embedding of the corresponding flow to be scheduled. After updating the link state of the flow to be scheduled, the agent extracts the flow feature embedding of the flow to be scheduled again to output the time slot location; wherein, for the flow to be scheduled... u The number of time slots is determined by the decision-making agent and assigned to the flow to be scheduled. u The number of time slots is K u ,cycle K u The set of time slot positions obtained after the output of the next time slot position is labeled as follows: O u : ; O u For the flow to be scheduled u The set of time slot locations within one period; For the flow to be scheduled u The first time slot position, For the flow to be scheduled u The second time slot position, For the flow to be scheduled u The K u Each time slot location; Step c4: Based on the number of time slots allocated to the corresponding scheduled flow by the time slot position decision agent and the time position of each time slot allocated to the scheduled flow within one cycle of the scheduled flow, a scheduling strategy is formed for the scheduled flow; wherein, for the current scheduled flow... u The scheduling policy is marked as s u : s u =[ K u , O u ].

[0029] In this embodiment, in step c2 above, the matrix feature representation of the current flow to be scheduled The methods for obtaining it are as follows: Step c21: For any candidate time slot allocated to all time slots of the current scheduled flow, traverse all possible starting time slot indices within the supercycle, and determine the set of associated time slots corresponding to that starting time slot on the flow route according to the hop-by-hop advancement rule of CQF; wherein, the set formed by all time slots allocated to the current scheduled flow is labeled as {1,2,..., G Any candidate time slot in this set is labeled as n ,1≤ n ≤ G ; G This indicates the total number of time slots in the associated time slot set; Step c22: Aggregate the current utilization of these determined associated time slot sets on each link to obtain the load metric corresponding to the starting time slot; Step c23: For any candidate time slot among all time slots allocated to the current scheduled flow, select the time slot from all starting time slots that best balances the aforementioned load metric, and use all selected time slots as the time slot pre-allocation scheme under this candidate size; wherein, the number of all selected time slots is marked as... , ; Step c24: Based on the obtained time slot pre-allocation scheme, calculate the comprehensive time slot load characteristics formed by each initial time slot triggering and propagating hop-by-hop along the route within the supercycle; wherein, for the determined Each time slot, the combined time slot load characteristics are denoted as : ; ; u 0 indicates the first n The 0th load component in the composite time slot load feature vector u 1 indicates the first n The first load component in the composite time slot load feature vector u λ-1 Indicates the first n The (λ-1)th load component in the comprehensive time slot load feature vector; R λ Represent the λ-dimensional real space; Step c25: All the obtained integrated time slot load characteristics are aggregated to obtain the aforementioned matrix feature representation; wherein, the matrix feature representation of the current flow to be scheduled... for: , Indicates by The matrix space consisting of rows and λ columns of real numbers, i.e., the matrix characteristic representation. The dimension is .

[0030] Regarding the aforementioned learning mechanism for the dual-agent adaptive time slot allocation strategy, this mechanism achieves dual optimization of burst flow scheduling at both the spatial (cross-node) and temporal (within-period) levels through the collaborative work of two complementary agents. Periodic traffic (or periodic flows) only needs to be allocated fixed time slots at each hop, while aperiodic burst flows (or burst flows) can obtain multiple dynamically distributed time slots, thus ensuring determinism while avoiding excessive bandwidth consumption. The reinforcement learning process uses end-to-end latency jitter, bandwidth utilization, and resource constraints as joint optimization objectives, and achieves adaptive learning and convergence of the time slot allocation strategy by executing a global reward convergence strategy.

[0031] This embodiment senses the high-frequency arrival times of burst flows and allocates time slots to burst flows near these high-frequency arrival times, avoiding the waste caused by allocating too many time slot resources, thereby meeting the transmission requirements of burst flows.

[0032] Periodic flows only require a fixed number of time slots to ensure transmission stability, while aperiodic burst flows, due to their irregular arrival, require dynamic adjustment of the number and location set of time slots allocated to them, in order to ensure that the aperiodic burst flows can still obtain deterministic delay guarantees in resource-constrained time-sensitive networks.

[0033] Of course, in step 3 above, the burst flow deterministic transmission method for computing power networks in this embodiment further includes: employing a global reward convergence strategy to optimize the scheduling strategy for each flow to be scheduled, thereby forming a stable scheduling strategy. The global reward convergence strategy is set as follows: R = α·R delay + β·R collision + c·R resource ; α + β + c =1; in, R This represents the global reward value, used to evaluate the overall optimization effect of the current scheduling strategy in terms of end-to-end latency, time slot conflicts, and resource consumption. R delay To determine whether the actual end-to-end latency of the scheduled flow is lower than the target value. R collision The penalty value for the scheduled flow that has a time slot conflict with other flows; R resource This is the resource usage penalty value, which is the resource waste caused by the actual transmission of the scheduled flow after allocating a time slot to the scheduled flow; α Representing numerical valuesR delay The corresponding weights β Indicates the penalty value R collision The corresponding weights c Indicates the resource usage penalty value R resource The corresponding weights.

[0034] More specifically, in step 4 of this embodiment, the process by which the centralized network controller generates a scheduling configuration table based on the real-time traffic status in the computing network includes the following steps: Step d1: The centralized network controller acquires the scheduling policies for all flows to be scheduled, and forms a scheduling policy set from all acquired scheduling policies; wherein, the scheduling policy set is labeled as... S , , ; The total number of all streams to be scheduled; Indicates the first The scheduling strategy for each flow to be scheduled; the scheduling strategy obtained by the centralized network controller here is the scheduling strategy for each flow formed in step c4. Step d2: The centralized network controller performs unified integration processing on the formed scheduling policy set to obtain a link-level scalable time slot-stream mapping structure; wherein, the unified integration processing includes: Cross-link consistency verification: The centralized network controller ensures that each hop of the end-to-end path of each scheduled flow uses a consistent periodic structure. The centralized network controller performs a consistency verification on the periodic structure used by each scheduled flow. If the verification passes, it indicates that the current periodic structure of the scheduled flow is coherent and executable on the end-to-end path. In this case, the current scheduling policy is retained, and the flow continues to the next step of capacity verification and conflict verification. Otherwise, it indicates that although the scheduled flow performs reasonably on a single hop, it is not coherent after being assembled into an end-to-end path and cannot be directly deployed. In this case, the centralized network controller performs consistency correction or determines that the current policy is invalid and triggers regeneration / rescheduling. Check if each hop can be K i Provide space for each time slot: if it can be used for K i The availability of time slots indicates that the proposed number of time slots for the scheduled flow is feasible, confirming the reserved time slot size for the scheduled flow, and proceeding to subsequent conflict checking / template generation; otherwise, it indicates that the current reserved size for the scheduled flow exceeds the carrying capacity of some links, so the number of time slots is reduced or the time slot positions are readjusted; if it still does not meet the requirements, the current scheduling strategy cannot be directly written into the link-level time slot-flow mapping structure and needs to be rolled back and replanned. Check whether the set of timeslot positions for each flow to be scheduled conflicts with the available timeslot matrix of the link: if there is no conflict, it means that the set of timeslot positions for the current flow to be scheduled can be directly written into the link-level timeslot-flow mapping structure and can continue to be used to generate the time-triggered gating list; otherwise, it means that some timeslot positions have been occupied by other flows or are incompatible with the current link idle matrix and cannot be directly written into the link-level timeslot-flow mapping structure.

[0035] Based on the stability index of each scheduled flow, the unavailable timeslot positions in the timeslot position set are subject to limited offset correction: if the stability index value is high, the offset correction is small to keep the original timeslot position unchanged as much as possible; if the stability index value is low, it is allowed to move within a larger local window to find the nearest available transmission timeslot. Convert to a link-level scheduling template: for each flow to be scheduled u Each link traversed l The centralized network controller converts the existing scheduling policy set into a link-level scalable time-slot-stream mapping structure; wherein, the converted link-level scalable time-slot-stream mapping structure is set as follows: ; T l Indicates link l The corresponding slot-stream mapping structure, f u Indicates the flow to be scheduled u Stream identifier, o uj Indicates the flow to be scheduled u The j Each time slot location; Step d3: The centralized network controller generates a time-triggered gating list based on the obtained link-level scalable time slot-stream mapping structure; wherein, the time-triggered gating list is tagged as... GCL l : GCL l ={( t , G t )| t∈[0, T c ]}; G t Indicates in time slot t A priority queue formed by the initiated or partially initiated scheduling flows; T c Indicates a unified scheduling cycle; Step d4: The centralized network controller performs queue binding and priority configuration for all scheduled flows, mapping each flow to a different priority queue according to the scheduling strategy formed in step c4, and binding non-periodic burst flows to a time-triggered gating list of high-priority queues; wherein, the centralized network controller configures queues and priorities for each scheduled flow. u The generated end-to-end stream record is denoted as Flow u ={ path u , K u , O u , q u}, path u Indicates the flow to be scheduled u The route, K u This is represented as the flow to be scheduled. u The number of time slots allocated in a cycle; q u For the flow to be scheduled u The corresponding priority or VLAN information.

[0036] In addition, in step 5 of this embodiment, the process of each switch performing time slot scheduling according to the scheduling configuration table issued by the centralized network controller includes the following steps: Step e1: The switch refreshes its own configured time-triggered gating list according to the newly received scheduling configuration table to update the on / off status of the corresponding time window. Step e2: The switch rebuilds the queue scheduler according to the received scheduling configuration table to ensure that high-priority periodic time-triggered gating can be opened on time. Step e3: The switch, according to the received scheduling configuration table and flow table, correctly classifies, queues, and schedules different aperiodic burst flows based on their paths. Specifically, the switch classifies, queues, and schedules the burst flows entering the switch based on the received flow table / rule table; where: Classification and processing: Identify which flow to be scheduled it belongs to based on the flow identifier, such as: determine the flow to be scheduled to which the packet belongs based on flow ID, VLAN, five-tuple, task category or control flow tag, etc. Queuing: Based on the priority / queue binding relationship of this flow in the scheduling configuration table, put it into the corresponding priority queue; Scheduling process: Based on the gating time window and time slot allocation results corresponding to the stream, allow it to be dequeued and sent within the specified time window; Step e4: The switch executes synchronization mode to synchronize the clock and period, and switches to the new scheduling plan. Here, clock synchronization instructs the switch to synchronize with the global time base in the network; period synchronization refers to boundary synchronization with the unified scheduling period or a super-period used in the current scheduling table. The new scheduling plan refers to the complete set of scheduling configuration results newly generated and issued by the centralized network controller in this round.

[0037] To enhance the self-learning capability of the burst flow deterministic transmission method in this embodiment and improve its performance for burst flow deterministic transmission, the burst flow deterministic transmission method for computing power networks in this embodiment further includes: The centralized network controller periodically collects the network status of the computing network and generates feedback information based on this network status. The collected network status includes time slot occupancy, actual flow latency, and burst density changes. The feedback information includes the current network link occupancy rate, time slot collision rate, end-to-end latency, and burst flow update statistics. The centralized network controller inputs the feedback information generated to the dual agent in real time to guide the dual agent to update the scheduling strategy in the execution of a new round of dual agent adaptive time slot allocation strategy learning mechanism.

[0038] This embodiment also provides a readable storage medium. Specifically, the readable storage medium stores a computer program, which, when executed by a processor, implements the aforementioned method for deterministic transmission of burst streams for computing power networks.

[0039] Although preferred embodiments of the present invention have been described in detail above, it should be clearly understood that various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for deterministic transmission of burst streams in computing power networks, characterized in that, Includes the following steps: Step 1: Obtain the real-time traffic set in the computing power network formed by the computing power center server cluster, and perform cluster analysis on all traffic in the real-time traffic set to classify periodic traffic and non-periodic burst traffic. Step 2: Analyze and process each periodic flow and non-periodic burst flow obtained from the classification to obtain the period of each flow; Step 3: Establish a dual-agent adaptive time slot allocation strategy learning mechanism for the flow to be scheduled, based on the decision agent for the number of time slots and the decision agent for the location of time slots. in; The time slot quantity decision agent is configured to: determine the number of time slots allocated to the flow to be scheduled in each hop based on the network status, link load, and characteristics of the flow to be scheduled in the computing power network; wherein the flow to be scheduled includes periodic traffic and aperiodic burst traffic; The time slot location decision agent is configured to: determine the time slot location of each time slot within the period of each scheduled flow, based on the time slot quantity decision agent having determined the number of time slots allocated to each scheduled flow; Step 4: Deploy the established dual-agent adaptive time slot allocation strategy learning mechanism to the centralized network controller. The centralized network controller generates a scheduling configuration table based on the real-time traffic status in the computing network. This table sets the forwarding priority of non-periodic burst traffic relative to periodic traffic for all types of traffic in the real-time traffic set. The scheduling configuration table is then distributed to each switch. Step 5: Each switch performs time slot scheduling according to the scheduling configuration table issued by the centralized network controller, and completes the joint scheduling and priority forwarding of periodic traffic and non-periodic burst traffic, so as to realize adaptive scheduling and deterministic transmission of burst traffic in mixed flow scenarios.

2. The method for deterministic transmission of burst streams in computing power networks according to claim 1, characterized in that, The real-time traffic set includes task scheduling stream, training data stream, heartbeat feedback stream, and model distribution stream.

3. The method for deterministic transmission of burst streams in computing power networks according to claim 2, characterized in that, In step 1, the process of performing cluster analysis on all traffic data within the real-time traffic set to obtain periodic traffic and non-periodic burst flows includes: Step a1: The multidimensional features of each traffic stream in the real-time traffic set are modeled using time-series feature extraction and cluster analysis methods, and the main statistical features of each traffic stream are calculated. Step a2: Based on the main statistical characteristics of each traffic stream, calculate the burst index of each traffic stream and the bandwidth requirement of that traffic stream. Step a3: Take the number of packets in each traffic flow within a unit time as the time series of that traffic flow, and perform autocorrelation and spectral analysis on each time series to obtain the autocorrelation function of each time series; Step a4: Perform fast Fourier transform on the autocorrelation function of the time series corresponding to each flow to extract the main spectral energy peak of each flow. Step a5: Based on the autocorrelation function and main spectral energy peaks of the time series corresponding to each flow, calculate the periodicity score index characterizing the periodicity of each flow. Step a6: Based on the obtained periodic scoring index representing the periodicity of each flow and the preset periodic scoring index threshold, determine whether each flow in the real-time flow set is a periodic flow or an aperiodic burst flow.

4. The method for deterministic transmission of burst streams in computing power networks according to claim 3, characterized in that, In step 2, the process of analyzing the aperiodic burst flow using the weighted quantile analysis method to obtain the period of the aperiodic burst flow includes the following steps: Step b1: Predefine the arrival interval sequence of the aperiodic burst flow within the time window; Step b2, define the set of quantiles that statistically describe the temporal distribution characteristics of aperiodic burst flows; Step b3: Calculate the dynamic weight of each quantile in the set of quantiles for the non-periodic burst flow. Step b4: Based on the obtained dynamic weights and using a weighted average method, calculate the representative interval for non-periodic bursts. Step b5: Based on the representative interval of the obtained aperiodic burst flow, calculate the real-time update estimate period of the aperiodic burst flow within the sliding window, and use the real-time update estimate period as the approximate period of the aperiodic burst flow, and use the approximate period as the period of the aperiodic burst flow. Alternatively, in the aforementioned burst flow deterministic transmission method for computing power networks, after step b5 is completed, the method further includes: Step b61: Calculate the variance of the arrival interval sequence of the non-periodic burst flow within the time window; Step b62: Determine whether the calculated variance exceeds a preset variance threshold. When the variance exceeds the preset variance threshold, a recalculation of the approximate period of the aperiodic burst is performed, and the recalculated approximate period of the aperiodic burst is used as the approximate period of the aperiodic burst, and the process proceeds to step b63; otherwise, the currently calculated approximate period of the aperiodic burst is maintained as the approximate period of the aperiodic burst, and the process proceeds to step b63. Step b63: Calculate a stability index that quantifies the confidence level of the periodicity of non-periodic bursts.

5. The method for deterministic transmission of burst streams in computing power networks according to claim 4, characterized in that, In step 3, the dual-agent adaptive time slot allocation strategy learning mechanism is established as follows: Step c1: Obtain the burst flow information and network status of each current flow to be scheduled; Step c2: Obtain the matrix feature representation of the current flow to be scheduled, and concatenate the matrix feature representation with the statistical feature vector of the flow to be scheduled to obtain the flow feature embedding of the time slot number decision agent corresponding to the flow to be scheduled. Step c3: After the time slot quantity decision agent allocates the number of time slots to the flow to be scheduled, the time slot location decision agent outputs the time slot location once according to the flow feature embedding of the corresponding flow to be scheduled. After updating the link state of the flow to be scheduled, the flow feature embedding of the flow to be scheduled is extracted again to output the time slot location. Step c4: Based on the number of time slots allocated to the corresponding scheduled flow by the time slot position decision agent and the time position of each time slot allocated to the scheduled flow within one cycle of the scheduled flow, a scheduling strategy for the scheduled flow is formed.

6. The method for deterministic transmission of burst streams in computing power networks according to claim 5, characterized in that, Step 3 also includes: adopting a global reward convergence strategy to optimize the scheduling strategy for each flow to be scheduled, thereby forming a stable scheduling strategy.

7. The method for deterministic transmission of burst streams in computing power networks according to claim 1, characterized in that, In step 4, the process by which the centralized network controller generates a scheduling configuration table based on the real-time traffic status in the computing network includes the following steps: Step d1: The centralized network controller obtains the scheduling policies of all flows to be scheduled, and forms a scheduling policy set from all the obtained scheduling policies; Step d2: The centralized network controller performs unified integration processing on the formed scheduling policy set to obtain a link-level scalable time slot-stream mapping structure. Step d3: The centralized network controller generates a time-triggered gating list based on the obtained link-level scalable time slot-stream mapping structure. Step d4: The centralized network controller performs queue binding and priority configuration for all scheduled flows, so as to map each scheduled flow to different priority queues according to the policy, and bind non-periodic burst flows to the time-triggered gating list of high-priority queues.

8. The method for deterministic transmission of burst streams in computing power networks according to claim 7, characterized in that, In step 5, the process of each switch performing time slot scheduling according to the scheduling configuration table issued by the centralized network controller includes the following steps: Step e1: The switch refreshes its own configured time-triggered gating list according to the newly received scheduling configuration table to update the on / off status of the corresponding time window. Step e2: The switch rebuilds the queue scheduler according to the received scheduling configuration table to ensure that high-priority periodic time-triggered gating can be opened on time. Step e3: The switch, according to the received scheduling configuration table and flow table, correctly classifies, queues and schedules different aperiodic burst flows based on the path. In step e4, the switch executes synchronization mode to synchronize the clock and period, and switches to the new scheduling plan.

9. The method for deterministic transmission of burst streams in computing power networks according to any one of claims 1 to 8, characterized in that, Also includes: The centralized network controller periodically collects the network status of the computing network and generates feedback information based on this network status. The collected network status includes time slot occupancy, actual flow latency, and burst density changes. The feedback information includes the current network link occupancy rate, time slot collision rate, end-to-end latency, and burst flow update statistics. The centralized network controller inputs the feedback information generated to the dual agent in real time to guide the dual agent to update the scheduling strategy in the execution of a new round of dual agent adaptive time slot allocation strategy learning mechanism.

10. A readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the burst flow deterministic transmission method for computing power networks as described in any one of claims 1 to 9.

Images