Data heat and topology-aware driven computing power network computing task scheduling method and system

Abstract

The application provides a data heat and topology-aware driving computing power network computing task scheduling method and system, which collects topology probe data sets and task access heat time series data sets between computing power nodes, constructs a dynamic topology delay distribution map based on the topology probe data sets, generates heat zone identification data according to the access frequency decay characteristics in the task access heat time series data sets, generates a weight combination strategy of the heat zone nodes based on the cooperative response operation of the dynamic topology delay distribution map and the heat zone identification data, constructs a heat zone scheduling channel between the computing power nodes through a topology-aware path selection operation, and migrates the computing task corresponding to the heat zone identification data to the optimal topology domain associated with the heat zone scheduling channel according to the weight combination strategy. The application can realize efficient scheduling of computing tasks and reasonable allocation of resources, effectively improve the overall performance and resource utilization of the computing power network, and meet the processing requirements of time-sensitive tasks.

Description

Technical Field

[0001] This invention relates to the field of computing task scheduling, and more specifically, to a method and system for scheduling computing tasks in a computing network driven by data popularity and topology awareness. Background Technology

[0002] With the rapid development of information technology, computing networks, as a key infrastructure supporting the efficient processing of various computing tasks, have seen increasing attention paid to task scheduling technology. Computing network task scheduling refers to the process of allocating computing tasks to appropriate computing nodes based on network status and task requirements, thereby achieving optimized resource allocation and efficient task processing. Existing technologies typically consider factors such as network topology or task access frequency, for example, selecting communication paths based on topology or allocating computing resources based on task access volume. However, traditional methods often struggle to deeply couple and collaboratively optimize the dynamic changes in network topology with the evolutionary characteristics of task access frequency, making it difficult for scheduling strategies to simultaneously meet the dual objectives of efficient network resource utilization and time-sensitive task processing requirements. Summary of the Invention

[0003] This invention provides a method and system for scheduling computing tasks in a computing network driven by data popularity and topology awareness.

[0004] In a first aspect, embodiments of the present invention provide a method for scheduling computing tasks in a computing power network driven by data popularity and topology awareness, comprising: collecting a set of topology probe data and a set of task access popularity time series data between computing power nodes, wherein the topology probe data set includes inter-node communication latency measurements and connection status identifiers; constructing a dynamic topology latency distribution map based on the topology probe data set, wherein the dynamic topology latency distribution map is used to reflect the real-time communication path cost between computing power nodes; generating hot zone identification data according to the access frequency attenuation characteristics in the task access popularity time series data set, wherein the hot zone identification data is used to mark the distribution of computing nodes for time-sensitive tasks; generating a weight combination strategy for hot zone nodes based on a collaborative response operation of the dynamic topology latency distribution map and the hot zone identification data; constructing a hot zone scheduling channel between computing power nodes through a topology-aware path selection operation; and migrating the computing tasks corresponding to the hot zone identification data to the optimal topology domain associated with the hot zone scheduling channel according to the weight combination strategy.

[0005] Secondly, embodiments of the present invention provide a computer system, including: a memory storing a computer program; and a processor for loading the computer program to implement the data heat and topology-aware driven computing network task scheduling method as described above.

[0006] This invention provides a data-driven and topology-aware computing network task scheduling method. By collecting topology probe data sets and task access heat time-series data sets between computing nodes, a dynamic topology delay distribution map reflecting the real-time communication path cost between computing nodes is constructed. Simultaneously, based on the access frequency attenuation characteristics in the task access heat time-series data set, hot zone identification data is generated to mark the distribution of computing nodes for time-sensitive tasks, achieving accurate perception of the computing network's physical state and task access characteristics. Through collaborative response operations using the dynamic topology delay distribution map and hot zone identification data, a weight combination strategy for hot zone nodes is generated, fully integrating network topology and task heat characteristics to make the scheduling strategy more aligned with the needs of practical application scenarios. By constructing hot zone scheduling channels between computing nodes through topology-aware path selection operations, and migrating the computing tasks corresponding to the hot zone identification data to the optimal topology domain associated with the hot zone scheduling channel according to the weight combination strategy, efficient scheduling of computing tasks and rational allocation of resources can be achieved, effectively improving the overall performance and resource utilization of the computing network and meeting the processing needs of time-sensitive tasks. Attached Figure Description

[0007] Figure 1 This is a flowchart of a data popularity and topology-aware computing network task scheduling method provided in an embodiment of the present invention.

[0008] Figure 2 This is a schematic diagram of the composition of a computer system provided in an embodiment of the present invention. Detailed Implementation

[0009] Please see Figure 1 The flowchart below illustrates a data popularity and topology-aware computing network task scheduling method provided by an embodiment of the present invention. This method can be executed by a computer system and includes the following steps: Step S100: Collect the topology probe data set and the task access heat time series data set between computing power nodes. The topology probe data set includes the communication latency measurement value between nodes and the connection status identifier.

[0010] The topology probe dataset describes the communication status between computing power nodes. The inter-node communication latency measurement refers to the time it takes for one node to send data to another in the computing power network, reflecting the speed of data transmission between nodes. Connection status identifiers indicate whether nodes are connected; for example, binary 0s and 1s can be used to represent disconnection and connection. The task access popularity time-series dataset records the access status of tasks at different points in time, reflecting the trend of task access popularity over time.

[0011] When collecting topology probe datasets, this can be achieved by deploying probe programs on each computing node. These probe programs periodically send probe packets to other nodes and record the time from sending to receiving a response, using this as a measurement of inter-node communication latency. Simultaneously, the presence or absence of a response determines the connection status; a response indicates a connection, while no response indicates a disconnection. For example, in a network with 10 computing nodes, node A's probe program sends a probe packet to node B every 10 seconds, records the communication latency, and determines the connection status.

[0012] To collect time-series data on task access popularity, a monitoring program can be set up at the task access entry point. When a task is accessed, the monitoring program records the access time and the task identifier. Over time, the number of times each task is accessed at different points in time can be obtained, thus forming a time-series data set on task access popularity.

[0013] Step S200: Construct a dynamic topology delay distribution map based on the topology probe dataset. This dynamic topology delay distribution map is used to reflect the real-time communication path cost between computing power nodes.

[0014] A dynamic topology latency distribution map is a graphical representation that can display the communication path cost between computing nodes in real time. The communication path cost can be understood as the cost incurred in transmitting data between nodes, which is mainly determined by the communication latency between nodes. By constructing a dynamic topology latency distribution map, the communication status between different nodes can be intuitively understood, providing a basis for task scheduling.

[0015] As one implementation method, step S200, constructing a dynamic topology delay distribution map based on the topology probe dataset, may specifically include the following steps S210~S260: Step S210: Perform spatiotemporal correlation mining on the topology probe dataset, extract the node pair communication delay fluctuation sequence in continuous time segments according to timestamp order, and convert the single time series data into a three-dimensional correlation tensor containing the time dimension through a sliding time window. Each time slice of the three-dimensional correlation tensor corresponds to the node communication state matrix of the preset time segment.

[0016] Spatiotemporal correlation mining refers to extracting temporal and spatial correlation information from topology probe datasets. Node pair communication delay fluctuation sequences refer to the changes in communication delay between two nodes within a continuous time segment. A sliding time window is a method for processing time-series data, sliding along the time axis to extract fixed-length time segments for processing each time. A three-dimensional correlation tensor is a data structure containing three dimensions: the time dimension represents different time segments, and the other two dimensions represent the communication relationships between nodes. A node communication state matrix is ​​a two-dimensional matrix used to represent the communication state between nodes within a preset time segment.

[0017] In spatiotemporal correlation mining, the topology probe dataset is first sorted according to timestamps, and then the communication delay fluctuation sequence of node pairs within consecutive time segments is extracted. For example, a 10-minute time segment is selected, and the communication delay changes between node A and node B are recorded within these 10 minutes. Next, a sliding time window is used to process these single time-series data. Assuming the sliding time window is 1 minute long and slides for 1 minute at a time, the 10-minute time segment can be divided into 10 1-minute sub-segments. For each sub-segment, the communication delay data of the node pairs is organized into a two-dimensional matrix, namely the node communication state matrix. Finally, these node communication state matrices are stacked in chronological order to form a three-dimensional correlation tensor.

[0018] Step S220: Perform dynamic time delay feature encoding on the three-dimensional correlation tensor. Extract multi-scale features from the time delay fluctuation sequence through dilated time convolution kernels with different dilation rates to generate a time delay feature tensor containing instantaneous fluctuation features and trend evolution features. The dilation rate of the dilated convolution kernel is adjusted in real time according to the fluctuation frequency of node communication time delay.

[0019] Dynamic delay feature encoding encodes data from a three-dimensional relational tensor to extract key features. Dilated temporal convolution kernels are a special type of convolution kernel that skips some elements during convolution, thus expanding the receptive field. Multi-scale feature extraction refers to extracting features from delay fluctuation sequences at different scales to obtain more comprehensive feature information. Instantaneous fluctuation features represent the rapid changes in node communication delay over a short period, while trend evolution features reflect the long-term trend of node communication delay over time.

[0020] When performing dynamic time-delay feature encoding, dilated temporal convolution kernels with different dilation rates are used to convolve the time-delay fluctuation sequence in the 3D associated tensor. For example, three different dilation rates are set: 1, 2, and 3 (for illustration only). A dilated temporal convolution kernel with a dilation rate of 1 can extract local features of the time-delay fluctuation sequence, a kernel with a dilation rate of 2 can extract features at slightly larger scales, and a kernel with a dilation rate of 3 can extract features at even larger scales. During convolution, the dilation rate of the dilated convolution kernel is adjusted in real time according to the fluctuation frequency of node communication delays. If the fluctuation frequency of node communication delays is high, a smaller dilation rate can be selected to capture instantaneous fluctuation features; if the fluctuation frequency is low, a larger dilation rate can be selected to extract trend evolution features. Finally, the extracted features are combined into a time-delay feature tensor.

[0021] Step S230: Input the time delay feature tensor into the node embedding layer of the temporal graph neural network, learn the real-time association strength between nodes through the dynamic graph attention mechanism, and generate dynamic association features of nodes. The attention coefficient of the dynamic graph attention mechanism is determined based on the joint calculation of the similarity of the current communication features of the nodes and the stability of historical association.

[0022] Temporal graph neural networks (TPNs) are neural networks used to process graph-structured data and time-series data. The node embedding layer is a component of a TPN, converting the input time-delay feature tensor into embedded representations of nodes. The dynamic graph attention mechanism is a mechanism that adaptively focuses on the strength of associations between different nodes in the graph. Node dynamic association features are feature vectors representing the real-time association between nodes. Current communication feature similarity refers to the degree of similarity between the communication features of two nodes at the current moment; historical association stability refers to the stability of the association strength between two nodes over a past period.

[0023] After inputting the time-delay feature tensor into the node embedding layer of the temporal graph neural network, the node embedding layer maps the information in the time-delay feature tensor into a low-dimensional vector space, obtaining the node's embedded representation. Then, a dynamic graph attention mechanism is used to learn the real-time association strength between nodes. Specifically, the attention coefficient of the dynamic graph attention mechanism is obtained by jointly calculating the similarity of the current communication features of nodes and the stability of historical associations. For example, cosine similarity can be used to calculate the similarity of the current communication features of nodes, and the variance of the association strength of nodes over a past period can be used to measure the stability of historical associations. The two metrics are weighted and summed to obtain the attention coefficient. Based on the attention coefficient, the features of neighboring nodes are weighted and aggregated to generate dynamic association features of nodes.

[0024] As one implementation method, in step S230, the real-time association strength between nodes is learned through a dynamic graph attention mechanism to generate dynamic association features of nodes, which may specifically include the following steps S231~S235: Step S231: Initialize the node feature matrix of the temporal graph neural network, compress the time dimension of the delay feature tensor into the channel dimension of the node feature vector through temporal attention pooling, and construct the initial dynamic representation of the node.

[0025] The node feature matrix is ​​a matrix used to represent the features of nodes, where each row corresponds to the feature vector of a node. Temporal attention pooling is a pooling method used to process time-series data. It can compress the time dimension according to the importance of the data. The initial dynamic representation of a node is the dynamic feature representation of the node after preliminary processing.

[0026] When initializing the node feature matrix of a temporal graph neural network, the initial features of each node can be set to all-zero vectors. Then, temporal attention pooling is used to compress the temporal dimension of the delay feature tensor. Temporal attention pooling assigns different weights based on the importance of each time step, weighted summing of the data in the time dimension, thereby compressing the time dimension information into the channel dimension of the node feature vector. For example, assuming the shape of the delay feature tensor is (number of time steps, number of nodes, feature dimension), after temporal attention pooling, the shape of the resulting node feature vector is (number of nodes, new feature dimension), where the new feature dimension includes the time dimension information. Finally, these node feature vectors are combined into a node feature matrix to construct the initial dynamic representation of the nodes.

[0027] Step S232: Calculate the real-time communication feature similarity between node pairs. The cosine similarity algorithm is used to measure the time delay feature tensor similarity between any two nodes. The similarity calculation window is dynamically adjusted according to the node communication frequency to generate a real-time similarity matrix.

[0028] Real-time communication feature similarity refers to the degree of similarity between the communication features of two nodes at the current moment. The cosine similarity algorithm is a general algorithm that can dynamically adjust the similarity calculation window according to the communication frequency of the nodes. The real-time similarity matrix is ​​a two-dimensional matrix used to represent the real-time communication feature similarity between any two nodes.

[0029] When calculating the real-time communication feature similarity between node pairs, the cosine similarity algorithm is used to calculate the time delay feature tensors of any two nodes. The cosine similarity algorithm dynamically adjusts the similarity calculation window based on the node communication frequency. For example, if the node communication frequency is high, the similarity calculation window can be set smaller to capture rapid changes in node communication features; if the communication frequency is low, the similarity calculation window can be set larger to obtain more comprehensive feature information. For each pair of nodes, the cosine similarity of their time delay feature tensors is calculated, and the result is stored in the corresponding position in the real-time similarity matrix.

[0030] Step S233: The real-time similarity matrix is ​​fused with the node historical association stability coefficient to generate a dynamic attention coefficient. The historical association stability coefficient is calculated by the variance of the association strength of the node within the past preset time segment.

[0031] The dynamic attention coefficient is a coefficient used to measure the strength of the association between nodes. It is obtained by fusing the real-time similarity matrix with the historical association stability coefficient of the nodes. The historical association stability coefficient is obtained by calculating the variance of the association strength of the node pair in the past preset time segment. The smaller the variance, the more stable the association strength.

[0032] When fusing the real-time similarity matrix and the node historical association stability coefficient, a weighted summation method can be used. For example, each element in the real-time similarity matrix is ​​multiplied by a weight, and the node historical association stability coefficient is also multiplied by a weight. These are then summed to obtain the dynamic attention coefficient. The weights can be adjusted according to the specific application scenario to balance the influence of real-time similarity and historical association stability. It can be understood that during the weighted summation, the node historical association stability coefficient can first be expanded into a matrix with the same dimension as the real-time similarity matrix (in similar cases in this invention, those skilled in the art can unify the dimensions based on common knowledge in the field before performing the weighted summation). For example, for each element position in the real-time similarity matrix, the historical association stability coefficient of that node pair is placed at the corresponding position, thus forming a stability coefficient matrix. Then, each element in the real-time similarity matrix is ​​multiplied by a weight, and each element in the stability coefficient matrix is ​​also multiplied by a weight. These two weights can be adjusted according to the specific application scenario to balance the influence of real-time similarity and historical association stability on the dynamic attention coefficient. Finally, the corresponding elements of the two weighted matrices are summed to obtain the dynamic attention coefficient matrix.

[0033] Step S234: Based on the dynamic attention coefficient, the features of neighboring nodes are weighted and aggregated. The updated feature of each node is represented by the weighted sum of the features of its neighboring nodes. The weight values ​​are dynamically updated with the real-time similarity matrix to generate node aggregated features.

[0034] Neighbor node features refer to the features of nodes adjacent to the current node. Weighted aggregation refers to the weighted summation of the features of neighbor nodes based on the dynamic attention coefficient. Node aggregation features are the node features obtained after weighted aggregation.

[0035] During weighted aggregation, for each node, the features of its neighboring nodes are summed in a weighted manner based on the dynamic attention coefficient. Since the dynamic attention coefficient is dynamically updated along with the real-time similarity matrix, the updated features of the nodes also change accordingly.

[0036] Step S235: Repeat the dynamic attention aggregation process for the preset number of iterations, strengthen the high-order dynamic association features between nodes through a multi-layer temporal graph attention network, and finally output the dynamic association features of nodes that include the real-time influence ranking of nodes. The element values ​​of the feature vector are positively correlated with the communication hub role of the node in the current topology.

[0037] Dynamic attention aggregation refers to the process of weighted aggregation of neighbor node features based on dynamic attention coefficients. Multi-layer temporal graph attention networks are networks composed of multiple temporal graph attention layers, which can strengthen higher-order dynamic association features between nodes. Real-time node influence ranking refers to ranking nodes based on their role as communication hubs in the current topology.

[0038] During the preset number of iterations of the dynamic attention aggregation process, the node features are updated with each iteration. Through a multi-layered temporal graph attention network, the higher-order dynamic association features between nodes are continuously strengthened. For example, after the first iteration, the node's features will include information about its neighboring nodes; after the second iteration, the node's features will include information about the neighboring nodes' neighboring nodes, and so on. Ultimately, the output node dynamic association features contain the real-time influence ranking of nodes; the larger the value of the feature vector element, the stronger the node's role as a communication hub in the current topology.

[0039] Step S240: Based on the dynamic association characteristics of nodes, nodes with similar communication evolution trends are aggregated into dynamic topological communities using an incremental density clustering algorithm. The boundary of each topological community is adjusted in real time as the communication characteristics of the nodes change, generating dynamic topological community partitioning results.

[0040] Incremental density clustering is an algorithm that can cluster data even when it is constantly being updated. A dynamic topological community refers to a community composed of nodes with similar communication evolution trends. The boundaries of the topological community are adjusted in real time as the communication characteristics of the nodes change to ensure that the community division reflects the real-time communication situation of the nodes.

[0041] When using incremental density clustering, the density index of each node is first calculated based on its dynamic association characteristics. This density index can be determined by counting the number of neighboring nodes whose communication feature similarity exceeds a preset threshold within the current time window. For example, setting a similarity threshold of 0.8, the number of neighboring nodes with a communication feature similarity higher than 0.8 with node A within the current time window is used as the density index for node A. Then, a high-density neighbor set is determined based on the density index, and the dynamic distance from each node to these high-density neighbors is calculated. This dynamic distance considers both the Euclidean distance of the node's current time delay feature tensor and the cosine distance of the node's dynamic association features. Next, a real-time decision graph is constructed using the dynamic density-distance feature pairs. The node with the largest product of dynamic density and dynamic distance is identified as the cluster center. Finally, incremental label propagation is performed based on the cluster centers. Newly added nodes or nodes whose communication feature changes exceed a threshold are assigned to the community of the cluster center with the closest dynamic distance, and the community center features are updated in real time. This process is repeated continuously as the node communication features change, adjusting the boundaries of the topological communities in real time to generate dynamic topological community partitioning results.

[0042] As one implementation method, based on the dynamic association characteristics of nodes, in step S240, nodes with similar communication evolution trends are aggregated into dynamic topological communities using an incremental density clustering algorithm. The boundary of each topological community is adjusted in real time as the communication characteristics of the nodes change, generating dynamic topological community partitioning results. Specifically, this may include the following steps S241~S245: Step S241: Calculate the dynamic density index of each node based on the dynamic association features of the nodes. The dynamic density index is determined by counting the number of neighboring nodes whose communication feature similarity is higher than a preset threshold within the current time window. The similarity threshold is dynamically adjusted according to the stability of node communication.

[0043] The dynamic density index is used to measure the density of neighboring nodes around a node. When calculating the dynamic density index, a current time window is first defined. Within this window, the similarity of communication characteristics between each node and other nodes is calculated. Similarity can be calculated using methods such as cosine similarity. Then, a preset threshold is set, and the number of neighboring nodes with communication characteristic similarities higher than this threshold is counted. This number is used as the node's dynamic density index. The similarity threshold is dynamically adjusted based on the node's communication stability. If a node's communication stability is high, the similarity threshold can be set higher to filter out neighboring nodes with more similar communication characteristics; if communication stability is low, the similarity threshold can be set lower to obtain more information about neighboring nodes.

[0044] Step S242: Determine the high-density neighbor set of a node based on the node's dynamic association features, calculate the dynamic distance from each node to its high-density neighbors, and combine the node's current delay feature tensor Euclidean distance with the node's dynamic association feature cosine distance. Here, a high-density neighbor is a node whose dynamic density index is greater than a preset density threshold.

[0045] A high-density neighbor set refers to a set of nodes whose dynamic density index is greater than a preset density threshold. Dynamic distance is a distance metric that comprehensively considers the Euclidean distance of the node's current time delay feature tensor and the cosine distance of the node's dynamic association feature.

[0046] When determining the high-density neighbor set of a node, firstly, based on the dynamic density index calculated in step S241, nodes with dynamic density indices greater than a preset density threshold are selected, and these nodes are grouped into a high-density neighbor set. Then, for each node, the dynamic distance to each node in the high-density neighbor set is calculated. The Euclidean distance of the node's current time delay feature tensor refers to the distance between the time delay feature tensors of two nodes in Euclidean space, reflecting the differences between nodes in terms of time delay features. The cosine distance of the node's dynamic association features refers to the cosine distance between the dynamic association feature vectors of two nodes, reflecting the differences between nodes in terms of association features. The dynamic distance is obtained by weighted summing of these two distances.

[0047] Step S243: Construct a real-time decision graph using dynamic density-distance feature pairs. Identify the node with the largest product of dynamic density and dynamic distance in the real-time decision graph as the cluster center. The number of cluster centers changes dynamically with the distribution of node communication features.

[0048] Real-time decision graphs are graphical tools used to assist in cluster center selection. They are constructed using dynamic density-distance feature pairs. The node with the largest product of dynamic density and dynamic distance is selected as the cluster center because such nodes typically have high density and relatively large distance, representing different cluster regions. The number of cluster centers changes dynamically with the distribution of node communication features. If the node communication features are more dispersed, the number of cluster centers may be larger; if the distribution is more concentrated, the number of cluster centers may be smaller.

[0049] When constructing the real-time decision graph, the dynamic density and dynamic distance of each node are plotted as a feature pair on a two-dimensional plane. Then, the product of the dynamic density and dynamic distance of each node is calculated, and the node with the largest product is selected as the cluster center. As the node communication characteristics change, the real-time decision graph is continuously updated, and cluster centers are reselected.

[0050] Step S244: Incremental label propagation is performed based on cluster centers. New nodes or nodes whose communication characteristics change beyond a threshold are assigned to the community where the nearest cluster center is located. Community center characteristics are updated in real time during label propagation.

[0051] Incremental label propagation is a method for assigning cluster labels in a continuously updated data environment. When propagating incremental labels based on cluster centers, for newly added nodes or nodes whose communication characteristics change beyond a threshold, the dynamic distance from the node to each cluster center is calculated, and the node is assigned to the community containing the cluster center with the closest dynamic distance. During label propagation, the community center characteristics are updated in real time. These characteristics can be obtained by calculating the average of the characteristics of all nodes within the community. As nodes are added or their communication characteristics change, the community center characteristics are recalculated to ensure that the community center represents the overall characteristics of the nodes within the community.

[0052] Step S245: Perform boundary optimization on the initial clustering results. By calculating the average communication feature similarity of nodes within a community and the dynamic association strength of boundary nodes between communities, boundary nodes with association strength higher than the average similarity within the community are reassigned until the rate of change of similarity within the community is lower than a preset stable threshold, generating a dynamic topological community partitioning result containing community identifiers and node affiliation relationships.

[0053] Boundary optimization is a process of further adjusting the initial clustering results to improve the quality of clustering. The average communication feature similarity among nodes within a community refers to the average similarity of communication features among all nodes within the community. The dynamic association strength of boundary nodes between communities refers to the association strength between boundary nodes and other community nodes.

[0054] During boundary optimization, the average communication feature similarity of nodes within a community and the dynamic association strength of boundary nodes between communities are first calculated. Then, boundary nodes with association strength higher than the average similarity within a community are reassigned to communities with even stronger association strength. This process is repeated until the rate of change in intra-community similarity falls below a preset stability threshold. This preset stability threshold can be set according to specific application scenarios. When the rate of change in intra-community similarity falls below this threshold, it indicates that the clustering results have stabilized, and at this point, a dynamic topological community partitioning result containing community identifiers and node affiliation relationships is generated.

[0055] Step S250: Perform community evolution tracking processing on the dynamic topological community division results, calculate the topological community association matrix for different time segments, and the matrix elements represent the node overlap between communities in adjacent time segments.

[0056] Community evolution tracking is the process of tracking and analyzing the evolution of dynamic topological communities. The topological community association matrix is ​​a matrix used to represent the association relationships between topological communities in different time segments. The matrix elements represent the node overlap between communities in adjacent time segments. Node overlap refers to the ratio of the number of nodes shared by two communities in adjacent time segments to the total number of nodes in both communities.

[0057] When performing community evolution tracking, the dynamic topological community partitioning results for different time segments are first obtained. Then, for two adjacent time segments, the node overlap between each community is calculated. The calculation results are stored in the corresponding positions of the topological community association matrix. By analyzing the topological community association matrix, the evolutionary trend of the topological communities can be understood, such as community merging and splitting.

[0058] Step S260: Merge the topological community association matrix with the node communication state matrix to generate a dynamic topological delay distribution map that includes node-level real-time communication costs and community-level evolution trends.

[0059] A dynamic topology delay distribution map is a graphical representation that simultaneously reflects both node-level real-time communication costs and community-level evolution trends. A weighted summation method can be used to fuse the topology community association matrix and the node communication state matrix. For example, each element in the topology community association matrix is ​​multiplied by a weight, and each element in the node communication state matrix is ​​also multiplied by a weight. These are then summed to obtain the fused matrix. The elements in the fused matrix contain both real-time communication cost information between nodes and evolution trend information between communities. Finally, the fused matrix is ​​visualized to generate the dynamic topology delay distribution map.

[0060] Step S300: Generate hot zone identification data based on the access frequency decay characteristics in the task access heat time series data set. This hot zone identification data is used to mark the distribution of computing nodes for time-sensitive tasks.

[0061] The access frequency decay characteristic refers to the characteristic that the access frequency of a task gradually decreases over time. Hot zone identification data is used to mark the distribution of computing nodes for time-sensitive tasks, which can help the scheduling system quickly locate the nodes where tasks that need to be prioritized are located.

[0062] As one implementation method, step S300, generating hot zone identification data based on the access frequency attenuation characteristics in the task access heat time series data set, may specifically include the following steps S310~S360: Step S310: Perform multi-scale temporal pattern decomposition on the task access popularity time series data set. Separate the original access frequency sequence into multiple modal components and residual trend components with different center frequencies through variational mode decomposition. Each modal component corresponds to the access fluctuation characteristics of a preset scale.

[0063] Variational mode decomposition is a method used to decompose a signal into multiple modal components. It can separate the original access frequency sequence into multiple modal components with different center frequencies and a residual trend component. The modal component refers to the component with a preset frequency and amplitude. Each modal component corresponds to the access fluctuation characteristics of a preset scale. The residual trend component represents the long-term trend in the original access frequency sequence.

[0064] When performing multi-scale temporal pattern decomposition, variational mode decomposition (VMD) is used to process the original access frequency sequence in the task access popularity time series dataset. VMD iteratively decomposes the original sequence into multiple modal components and one residual trend component. For example, an access frequency sequence containing 100 time points is decomposed into 5 modal components and 1 residual trend component. Each modal component has a different center frequency and amplitude, corresponding to access fluctuation characteristics at different scales.

[0065] Step S320: Extract the periodic characteristics of each modal component, calculate the instantaneous frequency and instantaneous amplitude of the modal component through adaptive Fourier decomposition, and generate a periodic feature vector containing the periodic change rate, amplitude modulation coefficient and phase drift.

[0066] Adaptive Fourier decomposition is a method that adaptively performs Fourier decomposition on a signal, calculating the instantaneous frequency and amplitude of the modal components. The rate of change of period represents how the period of the modal component changes over time, the amplitude modulation coefficient describes the modulation depth of the amplitude, and the phase shift represents the phase offset of the modal component relative to the reference mode.

[0067] As one implementation method, step S320 involves extracting the periodic characteristics of each modal component and calculating the instantaneous frequency and instantaneous amplitude of the modal components through adaptive Fourier decomposition. Specifically, this may include the following steps S321 to S325: Step S321: Perform Hilbert transform on each modal component to generate an analytic signal. Calculate the instantaneous phase of the modal component using the amplitude of the analytic signal. The derivative of the instantaneous phase with respect to time is the instantaneous frequency.

[0068] The Hilbert transform is a method for converting real signals into analytic signals. An analytic signal is a complex signal that contains both the amplitude and phase information of the original signal. The instantaneous phase is the argument of the analytic signal, and its change over time reflects the phase change of the signal. The instantaneous frequency is the derivative of the instantaneous phase with respect to time, representing the frequency of the signal at each instant.

[0069] When performing a Hilbert transform on each modal component, the modal component is used as input, and the analytic signal is obtained through the Hilbert transform. Then, the phase angle of the analytic signal is calculated to obtain the instantaneous phase of the modal component. The time derivative of the instantaneous phase is taken to obtain the instantaneous frequency. For example, for a modal component x(t), the analytic signal z(t) = x(t) + j × H[x(t)] is obtained after the Hilbert transform, where H[x(t)] is the Hilbert transform of x(t). The phase angle θ(t) of z(t) is calculated, then the instantaneous phase is θ(t), and the instantaneous frequency is dθ(t) / dt.

[0070] Step S322: Extract the amplitude envelope of the modal component using the sliding window peak detection algorithm. Generate the upper and lower envelopes by fitting the local maxima of the amplitude within the window. The mean of the upper and lower envelopes is used as the instantaneous amplitude of the modal component.

[0071] The sliding window peak detection algorithm is used to detect peak values ​​in a signal by sliding a fixed-length window to find local maxima within the window. The amplitude envelope refers to the curve showing the change in signal amplitude, and can be divided into an upper envelope and a lower envelope. The upper envelope is a curve fitted to the signal's local maxima, and the lower envelope is a curve fitted to the signal's local minima. The instantaneous amplitude is the average of the upper and lower envelopes, reflecting the magnitude of the signal's amplitude at each moment.

[0072] When extracting the amplitude envelope of modal components using the sliding window peak detection algorithm, a sliding window length is set, for example, 10 time points. Within each window, local maxima and local minima of the amplitude are found. All local maxima are connected to form an upper envelope; all local minima are connected to form a lower envelope. Then, the mean of the upper and lower envelopes is calculated to obtain the instantaneous amplitude of the modal component.

[0073] Step S323: Calculate the time derivative of the instantaneous frequency to obtain the frequency change rate. Identify the periodic stability of the modal components through the frequency change rate. The smaller the absolute value of the frequency change rate, the more stable the period.

[0074] The rate of change of frequency refers to the rate of change of instantaneous frequency over time, and it can be obtained by taking the time derivative of the instantaneous frequency. The smaller the absolute value of the rate of change of frequency, the more stable the period of the modal component, because the period changes less.

[0075] When calculating the rate of change of frequency, the instantaneous frequency obtained in step S321 is differentiated. For example, a numerical differentiation method, such as the finite difference method, is used to calculate the difference in instantaneous frequency between two adjacent time points, and then divided by the time interval to obtain the rate of change of frequency. The periodic stability of the modal component is determined based on the magnitude of the absolute value of the rate of change of frequency.

[0076] Step S324: Calculate the amplitude modulation coefficient based on the instantaneous amplitude. The amplitude modulation coefficient is the ratio of the instantaneous amplitude to the average amplitude and is used to describe the modulation depth of the amplitude.

[0077] The amplitude modulation coefficient is an indicator used to describe the depth of amplitude modulation. It is obtained by dividing the instantaneous amplitude by the average amplitude. The average amplitude can be obtained by calculating the average amplitude of the modal components over a period of time.

[0078] When calculating the amplitude modulation coefficient, the average amplitude of the modal component over a period of time is first calculated. Then, for the instantaneous amplitude at each moment, it is divided by the average amplitude to obtain the amplitude modulation coefficient. The larger the amplitude modulation coefficient, the greater the modulation depth of the amplitude.

[0079] Step S325: Align the instantaneous phases of different modal components, calculate the phase drift of each modal component relative to the reference mode, and generate a periodic feature vector containing the periodic change rate, amplitude modulation coefficient, and phase drift.

[0080] Aligning the instantaneous phases of different modal components is to make their phases comparable. The reference mode is a selected modal component used as a benchmark. Phase drift refers to the phase offset of each modal component relative to the reference mode.

[0081] When aligning the instantaneous phases of different modal components, a reference mode can be selected, and the instantaneous phases of the other modal components can be adjusted to align with the phase of the reference mode. Then, the phase drift of each modal component relative to the reference mode is calculated. The periodicity rate of change, amplitude modulation coefficient, and phase drift are combined into a vector, namely the periodic eigenvector.

[0082] Step S330: Input the periodic feature vector into the gated recurrent unit network to learn the nonlinear decay law of the access frequency over time. During the network training process, the decay parameters are dynamically updated by real-time access data, and the heat prediction features are output.

[0083] Gated recurrent unit (GRU) networks are a type of recurrent neural network capable of handling long-term dependencies in sequential data. When a periodic feature vector is input into the GRU, the network learns a non-linear decay pattern in access frequency over time. During network training, the decay parameters are dynamically updated using real-time access data to ensure the network adapts to changes in access frequency. Ultimately, the network outputs a popularity prediction feature, which can be used to predict the popularity of accesses for a given task.

[0084] The Gated Recurrent Unit (GRU) network consists of an input layer, hidden layers, and an output layer. The input layer receives periodic feature vectors, and the hidden layers contain multiple GRUs that update the current hidden state based on the input and the hidden state from the previous time step. The output layer outputs a heat prediction feature based on the hidden state. During training, the gradient of the loss function is calculated using the backpropagation algorithm, and the network parameters are updated based on the gradient.

[0085] Step S340: Calculate the dynamic coefficient of heat sensitivity based on the heat prediction characteristics. The coefficient is determined by a dynamic weighted combination of the cumulative absolute value of the deviation between the predicted heat and the actual heat and the rate of change of the deviation. The weights are dynamically adjusted with the prediction error.

[0086] The dynamic coefficient of heat sensitivity is a coefficient used to measure the sensitivity of a task to heat. It is determined by a dynamic weighted combination of the cumulative absolute value of the deviation between predicted and actual heat and the rate of change of the deviation. The predicted heat is obtained from the heat prediction features output by the gated recurrent unit network, while the actual heat is calculated based on the actual access data in the task access heat time series dataset. The weights are dynamically adjusted according to the prediction error. If the prediction error is large, the weights will be more biased towards the rate of change of the deviation; if the prediction error is small, the weights will be more biased towards the cumulative absolute value of the deviation. When calculating the dynamic coefficient of heat sensitivity, the cumulative absolute value of the deviation between predicted and actual heat and the rate of change of the deviation are first calculated. The cumulative absolute value of the deviation refers to the sum of the absolute values ​​of the deviations between predicted and actual heat over a period of time, and the rate of change of the deviation refers to the rate of change of the deviation between two adjacent time points. Then, the weights are dynamically adjusted according to the prediction error, and the cumulative absolute value of the deviation and the rate of change of the deviation are weighted and summed to obtain the dynamic coefficient of heat sensitivity.

[0087] Step S350: Construct a heat diffusion model based on the heat conduction equation, using the heat sensitivity dynamic coefficient as the heat source intensity and the inter-node communication delay as the heat conduction resistance, to simulate the real-time heat diffusion process between computing nodes and generate a heat diffusion feature map.

[0088] The heat conduction equation describes the process of heat conduction in a medium and can be used to simulate the diffusion of heat between computing nodes. The heat diffusion model is derived by applying the heat conduction equation to a computing network, using a heat-sensitive dynamic coefficient as the heat source intensity and inter-node communication delay as the heat conduction resistance. A heat diffusion feature map is a graphical representation used to illustrate the diffusion of heat between computing nodes.

[0089] As one implementation method, step S350 involves constructing a heat diffusion model based on the heat conduction equation, using the heat sensitivity dynamic coefficient as the heat source intensity and the inter-node communication delay as the heat conduction resistance, to simulate the real-time heat diffusion process between computing nodes and generate a heat diffusion feature map. Specifically, this may include the following steps S351~S355: Step S351: Abstract the computing power node as a discrete particle in the heat conduction medium, use the node's heat sensitivity dynamic coefficient as the initial heat source intensity, and use the node's computing resource utilization rate as the heat source attenuation coefficient.

[0090] When abstracting computing nodes as discrete particles in a heat conduction medium, each node is considered as a particle with a certain heat source intensity and attenuation coefficient. The node's heat sensitivity dynamic coefficient represents its ability to generate heat and is used as the initial heat source intensity. The node's computing resource utilization rate represents the proportion of resources consumed by the node during computation and is used as the heat source attenuation coefficient, because the higher the resource utilization rate, the faster the heat generated by the node may be consumed.

[0091] Step S352: Construct a heat conduction resistance matrix between nodes. The matrix elements are the product of the communication delay between nodes and the node load rate. The greater the communication delay or the higher the node load rate, the greater the heat conduction resistance.

[0092] The thermal conduction resistance matrix is ​​a matrix used to represent the resistance to thermal conduction between nodes. Matrix elements are the products of inter-node communication delay and node load rate. Inter-node communication delay represents the time required for data transmission between nodes, while node load rate represents the current load on a node. A higher communication delay or a higher node load rate indicates greater thermal conduction resistance, meaning that heat conduction between nodes is more difficult.

[0093] When constructing the heat conduction resistance matrix, for each pair of nodes, the product of their communication delay and node load rate is calculated, and the result is stored in the corresponding position of the matrix.

[0094] Step S353: Establish a partial differential equation for heat diffusion based on the heat conduction equation. The time term of the equation represents the decay of heat over time, and the spatial term represents the diffusion of heat between nodes. The equation coefficients are dynamically calibrated using historical heat diffusion data of the nodes.

[0095] The heat conduction equation is a partial differential equation that describes the heat conduction process in a medium. When establishing a partial differential equation for heat diffusion based on the heat conduction equation, the time term represents the decay of heat over time, and the spatial term represents the diffusion of heat between nodes. The equation coefficients are dynamically calibrated using historical heat diffusion data from the nodes to ensure that the equation accurately describes the heat diffusion process. The general form of the partial differential equation for heat diffusion is: Where u represents heat, t represents time, α is the thermal diffusivity coefficient, β is the heat source attenuation coefficient, and Q is the heat source intensity. The values ​​of α, β, and Q are dynamically calibrated using historical heat diffusion data of the nodes.

[0096] Step S354: Solve the partial differential equation of heat diffusion using the finite difference method to obtain the distribution of nodal heat values ​​at each time step. The time step is dynamically adjusted according to the heat diffusion rate.

[0097] The finite difference method (FDM) is a numerical method for solving partial differential equations. It approximates the derivatives of partial differential equations using finite differences, thus transforming the partial differential equations into algebraic equations for solution. In solving the partial differential equation for heat diffusion using the FDM, time and space are discretized, transforming the partial differential equation into a difference equation. Then, the difference equation is solved iteratively to obtain the nodal heat distribution at each time step.

[0098] The time step is dynamically adjusted based on the rate of heat diffusion. If the heat diffusion rate is fast, the time step can be set to a smaller value to ensure the accuracy of the solution; if the diffusion rate is slow, the time step can be set to a larger value to improve the solution efficiency.

[0099] Step S355: Monitor the diffusion process through the heat diffusion stability criterion. When the rate of change of heat value in adjacent time steps is lower than the preset stability threshold, stop the diffusion simulation and output a heat diffusion feature map containing the real-time heat value of each node and the diffusion direction.

[0100] The heat diffusion stability criterion is used to determine whether the heat diffusion process is stable. When the rate of change of heat values ​​between adjacent time steps is lower than a preset stability threshold, it indicates that the heat diffusion process has stabilized, and the diffusion simulation is stopped. The heat diffusion feature map contains the real-time heat values ​​and diffusion directions of each node, which can intuitively show the diffusion of heat among computing nodes. When monitoring the diffusion process, the rate of change of heat values ​​between adjacent time steps is calculated. If the rate of change is lower than the preset stability threshold, the simulation is stopped, and the current node heat values ​​and diffusion direction information are stored in the heat diffusion feature map.

[0101] Step S360: The heat zone boundary is identified by the heat diffusion feature map through the dynamic region growth algorithm. The nodes with heat values ​​higher than the preset dynamic threshold are used as seed points. The region is grown according to the heat gradient direction to generate heat zone identification data containing the real-time boundary of the heat zone, the heat diffusion speed and the core heat zone nodes.

[0102] Dynamic region growing (MRG) is an algorithm used for image segmentation that merges adjacent pixels into a single region according to certain rules. When identifying hot zone boundaries in a heat spread feature map, nodes with heat values ​​higher than a preset dynamic threshold are used as seed points, and regions are grown along the heat gradient direction. The heat gradient direction represents the direction of the fastest increase in heat; adjacent nodes are merged into the hot zone along this direction until a certain stopping condition is met.

[0103] When using the dynamic region growing algorithm, a preset dynamic threshold is first determined, and nodes with heat values ​​higher than this threshold are selected as seed points. Then, starting from the seed points, adjacent nodes are searched along the heat gradient direction, and nodes whose heat values ​​meet certain conditions are merged into the hot zone. This process is repeated until no nodes that meet the conditions can be merged. Finally, hot zone identification data containing real-time hot zone boundaries, heat diffusion rates, and core hot zone nodes is generated.

[0104] Step S400: Generate a weight combination strategy for hot zone nodes based on the collaborative response operation of the dynamic topology delay distribution map and hot zone identification data. Construct a hot zone scheduling channel between computing power nodes through topology-aware path selection operation. Then, according to the weight combination strategy, migrate the computing tasks corresponding to the hot zone identification data to the optimal topology domain associated with the hot zone scheduling channel.

[0105] Collaborative response operation refers to taking into account the dynamic topology delay distribution map and hot zone identification data, and performing corresponding operations to generate a weight combination strategy for hot zone nodes. Topology-aware path selection operation refers to selecting appropriate paths to construct hot zone scheduling channels between computing nodes based on the dynamic topology delay distribution map and hot zone identification data. Optimal topology domain refers to the optimal region within the topology domain associated with the hot zone scheduling channel that can meet the computing task requirements.

[0106] As one implementation method, step S400, which generates a weight combination strategy for hot zone nodes based on the collaborative response operation of the dynamic topology delay distribution map and hot zone identification data, may specifically include the following steps S410~S450: Step S410: Input the node communication characteristics of the dynamic topology delay distribution map and the heat diffusion characteristics of the hot zone identification data into the spatiotemporal fusion network, and learn the real-time correlation pattern between the two through the spatiotemporal cross-attention mechanism to generate topology-heat correlation features.

[0107] Spatiotemporal fusion networks are neural networks capable of fusing spatiotemporal information, integrating node communication characteristics from dynamic topology delay distribution maps with heat spread characteristics from hotspot identification data. The spatiotemporal cross-attention mechanism is a mechanism that focuses on the correlations between different features in both temporal and spatial dimensions, learning their real-time correlation patterns. The topology-heat correlation feature, learned through the spatiotemporal fusion network, represents the relationship between the dynamic topology delay distribution map and the hotspot identification data.

[0108] The spatiotemporal fusion network is used to fuse the node communication characteristics of a dynamic topology delay distribution map with the heat diffusion characteristics of hotspot identification data to learn the real-time correlation patterns between the two. In one possible network architecture, the input data is first preprocessed. The node communication characteristics of the dynamic topology delay distribution map are converted into a topology feature sequence. The feature vector at each time step contains the current communication cost and historical communication trend of the node. The change rate of communication cost data over a past period can be calculated using a sliding window as a historical communication trend. The heat diffusion characteristics of the hotspot identification data are converted into a heat feature sequence. The feature vector at each time step contains the real-time heat value and diffusion speed of the node. The diffusion speed is obtained by calculating the change in heat value between adjacent time steps. Next, the spatiotemporal cross-attention module is entered. This module calculates the cross-correlation coefficients of the topology feature sequence and the heat feature sequence in the time and spatial dimensions. For example, the Pearson correlation coefficient is used to measure the correlation between the two sequences at each time step and spatial location, generating a spatiotemporal attention weight tensor. This tensor is then used to weight and fuse the topology feature sequence and the heat feature sequence, enhancing the weights of features with high spatiotemporal correlation, and generating a preliminary fused feature sequence. The subsequent multi-scale feature extraction module uses combinations of different temporal and spatial convolutional kernels. For example, a temporal convolutional kernel of size 3 and a spatial convolutional kernel of size 2 are used to extract short-term local association features, while a temporal convolutional kernel of size 10 and a spatial convolutional kernel of size 5 are used to extract long-term global association features. Finally, the multi-scale association features are input into a bidirectional long short-term memory network for temporal modeling. This network consists of forward and backward long short-term memory units and can learn the evolution of topology-heat association patterns over time, outputting topology-heat association features that include real-time association strength.

[0109] As one implementation method, step S410 involves inputting the node communication characteristics of the dynamic topology delay distribution map and the heat diffusion characteristics of the hot zone identification data into the spatiotemporal fusion network, and learning the real-time correlation pattern between the two through a spatiotemporal cross-attention mechanism to generate topology-heat correlation features. Specifically, this may include the following steps S411~S416: Step S411: Convert the node communication characteristics of the dynamic topology delay distribution map into a topology feature sequence. The feature vector of each time step contains the current communication cost and historical communication trend of the node.

[0110] When converting the node communication characteristics of a dynamic topology delay distribution map into a topology feature sequence, the node communication characteristics at each time step are organized into a feature vector. The feature vector contains the current communication cost of a node and its historical communication trend. The current communication cost can be represented by the communication delay between nodes, and the historical communication trend can be represented by the changes in the node communication cost over a past period. For example, the average value and rate of change of the node communication cost over the past 10 time steps can be calculated, and this information can be used as part of the historical communication trend.

[0111] Step S412: Convert the heat diffusion characteristics of the hot zone identification data into a heat feature sequence. The feature vector of each time step contains the real-time heat value and diffusion speed of the node.

[0112] The heat dissipation characteristics of hotspot identification data include the real-time heat value and dissipation speed of nodes. When converting this into a heat feature sequence, the heat dissipation characteristics for each time step are organized into a feature vector. This feature vector contains the real-time heat value and dissipation speed of a node. The real-time heat value refers to the heat value of the node within the current time step, and the dissipation speed refers to the rate at which heat spreads between nodes.

[0113] Step S413: Calculate the cross-correlation coefficients of the topological feature sequence and the popularity feature sequence in the time and spatial dimensions, and generate the spatiotemporal attention weight tensor.

[0114] The cross-correlation coefficient is a metric used to measure the correlation between two sequences. It can calculate the correlation between a topological feature sequence and a popularity feature sequence in both time and spatial dimensions. The spatiotemporal attention weight tensor is a tensor used to represent the strength of the association between different features; the larger the element value, the stronger the association between the two features.

[0115] When calculating the cross-correlation coefficient, methods such as the Pearson correlation coefficient can be used to calculate the cross-correlation coefficient for each pair of time steps and nodes of the topological feature sequence and the heat feature sequence, and store the results in the corresponding position of the spatiotemporal attention weight tensor.

[0116] Step S414: The topological feature sequence and the popularity feature sequence are weighted and fused using the spatiotemporal attention weight tensor to generate a preliminary fused feature sequence, in which the weights of features with high spatiotemporal correlation are enhanced.

[0117] When weighted fusion of topological feature sequences and popularity feature sequences, elements in the spatiotemporal attention weight tensor are used as weights, and a weighted sum is performed on each element in both sequences. Features with high spatiotemporal correlation are weighted more strongly because they have larger element values ​​in the spatiotemporal attention weight tensor. Ultimately, a preliminary fused feature sequence is obtained.

[0118] Step S415: Perform multi-scale feature extraction on the preliminary fused feature sequence. Extract short-term local correlation features and long-term global correlation features by combining different temporal and spatial convolution kernels.

[0119] Multi-scale feature extraction refers to extracting features from a pre-fused feature sequence at different scales to obtain more comprehensive feature information. Combinations of different temporal and spatial convolutional kernels can extract short-term local correlation features and long-term global correlation features. Short-term local correlation features represent the correlation between features in a short time and a local region, while long-term global correlation features represent the correlation between features over a long time and on a global scale.

[0120] When performing multi-scale feature extraction, convolutional operations are performed on the initially fused feature sequence using temporal and spatial convolutional kernels of different sizes. For example, using a temporal convolutional kernel of size 3 and a spatial convolutional kernel of size 2 can extract short-term local correlation features; using a temporal convolutional kernel of size 10 and a spatial convolutional kernel of size 5 can extract long-term global correlation features.

[0121] Step S416: Input the multi-scale association features into the bidirectional long short-term memory network for time-series modeling, learn the evolution law of topology-heat association pattern over time, and output the topology-heat association features containing real-time association strength.

[0122] Bidirectional Long Short-Term Memory (BSSM) networks are neural networks capable of handling long-term dependencies in sequential data, learning the evolution of topological-heat association patterns over time. When multi-scale association features are input into a BSSM network, the network predicts future topological-heat association patterns based on the input feature sequence. Ultimately, the output is a topological-heat association feature containing real-time association strengths.

[0123] Bidirectional Long Short-Term Memory (LSTM) networks consist of forward and backward LSM units, which can process the forward and backward information of a sequence, respectively. During training, the gradient of the loss function is calculated using the backpropagation algorithm, and the network parameters are updated based on the gradient.

[0124] Step S420: With the optimization objectives of minimizing the communication path cost reflected by the dynamic topology delay distribution map and maximizing the heat response speed reflected by the hot zone identification data, a dual-objective dynamic optimization framework is constructed, and the optimization variable is the resource allocation weight of the hot zone node.

[0125] The dual-objective dynamic optimization framework is a framework for solving problems with two optimization objectives, allowing for adjustments to optimization variables in a dynamic environment. In constructing this framework, the optimization objectives are minimizing the communication path cost reflected in the dynamic topology delay distribution map and maximizing the heat response speed reflected in the hot zone identification data. The optimization variable is the resource allocation weight of the hot zone nodes, representing the proportion of resources allocated to them.

[0126] Step S430: Solve the bi-objective dynamic optimization framework using an adaptive multi-objective evolutionary algorithm to generate an initial weight allocation scheme. The crossover probability and mutation probability of the adaptive multi-objective evolutionary algorithm are dynamically adjusted according to the convergence state of the current population. The convergence state is determined by calculating the distribution entropy of the objective function values ​​of individuals in the population.

[0127] Adaptive multi-objective evolutionary algorithms are algorithms that adaptively adjust parameters during the evolutionary process, solving bi-objective dynamic optimization frameworks. Crossover probability and mutation probability are important parameters in evolutionary algorithms, dynamically adjusted according to the current convergence state of the population. The convergence state of the population is determined by calculating the entropy of the objective function values ​​of individuals in the population; the smaller the entropy value, the more convergent the population.

[0128] The adaptive multi-objective evolutionary algorithm improves upon the traditional multi-objective evolutionary algorithm by adaptively adjusting parameters based on the population state during evolution, thereby enhancing search efficiency and solution quality. It is used to solve bi-objective dynamic optimization frameworks to determine the resource allocation weights for hot zone nodes.

[0129] When using an adaptive multi-objective evolutionary algorithm to solve a bi-objective dynamic optimization framework, a set of initial solutions is first randomly generated as a population, with each solution representing a hot zone node resource allocation weight scheme. Then, for each individual in the population, the fitness value is calculated based on minimizing the communication path cost reflected in the dynamic topology delay distribution map and maximizing the heat response speed reflected in the hot zone identification data. Individuals are then sorted according to their fitness values, and tournament selection or similar methods are used to select individuals for the next generation. The crossover and mutation probabilities are adjusted based on the distribution entropy, and crossover and mutation operations are performed to generate new individuals. The crossover and mutation probabilities are dynamically adjusted according to the population convergence state, which is determined by calculating the distribution entropy of the objective function values ​​of the individuals in the population. A smaller entropy value indicates population convergence, at which point the crossover probability can be reduced and the mutation probability increased to enhance global search capabilities. Afterward, the new individuals are merged with their parent individuals, the fitness value is calculated again and sorted, and individuals with poor fitness are eliminated to maintain a stable population size. This process is repeated until the stopping condition is met (such as reaching the maximum number of iterations or the objective function value converging to a certain range). Finally, the optimal individual is selected as the initial weight allocation scheme for hot zone node resource allocation.

[0130] Step S440: Using the real-time task completion latency and resource waste rate of the computing network as reward signals, train a deep deterministic policy gradient agent to output the real-time adjustment of the weights of hot zone nodes for the initial weight allocation scheme.

[0131] A deep deterministic policy gradient agent is an agent based on deep reinforcement learning that can adjust its policy based on feedback from the environment. When training a deep deterministic policy gradient agent, the real-time task completion latency and resource waste rate of the computing network are used as reward signals. Real-time task completion latency represents the time required for a task to complete from start to finish, and resource waste rate represents the proportion of resources wasted during task execution. The reward signals guide the agent to adjust the real-time weights of hot-spot nodes to optimize task completion latency and resource waste rate.

[0132] The deep deterministic policy gradient agent consists of a policy network and a value network. The policy network outputs actions based on the current state, i.e., real-time adjustments to the weights of hot zone nodes; the value network evaluates the value based on the current state and actions. During training, the parameters of the policy network and the value network are updated using the deep deterministic policy gradient algorithm.

[0133] Specifically, in a deep deterministic policy gradient agent, the policy network and value network are used to learn and make decisions to optimize the real-time adjustment of the weights of hot zone nodes. The policy network is a neural network that receives current network state information, such as real-time task completion latency, resource waste rate, and topology-heat correlation features, as input. After a series of linear transformations and non-linear activation functions (e.g., ReLU), it outputs an action—the real-time adjustment of the hot zone node weights. The goal of the policy network is to learn an optimal policy that selects the action most beneficial to the optimization objective (such as reducing task completion latency and resource waste rate) under different states. The value network is also a neural network that receives the current state and the action output by the policy network as input, evaluating the value of the state-action pair. The value evaluation process involves passing and calculating the input information layer by layer through the network, ultimately outputting a scalar value representing an estimate of the long-term cumulative reward that can be obtained by taking this action in this state. During training, the policy network and the value network collaborate. The policy network continuously adjusts its output actions based on the evaluation results of the value network, while the value network continuously updates its evaluation of the value of state-action pairs based on the action choices made by the policy network. By constantly interacting with the environment, collecting reward signals, and using a deep deterministic policy gradient algorithm to calculate gradients and update network parameters, the policy network can learn better policies, and the value network can more accurately evaluate the value of state-action pairs, thereby achieving effective optimization of the real-time adjustment of the weights of hot zone nodes.

[0134] Step S450: Based on the initial weight allocation scheme and real-time adjustment amount, generate a hot zone node weight combination strategy that includes real-time node weight update rules, weight conflict arbitration mechanism and weight adjustment trigger conditions. The weight combination strategy is dynamically updated as the topology-heat correlation characteristics change.

[0135] The hot zone node weight combination strategy is a strategy that includes real-time node weight update rules, a weight conflict arbitration mechanism, and weight adjustment trigger conditions. The real-time node weight update rules specify how to update node weights based on real-time adjustments; the weight conflict arbitration mechanism is used to resolve conflicts that may arise during weight adjustments; and the weight adjustment trigger conditions specify under what circumstances weight adjustments are required.

[0136] When generating the hot zone node weight combination strategy, the initial weight allocation scheme is combined with real-time adjustments to formulate real-time node weight update rules. Simultaneously, a weight conflict arbitration mechanism is designed, which can arbitrate based on node importance and task urgency. Weight adjustment trigger conditions are determined, such as triggering weight adjustments when the topology-heat correlation characteristics change significantly. As the topology-heat correlation characteristics change, the hot zone node weight combination strategy is dynamically updated.

[0137] As one implementation method, in step S400, a hot zone scheduling channel is constructed between computing nodes through a topology-aware path selection operation, which may specifically include the following steps S460~S4100: Step S460: Using the communication path cost of the dynamic topology delay distribution map, the heat value of the hot zone identification data, and the real-time load rate of the nodes as input, construct a dynamic path evaluation function. The function output value comprehensively reflects the communication efficiency and heat adaptability of the path.

[0138] The dynamic path evaluation function is used to assess the quality of a path. Its inputs include the communication path cost from the dynamic topology delay distribution map, the heat value of the hotspot identification data, and the real-time node load rate. The function output comprehensively reflects the path's communication efficiency and heat suitability. Communication efficiency represents the speed and reliability of data transmission along the path, while heat suitability indicates whether the path can meet the heat requirements of hotspot tasks.

[0139] When constructing a dynamic path evaluation function, a weighted summation method can be used. For example, the communication path cost can be multiplied by a weight, the popularity value of the hot zone identification data can be multiplied by a weight, and the real-time load rate of the node can be multiplied by a weight. Then, these weights are added together to obtain the output value of the dynamic path evaluation function. The choice of weights can be adjusted according to the specific application scenario.

[0140] Step S470: Based on the dynamic path evaluation function, through... The algorithm searches for a set of candidate paths among hot nodes. During the search process, a dynamic heuristic function is used to accelerate the exploration of the optimal path. The weight of the heuristic function is dynamically adjusted according to the real-time communication cost.

[0141] The algorithm is used to search for the shortest path in a graph. It searches for a set of candidate paths between hot nodes based on a dynamic path evaluation function. The dynamic heuristic function is... One component of the algorithm is the heuristic function, which estimates the cost from the current node to the target node, accelerating the search for the optimal path. The heuristic function weights are dynamically adjusted based on real-time communication costs to ensure that the heuristic function can accurately estimate the cost.

[0142] In use When searching for a set of candidate paths, the algorithm first initializes an open list and a closed list. The starting node is added to the open list, and its evaluation function value is calculated. Then, the node with the smallest evaluation function value is selected from the open list, removed from the open list, and added to the closed list. The neighboring nodes of this node are expanded, and their evaluation function values ​​are calculated. Nodes that meet the criteria are added to the open list. This process is repeated until the target node is found or the open list is empty. During the search, the heuristic function weights are dynamically adjusted based on the real-time communication cost.

[0143] Step S480: Perform dynamic evaluation of path robustness on the candidate path set. Generate random perturbations of node communication delays through Monte Carlo simulation, and calculate the rate of change of the evaluation function value of the path under the perturbation. The smaller the rate of change, the higher the path robustness.

[0144] Path robustness dynamic evaluation is used to assess the stability of candidate paths under disturbance. Monte Carlo simulation is a method that estimates probability distributions through random sampling, generating random perturbations to node communication delays. When evaluating a set of candidate paths, Monte Carlo simulation is used to generate random perturbations to node communication delays, and the rate of change of the evaluation function value of the path under the perturbation is calculated. The smaller the rate of change, the more stable the path is under disturbance, i.e., the higher the path robustness.

[0145] When performing dynamic evaluation of path robustness, a candidate path is first selected. Monte Carlo simulation is used to generate random perturbations to node communication delays, which are then added to the node communication delays. The evaluation function value of the path is recalculated. This process is repeated multiple times, and the rate of change of the evaluation function value is calculated. All candidate paths are evaluated, and the path with the highest robustness is selected.

[0146] Step S490: Calculate the node overlap and similarity of evaluation function values ​​between candidate paths, and retain paths with overlap below a preset threshold and similarity within a preset range to ensure the diversity of the path set.

[0147] Node overlap represents the ratio of the number of nodes shared between two candidate paths to the total number of nodes in the path. Evaluation function value similarity represents the degree of similarity between the evaluation function values ​​of two candidate paths. By calculating the node overlap and evaluation function value similarity between candidate paths, paths with overlap below a preset threshold and similarity within a preset range are retained, ensuring the diversity of the path set.

[0148] When calculating node overlap, the number of nodes shared between two paths is counted and divided by the total number of nodes in the paths to obtain the node overlap. When calculating the similarity of evaluation function values, methods such as Euclidean distance can be used to calculate the distance between the evaluation function values ​​of two paths. Preset thresholds and ranges are set to filter out paths that meet the conditions, forming a new set of paths.

[0149] Step S4100: Based on the path robustness assessment results and diversity screening results, select the path with the highest robustness as the primary scheduling channel, and select a preset number of secondary optimal paths as backup channels. The primary and backup channels are dynamically switched through real-time health monitoring to construct a hot zone scheduling channel.

[0150] When constructing hot zone scheduling channels based on path robustness assessment results and diversity screening results, the path with the highest robustness is first selected as the primary scheduling channel. Then, a preset number of suboptimal paths are selected from the filtered path set as backup channels. The primary and backup channels are dynamically switched through real-time health monitoring. If the primary scheduling channel fails or its performance degrades, the system will automatically switch to the backup channel.

[0151] Real-time health monitoring can be achieved by monitoring indicators such as the communication status and load of nodes. When the health indicator of the main scheduling channel falls below a certain threshold, a switchover operation is triggered, and the task is switched to the backup channel.

[0152] As one implementation method, in step S400, the computational task corresponding to the hot zone identification data is migrated to the optimal topology domain associated with the hot zone scheduling channel according to the weight combination strategy. Specifically, this may include the following steps S4110~S4170: Step S4110: Parse the computational task attributes corresponding to the hot zone identification data, and extract the task feature vector containing the task computational complexity, data transmission volume, and real-time requirement level.

[0153] The computational task attributes include information such as task computational complexity, data transfer volume, and real-time requirement level. When parsing the computational task attributes corresponding to hotspot identification data, this information is extracted from the hotspot identification data and combined into a task feature vector. Task computational complexity represents the computational resources consumed during the task's computation process; data transfer volume represents the amount of data that needs to be transferred during the task's execution; and the real-time requirement level represents the task's time sensitivity.

[0154] When extracting task feature vectors, the computational complexity and data transmission volume can be determined by analyzing the task's code, input data, and other information. Real-time performance requirements can be categorized according to the task's business needs, for example, into high, medium, and low levels.

[0155] Step S4120: Determine the set of candidate topology domains associated with the hot zone scheduling channel based on the dynamic topology delay distribution map. Each candidate topology domain includes the nodes through which the channel passes and the surrounding associated nodes.

[0156] The candidate topology set is determined based on the dynamic topology delay distribution map and includes all possible topology domains associated with the hot zone scheduling channel. Each candidate topology domain consists of nodes traversed by the channel and surrounding associated nodes. Nodes traversed by the channel are those along the hot zone scheduling channel, and surrounding associated nodes are those adjacent to and related to the nodes traversed by the channel.

[0157] When determining the candidate topology domain set, the nodes traversed by the hot zone scheduling channel are identified based on the communication path information in the dynamic topology delay distribution map. Then, the nodes adjacent to these nodes are identified, and they are grouped into the candidate topology domain set.

[0158] Step S4130: Input the task feature vector and the resource feature vector of the candidate topology into the deep cross-network, learn the adaptation pattern between the task and the topology through the dynamic feature cross-layer, and output the task-topology matching degree.

[0159] Deep cross-connect networks are neural networks capable of learning the cross-relationships between features. They consist of an input layer, a dynamic feature cross-connect layer, and an output layer. The dynamic feature cross-connect layer learns the cross-relationships between the task feature vector and the resource feature vectors of the candidate topology, thus outputting the task-topology matching degree. The task-topology matching degree represents the degree of fit between the task and the topology; the higher the matching degree, the better the task performs in that topology.

[0160] When task feature vectors and resource feature vectors of candidate topologies are input into a deep cross-connect network, the input layer receives these feature vectors, and the dynamic feature cross-connect layer performs cross-connection operations on the feature vectors to learn the adaptation pattern between them. The output layer outputs the task-topology matching degree based on the adaptation pattern.

[0161] Step S4140: Dynamically calibrate the task-topology matching degree based on the hot zone node weight combination strategy to enhance the matching degree of the topology domain where the hot zone node with high weight is located. The calibration coefficient is adjusted in real time with the node weight.

[0162] The hot zone node weight combination strategy includes the weight information of hot zone nodes, and dynamically calibrates the task-topology matching degree based on this weight information. The calibration coefficient is adjusted in real time according to the node weights to ensure that the matching degree calibration can reflect the changes in node weights.

[0163] During dynamic calibration, for each candidate topology domain, the weights of hot zone nodes within that domain are determined according to a hot zone node weight combination strategy. The task-topology matching degree is then multiplied by a calibration coefficient, which is proportional to the weight of the hot zone nodes. This enhances the matching degree of topology domains containing high-weight hot zone nodes.

[0164] Step S4150: Using a task grouping dynamic programming algorithm, tasks with complementary resource requirements are grouped into task groups, and the resource conflict rate and collaborative completion time of the task groups in each candidate topology domain are calculated.

[0165] Task grouping dynamic programming is an algorithm for solving the task grouping problem, which combines tasks with complementary resource requirements into task groups. Complementarity in resource requirements means that the tasks' needs for computational resources, data transmission resources, etc., can complement each other. After grouping the tasks into task groups, the resource conflict rate and collaborative completion time of each task group in each candidate topology domain are calculated. The resource conflict rate represents the proportion of resource conflicts between the task group and the topology domain during execution, and the collaborative completion time represents the time required for all tasks in the task group to complete.

[0166] When using the task grouping dynamic programming algorithm, the resource requirements of each task are first analyzed to identify tasks with complementary resource requirements. Then, these tasks are grouped into task groups. For each task group, its resource conflict rate and collaborative completion time are calculated across candidate topologies.

[0167] Step S4160: Predict the resource load change curve of the candidate topology after the task group migration using a long short-term memory network to identify potential resource bottleneck moments.

[0168] Long Short-Term Memory (LSTM) networks are neural networks capable of handling long-term dependencies in sequential data. They can predict the resource load change curve of a candidate topology after a task group migration. The resource load change curve represents how the resource load of a topology changes over time after the task group migration. Potential resource bottleneck moments refer to the points where the resource load exceeds a certain threshold, at which point resource shortages may occur.

[0169] When using Long Short-Term Memory (LSTM) networks to predict resource load change curves, the information of the task group and the current resource load information of the candidate topology are first input into the LTM network. The LTM network consists of an input layer, hidden layers, and an output layer. The input layer receives feature information of the task group (such as task computation complexity and data transfer volume) and current resource load data of the candidate topology (such as computational resource utilization and storage resource occupancy). The LTM units in the hidden layer can remember long-term dependency information in the sequence data, gradually updating the hidden state based on the input data, and learning the impact of task group migration on the topology resource load. The output layer outputs the predicted resource load change curve based on the hidden state.

[0170] During the prediction process, the Long Short-Term Memory (LSTM) network is trained using the backpropagation algorithm. Historical task migration data and corresponding resource load change data are used as the training set, and the network's weight parameters are updated by minimizing the error (e.g., mean squared error) between the predicted and actual values. Optimization algorithms such as stochastic gradient descent can be employed during training to accelerate convergence.

[0171] By analyzing the predicted resource load change curves, potential resource bottleneck moments can be identified. A resource load threshold can be set; when the resource load value in the curve exceeds this threshold, the corresponding time point is marked as a potential resource bottleneck moment. For example, for calculating resource utilization, a threshold of 80% can be set; when the calculated resource utilization in the predicted curve exceeds 80%, it is a potential resource bottleneck moment.

[0172] Step S4170: Based on the comprehensive task-topology matching degree, resource conflict rate, collaborative completion time and load prediction results, select the optimal topology domain through a dynamic multi-attribute decision algorithm, and generate a task migration plan that includes task migration order, resource allocation time period and migration trigger conditions.

[0173] The dynamic multi-attribute decision algorithm is an algorithm that can make decisions by comprehensively considering multiple attributes. In this step, it is necessary to comprehensively consider multiple attributes such as task-topology matching degree, resource conflict rate, collaborative completion time and load prediction results to select the optimal topology.

[0174] As one implementation method, step S4170, selecting the optimal topological domain through a dynamic multi-attribute decision algorithm, may specifically include the following steps S4171~S4181: Step S4171: Perform dynamic feature alignment between the task feature vector and the resource feature vector of the candidate topology domain. Transform features of different dimensions into a unified feature space through a feature mapping network to generate an aligned feature matrix that includes task requirements and topology supply. The mapping parameters of the feature mapping network are updated in real time according to the computing power network resource status.

[0175] Task feature vectors and candidate topology resource feature vectors may have different dimensions and scales. To facilitate subsequent computation and comparison, dynamic feature alignment is required. The feature mapping network can adopt a multilayer perceptron (MLP) architecture, consisting of an input layer, multiple hidden layers, and an output layer. The input layer receives the task feature vector and the candidate topology resource feature vector. The hidden layers transform the input features through a series of linear transformations and nonlinear activation functions (such as the ReLU function). The output layer maps the transformed features to a unified feature space.

[0176] The mapping parameters of the feature mapping network are updated in real time according to the resource status of the computing power network. An online learning method can be adopted to collect new task feature vectors and resource feature vectors when the resource status of the computing power network changes. These new data can be used to fine-tune the feature mapping network and update the mapping parameters to ensure the accuracy of feature alignment.

[0177] During feature alignment, the task feature vector and resource feature vector are input into a feature mapping network. After transformation by the network, feature representations in a unified feature space are obtained. These feature representations are then combined into a matrix, namely the alignment feature matrix containing task requirements and topology supply.

[0178] Step S4172: Calculate the task-topology basic fit degree based on the alignment feature matrix. Measure the spatial similarity between the task feature vector and the topology resource feature vector using cosine similarity to generate an initial fit degree score. The score range is adjusted to a preset interval using a dynamic scaling factor.

[0179] After obtaining the alignment feature matrix, cosine similarity is used to calculate the task-topology-based fit. Cosine similarity is a method for measuring the cosine of the angle between two vectors, reflecting the degree of similarity between the two vectors in space. For the task feature vectors and topology resource feature vectors in the alignment feature matrix, their cosine similarity is calculated, and the result is used as the initial fit score.

[0180] To adjust the initial fit score to a preset range, a dynamic scaling factor can be used. The dynamic scaling factor adjusts based on the distribution of the initial fit score. For example, if the preset range is [0,1], and the initial fit score is in the range [0.2,0.8], a linear transformation can be used to map the score to the [0,1] range. The specific transformation formula is: Adjusted score = (Initial score - Minimum value) / (Maximum value - Minimum value), where the minimum and maximum values ​​are the minimum and maximum values ​​of the initial fit score.

[0181] Step S4173: Extract the real-time resource usage data sequence of the candidate topology domain, calculate the ratio of the number of resource request conflicts to the total number of requests per unit time as a dynamic indicator of resource conflict, and adaptively adjust the ratio calculation window length according to the current load intensity of the topology domain.

[0182] Real-time resource usage data sequences for candidate topologies can be obtained by monitoring resource usage within the topology, such as computing resource usage and storage resource usage. The number of resource request conflicts refers to the number of times a request is rejected due to insufficient resources during resource allocation. The total number of requests refers to all resource requests made.

[0183] To calculate dynamic indicators of resource conflicts, a ratio calculation window length needs to be set. The ratio calculation window length will adaptively adjust according to the current load intensity of the topology. If the load intensity of the topology is high, the ratio calculation window length can be set to be shorter in order to reflect the resource conflict situation more promptly; if the load intensity is low, the ratio calculation window length can be set to be longer to obtain more stable statistical results.

[0184] When calculating the dynamic index of resource conflict, the number of resource request conflicts and the total number of requests are counted in the ratio calculation window. The number of resource request conflicts is divided by the total number of requests to obtain the dynamic index of resource conflict.

[0185] Step S4174: Calculate the earliest start time and latest finish time of all tasks in the task group using the task dependency graph traversal algorithm. The difference between the two is used as the collaborative completion time indicator. The data transmission delay between tasks is considered during the traversal process.

[0186] A task dependency graph is a directed graph used to represent the dependencies between tasks. Nodes in the graph represent tasks, and edges represent the dependencies between tasks. Task dependency graph traversal algorithms can employ either depth-first search (DFS) or breadth-first search (BFS).

[0187] When traversing the task dependency graph, data transmission latency between tasks needs to be considered. For each task, the earliest start time is calculated based on the completion time of its predecessor tasks and the data transmission latency. The earliest start time refers to the earliest time the task can start under the condition that all predecessor tasks are completed and data transmission is satisfied. The latest finish time refers to the latest time the task can finish without affecting the overall task group's completion time.

[0188] By calculating the earliest start time and latest finish time of all tasks, the difference between the earliest start time and latest finish time in the task group is found and used as a collaborative completion time indicator.

[0189] Step S4175: Integrate the absolute value of the first derivative of the resource load prediction curve over time to generate a load prediction stability index. The smaller the integral value, the more stable the load fluctuation.

[0190] The first derivative of the resource load forecast curve represents the rate of change of resource load, and its absolute value represents the magnitude of the rate of change. Integrating the absolute value of the first derivative over time yields the total change in resource load over a period of time, which can be used as an indicator of load forecast stability.

[0191] The integration process can employ numerical integration methods, such as the trapezoidal rule. The resource load forecast curve is discretized, and the absolute value of the first derivative at each discrete point is multiplied by the corresponding time interval. These products are then summed to obtain the integral value. A smaller integral value indicates smaller fluctuations in resource load during this time period, meaning more stable load fluctuations.

[0192] Step S4176: Integrate basic adaptability, resource conflict dynamic indicators, collaborative completion time indicators, and load prediction stability indicators, and generate a comprehensive evaluation value through dynamic weighted summation. The weight values ​​are dynamically adjusted based on the discrimination of each indicator in recent decision-making. The discrimination is determined by the coefficient of variation of the indicator values.

[0193] To comprehensively consider basic adaptability, dynamic resource conflict indicators, collaborative completion time indicators, and load forecast stability indicators, a dynamic weighted summation method is used to generate a comprehensive evaluation value. The dynamic weighted summation process is as follows: Comprehensive Evaluation Value = Basic Adaptability × Weight 1 + Dynamic Resource Conflict Indicator × Weight 2 + Collaborative Completion Time Indicator × Weight 3 + Load Forecast Stability Indicator × Weight 4.

[0194] The weights are dynamically adjusted based on the discriminative power of each indicator in recent decision-making. Discriminative power can be calculated using the coefficient of variation of the indicator value, which is the ratio of the standard deviation to the mean of the indicator value. The larger the coefficient of variation, the greater the difference between the indicator and different candidate topologies, and the higher the discriminative power. Therefore, it should be given a higher weight in the calculation of the comprehensive evaluation value.

[0195] After each decision, the values ​​of each indicator are collected in all candidate topological domains, the coefficient of variation of each indicator is calculated, and the weight values ​​are adjusted according to the coefficient of variation.

[0196] Step S4177: Mark the candidate topology domain with the highest comprehensive evaluation value as the optimal topology domain, and record the contribution distribution of each indicator for subsequent indicator weight calibration in the decision-making process.

[0197] After calculating the comprehensive evaluation value of all candidate topologies, the candidate topology with the highest comprehensive evaluation value is selected as the optimal topology. Simultaneously, the contribution distribution of each indicator in the comprehensive evaluation value calculation is recorded. The contribution distribution can be represented by the proportion of each indicator in the comprehensive evaluation value after multiplying it by its corresponding weight.

[0198] Recording the contribution distribution of each indicator can provide a basis for calibrating indicator weights in subsequent decision-making processes. If the contribution of an indicator is consistently low, it indicates that the indicator plays a limited role in the current decision-making process, and its weight can be appropriately reduced; if the contribution of an indicator is high and it frequently has a significant impact on the decision outcome, its weight can be appropriately increased.

[0199] Step S4178: Identify the load trough period within the future preset time period of the optimal topology domain using the peak detection algorithm, and schedule the task migration process during the load trough period, ensuring that the time difference between the start time of the migration and the load trough time does not exceed a preset threshold.

[0200] Peak detection algorithms can employ either threshold-based or morphological methods. Threshold-based methods set a load threshold; when the resource load value falls below this threshold, it is considered a period of low load. Morphological methods identify low load periods by performing morphological operations (such as erosion or dilation) on the resource load curve.

[0201] After identifying the optimal topology domain's load slump periods within a preset future timeframe, the task migration process is scheduled to occur during these periods. To ensure the timeliness and effectiveness of the task migration, the time difference between the migration start time and the load slump period must not exceed a preset threshold. This preset threshold can be set based on the task's real-time requirements and the topology domain's load conditions.

[0202] Step S4179: Based on the real-time requirement level and resource requirement scale in the task feature vector, determine the task migration order through a priority sorting algorithm. The real-time requirement level is quantified by a heat-sensitive dynamic coefficient, and the resource requirement scale is represented by the product of the task computational complexity and the data transmission volume.

[0203] Priority ranking algorithms can employ rule-based or machine learning-based methods. Rule-based ranking methods can formulate ranking rules based on the level of real-time requirements and the scale of resource requirements. For example, tasks with higher real-time requirements are migrated first, and when real-time requirements are the same, tasks with smaller resource requirements are migrated first.

[0204] The real-time requirement level is quantified by a heat sensitivity dynamic coefficient; the higher the heat sensitivity dynamic coefficient, the higher the real-time requirement of the task. The resource requirement scale is represented by the product of the task's computational complexity and the amount of data transmitted; the larger the product, the larger the resource requirement scale.

[0205] When determining the task migration order, the real-time requirement level and resource requirement scale information in the task feature vector are input into the priority ranking algorithm. The algorithm sorts the tasks according to this information to determine the task migration order.

[0206] Step S4180: Calculate the resource reservation amount during the migration period. The reservation amount is the product of the task resource requirement and the load fluctuation safety factor. The safety factor is positively correlated with the load prediction stability index.

[0207] Task resource requirements can be determined based on the computational complexity and data transmission volume in the task feature vector. The load fluctuation safety factor is positively correlated with the load prediction stability index; the higher the load prediction stability index (i.e., the more stable the load fluctuation), the smaller the safety factor; conversely, the lower the load prediction stability index (i.e., the more drastic the load fluctuation), the larger the safety factor.

[0208] The positive correlation between the safety factor and the load forecast stability index can be determined by establishing a functional relationship, for example, safety factor = k / (load forecast stability index + ε), where k is a constant and ε is a small positive number used to avoid the case where the denominator is zero.

[0209] When calculating the resource reservation amount during the migration period, the task resource requirement is multiplied by the load fluctuation safety factor to obtain the resource reservation amount.

[0210] Step S4181: Combine the task migration order, resource reservation amount and migration time period into a task migration plan. The plan includes the resource allocation details and time period division of each task.

[0211] The task migration sequence determined in step S4179, the resource reservation amount calculated in step S4180, and the migration period determined in step S4178 are combined to form a task migration plan. The task migration plan will record the resource allocation details for each task, including the allocation of computing resources, data transmission resources, etc., as well as the migration period for each task, that is, when each task starts and ends its migration.

[0212] Through the above steps, the scheduling of computing tasks in the computing network driven by data popularity and topology awareness was completed. This enabled the reasonable scheduling and migration of computing tasks based on data popularity and topology information, thereby improving the resource utilization and task processing efficiency of the computing network.

[0213] It is understood that the various algorithms involved in the above descriptions of the embodiments of the present invention can all be obtained from relevant content in the prior art. To save space, they will not be elaborated on in the embodiments of the present invention. In addition, those skilled in the art can supplement the details based on common knowledge in the art when implementing the solutions of the present invention. For example, they can use normalization to eliminate dimensional conflicts before feature fusion, use interpolation to eliminate dimensional differences, reasonably set thresholds based on historical data, experience or business scenario requirements, train the model based on a general model training method, set the number of layers in the model structure based on actual needs, select activation functions, etc. The present invention will not provide redundant descriptions of overly detailed implementation processes here.

[0214] Please see Figure 2 , Figure 2 This is a schematic diagram of a computer system provided in an embodiment of the present invention. The computer system includes at least a processor 101, a communication interface 102, and a memory 103. The processor 101, communication interface 102, and memory 103 can be connected via a bus or other means. The processor 101 (or Central Processing Unit, CPU) is the computing and control core of the computer system, capable of parsing various instructions and processing various data within the computer system. The communication interface 102 may optionally include a standard wired interface or a wireless interface (such as Wi-Fi, mobile communication interface, etc.), and can be used to send and receive data under the control of the processor 101; the communication interface 102 can also be used for data transmission and interaction within the computer system. The memory 103 is a storage device in the computer system used to store programs and data. It is understood that the memory 103 here can include the computer system's built-in memory, or it can include extended memory supported by the computer system. The memory 103 provides storage space, which stores the computer system's operating system; this invention does not limit this storage space.

[0215] In one embodiment, the processor 101 executes the data heat and topology-aware computing power network computing task scheduling method provided in the above embodiments of the present invention by running a computer program in the memory 103.

Claims

1. A method for scheduling computing tasks in a computing network driven by data popularity and topology awareness, characterized in that, include: Collect a set of topology probe data and a set of task access heat time series data between computing power nodes. The topology probe data set includes inter-node communication latency measurement values ​​and connection status identifiers. A dynamic topology delay distribution map is constructed based on the topology probe data set. The dynamic topology delay distribution map is used to reflect the real-time communication path cost between computing power nodes. Based on the access frequency decay characteristics in the task access heat time series data set, hot zone identification data is generated, and the hot zone identification data is used to mark the distribution of computing nodes for time-sensitive tasks; Based on the collaborative response operation of the dynamic topology delay distribution map and hot zone identification data, a weight combination strategy for hot zone nodes is generated. A hot zone scheduling channel is constructed between computing power nodes through topology-aware path selection operation. The computing tasks corresponding to the hot zone identification data are migrated to the optimal topology domain associated with the hot zone scheduling channel according to the weight combination strategy.

2. The method according to claim 1, characterized in that, The construction of a dynamic topology delay distribution map based on the topology probe data set includes: The topology probe dataset is subjected to spatiotemporal correlation mining processing. The communication delay fluctuation sequence of node pairs within consecutive time segments is extracted in timestamp order. The single time series data is converted into a three-dimensional correlation tensor containing the time dimension through a sliding time window. Each time slice of the three-dimensional correlation tensor corresponds to the node communication state matrix of a preset time segment. The three-dimensional correlation tensor is dynamically encoded with time delay features. Multi-scale feature extraction is performed on the time delay fluctuation sequence through dilated time convolution kernels with different dilation rates to generate a time delay feature tensor containing instantaneous fluctuation features and trend evolution features. The dilation rate of the dilated convolution kernel is adjusted in real time according to the fluctuation frequency of node communication time delay. The time delay feature tensor is input into the node embedding layer of the temporal graph neural network, and the real-time association strength between nodes is learned through the dynamic graph attention mechanism to generate dynamic association features of nodes. The attention coefficient of the dynamic graph attention mechanism is determined based on the joint calculation of the similarity of the current communication features of the nodes and the stability of historical association. Based on the dynamic association characteristics of the nodes, nodes with similar communication evolution trends are aggregated into dynamic topological communities using an incremental density clustering algorithm. The boundary of each topological community is adjusted in real time as the communication characteristics of the nodes change, generating dynamic topological community partitioning results. The dynamic topological community division results are processed by community evolution tracking to calculate the topological community association matrix for different time segments. The matrix elements represent the node overlap between communities in adjacent time segments. The topological community association matrix is ​​fused with the node communication state matrix to generate a dynamic topological delay distribution map that includes node-level real-time communication costs and community-level evolution trends.

3. The method according to claim 2, characterized in that, The method of learning the real-time association strength between nodes through a dynamic graph attention mechanism to generate dynamic association features of nodes includes: Initialize the node feature matrix of the temporal graph neural network, and compress the time dimension of the delay feature tensor into the channel dimension of the node feature vector through temporal attention pooling to construct the initial dynamic representation of the node; The real-time communication feature similarity between node pairs is calculated. The cosine similarity algorithm is used to measure the time delay feature tensor similarity between any two nodes. The similarity calculation window is dynamically adjusted according to the node communication frequency to generate a real-time similarity matrix. The real-time similarity matrix is ​​fused with the node historical association stability coefficient to generate a dynamic attention coefficient. The historical association stability coefficient is calculated by the variance of the association strength of the node within a preset time segment in the past. The features of neighboring nodes are weighted and aggregated based on the dynamic attention coefficient. The updated features of each node are represented as the weighted sum of the features of its neighboring nodes. The weight values ​​are dynamically updated with the real-time similarity matrix to generate node aggregated features. The dynamic attention aggregation process is repeated a preset number of times. The high-order dynamic association features between nodes are enhanced through a multi-layer temporal graph attention network. The final output includes the dynamic association features of nodes that are ranked in real time by their influence. The element values ​​of the feature vector are positively correlated with the role of the node as a communication hub in the current topology.

4. The method according to claim 3, characterized in that, Based on the dynamic association characteristics of the nodes, an incremental density clustering algorithm is used to aggregate nodes with similar communication evolution trends into dynamic topological communities. The boundary of each topological community is adjusted in real time as the communication characteristics of the nodes change, generating dynamic topological community partitioning results, including: The dynamic density index of each node is calculated based on the dynamic association characteristics of the nodes. The dynamic density index is determined by counting the number of neighboring nodes whose communication feature similarity is higher than a preset threshold within the current time window. The similarity threshold is dynamically adjusted according to the stability of node communication. Based on the node dynamic association features, a high-density neighbor set of the node is determined, and the dynamic distance from each node to the high-density neighbor is calculated. The dynamic distance is a combination of the node's current time delay feature tensor Euclidean distance and the node's dynamic association feature cosine distance. The high-density neighbor is a node whose dynamic density index is greater than a preset density threshold. A real-time decision graph is constructed using dynamic density-distance features. The node with the largest product of dynamic density and dynamic distance is identified as the cluster center in the real-time decision graph. The number of cluster centers changes dynamically with the distribution of node communication features. Incremental label propagation is performed based on the cluster centers, and newly added nodes or nodes whose communication characteristics change beyond a threshold are assigned to the community where the nearest cluster center is located. The community center characteristics are updated in real time during the label propagation process. The initial clustering results are optimized by boundary optimization. By calculating the average communication feature similarity of nodes within a community and the dynamic association strength of boundary nodes between communities, boundary nodes with association strength higher than the average similarity within the community are reassigned until the rate of change of similarity within the community is lower than a preset stable threshold, thus generating a dynamic topological community partitioning result that includes community identifiers and node affiliation relationships.

5. The method according to claim 1, characterized in that, The step of generating hotspot identification data based on the access frequency decay characteristics in the task access popularity time-series data set includes: Multi-scale temporal pattern decomposition is performed on the task access popularity time series data set. The original access frequency sequence is separated into multiple modal components and residual trend components with different center frequencies through variational mode decomposition. Each modal component corresponds to the access fluctuation characteristics at a preset scale. Periodic characteristics are extracted from each modal component, and the instantaneous frequency and instantaneous amplitude of the modal component are calculated by adaptive Fourier decomposition to generate a periodic feature vector containing the periodic change rate, amplitude modulation coefficient and phase drift. The periodic feature vector is input into the gated recurrent unit network to learn the nonlinear decay law of the access frequency over time. During the network training process, the decay parameters are dynamically updated by real-time access data, and the heat prediction features are output. The heat sensitivity dynamic coefficient is calculated based on the heat prediction characteristics. The coefficient is determined by a dynamic weighted combination of the cumulative absolute value of the deviation between the predicted heat and the actual heat and the rate of change of the deviation. The weights are dynamically adjusted with the prediction error. A heat diffusion model is constructed based on the heat conduction equation. The heat sensitivity dynamic coefficient is used as the heat source intensity and the inter-node communication delay is used as the heat conduction resistance. The real-time heat diffusion process between computing nodes is simulated to generate a heat diffusion feature map. The heat diffusion feature map is subjected to heat zone boundary identification by a dynamic region growth algorithm. Nodes with heat values ​​higher than a preset dynamic threshold are used as seed points. Regions are grown in the direction of heat gradient to generate heat zone identification data that includes real-time heat zone boundaries, heat diffusion speed and core heat zone nodes.

6. The method according to claim 5, characterized in that, The step of extracting the periodic characteristics of each modal component and calculating the instantaneous frequency and instantaneous amplitude of the modal components through adaptive Fourier decomposition includes: Perform a Hilbert transform on each modal component to generate an analytic signal. Calculate the instantaneous phase of the modal component using the amplitude of the analytic signal. The derivative of the instantaneous phase with respect to time is the instantaneous frequency. The amplitude envelope of the modal component is extracted by the sliding window peak detection algorithm. The upper and lower envelopes are generated by fitting the local maxima of the amplitude within the window. The mean of the upper and lower envelopes is used as the instantaneous amplitude of the modal component. Calculate the time derivative of the instantaneous frequency to obtain the frequency change rate. The periodic stability of the modal components is identified by the frequency change rate. The smaller the absolute value of the frequency change rate, the more stable the period. The amplitude modulation coefficient is calculated based on the instantaneous amplitude. The amplitude modulation coefficient is the ratio of the instantaneous amplitude to the average amplitude and is used to describe the modulation depth of the amplitude. Align the instantaneous phases of different modal components, calculate the phase drift of each modal component relative to the reference mode, and generate a periodic feature vector containing the periodic rate of change, amplitude modulation coefficient, and phase drift.

7. The method according to claim 6, characterized in that, The heat diffusion model constructed based on the heat conduction equation uses the heat-sensitive dynamic coefficient as the heat source intensity and the inter-node communication delay as the heat conduction resistance to simulate the real-time heat diffusion process between computing nodes and generate a heat diffusion feature map, including: The computing nodes are abstracted as discrete particles in a heat conduction medium, the heat sensitivity dynamic coefficient of the node is used as the initial heat source intensity, and the node computing resource utilization rate is used as the heat source attenuation coefficient. Construct a heat conduction resistance matrix between nodes. The matrix elements are the product of the communication delay between nodes and the node load rate. The greater the communication delay or the higher the node load rate, the greater the heat conduction resistance. A partial differential equation for heat diffusion is established based on the heat conduction equation. The time term of the equation represents the decay of heat over time, and the spatial term represents the diffusion of heat between nodes. The equation coefficients are dynamically calibrated using historical heat diffusion data of nodes. The partial differential equation for heat diffusion is solved by the finite difference method to obtain the distribution of nodal heat values ​​at each time step. The time step is dynamically adjusted according to the heat diffusion rate. The diffusion process is monitored by a heat diffusion stability criterion. When the rate of change of heat value in adjacent time steps is lower than the preset stability threshold, the diffusion simulation is stopped, and a heat diffusion feature map containing the real-time heat value of each node and the diffusion direction is output.

8. The method according to claim 1, characterized in that, The weight combination strategy for generating hot zone nodes based on the collaborative response operation using the dynamic topology delay distribution map and hot zone identification data includes: The node communication characteristics of the dynamic topology delay distribution map and the heat diffusion characteristics of the hot zone identification data are input into the spatiotemporal fusion network. The real-time correlation pattern between the two is learned through the spatiotemporal cross-attention mechanism to generate topology-heat correlation features. With the optimization objectives of minimizing the communication path cost reflected by the dynamic topology delay distribution map and maximizing the heat response speed reflected by the hot zone identification data, a dual-objective dynamic optimization framework is constructed, with the resource allocation weight of the hot zone nodes as the optimization variable. The bi-objective dynamic optimization framework is solved by an adaptive multi-objective evolutionary algorithm to generate an initial weight allocation scheme. The crossover probability and mutation probability of the adaptive multi-objective evolutionary algorithm are dynamically adjusted according to the convergence state of the current population. The convergence state is determined by calculating the distribution entropy of the objective function values ​​of individuals in the population. Using the real-time task completion latency and resource waste rate of the computing network as reward signals, a deep deterministic policy gradient agent is trained to output the real-time adjustment amount of the hot zone node weights for the initial weight allocation scheme. Based on the initial weight allocation scheme and the real-time adjustment amount, a hot zone node weight combination strategy is generated, which includes real-time node weight update rules, weight conflict arbitration mechanism and weight adjustment trigger conditions. The weight combination strategy is dynamically updated as the topology-heat correlation characteristics change.

9. The method according to claim 8, characterized in that, The node communication characteristics of the dynamic topology delay distribution map and the heat diffusion characteristics of the hot zone identification data are input into the spatiotemporal fusion network. A spatiotemporal cross-attention mechanism is used to learn the real-time correlation pattern between the two, generating topology-heat correlation features, including: The node communication characteristics of the dynamic topology delay distribution map are converted into a topology feature sequence. The feature vector of each time step contains the current communication cost and historical communication trend of the node. The heat diffusion characteristics of the hot zone identification data are converted into a heat feature sequence, and the feature vector of each time step contains the real-time heat value and diffusion speed of the node; Calculate the cross-correlation coefficients between the topological feature sequence and the popularity feature sequence in the time and spatial dimensions, and generate a spatiotemporal attention weight tensor; The topological feature sequence and the popularity feature sequence are weighted and fused using the spatiotemporal attention weight tensor to generate a preliminary fused feature sequence; Multi-scale feature extraction is performed on the preliminary fused feature sequence. By combining different temporal and spatial convolution kernels, short-term local correlation features and long-term global correlation features are extracted. Multi-scale association features are input into a bidirectional long short-term memory network for time-series modeling to learn the evolution of topology-heat association patterns over time and output topology-heat association features containing real-time association strength.

10. A computer system, characterized in that, include: A memory, wherein a computer program is stored; A processor is configured to load the computer program to implement the data heat and topology-aware computing power network computing task scheduling method as described in any one of claims 1-9.

Images