Network task scheduling method and device, and electronic equipment

Abstract

The present disclosure relates to a network task scheduling method and device and electronic equipment, wherein the method comprises: collecting a network multi-modal data set, including multi-source link data, computing power index data, network state, and topology snapshot information representing network topology structure in the network; according to the network multi-modal data set, detecting available edge computing nodes and corresponding availability levels in the network, and evaluating the confidence levels of each type link in the network; according to the task type corresponding to the task request received by the network and the network state, the tasks in the task request are decomposed into multiple sub-tasks with priority marks; according to the available edge computing nodes and the availability levels, the confidence levels and the multiple sub-tasks, a feasible task scheduling table is generated and network task scheduling is performed. Thus, the optimal resource allocation scheme for tasks in a high-dynamic, multi-modal network environment is realized, and the overall throughput of the network and the stability of network task execution are improved.

Description

Technical Field

[0001] This disclosure relates to the field of computer network technology, and in particular to a network task scheduling method, apparatus and electronic device. Background Technology

[0002] Currently, with the popularization of 5G, 6G and low-Earth orbit satellite networks, edge computing is widely used in modern communication networks. In related technologies, task scheduling methods usually assume that the network environment is static and stable, rely on a single resource parameter for scheduling, and lack effective integration of multimodal network information. This results in poor adaptability and low resource allocation efficiency in dynamic network environments such as bandwidth fluctuations, latency changes, and link instability. At the same time, network task scheduling lacks a flexible adjustment mechanism, making it difficult to meet the requirements for accuracy and reliability of task scheduling in highly dynamic and multimodal network environments. Summary of the Invention

[0003] This disclosure provides a network task scheduling method, apparatus, and electronic device.

[0004] According to a first aspect of the present disclosure, a network task scheduling method is provided. The method includes: collecting a network multimodal dataset, the multimodal dataset including multi-source link data of multi-mode links in the network, computing power index data of edge computing nodes, real-time network status, and topology snapshots representing network topology information; performing anomaly detection and availability level assessment on the operating status of each edge computing node based on the computing power index data, generating available edge computing nodes and their corresponding availability levels; assessing the confidence level of each mode link based on the multi-source link data and the topology snapshots, generating confidence levels for each mode link; decomposing the task in the task request into multiple sub-tasks with priority labels based on the task type corresponding to the task request received by the network and the network status; generating a feasible task scheduling table based on the available edge computing nodes and their corresponding availability levels, the confidence levels, and the multiple sub-tasks, and performing network task scheduling according to the task scheduling table.

[0005] In one embodiment of this disclosure, the acquisition of the network multimodal dataset includes: acquiring radio frequency measurement parameters of the multi-standard links through radio frequency measurement, and obtaining latency, bandwidth, and jitter parameters of each standard link through link probing to form the multi-source link data; acquiring processor utilization, memory usage, task queue length, and cache status of each edge computing node through edge computing node resource monitoring to form the computing power index data; aligning the multi-source link data and the computing power index data with timestamps to eliminate asynchronous deviations, thereby obtaining aligned multi-source link data; obtaining the available bandwidth and latency of each standard link based on the aligned multi-source link data; obtaining the available computing power and load status of each edge computing node based on the aligned computing power index data, thereby generating a real-time network status; constructing a topology snapshot representing network topology information using a network topology adjacency matrix; and integrating the aligned multi-source link data, computing power index data, the real-time network status, and the topology snapshot to generate a multimodal dataset.

[0006] In one embodiment of this disclosure, the step of performing anomaly detection and availability level assessment on the operating status of each edge computing node based on the computing power index data, and generating available edge computing nodes and their corresponding availability levels, includes: constructing a state time series of each edge computing node based on the computing power index data of each edge computing node; performing anomaly detection on the state time series of each edge computing node using an anomaly detection model, and calculating anomaly scores for each edge computing node; selecting available edge computing nodes from among the edge computing nodes based on the anomaly scores; calculating a fluctuation index of the state index of each available edge computing node, wherein the fluctuation index is used to characterize the stability of the operating status of the edge computing node; and determining the availability level of each available edge computing node based on the anomaly scores and the fluctuation index.

[0007] In one embodiment of this disclosure, the step of using an anomaly detection model to perform anomaly detection on the state time series of each edge computing node and calculating the anomaly score of each edge computing node includes: constructing a random segmentation tree to isolate each state index in the state time series of each edge computing node to obtain the isolated path length of each state index in the random segmentation tree; and calculating the anomaly score of each edge computing node based on the isolated path length.

[0008] In one embodiment of this disclosure, the step of evaluating the confidence level of each link based on the multi-source link data and topology snapshots, and generating a confidence level for each link, includes: weighting the link connections between edge computing nodes in the network using the multi-source link data, with bandwidth as a positive weighting factor and latency as a negative weighting factor, to construct a weighted adjacency matrix; aligning the topology snapshots with the weighted adjacency matrix in a time sequence, and then performing time slicing on the aligned weighted adjacency matrix to obtain a topology snapshot sequence that changes over time; determining the topology influence level of each link based on the connection relationships and edge attributes of each link in the topology snapshot sequence; performing regression fitting on the packet loss rate of each link in the next time window using a time-series prediction model based on historical and real-time packet loss rate data for each link to obtain a predicted packet loss rate value; and generating a confidence level for each link based on its topology influence level and the predicted packet loss rate value.

[0009] In one embodiment of this disclosure, the step of decomposing the task in the task request into multiple subtasks with priority tags according to the task type corresponding to the task request received from the network and the network state includes: classifying the descriptive text of the task request using a semantic parsing model to identify the task type; matching a corresponding basic decomposition template according to the task type, wherein the basic decomposition template defines the subtask composition of each task type under standard conditions and the dependencies between each subtask; obtaining the computational load and bandwidth requirements of the task at the current moment according to the network state, and calculating the interaction factor between the computational load and the bandwidth requirements; wherein the interaction factor is determined through the nonlinear coupling relationship between the computational load and the bandwidth requirements; determining the decomposition granularity according to the magnitude of the interaction factor; and merging or splitting each subtask in the basic decomposition template according to the determined decomposition granularity based on the subtask composition and dependencies defined by the basic decomposition template to generate multiple subtasks adapted to the current network state, and marking each subtask with an execution priority.

[0010] In one embodiment of this disclosure, the method further includes: calculating the task scheduling delay caused by the interaction between computing load and bandwidth demand based on the interaction factor; calculating the execution failure probability of the task at the current moment based on the interaction factor and the task scheduling delay; and adjusting the decomposition granularity when the execution failure probability exceeds a preset probability threshold.

[0011] In one embodiment of this disclosure, the step of generating a feasible task scheduling table based on the available edge computing nodes and their corresponding availability levels, the confidence level, and multiple sub-tasks, and then performing network task scheduling according to the task scheduling table, includes: selecting a set of links and a set of nodes from the network's standard links and available edge computing nodes that meet the requirements of each sub-task, based on the latency constraints, bandwidth requirements, and computational load of each sub-task; using the availability level and the confidence level as constraints, matching each sub-task with the standard links and available edge computing nodes in the link set and node set, with the optimization objective of minimizing the combined cost of latency, bandwidth utilization, and computational load of each sub-task, and solving for a resource allocation scheme for each sub-task; verifying the feasibility of the resource allocation scheme, using the verified resource allocation scheme as a feasible task scheduling table, and performing network task scheduling according to the task scheduling table.

[0012] In one embodiment of this disclosure, the process of solving for the resource allocation scheme of each subtask further includes: for critical subtasks with a priority higher than a preset threshold, a multi-path redundancy allocation strategy is adopted to simultaneously allocate the critical subtasks to multiple standard links with confidence levels higher than the preset threshold for parallel transmission; for ordinary subtasks with a priority lower than the preset threshold, allocation is performed sequentially according to the remaining resources in the link set and node set to obtain the resource allocation scheme of each subtask.

[0013] In one embodiment of this disclosure, the method further includes: after generating a feasible task scheduling table and performing network task scheduling according to the task scheduling table, collecting real-time feedback data of each subtask during execution, the real-time feedback data including network task scheduling completion latency, packet loss rate, bandwidth utilization rate, and task completion rate; when the real-time feedback data of any subtask does not meet the preset threshold requirement, triggering feedback adjustment; repeating the feedback adjustment until the real-time feedback data of all subtasks meet the preset threshold requirement, and outputting an optimized task scheduling table.

[0014] In one embodiment of this disclosure, the feedback adjustment includes at least one of the following: recalculating the interaction factor based on real-time feedback data, adjusting the decomposition granularity, and re-dividing subtasks; reallocating subtasks whose real-time feedback data does not meet the preset threshold requirements to backup standard links; and migrating subtasks whose real-time feedback data does not meet the preset threshold requirements to available edge computing nodes with higher availability levels.

[0015] In one embodiment of this disclosure, the method further includes: storing the real-time feedback data into a historical database; and using the real-time feedback data accumulated in the historical database to perform incremental training or parameter updates on the anomaly detection model, the time series prediction model for predicting link packet loss rate, and the semantic parsing model for identifying task categories.

[0016] According to a second aspect of the present disclosure, a network task scheduling apparatus is also provided. The apparatus includes: a collection module, configured to collect a network multimodal dataset, the multimodal dataset including multi-source link data of multi-mode links in the network, computing power index data of edge computing nodes, real-time network status, and topology snapshots representing network topology information; a detection module, configured to perform anomaly detection and availability level assessment on the operating status of each edge computing node based on the computing power index data, and generate available edge computing nodes and their corresponding availability levels; an evaluation module, configured to evaluate the confidence level of each mode link based on the multi-source link data and the topology snapshots, and generate confidence level for each mode link; a decomposition module, configured to decompose the task in the task request into multiple sub-tasks with priority tags based on the task type corresponding to the task request received by the network and the network status; and a scheduling module, configured to generate a feasible task scheduling table based on the available edge computing nodes and their corresponding availability levels, the confidence levels, and the multiple sub-tasks, and perform network task scheduling according to the task scheduling table.

[0017] According to a third aspect of the present disclosure, an electronic device is also provided, comprising: a processor; a memory for storing processor-executable instructions; wherein the processor is configured to: implement the steps of the network task scheduling method as described above.

[0018] According to a fourth aspect of the present disclosure, a non-transitory computer-readable storage medium is also provided, which, when the instructions in the storage medium are executed by a processor, enables the processor to perform the network task scheduling method as described above.

[0019] The technical solutions provided by the embodiments of this disclosure bring at least the following beneficial effects: A multimodal network dataset is collected, including multi-source link data, computing power metrics, network status, and topology snapshots representing the network topology. Based on this dataset, available edge computing nodes and their corresponding availability levels, as well as the confidence levels of links of various network types, are detected. According to the task type and network status of received task requests, the tasks in the requests are decomposed into multiple subtasks with priority tags. Based on the available edge computing nodes, their availability levels, confidence levels, and the multiple subtasks, a feasible task scheduling table is generated, and network task scheduling is performed. This enables the provision of optimal resource allocation schemes for tasks in a highly dynamic, multimodal network environment, improving the overall network throughput and the stability of network task execution.

[0020] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and are not intended to limit this disclosure. Attached Figure Description

[0021] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this disclosure and, together with the description, serve to explain the principles of this disclosure, and are not intended to unduly limit this disclosure.

[0022] Figure 1 This is a flowchart of a network task scheduling method according to an embodiment of the present disclosure; Figure 2 This is a flowchart of a network task scheduling method according to an embodiment of the present disclosure; Figure 3 This is a flowchart of a network task scheduling method according to an embodiment of the present disclosure; Figure 4 This is a flowchart of a network task scheduling method according to an embodiment of the present disclosure; Figure 5 This is a schematic diagram of the structure of a network task scheduling device according to an embodiment of the present disclosure; Figure 6 This is a structural block diagram of an electronic device according to an exemplary embodiment of the present disclosure. Detailed Implementation

[0023] To enable those skilled in the art to better understand the technical solutions of this disclosure, the technical solutions in the embodiments of this disclosure will be clearly and completely described below with reference to the accompanying drawings.

[0024] It should be noted that the terms "first," "second," etc., used in the specification, claims, and accompanying drawings of this disclosure are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this disclosure described herein can be implemented in orders other than those illustrated or described herein. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this disclosure. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this disclosure as detailed in the appended claims.

[0025] Currently, with the popularization of 5G, 6G, and low-Earth orbit satellite networks, edge computing is widely used in modern communication networks. However, existing task scheduling methods usually assume that the network environment is statically stable and rely on a single resource parameter for scheduling. They lack effective integration of multimodal network information, resulting in poor adaptability and low resource allocation efficiency in dynamic network environments such as bandwidth fluctuations, latency changes, and link instability. At the same time, in related technologies, the anomaly detection of edge computing nodes and standard links relies on static thresholds, and the importance assessment of edge computing nodes and standard links lacks a flexible adjustment mechanism, making it difficult to meet the requirements for accuracy and reliability of task scheduling in highly dynamic and multimodal network environments.

[0026] Figure 1 This is a flowchart of a network task scheduling method according to an embodiment of the present disclosure. It should be noted that the network task scheduling method of this embodiment can be applied to a network task scheduling device, which can be configured in an electronic device to enable the electronic device to perform network task scheduling functions.

[0027] The electronic device can be any device with computing capabilities, such as a terminal device or a server. The server can be a server used in network task scheduling, etc. The following embodiments use a server that performs network task scheduling as an example.

[0028] like Figure 1 As shown, the method includes the following steps: Step 101: Collect a network multimodal dataset. The multimodal dataset includes multi-source link data of multi-standard links in the network, computing power index data of edge computing nodes, real-time network status, and topology snapshots that characterize network topology information.

[0029] In this embodiment, one method for acquiring a network multimodal dataset is as follows: Radio frequency measurement parameters of multi-standard links are collected via radio frequency measurement, and latency, bandwidth, and jitter parameters of each standard link are obtained via link probing to form multi-source link data; processor utilization, memory usage, task queue length, and cache status of each edge computing node are collected via edge computing node resource monitoring to form computing power index data; timestamp alignment is performed on the multi-source link data and computing power index data to eliminate asynchronous deviations, resulting in aligned multi-source link data; available bandwidth and latency of each standard link are obtained based on the aligned multi-source link data; available computing power and load status of each edge computing node are obtained based on the aligned computing power index data, generating real-time network status; a topology snapshot representing network topology information is constructed using a network topology adjacency matrix; and the aligned multi-source link data, computing power index data, real-time network status, and topology snapshot are integrated to generate a multimodal dataset. By integrating multi-source link data from multiple standards, edge computing power indicators, network topology information, and network status, a multi-dimensional and real-time perception of the global network status is achieved, providing a comprehensive and accurate data foundation for subsequent edge computing node anomaly detection, standard link evaluation, and task scheduling.

[0030] The multi-standard links include 5G links, 6G links, Wireless Fidelity (WiFi) links, and low-Earth orbit satellite links, but are not limited to these. Radio frequency measurement parameters include signal strength, signal-to-noise ratio, and packet loss rate. Link detection can be achieved by sending link detection packets via a link detection protocol, including Internet Control Message Protocol (ICMP) detection packets and custom detection messages. Edge computing power metrics include processor utilization, memory usage, task queue length, and cache status.

[0031] In this embodiment, to eliminate asynchronous bias in multi-source link data, a unified timestamp is used to align the multi-source link data, resulting in time-aligned multi-source link data after eliminating asynchronous bias. Further, the aligned multi-source link data signal can be processed using Fast Fourier Transform to obtain periodic interference characteristics, and the measured time-series multi-source link data can be processed using a sliding window averaging method to obtain the smoothed channel state. The periodic interference characteristics include the main interference frequency components, interference intensity, interference energy distribution spectrum, and the high-frequency to low-frequency energy ratio.

[0032] In this embodiment of the disclosure, a network topology snapshot can be constructed through a network topology adjacency matrix to obtain multiple network coverage relationships and perform sparse compression storage.

[0033] For example, after receiving a task request from a user terminal, an edge computing node begins a series of radio frequency (RF) probing operations. It uses RF probes to perform RF measurements on 5G, WiFi, and LEO satellite networks, recording information such as signal strength, signal-to-noise ratio (SNR), and packet loss rate for each type of link. When performing RF measurements on the 5G network, the RF probe returned a signal strength of -75 dBm, an SNR of 30 dB, and a packet loss rate of 2%. For the WiFi link, the signal strength was -65 dBm, the SNR was 40 dB, and the packet loss rate was 1%. For the LEO satellite link, the signal strength was -80 dBm, the SNR was 25 dB, and the packet loss rate was 5%. By sending ICMP probe packets and custom probe messages through the link probe protocol, the latency, bandwidth, and jitter parameters of each link are obtained: the latency of the 5G link is 10ms, the bandwidth is 500Mbps, and the jitter parameter is 2ms; the latency of the WiFi link is 15ms, the bandwidth is 200Mbps, and the jitter parameter is 4ms; the latency of the low-Earth orbit satellite link is 100ms, the bandwidth is 50Mbps, and the jitter parameter is 10ms.

[0034] Based on this, edge computing power metrics, including processor utilization, memory usage, task queue length, and cache status, are obtained through the system monitoring agent of the edge computing node. If the node's processor utilization is 80%, memory usage is 70%, task queue length is 15, and cache status is at full load, this indicates that the node's current computing resources are under pressure, which may affect task execution efficiency.

[0035] To ensure the synchronization of multi-source data, the multi-source link data is aligned using a unified timestamp to eliminate deviations caused by asynchronicity, resulting in aligned multi-source link data. This data is then processed using Fast Fourier Transform (FFT) to extract the frequency characteristics of periodic interference: the main interference frequency component for the 5G link is 2.4GHz, with an interference intensity of -50dBm and a high-to-low frequency energy ratio of 1.2, indicating relatively balanced interference in this frequency band; the interference frequency for the WiFi link is 5GHz, with an interference intensity of -60dBm and an energy ratio of 1.5, indicating that interference is slightly dominant in the high-frequency range; the interference frequency band for the low-Earth orbit satellite link is 12GHz, with an interference intensity of -75dBm and an energy ratio of 2, indicating that interference is mainly concentrated in the low-frequency band.

[0036] A topological snapshot of the network is constructed using a network topology adjacency matrix to obtain detailed data on different network coverage relationships. For example, assuming that the coverage areas of 5G, WiFi, and LEO satellite networks overlap within a given region, WiFi links have a smaller coverage area but higher bandwidth, 5G links offer lower bandwidth but lower latency, and LEO satellite links offer wide coverage but higher latency. To optimize network utilization, sparse compression storage is performed to reduce storage pressure while preserving key characteristics of the network topology.

[0037] Finally, by integrating multi-source link data, topology snapshots, task request descriptions, and computing power metrics data to represent network status, a unified multimodal state representation is obtained. Assuming a task request description includes a task size of 5GB, a deadline of 30 minutes, and a fault tolerance requirement of 3%, the network conditions, link performance, and edge node resources are comprehensively considered to assess whether the task requirements can be met. For this task, analysis reveals that 5G links can provide low latency and sufficient bandwidth, WiFi links offer greater bandwidth but have limited coverage, and low-Earth orbit (LEO) satellite links have wide coverage but higher latency. Therefore, during resource scheduling, priority is given to using 5G and WiFi links for data transmission, while LEO satellite links serve as redundant links for task backup.

[0038] Step 102: Based on the computing power index data, perform anomaly detection and availability level assessment on the operating status of each edge computing node, and generate available edge computing nodes and their corresponding availability levels.

[0039] In this embodiment, one method for evaluating available edge computing nodes and their corresponding availability levels is as follows: Construct a state time series for each edge computing node based on its computing power metrics; perform anomaly detection on the state time series of each edge computing node using an anomaly detection model, and calculate anomaly scores for each edge computing node; based on the anomaly scores, select available edge computing nodes from among them; calculate the fluctuation index of the state metrics for each available edge computing node, where the fluctuation index characterizes the stability of the edge computing node's operating state; and determine the availability level of each available edge computing node based on the anomaly scores and fluctuation indexes. Thus, based on the aforementioned available edge computing nodes and their corresponding availability levels, through a dual evaluation mechanism of anomaly detection models and fluctuation indices, accurate identification of abnormal states of edge computing nodes and quantitative classification of their availability are achieved, providing a reliable basis for node selection in subsequent task scheduling.

[0040] Step 103: Based on multi-source link data and topology snapshots, evaluate the confidence level of each type of link and generate the confidence level of each type of link.

[0041] In this embodiment of the disclosure, one method for evaluating the confidence level of each standard link can be as follows: Based on multi-source link data, the bandwidth of each standard link is used as a positive weighting factor, and the latency of each standard link is used as a negative weighting factor to perform weighted calculation on the standard link connections between edge computing nodes in the network, and a weighted adjacency matrix is ​​constructed; the topology snapshot is time-series aligned with the weighted adjacency matrix, and the aligned weighted adjacency matrix is ​​time-sliced ​​to obtain a topology snapshot sequence that changes over time; the topology influence level of each standard link is determined based on the connection relationship and edge attributes of each standard link in the topology snapshot sequence; based on the historical packet loss rate data and real-time packet loss rate data of each standard link, the packet loss rate of each standard link in the next time window is regression-fitted using a time-series prediction model to obtain a packet loss rate prediction value; and the confidence level of each standard link is generated based on the topology influence level and the packet loss rate prediction value. Therefore, based on the confidence levels of the aforementioned links of various standards, a quantitative assessment of the reliability of multi-standard links was achieved through the graph topology centrality measure of the weighted adjacency matrix and topology snapshot sequence, as well as the packet loss rate prediction of the time series prediction model, providing a confidence basis for link selection in subsequent task scheduling.

[0042] Step 104: Based on the task type corresponding to the task request received from the network and the network status, decompose the task in the task request into multiple sub-tasks with priority tags.

[0043] In this embodiment, one method for decomposing multiple subtasks with priority tags is as follows: The descriptive text of the task request is classified using a semantic parsing model to identify the task type; a corresponding basic decomposition template is matched according to the task type, the basic decomposition template defining the subtask composition of each task type under standard conditions and the dependencies between subtasks; the computational load and bandwidth requirements of the task at the current moment are obtained based on the network state, and the interaction factor between the computational load and bandwidth requirements is calculated; wherein, the interaction factor is determined based on the nonlinear coupling relationship between the computational load and bandwidth requirements; the decomposition granularity is determined based on the magnitude of the interaction factor; based on the subtask composition and dependencies defined in the basic decomposition template, each subtask in the basic decomposition template is merged or split according to the determined decomposition granularity to generate multiple subtasks adapted to the current network state, and an execution priority is marked for each subtask. This achieves fine-grained task decomposition and priority marking, improving the flexibility and efficiency of task scheduling.

[0044] It should be noted that the system can also calculate the task scheduling latency caused by the interaction between computational load and bandwidth requirements based on the interaction factor; calculate the task's failure probability at the current moment based on the interaction factor and task scheduling latency; and adjust the decomposition granularity when the failure probability exceeds a preset probability threshold. This achieves closed-loop optimization of the task decomposition strategy, effectively reducing the risk of task execution failure due to the mismatch between computational load and bandwidth requirements.

[0045] Task requests can be received via an upper-level scheduling controller or user terminal. These requests are input as structured descriptive text through a task description interface, including the task name, functional description, and performance constraints. The semantic parsing model's processing includes: reading the input task description text through the task semantic parsing unit; performing preprocessing operations, including word segmentation, stop word removal, and vectorization encoding, to obtain a standardized semantic input sequence; and then calling a pre-trained language model (Bidirectional Encoder Representations from Transformers, BERT) to perform multi-classification inference on the standardized semantic input sequence to identify the task's category label, i.e., the task type. Task category labels include control command type, high-definition video type, status monitoring type, and hybrid type.

[0046] In this embodiment, the basic decomposition template is constructed as follows: based on the task type labels and task attribute parameters in historical task execution samples, a recurrent neural network is used for model training to construct a task decomposition prediction model. The optimal decomposition granularity parameters for different task types are obtained, and the composition (candidate size) and dependencies of each subtask are determined. The optimal decomposition granularity parameters and subtask dependency constraints output by the task decomposition prediction model constitute the basic decomposition template for that type of task. The task attribute parameters can be extracted from the task description text using keyword matching and dependency parsing methods, extracting maximum latency, bandwidth requirements, and computational load to form a set of task attribute parameters.

[0047] In this embodiment of the disclosure, the interaction factor calculation formula is used to calculate the interaction factor between the computational load and bandwidth requirement of a task at a certain moment. The calculation formula is as follows: Where C(t) represents the task computation load at time t, B(t) represents the bandwidth requirement at time t, and k is a nonlinear exponent that reflects the interaction between computation load and bandwidth, obtained by fitting historical data.

[0048] Furthermore, the task scheduling delay can be calculated based on the interaction factor, using the following formula: Where D(t) represents the scheduling delay at time t, and T(t) represents the remaining execution time of the task at time t. It is a moderating factor that reflects the degree of impact of the interaction between computing load and bandwidth demand on task execution latency. It is obtained by fitting historical data and the task priority is sorted according to the calculated task scheduling latency.

[0049] Furthermore, the probability of task execution failure can be calculated based on the interaction factor and task scheduling delay, using the following formula: ,in, This is a coefficient used to adjust the impact of scheduling latency on the probability of task failure, obtained through fitting historical data. When the probability of execution failure exceeds a preset probability threshold, the decomposition granularity is adjusted. This includes using a coarser decomposition granularity for tasks with high interaction factors and a finer granularity for tasks with low interaction factors. By recursively adjusting the nodes in the task decomposition tree, the execution path of the task is flexibly adjusted according to the current network status and resource availability, ensuring that the task achieves an optimal balance between bandwidth and computational load.

[0050] In this embodiment, the task decomposition tree is constructed as follows: the root node represents the overall task, and the leaf nodes represent the smallest decomposable computational units, obtaining the set of subtasks and their corresponding data sizes; the task input data and computational load are recursively divided, and the decomposition level is dynamically determined according to the task size and parallelism threshold to form a multi-level task decomposition tree structure, where the number of tree levels is defined as the decomposition depth. For tasks with high interaction factors, a coarser decomposition granularity is used; for tasks with low interaction factors, a finer granularity is used.

[0051] In this embodiment, the optimal decomposition depth is determined as follows: A decomposition depth range is set based on the task decomposition tree structure; at each decomposition depth, the corresponding communication overhead and latency cost are calculated based on the number of subtasks, data dependencies, and transmission bandwidth; and a dynamic programming method is used to recursively calculate the comprehensive cost function at different decomposition depths. The process involves determining whether the decomposition level is too coarse or too fine; based on the obtained cost distribution sequence, an optimization search is performed in the decomposition depth and communication cost space to determine the optimal decomposition level; and the task decomposition tree is reconstructed based on the optimal decomposition level to obtain a set of subtasks with priority labels.

[0052] For example, a user submits a task request through the upper-level scheduling controller. This request includes a task name, a functional description, and performance constraints. The task description is a high-definition video processing task, requiring completion within 30 minutes and demanding high bandwidth, at least 10 Mbps. Simultaneously, the task has a high computational load requirement, involving complex calculations such as video decoding and image processing. The task request also mentions latency requirements, demanding that the processing latency for each video frame not exceed 200 milliseconds.

[0053] After receiving a task request, the system first parses the task text through the task description interface to extract task attribute parameters. After word segmentation, stop word removal, and vectorization encoding by the task semantic parsing unit, a standardized semantic input sequence is obtained. A pre-trained BERT model is then invoked to perform multi-class classification reasoning on the task, identifying it as belonging to the high-definition video category. Specific task attribute parameters are extracted from the task description, including a maximum latency of 200 milliseconds, bandwidth requirement of 10 Mbps, and high computational load; these parameters are stored as a task attribute set.

[0054] Based on data from historical task execution samples, a recurrent neural network was used to train the model. During training, the optimal decomposition granularity for the video processing task was predicted according to different task labels and task attribute parameters. It was found that a decomposition level of depth 3 was most suitable for the video processing task, the dependencies between subtasks were reasonably divided, and the computational load was effectively distributed.

[0055] The interaction factor is calculated based on the real-time computational load and bandwidth requirements of the task. Assuming at a certain time t, C(t) = 80, representing the computational resource requirement as a percentage, and B(t) = 100 Mbps, and based on historical data, the system pre-sets a nonlinear index k = 2, then the interaction factor... The value of 0.64 indicates a certain degree of coupling between computational load and bandwidth requirements. A function of task scheduling latency is used to evaluate the impact of computational load and bandwidth requirements on task execution latency. Assuming the remaining execution time of the task is T(t) = 120ms, representing the remaining processing time of the task at the current time, and an adjustment factor is obtained by fitting historical data. The value is 0.5, using the function for task scheduling delay. The task execution latency caused by the interaction of computational load and bandwidth requirements is determined to be 38.4 ms. Therefore, the task's execution latency at the current moment is 38.4 milliseconds, reflecting the additional latency caused by the interaction between computational load and bandwidth requirements. Based on the interaction factor between task computational load and bandwidth requirements and a function of task scheduling latency, the task failure prediction factor formula is used. Calculate the probability that the task will fail at time t. If the coefficients are obtained by fitting historical data If the interaction factor is 0.3, the calculated failure prediction factor for the task is 0.17312, meaning the probability of task failure is 17.3%, which is greater than the preset probability threshold of 15%. This indicates a low likelihood of timely task completion and may lead to failure. Based on the calculation results, the task decomposition granularity is adjusted. Since the task's interaction factor is high (0.64), a coarser decomposition granularity is adopted. This means the task will be decomposed into fewer subtasks to reduce conflicts and dependencies between computational load and bandwidth requirements, thereby improving the task's success rate and execution efficiency. Finally, the task decomposition tree is adjusted according to the current network status and resource availability. Due to the limited bandwidth and computational resources for the task, edge computing nodes with higher bandwidth and lower latency are prioritized for processing, ensuring the task executes in an optimal resource environment and minimizing the risk of latency and failure.

[0056] By constructing a task decomposition tree, the root node represents the entire video processing task, and the leaf nodes represent the smallest decomposable computational units. For example, one leaf node represents the video decoding operation, and another leaf node represents the image processing operation of a video frame. The task input data and computational load are recursively divided, and the decomposition level is dynamically determined based on the task size and parallelism thresholds. The depth of the task decomposition tree is set to 3.

[0057] To evaluate the cost of each decomposition level, communication overhead and latency costs are calculated based on the number of subtasks, data dependency, and transmission bandwidth at each level. Decomposition level 1 has 5 subtasks, each with an average bandwidth requirement of 2 Mbps, low data dependency, and a latency cost of 50 milliseconds. Decomposition level 2 has 10 subtasks, each with an average bandwidth requirement of 1 Mbps, high data dependency, and a latency cost of 80 milliseconds. A dynamic programming approach is used to recursively calculate the comprehensive cost function at different decomposition depths. The integrated cost function includes computational and communication costs, and its goal is to find a balance point that allows the task to be completed within a reasonable time while maximizing resource utilization. The system found that a decomposition level with a depth of 3 provides the minimum integrated cost.

[0058] Ultimately, the system finds the optimal task decomposition scheme within the decomposition depth and communication cost space through optimization search, and reconstructs the task decomposition tree based on this optimal decomposition level, obtaining a set of subtasks with priority labels. Some of these subtasks are executed first to ensure the timely completion of the entire task. At this point, the task decomposition strategy has been optimized. The system will then allocate the subtasks to different edge computing nodes based on the specific requirements of the task and the resource status of the edge computing nodes, ensuring that the task can be completed on time and meets the user's performance requirements.

[0059] Step 105: Based on the available edge computing nodes and their corresponding availability level, confidence level, and multiple sub-tasks, generate a feasible task scheduling table and schedule network tasks according to the task scheduling table.

[0060] In this embodiment, one method for generating a feasible task scheduling table is as follows: Based on the latency constraints, bandwidth requirements, and computational load of each subtask, a set of links and a set of nodes that meet the requirements of each subtask are selected from the network's standard links and available edge computing nodes; using availability level and confidence level as constraints, each subtask and the set of links are matched with the standard links and available edge computing nodes in the node set, with the optimization objective of minimizing the combined cost of latency, bandwidth utilization, and computational load of each subtask, to obtain the resource allocation scheme for each subtask; the feasibility of the resource allocation scheme is verified, and the verified resource allocation scheme is used as a feasible task scheduling table, and network task scheduling is performed according to the task scheduling table. This ensures that the task scheduling scheme achieves optimized resource allocation while meeting the requirements of each subtask, effectively improving network resource utilization and task execution efficiency.

[0061] Specifically, the resource scanning module can perform status detection on current network links and available edge computing nodes, and filter the set of links that meet latency constraints, bandwidth requirements, and computing load. With node set Record the available bandwidth for each type of link. and average latency And record the available computing power for each available edge computing node. With load rate This process generates an available resource table. By constructing a constraint matrix, the subtask requirement parameters are mapped to latency, bandwidth utilization, and computational load, generating a task resource matching condition matrix. This matrix represents the feasible mapping position of each subtask on each candidate resource. Based on the network topology snapshot, a set of end-to-end candidate transmission paths is generated for each subtask. Each path consists of a custom link and a sequence of available edge computing nodes.

[0062] In this embodiment of the disclosure, the resource allocation scheme is solved by modeling the task resource matching problem as a 0-1 integer programming problem. ,in, This represents the time delay from available edge computing node i to j. Indicates task Assigned to resource j These are weighting coefficients, obtained by fitting historical data. The objective function is to minimize the combined cost of latency constraints, bandwidth requirements, and computational load. The constraints are: , , .in, The bandwidth requirement for subtask i. The computational power requirement for subtask i.

[0063] Furthermore, for critical subtasks with a priority higher than a preset threshold, a multi-path redundancy allocation strategy is adopted, simultaneously assigning critical subtasks to multiple standard links with confidence levels higher than the preset threshold for parallel transmission. For ordinary subtasks with a priority lower than the preset threshold, allocation is performed sequentially according to the remaining resources in the link set and node set, prioritizing the nodes with the lowest load rate and idle links to obtain a resource allocation scheme for each subtask. By establishing a task path mapping table, a multi-path scheduling scheme for critical tasks and a complete task scheduling table are generated.

[0064] By verifying the feasibility of the task scheduling table, it is determined whether the bandwidth and latency constraints are met, and a feasible task scheduling table is determined to obtain the mapping relationship and bandwidth allocation ratio between tasks and network standard links and available edge computing nodes.

[0065] For example, a user submits a task involving high-definition video processing, which includes multiple subtasks, each with specific bandwidth requirements and computational load. Upon system startup, the resource scanning module first performs status checks on various standard links and edge computing nodes in the current network, filtering for sets of links and nodes that meet latency constraints, bandwidth requirements, and computational load. For each standard link, the available bandwidth and average latency are recorded; for each available edge computing node, the available computing power and load rate are recorded. It is found that standard link 1 has a bandwidth requirement of 200Mbps and a latency of 30ms; standard link 2 has a bandwidth of 100Mbps and a latency of 50ms; and standard link 3 has a bandwidth requirement of 300Mbps and a latency of 20ms. The status information of the available edge computing nodes is as follows: available edge computing node A has an available computing power of 1000MIPS and a load rate of 70%; available edge computing node B has an available computing power of 500MIPS and a load rate of 60%; and available edge computing node C has an available computing power of 2000MIPS and a load rate of 50%.

[0066] Based on this information, an available resource table is created, recording the bandwidth, latency, computing power, and load rate of each link. Task descriptions include: Subtask 1 (video decoding) requires 150Mbps bandwidth and has a computational load of 500MIPS; Subtask 2 (image effects processing) requires 100Mbps bandwidth and has a computational load of 400MIPS; Subtask 3 (video encoding) requires 80Mbps bandwidth and has a computational load of 300MIPS. Based on the network topology snapshot, a set of end-to-end candidate transmission paths is generated for each subtask. Each path consists of a specific standard link and a sequence of available edge computing nodes.

[0067] A constraint matrix is ​​constructed to match the requirements of each subtask with the network's standard links and available edge computing nodes. The goal is to allocate each subtask to appropriate resources while satisfying bandwidth requirements and computational load constraints. The task resource matching problem is modeled as a 0-1 integer programming problem, with the optimization objective being to minimize the overall cost. This includes latency, bandwidth utilization, and computational load. Using the branch-and-bound method, a preliminary scheduling solution is obtained: subtask 1 is assigned to standard link 1 and available edge computing node A; subtask 2 is assigned to standard link 2 and available edge computing node B; and subtask 3 is assigned to standard link 3 and available edge computing node C.

[0068] According to the scheduling scheme, a multi-path redundancy allocation strategy is adopted based on task priority. Subtask 1 is a critical task, requiring high bandwidth and low latency. Therefore, the system prioritizes allocating it to link 1 with higher bandwidth and assigns the task to node A with lower computational load. If the bandwidth of link 1 is insufficient, the system selects link 3 as a backup path.

[0069] For the remaining low-priority subtasks, based on the remaining bandwidth requirements and the computing power of available edge computing nodes, the available edge computing node with the lowest load and the idle standard link are selected first. Subtask 3 has a smaller bandwidth requirement, and the available edge computing node C has a lower load; therefore, subtask 3 is assigned to available edge computing node C and standard link 3.

[0070] Feasibility was verified using a task scheduling table to ensure that the selected resource allocation scheme met bandwidth requirements and latency constraints. After verification, the system determined the mapping relationship between tasks and network standard links and available edge computing nodes, as well as the bandwidth allocation ratio for each standard link: Standard link 1 provides 150Mbps of bandwidth to subtask 1, standard link 2 provides 100Mbps to subtask 2, and standard link 3 provides 80Mbps to subtask 3.

[0071] In the network task scheduling method of this disclosure, a network multimodal dataset is collected, including multi-source link data, computing power index data, network status, and topology snapshots representing network topology information. Based on the network multimodal dataset, available edge computing nodes and their corresponding availability levels, as well as the confidence levels of links of various standards in the network, are detected. According to the task type corresponding to the task request received by the network and the network status, the task in the task request is decomposed into multiple sub-tasks with priority labels. Based on the available edge computing nodes and their availability levels, confidence levels, and multiple sub-tasks, a feasible task scheduling table is generated and network task scheduling is performed. Thus, an optimal resource allocation scheme for tasks is provided in a highly dynamic, multimodal network environment, improving the overall network throughput and the stability of network task execution.

[0072] To explain how anomaly detection and availability level assessment of the operational status of each edge computing node are performed based on computing power metrics data, generating available edge computing nodes and their corresponding availability levels, and generating audit task instructions based on the target audit intent, Figure 2 This is a flowchart of a network task scheduling method according to an embodiment of the present disclosure. It should be noted that the network task scheduling method of this embodiment can be applied to a network task scheduling device, which can be configured in an electronic device to enable the electronic device to perform network task scheduling functions.

[0073] Among them, electronic devices can be any device with computing capabilities, such as terminal devices, servers, etc.

[0074] like Figure 2 As shown, the method includes the following steps: Step 201: Collect a network multimodal dataset. The multimodal dataset includes multi-source link data of multi-standard links in the network, computing power index data of edge computing nodes, real-time network status, and topology snapshots that characterize network topology information.

[0075] Step 202: Construct the state time series of each edge computing node based on the computing power index data of each edge computing node.

[0076] In this embodiment, the processor utilization, memory usage, task queue length, and processing latency of each edge node can be continuously collected using radio frequency probes, and stored according to timestamps to form a state time series of the edge computing nodes. Specifically, the moving average and standard deviation of the state indicators are calculated for each edge computing node. If the current observation value exceeds the dynamic threshold range, it is determined as a deviation point, and the same detection is performed on all state indicators to obtain the cumulative deviation value and generate a set of edge computing nodes that exceed the limit.

[0077] For example, in an edge computing network, the resource usage of each edge computing node is continuously collected using radio frequency probes, and the data is recorded by timestamp to form a time series of edge computing node status. Using processor utilization, memory usage, task queue length, and processing latency as the main monitoring indicators, the resource data of edge computing node A over five consecutive time windows from T1 to T5 are as follows: processor utilization is 80%, 82%, 85%, 87%, and 90%; memory usage is 70%, 71%, 72%, 73%, and 75%; task queue length is 15, 16, 18, 20, and 22; and processing latency is 100ms, 105ms, 110ms, 115ms, and 120ms. The fluctuation of each indicator is monitored by calculating its moving average and standard deviation. The moving average of processor utilization is 85%, and the standard deviation is 3.4%.

[0078] If the preset dynamic threshold ranges are: processor utilization upper limit of 90%, lower limit of 80%, and memory utilization upper limit of 75%, lower limit of 65%, then at time T5, edge computing node A's processor utilization is 90%, exactly equal to the upper limit and within the threshold; however, its memory utilization is 75%, reaching the upper limit; and its task queue length is 22, exceeding the preset normal range. Therefore, edge computing node A is determined to be an deviation point, and the cumulative deviation value is recorded.

[0079] Step 203: Use the anomaly detection model to perform anomaly detection on the state time series of each edge computing node and calculate the anomaly score of each edge computing node.

[0080] In this embodiment of the disclosure, one method for detecting anomaly scores is as follows: by constructing a random segmentation tree to isolate each state index in the state time series of each edge computing node, so as to obtain the isolated path length of each state index in the random segmentation tree; and by calculating the anomaly score of each edge computing node based on the isolated path length.

[0081] Specifically, an isolation forest model can be used for anomaly detection. This model constructs a randomized split tree to calculate the length of isolated paths in a sample, obtaining anomaly scores for nodes. The nodes are then sorted in descending order of their anomaly scores to create a ranking table of anomaly nodes in edge computing. Based on the anomaly scores and a preset anomaly threshold, the anomaly types of edge computing nodes can be identified, including normal edge computing nodes, mildly anomalous edge computing nodes, and severely anomalous edge computing nodes.

[0082] For example, multiple random split trees are constructed using an isolated forest model, and the isolated path length for each sample is calculated. The anomaly score for edge computing node A is calculated as 0.8, the anomaly score for edge computing node B is 0.5, and the anomaly score for edge computing node C is 0.9. Based on the anomaly scores, edge computing node A is ranked first in descending order.

[0083] Step 204: Based on the anomaly scores, select available edge computing nodes from each edge computing node.

[0084] In this embodiment of the disclosure, nodes with anomaly scores lower than a preset severe anomaly threshold are selected as usable edge computing nodes by comparing with a preset anomaly threshold. For example, by comparing with the preset anomaly threshold, edge computing node A is identified as a severely abnormal edge computing node, edge computing node B is a normal edge computing node, and edge computing node C is a slightly abnormal edge computing node. Therefore, edge computing nodes B and C are selected as usable edge computing nodes.

[0085] Step 205: Calculate the volatility index of the status indicators of each available edge computing node. The volatility index is used to characterize the stability of the operating status of the edge computing node.

[0086] In this embodiment, one method for calculating the volatility index is as follows: calculate the standard deviation of each indicator within the time window, and then calculate the volatility index of node n using the volatility index formula. The volatility index formula is: ,in, The standard deviation of the i-th state index Let be the average value of the i-th state index. The weight of the i-th state index is obtained by fitting historical data.

[0087] For example, to evaluate the stability of edge computing node A, the standard deviation of each state indicator within the time window is calculated, and its volatility index is calculated using the volatility index formula. If the calculated volatility index is 0.2, it indicates that the state of edge computing node A fluctuates significantly.

[0088] Step 206: Determine the availability level of each available edge computing node based on the anomaly score and volatility index.

[0089] In this embodiment, a node availability level assessment model can be constructed by training a model using a decision tree algorithm based on historical data of anomaly scores and volatility indices. Availability levels include high, medium, and low. The availability level of each edge computing node is assessed using the node availability level assessment model based on real-time acquired anomaly scores and volatility indices over a preset time period.

[0090] For example, based on the volatility index and anomaly score, the node availability rating model trained using a decision tree algorithm classifies edge computing node A as low availability because of its high volatility index and large anomaly score. Edge computing node B has an anomaly score of 0.5 and a low volatility index, and is classified as high availability. Edge computing node C has an anomaly score of 0.9 but a moderate volatility index, and is classified as medium availability. For edge computing node A, the load on it is reduced or some tasks are transferred to other more stable edge computing nodes to ensure the efficient operation of the overall network and the timely completion of tasks.

[0091] Step 207: Based on multi-source link data and topology snapshots, evaluate the confidence level of each type of link and generate the confidence level of each type of link.

[0092] Step 208: Based on the task type corresponding to the task request received from the network and the network status, decompose the task in the task request into multiple sub-tasks with priority tags.

[0093] Step 209: Based on the available edge computing nodes and their corresponding availability levels, confidence levels, and multiple sub-tasks, generate a feasible task scheduling table and schedule network tasks according to the task scheduling table.

[0094] It should be noted that the execution processes of steps 201 and 207 to 209 can be implemented in any of the embodiments of this disclosure. This disclosure does not limit these processes and will not elaborate further.

[0095] In the network task scheduling method of this disclosure, a network multimodal dataset is collected. This multimodal dataset includes multi-source link data of multi-standard links in the network, computing power index data of edge computing nodes, real-time network status, and topology snapshots representing network topology information. Based on the computing power index data of each edge computing node, a state time series of each edge computing node is constructed. An anomaly detection model is used to detect anomalies in the state time series of each edge computing node, and an anomaly score is calculated for each edge computing node. Based on the anomaly scores, usable edge computing nodes are selected from among the edge computing nodes. Finally, the state index of each usable edge computing node is calculated. A volatility index characterizes the stability of edge computing node operation. Based on anomaly scores and volatility indices, the availability level of each available edge computing node is determined. Confidence levels for each link type are evaluated using multi-source link data and topology snapshots. Tasks in network requests are decomposed into multiple sub-tasks with priority tags based on the task type and network status. A feasible task scheduling table is generated based on available edge computing nodes, their corresponding availability and confidence levels, and the multiple sub-tasks, and network task scheduling is performed according to this table. This achieves accurate identification of abnormal edge computing node states and quantitative grading of availability, effectively improving the reliability and accuracy of subsequent task scheduling.

[0096] To explain how the confidence level of each link type is evaluated and generated based on multi-source link data and topology snapshots, Figure 3 This is a flowchart of a network task scheduling method according to an embodiment of the present disclosure. It should be noted that the network task scheduling method of this embodiment can be applied to a network task scheduling device, which can be configured in an electronic device to enable the electronic device to perform network task scheduling functions.

[0097] Among them, electronic devices can be any device with computing capabilities, such as terminal devices, servers, etc.

[0098] like Figure 3 As shown, the method includes the following steps: Step 301: Collect a network multimodal dataset. The multimodal dataset includes multi-source link data of multi-standard links in the network, computing power index data of edge computing nodes, real-time network status, and topology snapshots that characterize network topology information.

[0099] Step 302: Based on the computing power index data, perform anomaly detection and availability level assessment on the operating status of each edge computing node, and generate available edge computing nodes and their corresponding availability levels.

[0100] Step 303: Based on the multi-source link data, the bandwidth of each standard link is used as the positive weighting factor and the latency of each standard link is used as the negative weighting factor to perform weighted calculation on the standard link connections between edge computing nodes in the network, and a weighted adjacency matrix is ​​constructed.

[0101] In this embodiment, a node identification method can be used to add attributes to base stations, satellites, WiFi access points, edge servers, and terminals to obtain the bandwidth limit, historical packet loss rate, and available computing power of each edge computing node. For any two edge computing nodes in the network, the bandwidth value of the connection is used as a positive weighting factor, and the latency value is used as a negative weighting factor to calculate the bandwidth weight and latency weight. The bandwidth weight and latency weight are then used as edge attributes of the connection to form a weighted adjacency matrix. This weighted adjacency matrix uses edge computing nodes as nodes and standard link connections as edges, with the weight of each edge comprehensively reflecting the bandwidth capacity and latency cost of the connection.

[0102] For example, the network includes various types of edge computing nodes, including base stations, satellites, WiFi access points, edge servers, LEO satellites, and terminal devices. Each edge computing node has different attributes: the base station has a bandwidth limit of 1Gbps, a historical packet loss rate of 2%, and an available computing power of 1000MIPS; the WiFi access point has a bandwidth limit of 500Mbps, a historical packet loss rate of 1%, and an available computing power of 300MIPS; the edge server has a bandwidth limit of 10Gbps, a historical packet loss rate of 1.5%, and an available computing power of 5000MIPS; the LEO satellite has a bandwidth limit of 50Mbps, a historical packet loss rate of 5%, and an available computing power of 200MIPS; and the terminal device has a bandwidth limit of 100Mbps, a historical packet loss rate of 3%, and an available computing power of 50MIPS. The bandwidth of the link connection between the base station and the WiFi access point is 1Gbps, and the latency is 50ms; therefore, the bandwidth weight of this connection is 1Gbps, and the latency weight is 50ms. The bandwidth of the standard link connection between the base station and the satellite is 50Mbps, and the latency is 300ms. Therefore, the bandwidth weight of this connection is 50Mbps, and the latency weight is 300ms. Through this weighted calculation, edge weights are assigned to the standard link connection between any two edge computing nodes in the network, forming a weighted adjacency matrix.

[0103] Therefore, by constructing a weighted adjacency matrix with bandwidth as the positive weighting factor and latency as the negative weighting factor, the connection relationships between edge computing nodes in the network are transformed into a quantifiable and computable structured representation, providing a data foundation for subsequent topology snapshot slicing and graph topology analysis.

[0104] Step 304: Time-series align the topology snapshots with the weighted adjacency matrix, and perform time slicing on the aligned weighted adjacency matrix to obtain a sequence of topology snapshots that change over time.

[0105] In this embodiment, to capture the dynamic changes in network topology over time, the topology snapshots collected in step 101 and the weighted adjacency matrix constructed in step 303 are time-sequentially aligned according to timestamps, ensuring that the topological connectivity relationships and corresponding edge weights at each time point remain consistent. Based on this, a fixed time window method is used to time-slice the aligned weighted adjacency matrix, dividing the continuous network state into multiple discrete time windows. Each time window corresponds to a slice of the weighted adjacency matrix, thereby obtaining a sequence of topology snapshots that change over time. Each topology snapshot records the connectivity relationships and corresponding edge weights between edge computing nodes in the network within that time window.

[0106] For example, a 5-minute time window is selected to record the changes in network topology within that time period. Within each 5-minute time window, a corresponding weighted adjacency matrix slice is generated, and multiple consecutive time windows form a topology snapshot sequence. This topology snapshot sequence reflects the temporal evolution of the connection relationships between edge computing nodes and the weights of standard links in the network.

[0107] Therefore, by temporally aligning the topology snapshot with the weighted adjacency matrix and performing time slicing, the static network topology representation is transformed into a dynamic temporal topology sequence, enabling subsequent graph topology analysis to capture the time-varying characteristics of the network structure.

[0108] Step 305: Determine the topology influence level of each link based on the connection relationship and edge attributes of each link in the topology snapshot sequence.

[0109] In this embodiment of the disclosure, based on the topology snapshot sequence obtained in step 304, within each time window, the network topology is analyzed using a graph neural network algorithm with weighted adjacency matrix slices as input to determine the topological influence level of each type of link. As an example, a graph sampling and aggregation algorithm (GraphSAGE) can be used to perform neighbor sampling and edge attribute aggregation operations within each time window. Specifically, for each edge computing node, its neighboring nodes in the topology snapshot (i.e., other nodes connected to it by a type of link) are sampled, and the attribute features of the neighboring nodes and the weight information of the connecting edges are aggregated to generate a node embedding vector for that edge computing node. This node embedding vector integrates the attribute features of the edge computing node itself and its structural position information in the network topology, and can characterize the topological importance of the edge computing node in the network.

[0110] Furthermore, based on the node embedding vectors of each edge computing node output by the GraphSAGE algorithm, and combined with performance metrics such as bandwidth contribution, latency stability, task success rate, and availability of each edge computing node evaluated in historical task execution records, the topology impact of the links between edge computing nodes is assessed. The evaluation logic for topology impact is: the higher the topological importance and the better the performance metrics of the edge computing nodes at both ends of the link, the higher the topology impact level of that link. By fusing the node embedding vectors with performance metrics, the topology impact level of each link can be comprehensively determined.

[0111] For example, in the past 10 task executions, the base station contributed 8Gbps of bandwidth, with high latency stability and a task success rate of 95%; the WiFi access point contributed 2Gbps of bandwidth, with relatively stable latency and a task success rate of 90%; and the low-Earth orbit satellite contributed 30Mbps of bandwidth, with large latency fluctuations and a low task success rate of 70%. Based on the node embedding vectors output by the GraphSAGE algorithm, and combined with the above performance indicators, the base station, due to its critical topological location and excellent performance indicators, is rated as having a high topological impact level for the standard link connected to it; while the low-Earth orbit satellite, due to its poor performance indicators, is rated as having a low topological impact level for the standard link connected to it.

[0112] Furthermore, decision tree algorithms can be used to train models using node embedding vectors and performance metrics as input features, constructing a node importance assessment model that automatically identifies the topological influence level of each edge computing node, thereby improving the efficiency of task scheduling and resource allocation.

[0113] Therefore, by using a graph neural network algorithm to learn node embeddings in the topology snapshot sequence and combining historical performance indicators to evaluate the link topology impact, a quantitative classification of the importance of standard links in the network topology is achieved.

[0114] Step 306: Based on the historical and real-time packet loss rate data of each type of link, the packet loss rate of each type of link in the next time window is regression-fitted using a time-series prediction model to obtain the predicted packet loss rate.

[0115] In this embodiment, the time-series prediction model is used to predict the packet loss rate trend of each link in the future time period based on the historical packet loss rate time series and the current real-time packet loss rate data. As an example, the Autoregressive Integrated Moving Average (ARIMA) model can be used as the time-series prediction model to model and predict the packet loss rate time series of each link. The ARIMA model captures the historical dependencies of the packet loss rate series through the autoregressive part, transforms the non-stationary packet loss rate series into a stationary series through the differential integration part, and captures random fluctuations through the moving average part, thereby outputting the predicted packet loss rate value for the next time window.

[0116] Furthermore, after obtaining the predicted packet loss rate for each type of link, a linear regression algorithm can be used to combine the predicted packet loss rate with the historical availability data of the link to calculate the short-term availability distribution of the link, providing input for subsequent confidence assessment of the link.

[0117] For example, based on historical packet loss rate data, the ARIMA model predicts that the packet loss rate of the standard link between the base station and the WiFi access point will remain at around 1.5% in the next hour, indicating that the standard link is relatively stable; while the packet loss rate of the standard link between the base station and the satellite may rise to 6%, indicating that the standard link has a greater risk of quality degradation.

[0118] Therefore, by using a time-series prediction model to regress and fit the future packet loss rate of each standard link, the transmission quality of the standard link can be predicted in advance, providing forward-looking data support for the confidence assessment of the standard link.

[0119] Step 307: Generate the confidence level of each type of link based on the topology impact level and packet loss rate prediction value of each type of link.

[0120] In this embodiment, the overall reliability of each link is quantitatively evaluated by combining the topology impact level of each link determined in step 305 and the packet loss rate prediction value of each link obtained in step 306, generating a confidence level. Specifically, the availability prediction result of the link (obtained through a linear regression algorithm based on the packet loss rate prediction value) is weighted and fused with the historical average availability of the link to obtain a link confidence matrix. The link confidence matrix uses edge computing nodes as indices, and each element in the link confidence matrix represents the confidence value of the link between two corresponding edge computing nodes. This confidence value comprehensively reflects the topological importance and transmission quality reliability of the link.

[0121] Based on the link confidence matrix, the ranking of each edge computing node and its corresponding link type can be determined, and the confidence distribution of each edge computing node can be evaluated. A higher confidence level indicates that the link type is more critical in the topology and has more reliable transmission quality, and should be prioritized during task scheduling; a lower confidence level indicates that the link type is more peripheral in the topology or has unstable transmission quality, and should be considered as an alternative or redundant path during task scheduling.

[0122] For example, based on the analysis results of the link confidence matrix, the base station is located in the core position of the topology and has a low predicted packet loss rate. It is rated as the most reliable edge computing node in the network, and the standard link connected to it is given a high confidence level. On the other hand, the low-orbit satellite has a high predicted packet loss rate and large fluctuations in performance indicators. Its confidence level is low, and the standard link connected to it is given a low confidence level. It is given priority as a redundant path for task allocation during task scheduling.

[0123] Therefore, by generating confidence levels based on the topology impact level and packet loss rate prediction of the integrated links, a multi-dimensional quantitative assessment of the reliability of multi-standard links is achieved, providing an accurate confidence basis for link selection in subsequent task scheduling. In this way, the importance of edge computing nodes and the reliability of standard links can be dynamically evaluated in various network environments, optimizing task scheduling and resource allocation, ensuring that various tasks in the network can be completed efficiently, while reducing potential resource bottlenecks and latency risks.

[0124] Step 308: Based on the task type corresponding to the task request received from the network and the network status, decompose the task in the task request into multiple sub-tasks with priority tags.

[0125] Step 309: Based on the available edge computing nodes and their corresponding availability level, confidence level, and multiple sub-tasks, generate a feasible task scheduling table and schedule network tasks according to the task scheduling table.

[0126] It should be noted that the execution processes of steps 301, 302 and steps 308, 309 can be implemented in any of the embodiments of this disclosure. This disclosure does not limit these processes and will not elaborate further.

[0127] In the network task scheduling method of this disclosure embodiment, a network multimodal dataset is collected. The multimodal dataset includes multi-source link data of multi-standard links in the network, computing power index data of edge computing nodes, real-time network status, and topology snapshots representing network topology information. Based on the computing power index data, anomaly detection and availability level assessment are performed on the operating status of each edge computing node to generate available edge computing nodes and their corresponding availability levels. Based on the multi-source link data, the bandwidth of each standard link is used as a positive weighting factor, and the latency of each standard link is used as a negative weighting factor to perform weighted calculation on the standard link connections between edge computing nodes in the network to construct a weighted adjacency matrix. The topology snapshot and the weighted adjacency matrix are time-aligned, and the aligned weighted adjacency matrix is ​​time-sliced ​​to obtain the time-series data. The system generates a time-varying topology snapshot sequence. Based on the connection relationships and edge attributes of each link type in the topology snapshot sequence, it determines the topology impact level of each link type. Using historical and real-time packet loss rate data for each link type, it performs regression fitting on the packet loss rate of each link type in the next time window through a time-series prediction model to obtain predicted packet loss rates. Based on the topology impact level and predicted packet loss rates of each link type, it generates a confidence level for each link type. According to the task type corresponding to the task request received by the network and the network status, the tasks in the task request are decomposed into multiple sub-tasks with priority labels. Based on the available edge computing nodes and their corresponding availability levels, confidence levels, and multiple sub-tasks, a feasible task scheduling table is generated, and network task scheduling is performed according to the task scheduling table. This achieves a quantitative assessment of the reliability of multi-link systems, effectively improving the accuracy of link type selection in subsequent task scheduling.

[0128] Figure 4 This is a flowchart illustrating another embodiment of a network task scheduling method according to the present disclosure. It should be noted that the network task scheduling method of this embodiment can be applied to a network task scheduling device, which can be configured in an electronic device to enable the electronic device to perform network task scheduling functions.

[0129] Among them, electronic devices can be any device with computing capabilities, such as terminal devices, servers, etc.

[0130] like Figure 4 As shown, the method includes the following steps: Step 401: Collect a network multimodal dataset. The multimodal dataset includes multi-source link data of multi-standard links in the network, computing power index data of edge computing nodes, real-time network status, and topology snapshots that characterize network topology information.

[0131] Step 402: Based on the computing power index data, perform anomaly detection and availability level assessment on the operating status of each edge computing node, and generate available edge computing nodes and their corresponding availability levels.

[0132] Step 403: Based on multi-source link data and topology snapshots, evaluate the confidence level of each type of link and generate the confidence level of each type of link.

[0133] Step 404: Based on the task type corresponding to the task request received from the network and the network status, decompose the task in the task request into multiple sub-tasks with priority tags.

[0134] Step 405: Based on the available edge computing nodes and their corresponding availability levels, confidence levels, and multiple subtasks, generate a feasible task scheduling table and schedule network tasks according to the task scheduling table.

[0135] It should be noted that the execution process of steps 401 to 405 can be implemented in any of the embodiments of this disclosure. This disclosure does not limit this and will not elaborate further.

[0136] Step 406: After generating a feasible task scheduling table and performing network task scheduling according to the task scheduling table, collect real-time feedback data of each subtask during execution. The real-time feedback data includes the latency, packet loss rate, bandwidth utilization, and task completion rate of network task scheduling.

[0137] In this embodiment, after the feasible task scheduling table is sent to the execution layer via the control channel, each subtask begins execution on its assigned standard link and edge computing node. During task execution, the task feedback unit deployed on each edge computing node collects real-time feedback data on the execution of each subtask, including: completion latency (the time consumed from the start of execution to completion of the subtask), packet loss rate (the proportion of data packets lost during the data transmission process of the subtask), bandwidth utilization (the ratio of the actual bandwidth occupied by the subtask to the allocated bandwidth), and task completion rate (the ratio of the amount of computation completed by the subtask to the total amount of computation). This real-time feedback data is transmitted back to the network task scheduling server via the control channel to determine the execution effect of the current scheduling scheme.

[0138] For example, a user submits a video processing task involving multiple subtasks, including video decoding, image effects processing, and video encoding. The system has performed preliminary scheduling based on resource availability and task requirements: Subtask 1 (video decoding) is assigned to a standard link 1 with a bandwidth of 200Mbps and a latency of 30ms, and an available edge computing node A with a computing power of 1000MIPS; Subtask 2 (image effects processing) is assigned to a standard link 2 with a bandwidth of 150Mbps and a latency of 40ms, and an available edge computing node B with a computing power of 500MIPS; Subtask 3 (video encoding) is assigned to a standard link 3 with a bandwidth of 100Mbps and a latency of 50ms, and an available edge computing node C with a computing power of 1500MIPS. After the task scheduling table is sent to the execution layer through the control channel, the task feedback unit collects the real-time feedback data of the execution of each subtask in real time as follows: the completion latency of subtask 1 is 350ms, the packet loss rate is 2%, the bandwidth utilization rate is 85%, and the task completion rate is 95%; the completion latency of subtask 2 is 200ms, the packet loss rate is 3%, the bandwidth utilization rate is 90%, and the task completion rate is 90%; the completion latency of subtask 3 is 250ms, the packet loss rate is 1%, the bandwidth utilization rate is 80%, and the task completion rate is 100%.

[0139] Therefore, by collecting real-time feedback data of each subtask during task execution, a timely and accurate closed-loop data foundation is provided for subsequent execution status determination and feedback adjustment.

[0140] Step 407: When the real-time feedback data of any subtask does not meet the preset threshold requirements, feedback adjustment is triggered.

[0141] In this embodiment of the disclosure, the feedback adjustment includes at least one of the following: recalculating the interaction factor based on the real-time feedback data, adjusting the decomposition granularity, and re-dividing the subtasks; reallocating the subtasks whose real-time feedback data does not meet the preset threshold requirements to the backup standard link; and migrating the subtasks whose real-time feedback data does not meet the preset threshold requirements to the available edge computing nodes with higher availability levels.

[0142] In this embodiment, the preset threshold requirements are pre-set based on the performance constraint fields in the task request, such as a maximum latency of no more than 300ms, a packet loss rate of no more than 1%, and a task completion rate of no less than 98%. The system compares the real-time feedback data of each subtask collected in step 406 with the corresponding preset thresholds one by one: if all indicators meet the threshold requirements, the current scheduling scheme is deemed to be in a qualified execution state and no adjustment is required; if any indicator of any subtask does not meet the threshold requirements, the execution state of the subtask is deemed to have not met expectations, triggering feedback adjustment.

[0143] The specific methods for feedback and adjustment include at least one of the following: First, adjust the decomposition granularity. Recalculate the interaction factor based on real-time feedback data, and redetermine the decomposition granularity based on the updated interaction factor, then re-divide subtasks that fail to meet the targets. For example, when a subtask's completion latency exceeds the limit and its computational load is high, the current subtask is further decomposed into smaller, finer-grained tasks to reduce the computational load and bandwidth requirements of individual subtasks, thereby accelerating execution and reducing latency.

[0144] Secondly, subtasks are reallocated to backup standard links. Subtasks whose real-time feedback data does not meet preset threshold requirements are reallocated to backup standard links. For example, when the bandwidth utilization of a link where a subtask is located is close to saturation, while other standard links still have remaining bandwidth, the optimal standard link reallocation scheme is solved, some computing tasks are migrated to backup standard links, and the task resource mapping matrix is ​​modified to reflect the updated allocation relationship.

[0145] Third, migrate subtasks to available edge computing nodes with higher availability levels. Subtasks whose real-time feedback data does not meet preset threshold requirements are migrated to available edge computing nodes with higher availability levels, and nodes that show fluctuations in indicators during the feedback period are filtered out, and new tasks are temporarily suspended from being assigned to them until their status returns to stability.

[0146] For example, the real-time feedback data collected in step 406 is compared with a preset threshold, where the preset maximum latency is 300ms. The comparison reveals that the completion latency of subtask 1 is 350ms, exceeding the preset maximum latency of 300ms, and the packet loss rate is 2% and the bandwidth utilization rate is 85%, both of which fail to meet the ideal values. Therefore, it is determined that the execution status of subtask 1 has not met the expected requirements, triggering feedback adjustment. Based on the feedback adjustment strategy: First, the task decomposition granularity is adjusted. Since subtask 1 has a heavy computational load and high bandwidth requirements, it is further decomposed into more smaller tasks. Second, some subtasks are reallocated to backup link standards. Link 1's bandwidth utilization has reached 85%, while link 3 still has remaining bandwidth (100Mbps). Therefore, some computation of subtask 1 is allocated to link 3, while a less loaded available edge computing node C is selected for processing. Third, subtasks on high-load available edge computing nodes are migrated. Specifically, if the load on available edge computing node B is detected to be high, subtask 2 is migrated from available edge computing node B to available edge computing node C to balance the computational load. Available edge computing node B is temporarily suspended until its load stabilizes. Thus, by comparing real-time feedback data with preset thresholds and triggering various feedback adjustments, the scheduling scheme is dynamically corrected, ensuring that the subtask execution status can be restored to the required level in a timely manner.

[0147] Step 408: Repeat the feedback adjustment until the real-time feedback data of all subtasks meets the preset threshold requirements, and output the optimized task scheduling table.

[0148] In this embodiment of the disclosure, real-time feedback data can also be stored in a historical database; and the real-time feedback data accumulated in the historical database can be used to perform incremental training or parameter updates on the anomaly detection model, the time series prediction model for predicting link packet loss rate, and the semantic parsing model for identifying task categories.

[0149] In this embodiment, the feedback adjustment in step 407 is not performed all at once, but rather uses an iterative closed-loop mechanism: after each adjustment, the real-time feedback data of each subtask's execution is collected again through the task feedback unit and compared with the preset threshold. If any subtask still does not meet the requirements, the next round of feedback adjustment is triggered until the real-time feedback data of all subtasks meets the preset threshold requirements. At this point, the optimized task scheduling table is output. This optimized task scheduling table is redeployed to the execution layer through the control channel to replace the initial scheduling scheme and guide subsequent task execution.

[0150] Furthermore, during each feedback adjustment process, the real-time feedback data from each round is stored in a historical database, forming a time-series accumulation of real-time feedback data. Using the real-time feedback data accumulated in the historical database, incremental training or parameter updates can be performed on relevant models, including: incremental training of anomaly detection models (such as the Isolation Forest model) to improve anomaly detection accuracy; parameter updates for time-series prediction models used to predict packet loss rates in standard links (such as the ARIMA model) to improve packet loss rate prediction accuracy; and incremental training of semantic parsing models used to identify task categories (such as the BERT model) to improve task classification accuracy. Through continuous optimization of these models, the system's adaptive capabilities and the intelligence level of scheduling decisions are continuously enhanced.

[0151] For example, after the feedback adjustment in step 407, an optimized task scheduling scheme is generated: subtask 1 is assigned to standard link 3 and available edge computing node C, subtask 2 is adjusted to be executed on available edge computing node C, and subtask 3 remains unchanged. After the system sends the optimized scheduling scheme to the execution layer through the control channel, it continues to monitor the real-time feedback data of each subtask's execution. Ultimately, the completion latency of subtask 1 is shortened to less than 300ms, the packet loss rate is reduced to 1%, the bandwidth utilization rate is increased to 90%, the task completion rate is increased to 98%, and the real-time feedback data of all subtasks meets the preset threshold requirements, outputting the optimized task scheduling table. At the same time, the real-time feedback data generated during this round of feedback is recorded in the historical database for subsequent incremental training of the anomaly detection model, packet loss rate prediction model, and semantic parsing model. Thus, continuous dynamic optimization of the task scheduling scheme is achieved, effectively improving the stability and reliability of task execution. Furthermore, by combining the accumulation of real-time feedback data with incremental model training, the system can continuously learn from historical scheduling experience, constantly enhancing its adaptive capability for future task scheduling.

[0152] In the network task scheduling method of this disclosure embodiment, a network multimodal dataset is collected. The multimodal dataset includes multi-source link data of multi-standard links in the network, computing power index data of edge computing nodes, real-time network status, and topology snapshots representing network topology information. Based on the computing power index data, anomaly detection and availability level assessment are performed on the operating status of each edge computing node to generate available edge computing nodes and their corresponding availability levels. Based on the multi-source link data and topology snapshots, the confidence level of each standard link is assessed to generate the confidence level of each standard link. Based on the task type corresponding to the task request received by the network and the network status, the task in the task request is decomposed into multiple... The system identifies several subtasks with priority tags. Based on available edge computing nodes, their corresponding availability and confidence levels, and the multiple subtasks, a feasible task scheduling table is generated, and network task scheduling is performed according to this table. After generating the feasible task scheduling table and performing network task scheduling, real-time feedback data is collected for each subtask during execution. This real-time feedback data includes latency, packet loss rate, bandwidth utilization, and task completion rate. When the real-time feedback data of any subtask fails to meet a preset threshold, feedback adjustment is triggered. This feedback adjustment is repeated until the real-time feedback data of all subtasks meets the preset threshold, and an optimized task scheduling table is output. Thus, through the collection of real-time feedback data and a closed-loop adjustment mechanism, dynamic optimization of the task scheduling scheme is achieved, effectively improving the stability of task execution and the system's adaptability.

[0153] Figure 5 This is a schematic diagram of the structure of a network task scheduling device according to an embodiment of the present disclosure.

[0154] like Figure 5 As shown, the network task scheduling device includes: a data acquisition module 501, a detection module 502, an evaluation module 503, a decomposition module 504, and a scheduling module 505.

[0155] The system comprises the following modules: a first acquisition module 501, used to acquire a network multimodal dataset, which includes multi-source link data of multi-mode links in the network, computing power index data of edge computing nodes, real-time network status, and topology snapshots representing network topology information; a detection module 502, used to perform anomaly detection and availability level assessment on the operating status of each edge computing node based on the computing power index data, and generate available edge computing nodes and their corresponding availability levels; an evaluation module 503, used to evaluate the confidence level of each mode link based on the multi-source link data and topology snapshots, and generate confidence level for each mode link; a decomposition module 504, used to decompose the task in the task request into multiple sub-tasks with priority tags based on the task type corresponding to the task request received by the network and the network status; and a scheduling module 505, used to generate a feasible task scheduling table based on the available edge computing nodes and their corresponding availability levels, the confidence levels, and the multiple sub-tasks, and to perform network task scheduling according to the task scheduling table.

[0156] In one embodiment of this disclosure, the first acquisition module 501 is specifically configured to: acquire radio frequency measurement parameters of the multi-standard links through radio frequency measurement, and obtain latency, bandwidth, and jitter parameters of each standard link through link detection to form the multi-source link data; acquire processor utilization, memory occupancy, task queue length, and cache status of each edge computing node through edge computing node resource monitoring to form the computing power index data; perform timestamp alignment on the multi-source link data and the computing power index data to eliminate asynchronous deviations, so as to obtain aligned multi-source link data; obtain the available bandwidth and latency of each standard link based on the aligned multi-source link data; obtain the available computing power and load status of each edge computing node based on the aligned computing power index data, and generate real-time network status; construct a topology snapshot representing network topology information using a network topology adjacency matrix; and integrate the aligned multi-source link data, computing power index data, real-time network status, and topology snapshot to generate a multimodal dataset.

[0157] In one embodiment of this disclosure, the detection module 502 is specifically configured to: construct a state time series of each edge computing node based on the computing power index data of each edge computing node; perform anomaly detection on the state time series of each edge computing node using an anomaly detection model, and calculate the anomaly score of each edge computing node; filter out available edge computing nodes from each edge computing node based on the anomaly score; calculate the fluctuation index of the state index of each available edge computing node, the fluctuation index being used to characterize the stability of the operating state of the edge computing node; and determine the availability level of each available edge computing node based on the anomaly score and the fluctuation index.

[0158] In one embodiment of this disclosure, the detection module 502 is further configured to: isolate and divide each state index in the state time series of each edge computing node by constructing a random segmentation tree to obtain the isolated path length of each state index in the random segmentation tree; and calculate the anomaly score of each edge computing node based on the isolated path length.

[0159] In one embodiment of this disclosure, the evaluation module 503 is specifically configured to: construct a weighted adjacency matrix by weighting the connections between edge computing nodes in the network using the bandwidth of each standard link as a positive weighting factor and the latency of each standard link as a negative weighting factor, based on the multi-source link data; align the topology snapshot with the weighted adjacency matrix in a time sequence, and perform time slicing on the aligned weighted adjacency matrix to obtain a topology snapshot sequence that changes over time; determine the topology influence level of each standard link based on the connection relationship and edge attributes of each standard link in the topology snapshot sequence; perform regression fitting on the packet loss rate of each standard link in the next time window using a time-series prediction model based on the historical packet loss rate data and real-time packet loss rate data of each standard link to obtain a predicted packet loss rate value; and generate a confidence level for each standard link based on the topology influence level and the predicted packet loss rate value.

[0160] In one embodiment of this disclosure, the decomposition module 504 is specifically configured to: classify the descriptive text of the task request using a semantic parsing model to identify the task type; match a corresponding basic decomposition template according to the task type, wherein the basic decomposition template defines the subtask composition of each task type under standard conditions and the dependencies between each subtask; obtain the computational load and bandwidth requirements of the task at the current moment according to the network state, and calculate the interaction factor between the computational load and the bandwidth requirements; wherein the interaction factor is determined through the nonlinear coupling relationship between the computational load and the bandwidth requirements; determine the decomposition granularity according to the magnitude of the interaction factor; based on the subtask composition and dependencies defined by the basic decomposition template, merge or split each subtask in the basic decomposition template according to the determined decomposition granularity to generate multiple subtasks adapted to the current network state, and mark the execution priority of each subtask.

[0161] In one embodiment of this disclosure, the apparatus further includes: a first calculation module, configured to calculate the task scheduling delay caused by the interaction between computing load and bandwidth demand based on the interaction factor; a second calculation module, configured to calculate the execution failure probability of the task at the current moment based on the interaction factor and the task scheduling delay; and an adjustment module, configured to adjust the decomposition granularity when the execution failure probability exceeds a preset probability threshold.

[0162] In one embodiment of this disclosure, the scheduling module 505 is specifically configured to: based on the latency constraints, bandwidth requirements, and computational load of each subtask, select a set of links and a set of nodes from the network's standard links and available edge computing nodes that meet the requirements of each subtask; using the availability level and the confidence level as constraints, match each subtask with the standard links and available edge computing nodes in the link set and node set, with the optimization objective of minimizing the combined cost of latency, bandwidth utilization, and computational load of each subtask, and solve for a resource allocation scheme for each subtask; verify the feasibility of the resource allocation scheme, and use the verified resource allocation scheme as a feasible task scheduling table, and perform network task scheduling according to the task scheduling table.

[0163] In one embodiment of this disclosure, the scheduling module 505 is further configured to: for critical subtasks with a priority higher than a preset threshold, adopt a multi-path redundancy allocation strategy to simultaneously allocate the critical subtasks to multiple standard links with confidence levels higher than the preset threshold for parallel transmission; for ordinary subtasks with a priority lower than the preset threshold, allocate them sequentially according to the remaining resources in the link set and node set to obtain a resource allocation scheme for each subtask.

[0164] In one embodiment of this disclosure, the apparatus further includes: a second acquisition module, configured to acquire real-time feedback data of each subtask during execution after generating a feasible task scheduling table and performing network task scheduling according to the task scheduling table, the real-time feedback data including network task scheduling completion latency, packet loss rate, bandwidth utilization rate, and task completion rate; a feedback module, configured to trigger feedback adjustment when the real-time feedback data of any subtask does not meet a preset threshold requirement; and an output module, configured to repeatedly execute the feedback adjustment until the real-time feedback data of all subtasks meets the preset threshold requirement, and output an optimized task scheduling table.

[0165] In one embodiment of this disclosure, the feedback adjustment includes at least one of the following: recalculating the interaction factor based on real-time feedback data, adjusting the decomposition granularity, and re-dividing subtasks; reallocating subtasks whose real-time feedback data does not meet the preset threshold requirements to backup standard links; and migrating subtasks whose real-time feedback data does not meet the preset threshold requirements to available edge computing nodes with higher availability levels.

[0166] In one embodiment of this disclosure, the apparatus further includes: a storage module for storing the real-time feedback data into a historical database; and an update module for incrementally training or updating the parameters of the anomaly detection model, the time series prediction model for predicting link packet loss rate, and the semantic parsing model for identifying task categories using the real-time feedback data accumulated in the historical database.

[0167] In the network task scheduling device of this embodiment, a multimodal network dataset is collected, including multi-source link data, computing power index data, network status, and topology snapshots representing network topology information. Based on the multimodal network dataset, available edge computing nodes and their corresponding availability levels, as well as the confidence levels of links of various standards in the network, are detected. According to the task type corresponding to the task request received by the network and the network status, the task in the task request is decomposed into multiple sub-tasks with priority labels. Based on the available edge computing nodes and their availability levels, confidence levels, and multiple sub-tasks, a feasible task scheduling table is generated and network task scheduling is performed. Thus, an optimal resource allocation scheme for tasks is provided in a highly dynamic, multimodal network environment, improving the overall network throughput and the stability of network task execution.

[0168] According to a third aspect of the present disclosure, an electronic device is also provided, comprising: a processor; and a memory for storing processor-executable instructions, wherein the processor is configured to implement the network task scheduling method as described above.

[0169] To implement the above embodiments, this disclosure also proposes a storage medium.

[0170] When the instructions in the storage medium are executed by the processor, the processor is able to execute the network task scheduling method described above.

[0171] To implement the above embodiments, this disclosure also provides a computer program product.

[0172] When the computer program product is executed by the processor of the electronic device, it enables the electronic device to perform the above-described method.

[0173] Figure 6 This is a structural block diagram of an electronic device according to an exemplary embodiment. Figure 6 The electronic device shown is merely an example and should not be construed as limiting the functionality and scope of the embodiments disclosed herein.

[0174] like Figure 6 As shown, the electronic device 600 includes a processor 611, which can perform various appropriate actions and processes according to a program stored in read-only memory (ROM) 612 or a program loaded from memory 616 into random access memory (RAM) 613. The RAM 613 also stores various programs and data required for the operation of the electronic device 600. The processor 611, ROM 612, and RAM 613 are interconnected via a bus 614. An input / output (I / O) interface 615 is also connected to the bus 614.

[0175] The following components are connected to I / O interface 615: memory 616 including hard disk; and communication section 617 including network interface card such as local area network (LAN) card, modem, etc., communication section 617 performs communication processing via a network such as the Internet; drive 618 is also connected to I / O interface 615 as needed.

[0176] In particular, according to embodiments of this disclosure, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments of this disclosure include a computer program carried on a computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via communication section 617. When the computer program is executed by processor 611, it performs the functions defined in the methods of this disclosure.

[0177] In an exemplary embodiment, a storage medium including instructions is also provided, such as a memory including instructions, which can be executed by the processor 611 of the electronic device 600 to perform the above-described method. Optionally, the storage medium may be a non-transitory computer-readable storage medium, such as a ROM, random access memory (RAM), CD-ROM, magnetic tape, floppy disk, and optical data storage device.

[0178] In this disclosure, a computer-readable storage medium can be any tangible medium containing or storing a program that can be used by or in connection with an instruction execution system, apparatus, or device. In this disclosure, a computer-readable signal medium can include a data signal propagated in baseband or as part of a carrier wave, carrying computer-readable program code. Such propagated data signals can take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. A computer-readable signal medium can also be any computer-readable medium other than a computer-readable storage medium, which can transmit, propagate, or transfer a program for use by or in connection with an instruction execution system, apparatus, or device. The program code contained on the computer-readable medium can be transmitted using any suitable medium, including but not limited to: wireless, wireline, optical fiber, RF, etc., or any suitable combination thereof.

[0179] Furthermore, the term “exemplary” is used herein to mean serving as an example, instance, or illustration. Any aspect or design described herein as “exemplary” is not necessarily to be construed as advantageous compared to other aspects or designs. Rather, the use of the term “exemplary” is intended to present the concept in a concrete manner. As used herein, the term “or” is intended to mean an inclusive “or” rather than an exclusive “or.” That is, unless otherwise specified or clear from the context, “X applies A or B” is intended to mean any of the natural inclusive arrangements. That is, “X applies A or B” satisfies any of the foregoing instances if X applies A; X applies B; or both X applies A and B. Additionally, unless otherwise specified or clear from the context to refer to the singular form, the articles “a” and “an” as used in this application and the appended claims are generally understood to mean “one or more.”

[0180] Similarly, although this disclosure has been shown and described with respect to one or more implementations, equivalent variations and modifications will occur to those skilled in the art upon reading and understanding the specification and drawings. This disclosure includes all such modifications and variations and is limited only by the scope of the claims. In particular, with respect to the various functions performed by the components described above (e.g., elements, resources, etc.), unless otherwise indicated, the terminology used to describe such components is intended to correspond to any component (functionally equivalent) that performs the specific function of the described component, even if structurally not equivalent to the disclosed structure. Furthermore, although specific features of this disclosure may have been disclosed with respect to only one of several implementations, such features may be combined with one or more other features of other implementations, as may be desired and advantageous for any given or particular application. Moreover, with regard to the terms “comprising,” “owning,” “having,” “having,” or variations thereof as used in the detailed description or claims, such terms are intended to be inclusive in a manner similar to the term “including.”

[0181] Other embodiments of this disclosure will readily occur to those skilled in the art upon consideration of the specification and practice of the invention disclosed herein. This disclosure is intended to cover any variations, uses, or adaptations of this disclosure that follow the general principles of this disclosure and include common knowledge or customary techniques in the art not disclosed herein. The specification and examples are to be considered exemplary only, and the true scope and spirit of this disclosure are indicated by the following claims.

[0182] It should be understood that this disclosure is not limited to the precise structures described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The scope of this disclosure is limited only by the appended claims.

Claims

1. A network task scheduling method, characterized in that, The method includes: Collect a network multimodal dataset, which includes multi-source link data of multi-standard links in the network, computing power index data of edge computing nodes, real-time network status, and topology snapshots that characterize network topology information; Based on the computing power index data, anomaly detection and availability level assessment are performed on the operating status of each edge computing node to generate available edge computing nodes and their corresponding availability levels. Based on the multi-source link data and topology snapshots, the confidence level of each type of link is evaluated, and a confidence level of each type of link is generated. Based on the task type corresponding to the task request received from the network and the network status, the task in the task request is decomposed into multiple sub-tasks with priority tags. Based on the available edge computing nodes and their corresponding availability levels, the confidence levels, and the multiple subtasks, a feasible task scheduling table is generated, and network task scheduling is performed according to the task scheduling table.

2. The method according to claim 1, characterized in that, The acquired network multimodal dataset includes: The radio frequency measurement parameters of the multi-standard links are collected by radio frequency measurement, and the latency, bandwidth and jitter parameters of each standard link are obtained by link detection to form the multi-source link data; The processor utilization, memory usage, task queue length, and cache status of each edge computing node are collected through resource monitoring methods to form the computing power index data. The multi-source link data and the computing power index data are timestamped to eliminate asynchronous deviations and obtain aligned multi-source link data. Based on the aligned multi-source link data, the available bandwidth and latency of each standard link are obtained. Based on the aligned computing power index data, the available computing power and load status of each edge computing node are obtained, and real-time network status is generated. A topological snapshot representing network topology information is constructed using a network topology adjacency matrix. The aligned multi-source link data, computing power index data, real-time network status, and topology snapshot are integrated to generate a multimodal dataset.

3. The method according to claim 1, characterized in that, The step of performing anomaly detection and availability level assessment on the operating status of each edge computing node based on the computing power index data, and generating available edge computing nodes and their corresponding availability levels, includes: Based on the computing power index data of each edge computing node, construct the state time series of each edge computing node; An anomaly detection model is used to detect anomalies in the state time series of each edge computing node, and an anomaly score is calculated for each edge computing node. Based on the anomaly scores, usable edge computing nodes are selected from each edge computing node; Calculate the fluctuation index of the state indicators of each available edge computing node, wherein the fluctuation index is used to characterize the stability of the operating state of the edge computing node; The availability level of each available edge computing node is determined based on the anomaly score and the volatility index.

4. The method according to claim 3, characterized in that, The process of using an anomaly detection model to detect anomalies in the state time series of each edge computing node and calculating anomaly scores for each edge computing node includes: By constructing a random partitioning tree, each state index in the state time series of each edge computing node is isolated and partitioned to obtain the isolated path length of each state index in the random partitioning tree. Based on the isolated path length, calculate the anomaly score for each edge computing node.

5. The method according to claim 1, characterized in that, The step of evaluating the confidence level of each type of link based on the multi-source link data and topology snapshot, and generating a confidence level for each type of link, includes: Based on the multi-source link data, the bandwidth of each type of link is used as a positive weighting factor and the latency of each type of link is used as a negative weighting factor to perform weighted calculation on the type of link connections between edge computing nodes in the network, and a weighted adjacency matrix is ​​constructed. The topology snapshots are time-series aligned with the weighted adjacency matrix, and the aligned weighted adjacency matrix is ​​time-sliced ​​to obtain a sequence of topology snapshots that change over time. Based on the connection relationships and edge attributes of each type of link in the topology snapshot sequence, determine the topology influence level of each type of link; Based on the historical and real-time packet loss rate data of each type of link, the packet loss rate of each type of link in the next time window is regression-fitted using a time series prediction model to obtain the predicted packet loss rate. Based on the topology impact level and the predicted packet loss rate of each link, a confidence level is generated for each link.

6. The method according to claim 1, characterized in that, The step of decomposing the task in the task request into multiple sub-tasks with priority tags according to the task type corresponding to the task request received from the network and the network status includes: The semantic parsing model is used to classify the descriptive text of task requests and identify the task type. The corresponding basic decomposition template is matched according to the task type. The basic decomposition template defines the subtask composition of each task type under standard conditions and the dependency relationship between each subtask. The computational load and bandwidth requirements of the task at the current moment are obtained based on the network status, and the interaction factor between the computational load and the bandwidth requirements is calculated; wherein, the interaction factor is determined through the nonlinear coupling relationship between the computational load and the bandwidth requirements; The decomposition granularity is determined based on the magnitude of the interaction factor. Based on the subtask composition and dependencies defined in the basic decomposition template, each subtask in the basic decomposition template is merged or split according to the determined decomposition granularity to generate multiple subtasks adapted to the current network state, and each subtask is marked with an execution priority.

7. The method according to claim 6, characterized in that, The method further includes: The task scheduling delay caused by the interaction between computing load and bandwidth demand is calculated based on the interaction factor. Based on the interaction factor and the task scheduling delay, calculate the probability of task execution failure at the current moment; When the execution failure probability exceeds a preset probability threshold, the decomposition granularity is adjusted.

8. The method according to claim 1, characterized in that, The step of generating a feasible task scheduling table based on the available edge computing nodes and their corresponding availability levels, the confidence levels, and multiple subtasks, and then scheduling network tasks according to the task scheduling table, includes: Based on the latency constraints, bandwidth requirements, and computational load of each subtask, a set of links and a set of nodes that meet the requirements of each subtask are selected from the standard links and available edge computing nodes in the network. Using the availability level and the confidence level as constraints, each subtask is matched with the standard links and available edge computing nodes in the link set and node set. The optimization objective is to minimize the combined cost of latency, bandwidth utilization and computing load of each subtask, and the resource allocation scheme of each subtask is obtained. The resource allocation scheme is verified for feasibility. The verified resource allocation scheme is used as a feasible task scheduling table, and network task scheduling is performed according to the task scheduling table.

9. The method according to claim 8, characterized in that, The solution for obtaining the resource allocation scheme for each subtask also includes: For critical subtasks with a priority higher than a preset threshold, a multi-path redundancy allocation strategy is adopted to simultaneously allocate the critical subtasks to multiple standard links with a confidence level higher than the preset threshold for parallel transmission; for ordinary subtasks with a priority lower than the preset threshold, they are allocated sequentially according to the remaining resources in the link set and node set to obtain the resource allocation scheme for each subtask.

10. The method according to claim 1, characterized in that, The method further includes: After generating a feasible task scheduling table and performing network task scheduling according to the task scheduling table, real-time feedback data of each subtask during execution is collected. The real-time feedback data includes the latency, packet loss rate, bandwidth utilization rate and task completion rate of network task scheduling. When the real-time feedback data of any subtask does not meet the preset threshold requirements, feedback adjustment is triggered. Repeat the feedback adjustment until the real-time feedback data of all subtasks meets the preset threshold requirements, and output the optimized task scheduling table.

11. The method according to claim 10, characterized in that, The feedback adjustment includes at least one of the following: Based on real-time feedback data, the interaction factor is recalculated, the decomposition granularity is adjusted, and the subtasks are re-divided. Subtasks whose real-time feedback data does not meet the preset threshold requirements will be reassigned to backup standard links. Subtasks whose real-time feedback data does not meet the preset threshold requirements will be migrated to available edge computing nodes with higher availability levels.

12. The method according to claim 10, characterized in that, The method further includes: The real-time feedback data is stored in a historical database; Using the real-time feedback data accumulated in the historical database, incremental training or parameter updates are performed on the anomaly detection model, the time series prediction model for predicting link packet loss rate, and the semantic parsing model for identifying task categories.

13. A network task scheduling device, characterized in that, The device includes: The acquisition module is used to acquire network multimodal datasets, which include multi-source link data of multi-mode links in the network, computing power index data of edge computing nodes, real-time network status, and topology snapshots that characterize network topology information. The detection module is used to perform anomaly detection and availability level assessment on the operating status of each edge computing node based on the computing power index data, and generate available edge computing nodes and their corresponding availability levels. The evaluation module is used to evaluate the confidence level of each type of link based on the multi-source link data and topology snapshot, and generate the confidence level of each type of link. The decomposition module is used to decompose the task in the task request into multiple sub-tasks with priority tags according to the task type corresponding to the task request received from the network and the network status. The scheduling module is used to generate a feasible task scheduling table based on the available edge computing nodes and their corresponding availability levels, the confidence levels, and multiple sub-tasks, and to perform network task scheduling according to the task scheduling table.

14. An electronic device, characterized in that, include: processor; Memory used to store the processor's executable instructions; The processor is configured as follows: The steps of implementing the network task scheduling method as described in any one of claims 1 to 12.

15. A non-transitory computer-readable storage medium, characterized in that, When the instructions in the storage medium are executed by the processor, the processor is able to perform the network task scheduling method as described in any one of claims 1 to 12.

Images