An edge computing management system that supports dynamic migration of AI models

Abstract

The present invention relates to the field of dynamic migration technology, and in particular to an edge computing management system that supports dynamic migration of AI models. The system includes the state mining module, which is used to identify available edge computing devices; perform multi-frequency information monitoring on available edge computing devices, and perform equipment comprehensive state mining to generate comprehensive state characteristics of edge devices; the edge topology module is used to locate the position of each node based on the comprehensive state characteristics of edge devices, and perform real-time state mapping modeling to build an edge computing node topology model; the task traffic trend module is used to obtain the AI ​​model task operation log; perform task queue analysis on the AI ​​model task operation log, and then perform model task traffic trend analysis to obtain AI model task traffic trend data. The present invention realizes edge computing management with fast response and high processing efficiency.

Description

Technical Field

[0001] The present invention relates to the field of dynamic migration technology, and in particular to an edge computing management system that supports dynamic migration of AI models. Background Art

[0002] With the rapid development of Internet of Things (IoT) and artificial intelligence (AI) technologies, edge computing, as an emerging computing architecture, has been widely used in many fields. Edge computing pushes data processing to the edge of the network, where the data is generated, thereby reducing data transmission latency and improving system response speed and processing efficiency. However, with the increasing number of devices and the increasing complexity of application scenarios, how to effectively manage and optimize AI models distributed on various edge nodes has become an important challenge facing current edge computing systems.

[0003] Traditional edge computing management systems rely on static model deployment and fixed computing resource configuration, and are often unable to cope with changing loads and demands. For example, in applications such as intelligent transportation, industrial automation, and security monitoring, the reasoning tasks of AI models need to be dynamically adjusted and optimized according to changes in real-time data. Static deployment often fails to fully utilize the computing resources of edge nodes, nor can it adjust AI models in a timely manner according to changes in different application scenarios, resulting in waste of resources and low processing efficiency. Therefore, in the case of limited resources, how to prioritize the most critical computing tasks and ensure the efficiency and real-time performance of model migration is also a technical challenge currently faced by edge computing management. Summary of the invention

[0004] In order to solve the above-mentioned technical problems, the present invention proposes an edge computing management system that supports dynamic migration of AI models to solve at least one of the above-mentioned technical problems.

[0005] To achieve the above objectives, the present invention provides an edge computing management system that supports dynamic migration of AI models, and the edge computing management system that supports dynamic migration of AI models includes a state mining module, an edge topology module, a task traffic trend module, a rolling prediction module, an optimal migration time point module, and a migration decision module:

[0006] The state mining module is used to identify available edge computing devices; perform multi-frequency information monitoring on available edge computing devices, and perform device comprehensive state mining to generate comprehensive state features of edge devices;

[0007] The edge topology module is used to locate the position of each node based on the comprehensive state characteristics of the edge device, and perform real-time state mapping modeling, so as to build an edge computing node topology model;

[0008] The task traffic trend module is used to obtain the AI ​​model task operation log; perform task queue analysis on the AI ​​model task operation log, and then perform model task traffic trend analysis to obtain AI model task traffic trend data;

[0009] The rolling prediction module is used to perform multi-time point rolling prediction on the AI ​​model task traffic trend data and perform model computing power load demand analysis to generate multi-time point computing power load demand data;

[0010] The optimal migration time point module is used to perform edge computing simulation and optimal migration time point analysis on the edge computing node topology model based on the multi-time point computing load demand data, so as to obtain the optimal migration time point data;

[0011] The migration decision module is used to make distributed edge migration decisions based on the optimal migration time point data, perform dynamic cloud-based collaborative management, and build an edge migration collaborative management model.

[0012] The present invention identifies available edge computing devices and establishes a list of edge computing resources to provide basic data support for subsequent task scheduling and resource management. It monitors the multi-frequency information of edge computing devices to fully understand the operating status of the devices, and provides multi-dimensional data support for device status evaluation. The generation of comprehensive status characteristics of edge devices is helpful for comprehensive analysis of device status, providing a basis for subsequent task scheduling and resource planning. Through node location positioning and status mapping modeling, the location and status of edge computing nodes are accurately understood, and detailed information is provided for building an edge computing topology model. Building an edge computing node topology model is helpful for realizing the spatial layout and correlation analysis of edge computing resources, providing support for task scheduling and migration decisions, obtaining AI model task operation logs to record task execution status, providing log data support for task performance evaluation and troubleshooting, and performing task queue analysis is helpful for understanding the task execution order and scheduling status, and providing support for subsequent It provides a data basis for task traffic trend analysis, predicts the changing trend of computing demand for AI model tasks through multi-time point rolling prediction and computing load demand analysis, and provides a prediction basis for resource planning and performance optimization. Generating multi-time point computing load demand data helps to realize dynamic allocation and optimal configuration of computing resources, improve the efficiency and performance of edge computing systems, and evaluate the effects of different migration strategies through edge computing simulation and optimal migration time point analysis to provide data support for migration decisions. Obtaining optimal migration time point data helps to achieve balanced allocation and efficient migration of tasks in edge computing environments, improve the overall performance of the system, and make distributed edge migration decisions based on optimal migration time point data to realize intelligent scheduling of tasks and resource optimization, improve the overall efficiency of the system, and build an edge migration collaborative management model to help realize collaborative management of edge computing and cloud, optimize resource utilization and task response speed, and improve the flexibility and performance of the system.

[0013] Preferably, the state mining module is used to identify available edge computing devices; perform multi-frequency information monitoring on available edge computing devices, and perform device comprehensive state mining to generate edge device comprehensive state features, specifically for:

[0014] Perform a global scan of the edge computing network available to the AI ​​model and extract available edge computing devices;

[0015] Perform multi-frequency information monitoring on available edge computing devices and extract edge computing device status monitoring parameters;

[0016] Calculate the GPU utilization of the edge computing device status monitoring parameters to obtain the device GPU utilization;

[0017] Perform device load analysis based on device GPU utilization to generate edge device load data;

[0018] Quantify the edge computing power of available edge computing devices to obtain a comprehensive computing power quantification value;

[0019] The comprehensive status of the equipment is mined on the real-time computing power quantization value and edge device load data to generate the comprehensive status characteristics of the edge device.

[0020] The present invention extracts available edge computing devices through global scanning, establishes a device list, provides detailed information of available devices for subsequent task scheduling and resource allocation, monitors multi-frequency information of edge computing devices and extracts monitoring parameters to fully understand device performance and status, and provides data support for subsequent analysis and optimization. Calculating the GPU utilization of the device helps to understand the usage of GPU resources, providing a basis for task scheduling and performance optimization. Generating edge device load data through device load analysis helps to balance device load and improve system stability and efficiency. Quantifying the computing power of edge devices helps to quantify the computing power of the device, providing a reference for resource management and task allocation. Obtaining a comprehensive computing power quantification value comprehensively considers various computing power indicators of the device, and provides a basis for the overall performance evaluation of the device. Combining real-time computing power quantification values ​​and device load data to conduct comprehensive device status mining to comprehensively evaluate the status and performance of the device, provide comprehensive data support for decision-making, and generate comprehensive edge device status characteristics to help understand the performance of the device under different load conditions, and provide a basis for dynamic migration decisions and resource optimization.

[0021] Preferably, the specific steps of quantifying the edge computing power of the available edge computing devices to obtain the comprehensive computing power quantization value are:

[0022] Extracting hardware configuration information of the available edge computing device;

[0023] Performing performance parameter identification on the hardware configuration information and extracting hardware performance parameters, wherein the hardware performance parameters include the number of CPU cores, CPU main frequency, the number of GPU cores, clock frequency and architecture floating-point performance parameters;

[0024] Calculate the CPU floating point performance based on the number of CPU cores and CPU main frequency to generate the CPU floating point performance;

[0025] Quantify the theoretical floating-point operation peak value of the GPU core number, clock frequency and architecture floating-point performance parameters to obtain the theoretical floating-point operation peak value;

[0026] Comprehensive computing performance evaluation is performed on the CPU floating-point computing performance and theoretical floating-point computing peak value to obtain the comprehensive computing power quantification value.

[0027] The present invention extracts hardware configuration information to fully understand the hardware composition of edge computing devices, provide basic data for performance evaluation and resource allocation, identify and extract hardware performance parameters such as CPU core number, CPU main frequency, GPU core number, clock frequency, etc., which is helpful to quantify the computing power of the device and provide data support for subsequent performance evaluation. The floating-point operation performance of the CPU is calculated to help evaluate the performance of the CPU when processing computing-intensive tasks, and provide a basis for task scheduling and performance optimization. By quantifying the theoretical floating-point operation peak value of the GPU, the computing power of the device under ideal conditions is estimated, and a reference for the upper limit of the device performance is provided. A comprehensive computing performance evaluation is performed in combination with the CPU floating-point operation performance and the theoretical floating-point operation peak value of the GPU, and the comprehensive computing power of the device is obtained, providing decision support for device selection and task allocation.

[0028] Preferably, the edge topology module is used to locate the position of each node based on the comprehensive state characteristics of the edge device, and to perform real-time state mapping modeling, so as to construct an edge computing node topology model, specifically for:

[0029] Sequentially encode and mark the available edge computing devices to obtain multiple edge computing device encoding nodes;

[0030] Locate multiple edge computing device encoding nodes one by one to generate the location coordinates of each node;

[0031] Perform spatial distribution analysis on the location coordinates of each node to generate spatial distribution data of edge nodes;

[0032] Based on the comprehensive state characteristics of edge devices, the spatial distribution data of edge nodes is modeled in real time to construct an edge computing node topology model.

[0033] The present invention sequentially codes and marks edge computing devices, generates a unique coding node for each device, facilitates the identification and management of each device, and determines the specific position of each device in space by locating the position of each node one by one, providing an accurate spatial reference for subsequent spatial distribution analysis and topological model construction. Spatial distribution analysis reveals the distribution of edge nodes in physical space, helps to understand the correlation and layout structure between devices, and provides basic data for the establishment of node topological models. Instant state mapping modeling is performed based on the comprehensive state characteristics of edge devices, and the device state is mapped to the spatial distribution in real time, which helps to monitor and manage edge nodes in real time. An edge computing node topological model is constructed to display the connection relationship and state information between edge nodes, providing visual support for dynamic migration decisions and resource optimization.

[0034] Preferably, the specific steps of performing instant state mapping modeling on the spatial distribution data of edge nodes based on the comprehensive state characteristics of edge devices to construct the edge computing node topology model are:

[0035] Perform physical link interaction identification on multiple edge computing device encoding nodes and extract the interaction paths between nodes;

[0036] Mining the interactive connection logic of the interactive paths between nodes to obtain the connection logic between nodes;

[0037] Perform dynamic topological association evolution on the connection logic between nodes to generate dynamic topological association features;

[0038] Based on the dynamic interactive topological association characteristics, the spatial distribution data of edge nodes is fitted with topological distribution to build a topological distribution network of edge devices.

[0039] According to the comprehensive state characteristics of edge devices, the edge device topology distribution network is modeled in real time to build an edge computing node topology model.

[0040] The present invention identifies the connection relationship between edge computing devices by identifying physical link interactions, and provides basic data for establishing interaction paths between nodes. Mining the interactive connection logic between nodes helps to understand the communication mode and interaction method between devices, and provides a basis for subsequent topological association analysis. Generating dynamic topological association features reflects the evolution and changes of the connection relationship between edge computing devices, and provides data support for building a dynamic topological model. Topological distribution fitting of edge node spatial distribution data based on dynamic interactive topological association features helps to build an edge device topological distribution network, providing a basis for optimizing and managing communication paths between devices. According to the comprehensive state characteristics of the edge device, the edge device topological distribution network is modeled for real-time state mapping, and an edge computing node topology model that reflects the device connection status in real time is constructed to provide support for dynamic migration management and resource optimization.

[0041] Preferably, the task traffic trend module is used to obtain the AI ​​model task operation log; perform task queue analysis on the AI ​​model task operation log, and then perform model task traffic trend analysis to obtain AI model task traffic trend data, which is specifically used for:

[0042] Get AI model task running log;

[0043] Perform task queue analysis on AI model task running logs to identify real-time processing task data and unprocessed task data;

[0044] Perform task processing efficiency analysis on real-time processing task data and generate a model for real-time task processing efficiency;

[0045] Calculate the task processing throughput of the real-time processing task data to generate the model real-time task throughput;

[0046] Perform unprocessed task time series change analysis on unprocessed task data to generate unprocessed task change data;

[0047] Based on the model's real-time task processing efficiency and real-time task throughput, the model task traffic trend analysis is performed on the unprocessed task change data to obtain the AI ​​model task traffic trend data.

[0048] The present invention obtains the AI ​​model task running log, records the running status, data processing and abnormal situation of the model task, and provides basic data for subsequent analysis. The running log record is helpful to monitor the running status of the model task in real time, find problems in time and make adjustments and optimizations. Through task queue parsing, real-time processing task data and unprocessed task data are accurately identified to realize dynamic management and analysis of the task queue. Distinguishing between real-time processing tasks and unprocessed task data is helpful to optimize task scheduling and resource allocation, improve task processing efficiency and response speed, and evaluate the processing speed and efficiency of model tasks through efficiency analysis of real-time processing task data, identify potential bottlenecks and improvement space, generate model real-time task processing efficiency, help monitor task execution in real time, adjust resource allocation and optimization algorithm in time, improve task processing efficiency, calculate the throughput of real-time processing tasks to quantify the processing capacity of tasks, help evaluate the performance and resource utilization of the system, generate model real-time task throughput, help adjust task processing strategies and optimize task scheduling, improve the overall efficiency and performance of the system, analyze the changes in unprocessed task data to find the change trend and characteristics of the task queue, and provide a reference basis for subsequent resource planning and task scheduling. Generating unprocessed task change data is helpful to predict the load of the system, adjust resource allocation and optimize task allocation in time.

[0049] Preferably, the rolling prediction module is used to perform multi-time point rolling prediction on the AI ​​model task traffic trend data and perform model computing power load demand analysis, thereby generating multi-time point computing power load demand data, specifically for:

[0050] Define multiple prediction time points, and perform multi-time point rolling predictions on the AI ​​model task traffic trend data based on the multiple prediction time points to generate task traffic trend prediction data at each time point;

[0051] Serialize the task traffic trend forecast data at each time point and build a task trend forecast sequence;

[0052] Perform model computing power load demand analysis on the task trend prediction sequence to generate computing power load demand data at multiple time points.

[0053] The present invention defines multiple prediction time points to predict the task flow trend in different time periods, thereby providing a more comprehensive reference for system resource planning and task scheduling. Multi-time point rolling prediction helps to capture the changing trend of task flow and helps the system better adapt to different workload conditions. The task flow trend prediction data at each time point is serialized to construct a task trend prediction sequence to form a continuous prediction trend, providing a basis for computing power load demand analysis. Task trend prediction serialization helps to understand the laws and trends of task flow changes and provide a basis for subsequent decision-making and planning. By performing computing power load demand analysis on the task trend prediction sequence, the computing resources required by the system at different time points are determined to achieve dynamic adjustment and optimization of resources. Generating computing power load demand data at multiple time points helps to reasonably plan the computing power configuration of the system and improve the performance and efficiency of the system.

[0054] Preferably, the optimal migration time point module is used to perform edge computing simulation and optimal migration time point analysis on the edge computing node topology model based on multi-time point computing load demand data, so as to obtain optimal migration time point data, specifically for:

[0055] Perform edge computing simulation on the edge computing node topology model based on the computing power load demand data at multiple time points to generate edge computing simulation response data;

[0056] The edge computing simulation response data is subjected to node-by-node response delay calculation, and the node calculation response delay of each node is extracted;

[0057] Perform real-time computing performance evaluation on the node computing response delay of each node, using the real-time computing performance value of each node;

[0058] The optimal migration time point analysis is performed based on the real-time computing power performance value of each node to obtain the optimal migration time point data.

[0059] The present invention performs edge computing simulation based on multi-time point computing load demand data, simulates the working conditions of edge computing nodes under different conditions, and generates simulated response data. The beneficial effect of edge computing simulation is that it can evaluate the load conditions of edge nodes and provide a basis for subsequent decision-making. By calculating the computing response delay of each node, the computing performance and response speed of different nodes are evaluated, and performance bottlenecks and optimization space are found. Extracting the computing response delay of the node is helpful to understand the working efficiency of each node, and provides a reference for node scheduling and task allocation. The computing response delay of each node is evaluated in real time. The running status of the node is monitored in time, and performance anomalies or changes are identified. The real-time computing power performance evaluation is helpful to adjust the node resource configuration and optimize the computing task allocation, improve the overall performance and stability of the system, and perform the optimal migration time point analysis based on the real-time computing power performance value of each node to determine the best migration time, realize efficient task migration and resource optimization, and the optimal migration time point data helps to reduce the waste of computing resources, improve the task processing efficiency, and ensure the stability and reliability of the system.

[0060] Preferably, the migration decision module is used to make distributed edge migration decisions based on the optimal migration time point data, and to perform dynamic cloud collaborative management to build an edge migration collaborative management model, specifically for:

[0061] Identify current unprocessed tasks based on the AI ​​model task running log and extract all current unprocessed tasks;

[0062] Performing data lossless encapsulation processing on the current unprocessed task to obtain multiple lossless encapsulation task packages;

[0063] Make distributed edge migration decisions for multiple lossless encapsulation task packages based on the optimal migration time point data and build a distributed edge migration strategy;

[0064] According to the distributed edge migration strategy, the edge computing node topology model is dynamically managed in the cloud and an edge migration collaborative management model is constructed.

[0065] The present invention analyzes the AI ​​model task operation log, identifies and extracts the current unprocessed tasks, promptly discovers the unfinished tasks in the system, and prepares for subsequent processing. Extracting the current unprocessed tasks helps to understand the system load and task execution status, and makes reasonable arrangements for task scheduling and migration. The data of the current unprocessed tasks is losslessly encapsulated to ensure that the task data will not be lost or damaged during the migration process, and the integrity and reliability of the task data are guaranteed. The generation of multiple losslessly encapsulated task packages helps to structure the task data and prepare for migration, simplify the migration process and reduce data risks. Based on the optimal migration time point data, distributed edge migration decisions are made for the losslessly encapsulated task packages. Appropriate migration strategies are formulated according to system status and requirements. Constructing a distributed edge migration strategy helps to optimize the task migration process, improve system efficiency and performance, and ensure that the task is smoothly migrated to the target node. According to the distributed edge migration strategy, the edge computing node topology model is dynamically cloud-coordinated to achieve coordination and optimization of edge migration. Constructing an edge migration collaborative management model helps to improve the elasticity and flexibility of the system, effectively allocate computing resources and improve task processing efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0066] Figure 1 This is a structural schematic diagram of an edge computing management system that supports dynamic migration of AI models in the present invention. DETAILED DESCRIPTION

[0067] It should be understood that the specific embodiments described herein are only used to explain the present invention, and are not used to limit the present invention.

[0068] This application example provides an edge computing management system that supports dynamic migration of AI models. The execution subject of the edge computing management system that supports dynamic migration of AI models includes but is not limited to: mechanical equipment, data processing platform, cloud server node, network upload device, etc. that are equipped with the system can be regarded as the general computing node of this application, and the data processing platform includes but is not limited to: audio image management system, information management system, cloud data management system at least one.

[0069] See also Figure 1 The present invention provides an edge computing management system that supports dynamic migration of AI models. The edge computing management system that supports dynamic migration of AI models includes a state mining module, an edge topology module, a task traffic trend module, a rolling prediction module, an optimal migration time point module, and a migration decision module:

[0070] The state mining module is used to identify available edge computing devices; perform multi-frequency information monitoring on available edge computing devices, and perform device comprehensive state mining to generate comprehensive state features of edge devices;

[0071] The edge topology module is used to locate the position of each node based on the comprehensive state characteristics of the edge device, and perform real-time state mapping modeling, so as to build an edge computing node topology model;

[0072] The task traffic trend module is used to obtain the AI ​​model task operation log; perform task queue analysis on the AI ​​model task operation log, and then perform model task traffic trend analysis to obtain AI model task traffic trend data;

[0073] The rolling prediction module is used to perform multi-time point rolling prediction on the AI ​​model task traffic trend data and perform model computing power load demand analysis to generate multi-time point computing power load demand data;

[0074] The optimal migration time point module is used to perform edge computing simulation and optimal migration time point analysis on the edge computing node topology model based on the multi-time point computing load demand data, so as to obtain the optimal migration time point data;

[0075] The migration decision module is used to make distributed edge migration decisions based on the optimal migration time point data, perform dynamic cloud-based collaborative management, and build an edge migration collaborative management model.

[0076] The present invention identifies available edge computing devices and establishes a list of edge computing resources to provide basic data support for subsequent task scheduling and resource management. It monitors the multi-frequency information of edge computing devices to fully understand the operating status of the devices, and provides multi-dimensional data support for device status evaluation. The generation of comprehensive status characteristics of edge devices is helpful for comprehensive analysis of device status, providing a basis for subsequent task scheduling and resource planning. Through node location positioning and status mapping modeling, the location and status of edge computing nodes are accurately understood, and detailed information is provided for building an edge computing topology model. Building an edge computing node topology model is helpful for realizing the spatial layout and correlation analysis of edge computing resources, providing support for task scheduling and migration decisions, obtaining AI model task operation logs to record task execution status, providing log data support for task performance evaluation and troubleshooting, and performing task queue analysis is helpful for understanding the task execution order and scheduling status, and providing support for subsequent It provides a data basis for task traffic trend analysis, predicts the changing trend of computing demand for AI model tasks through multi-time point rolling prediction and computing load demand analysis, and provides a prediction basis for resource planning and performance optimization. Generating multi-time point computing load demand data helps to realize dynamic allocation and optimal configuration of computing resources, improve the efficiency and performance of edge computing systems, and evaluate the effects of different migration strategies through edge computing simulation and optimal migration time point analysis to provide data support for migration decisions. Obtaining optimal migration time point data helps to achieve balanced allocation and efficient migration of tasks in edge computing environments, improve the overall performance of the system, and make distributed edge migration decisions based on optimal migration time point data to realize intelligent scheduling of tasks and resource optimization, improve the overall efficiency of the system, and build an edge migration collaborative management model to help realize collaborative management of edge computing and cloud, optimize resource utilization and task response speed, and improve the flexibility and performance of the system.

[0077] In the embodiment of the present invention, refer to Figure 1 , is a schematic diagram of the structure of an edge computing management system supporting dynamic migration of AI models in the present invention. In this example, the edge computing management system supporting dynamic migration of AI models includes:

[0078] The state mining module is used to identify available edge computing devices; perform multi-frequency information monitoring on available edge computing devices, and perform device comprehensive state mining to generate comprehensive state features of edge devices;

[0079] In this embodiment, a network scanning tool (such as Nmap) or a service discovery protocol (such as mDNS, SSDP) is used to identify available edge computing devices in the network. Ensure that basic information of the device is collected, including device ID, IP address, MAC address, device type, etc. The identified device information is stored in a centralized database to form a directory of available devices. Use structured data storage (such as SQL database, NoSQL database) to facilitate subsequent query and analysis. Perform a preliminary status detection on each identified device to obtain its online status, basic performance indicators (such as CPU, memory, storage), etc. as a baseline for subsequent monitoring. Select appropriate monitoring tools and technologies (such as Zabbix, Prometheus, Nagios) to regularly monitor the device status. Determine the frequency of monitoring, such as status collection every minute, every hour, or at regular intervals. Configure monitoring parameters in the monitoring tool to ensure that the following information is collected: Performance indicators: CPU usage, memory usage, disk I / O, network bandwidth, etc. Health status: device online status, runtime exceptions, error logs, etc. Obtain this information regularly through SNMP, RESTAPI, or custom scripts. Store the collected monitoring data in a data warehouse or time series database (such as InfluxDB, TimescaleDB) for subsequent analysis. Ensure that the data structure is designed reasonably for fast query and analysis. Clean and preprocess the stored monitoring data to remove outliers, fill in missing data, and standardize. Use data processing tools (such as Pandas) for data processing and conversion. Based on the preprocessed data, extract comprehensive status features, such as: Average load: calculate the average CPU and memory usage over a period of time. Response time: analyze network latency and request response time. Health index: generate a health score based on the device's error log and status information. Use machine learning algorithms (such as decision trees and random forests) to identify and build important features. Aggregate and analyze the extracted comprehensive status features to generate a comprehensive status report for each device. Use visualization tools (such as Matplotlib and Seaborn) to show the changing trend of device status features. Integrate the status features of all devices into a comprehensive feature set to form a unified status description. Ensure that the feature set includes the identification information of each device and the corresponding status feature value. Store the comprehensive status feature data in the database for subsequent query and analysis. Design a reasonable data model to ensure fast retrieval and update of device status. Set up a regular update mechanism to regularly recalculate and update the comprehensive status characteristics of the device to reflect the real-time status changes of the device. Use scheduled tasks (such as cron jobs) or event-driven methods to achieve regular updates.

[0080] The edge topology module is used to locate the position of each node based on the comprehensive state characteristics of the edge device, and perform real-time state mapping modeling, so as to build an edge computing node topology model;

[0081] In this embodiment, the comprehensive state characteristics of each edge computing device are extracted from the previous state mining module, including device ID, health status, load status, etc. The positioning method for determining the node position uses GPS, Wi-Fi signal strength, network topology information, etc. If the device has a known IP address in the network, use network management tools (such as Traceroute, Ping) to determine the position. For each edge computing device, the selected positioning method is applied to locate the position, and the coordinate information of each device (such as latitude and longitude, network topology coordinates) is obtained. The positioning coordinates of each node are associated with its comprehensive state characteristics and stored in a database to form structured node information. The structure of the state mapping model is defined, including nodes, connection relationships and state characteristics. A graph structure is used to represent the edge computing node. Each node contains the following information: node ID, location coordinates, real-time state characteristics (such as load, delay, health index, etc.). The previously collected device location and state characteristic data are integrated into the model to ensure that the information of each node is complete, forming a topological structure containing all edge computing nodes, and recording the connection relationship between nodes (such as network bandwidth, delay, etc.). Graph theory or network analysis tools ( Use tools such as NetworkX and Graph-tool to build edge computing node topology models, represent nodes and connection relationships in the form of graphs, ensure easy access and update of node status information, verify the accuracy of the built topology model, and compare with the actual network conditions to ensure that the node status and connection relationships reflected by the model are correct. Use network simulation tools (such as NS-3) to conduct simulation experiments to verify the performance of the model under different loads and fault conditions. Based on the verification results, optimize the topology model, such as adjusting the connection weights between nodes, updating node status characteristics, etc. Introduce a dynamic adjustment mechanism to update the topology model based on real-time monitoring data to adapt to the changing network environment. Establish a monitoring mechanism to continuously monitor the status and location changes of each node. Use automated tools (such as Prometheus and Grafana) to monitor the performance indicators of the nodes to ensure timely detection of anomalies. Regularly re-evaluate and update the topology model to ensure the accuracy and effectiveness of the model. Set scheduled tasks (such as cron jobs) or event-driven mechanisms to automatically update the model under specific conditions. Establish a feedback mechanism to collect feedback from users and the system, and adjust and optimize the model based on these feedbacks.

[0082] The task traffic trend module is used to obtain the AI ​​model task operation log; perform task queue analysis on the AI ​​model task operation log, and then perform model task traffic trend analysis to obtain AI model task traffic trend data;

[0083] In this embodiment, the source of the AI ​​model operation log is determined, which usually includes the model training, reasoning process and related system logs. An appropriate log collection tool (such as ELKStack, Fluentd, Logstash, etc.) is selected to centrally manage and collect log information. The log collection tool is configured to set the log file path, format and collection frequency to ensure that all relevant task operation logs are captured. The collected log data is stored in a database or data lake for subsequent analysis. A relational database (such as MySQL) or a non-relational database (such as MongoDB) is used to analyze the collected log format, including timestamp, task ID, task status, execution time and other fields. An appropriate log parsing method is selected, and regular expressions, JSON parsing libraries or log parsers (such as Apache Flink, Spark Streaming) are used to extract key information. Log parsing is implemented to extract and convert each record in the task operation log into structured data, and the following information of each task is recorded: Task ID, start time, end time, task status (success, failure, etc.), resource usage (such as CPU, memory), store the parsed structured data in the database to form task queue data for subsequent processing and analysis, determine the traffic trend indicators that need to be analyzed, such as: the number of tasks per unit time, the ratio of successful to failed tasks, the distribution of task execution time, extract the parsed task queue data from the database, prepare for traffic trend analysis, select appropriate analysis methods, such as time series analysis, sliding window analysis, or trend analysis algorithms (such as ARIMA, exponential smoothing, etc.), analyze the task queue data according to the selected analysis method, generate task traffic trend data, use the Pandas library in Python or R language for data analysis, generate time series graphs, and show the trend of task traffic over time, store the analysis results in the database, and use visualization tools (such as Matplotlib, Grafana, Tableau, etc.) to generate charts to facilitate understanding and display of task traffic trends.

[0084] The rolling prediction module is used to perform multi-time point rolling prediction on the AI ​​model task traffic trend data and perform model computing power load demand analysis to generate multi-time point computing power load demand data;

[0085] In this embodiment, previously generated task traffic trend data, including the number of tasks per unit time, success rate, failure rate, etc., are extracted from the database, the extracted data is cleaned, missing values, outliers and noise are processed, and missing values ​​are filled using interpolation, mean filling and other methods to ensure that the time format of the data is unified and arranged in chronological order. According to the task traffic trend data, relevant features are extracted, such as time features (such as hours, days, weeks, etc.), lag features (such as the number of tasks at the previous moment), and a suitable rolling prediction model is selected according to demand, which may be a simple time series model (such as ARIMA, SARIMA), a machine learning model (such as random forest, XGBoost) or a deep learning model (such as LSTM). The selected model is trained using the prepared task traffic trend data, and the data is divided into a training set and a test set. Usually 80% is used as a training set and 20% is used as a test set. The model is cross-validated, and hyperparameters are tuned to improve the prediction accuracy. Multi-time point rolling prediction is performed. The specific steps are as follows :Set the prediction window (such as the next hour, the next day, etc.), use the trained model to predict the data at the current moment, and store the prediction results, move the rolling window forward, use the latest available data to predict the next moment, until all required time points are covered, determine the key indicators of computing load demand, such as: the average execution time of each task, the CPU and memory resources required for each task, calculate the computing load demand at each point in the future based on the number of tasks in the rolling prediction, and generate the overall computing load demand data for each point in time based on the resource demand characteristics of the tasks. Store the computing load demand data in the database and use visualization tools (such as Grafana, Matplotlib) to generate charts to show the changing trend of future computing load demand. Analyze the generated computing load demand data, identify peak and trough periods, and evaluate whether the current resource configuration meets the demand. Based on the analysis results, generate optimization suggestions, such as: increase computing resources during peak periods and adjust task scheduling strategies to balance the load.

[0086] The optimal migration time point module is used to perform edge computing simulation and optimal migration time point analysis on the edge computing node topology model based on the multi-time point computing load demand data, so as to obtain the optimal migration time point data;

[0087] In this embodiment, the computing load demand data at multiple time points is obtained from the rolling prediction module to ensure that the data contains the load demand information at each time point and the corresponding edge computing node status. The current edge computing node topology model is obtained, including node ID, location, performance indicators (such as computing power, storage capacity), current load status, etc., to ensure that the topology information is complete and can reflect the connection relationship and performance between nodes. The edge computing environment is built using simulation tools (such as NS-3, OMNeT++, or a custom Python simulator), the performance parameters and network configuration of the edge computing nodes are set, the collected computing load demand data and edge node topology information are input into the simulation environment, the simulation time range and time step are determined to perform time-by-time simulation, run the edge computing simulation, observe the response and performance of the nodes under different load conditions, and collect The simulation results at each time point include the load situation and task processing capacity of each node. According to the simulation results and load demand data, a decision model for the optimal migration time point is established based on the following indicators: Node load: the load situation of the current node, Computing power demand: the upcoming computing power demand, Delay and bandwidth: the connection delay and bandwidth between nodes, select the appropriate algorithm for optimal migration time point analysis, for example: Heuristic algorithm: such as genetic algorithm, particle swarm optimization, etc., suitable for complex optimization problems, dynamic programming: suitable for problems with clear stages and decision-making processes, linear programming: used when a linear relationship is established between demand and resources, use the selected algorithm to perform optimal migration time point analysis for each time point, calculate which tasks should be migrated to which nodes and the best time for migration, record the migration decision results at each predicted time point, and form an optimal migration time point data set.

[0088] The migration decision module is used to make distributed edge migration decisions based on the optimal migration time point data, perform dynamic cloud-based collaborative management, and build an edge migration collaborative management model.

[0089] In this embodiment, the optimal migration time point data is obtained from the optimal migration time point module, including the migration source node, target node, migration time and other information of each task, to ensure that the data is complete and in a consistent format for subsequent processing, and a structure suitable for storing and processing migration decision data is designed, such as using a data frame (DataFrame) or a custom object to encapsulate the migration information. The data structure should include the following fields: task ID, source node ID, target node ID, migration time, migration status (to be migrated, migrating, migrated). Based on the optimal migration time point data, a distributed edge migration decision model is constructed. The model should consider the following factors: current load: real-time load conditions of the source node and the target node, network status: latency, bandwidth and availability between nodes, task priority: prioritize tasks according to their importance and urgency, and select appropriate decision algorithms, such as: rule-based decision-making: make simple decisions based on set rules (such as load thresholds, latency limits), machine learning models: use classification or regression algorithms (such as random forests, support vector machines) to predict the best migration decisions, and optimize algorithms. Methods: such as genetic algorithms and particle swarm optimization, to find the global optimal solution. According to the constructed decision model and the selected algorithm, each task to be migrated is evaluated one by one, and its migration decision is determined. The migration status is updated, the task is marked as "migrating", and the start time of the migration is recorded. A dynamic cloud collaborative management model is designed to achieve coordination between edge devices and the cloud. The model should include the following components: Monitoring module: real-time monitoring of the status of edge nodes and task execution, Scheduling module: responsible for task scheduling and migration decision execution, Resource management module: manages the allocation and usage of computing resources, Use monitoring tools (such as Prometheus, Grafana) to collect performance data of edge nodes in real time, including CPU, memory, network bandwidth, etc., to ensure that the monitoring data can be fed back to the decision module so as to dynamically adjust the migration strategy, and execute task migration according to the results of the migration decision. Message queues (such as RabbitMQ, Kafka) can be used for task scheduling to ensure the reliability of task migration. During the migration process, the task status is continuously monitored to ensure that the task is executed smoothly on the target node.

[0090] In this embodiment, the state mining module is used to identify available edge computing devices; perform multi-frequency information monitoring on available edge computing devices, and perform device comprehensive state mining to generate edge device comprehensive state features, specifically for:

[0091] Perform a global scan of the edge computing network available to the AI ​​model and extract available edge computing devices;

[0092] Perform multi-frequency information monitoring on available edge computing devices and extract edge computing device status monitoring parameters;

[0093] Calculate the GPU utilization of the edge computing device status monitoring parameters to obtain the device GPU utilization;

[0094] Perform device load analysis based on device GPU utilization to generate edge device load data;

[0095] Quantify the edge computing power of available edge computing devices to obtain a comprehensive computing power quantification value;

[0096] The comprehensive status of the equipment is mined on the real-time computing power quantization value and edge device load data to generate the comprehensive status characteristics of the edge device.

[0097] In this embodiment, a network scanning tool (such as Nmap, AngryIPScanner) is used to perform a global scan on the edge computing network to obtain information about all nodes in the network, record the IP address, MAC address, device type and running status of the device, filter out available edge computing devices (such as edge servers, IoT devices, etc.) according to the scanning results, ensure that the device is online and accessible, and confirm the device status through a ping command or other health check mechanism. Select a suitable monitoring tool (such as Prometheus, Grafana or Zabbix) to perform real-time monitoring of edge computing devices, configure a monitoring agent to ensure that the required data can be collected from each device, monitor information at multiple frequencies, including CPU load, memory usage, network traffic and GPU status, set a monitoring time interval (such as every minute, every hour), and record relevant status parameters. The GPU status data of the edge computing device is extracted through the monitoring tool, and a command line tool (such as nvidia-smi) is used to obtain GPU utilization, ensuring that the monitoring data contains information such as the current utilization, temperature and power consumption of the GPU, and using the collected GPU data, calculate the GPU utilization (such as utilization = (current usage time / total time)*1 00%), record the calculation results in the database for subsequent analysis, integrate the CPU, memory and GPU utilization data of the device to form a comprehensive device load data set, calculate the comprehensive load index of each device, consider the weights of different resources, use statistical analysis methods (such as linear regression, cluster analysis) to analyze the device load, identify devices with too high or too low load, define load thresholds, identify normal and abnormal load states, define the computing power evaluation indicators of edge computing devices, including the number of CPU cores, GPU performance, memory bandwidth, etc., determine the weight of each indicator, so as to calculate the comprehensive computing power quantization value, and calculate the value of the evaluation index according to the evaluation index. The system uses the computing power of each device to generate a comprehensive computing power quantification value, record the computing power quantification value, and associate it with the load data for comprehensive analysis. It extracts features based on the real-time computing power quantification value and device load data to generate comprehensive status features. It uses data mining techniques (such as principal component analysis and cluster analysis) to analyze the device status, identify potential performance bottlenecks and abnormal conditions, and use visualization tools (such as Grafana and Tableau) to display the comprehensive status features of the device, provide an intuitive data view, generate reports, summarize the device status, load analysis, and computing power evaluation results, and provide a basis for subsequent decision-making.

[0098] In this embodiment, the specific steps of quantifying the edge computing power of the available edge computing devices to obtain the comprehensive computing power quantization value are:

[0099] Extracting hardware configuration information of the available edge computing device;

[0100] Performing performance parameter identification on the hardware configuration information and extracting hardware performance parameters, wherein the hardware performance parameters include the number of CPU cores, CPU main frequency, the number of GPU cores, clock frequency and architecture floating-point performance parameters;

[0101] Calculate the CPU floating point performance based on the number of CPU cores and CPU main frequency to generate the CPU floating point performance;

[0102] Quantify the theoretical floating-point operation peak value of the GPU core number, clock frequency and architecture floating-point performance parameters to obtain the theoretical floating-point operation peak value;

[0103] Comprehensive computing performance evaluation is performed on the CPU floating-point computing performance and theoretical floating-point computing peak value to obtain the comprehensive computing power quantification value.

[0104] In this embodiment, a suitable hardware information extraction tool is selected, such as lscpu, nvidia-smi (for GPU), dmidecode, or hwinfo. These tools provide detailed hardware configuration information, including detailed specifications of the CPU and GPU. Run the selected tool on each edge computing device to extract hardware configuration data. Record the following information: number of CPU cores, CPU main frequency, number of GPU cores, GPU clock frequency, GPU architecture floating-point performance parameters (such as floating-point computing power per cycle), organize the extracted hardware configuration information, and store it in a structured data format (such as JSON or a database). Ensure that the data is easily accessible and analyzed. Extract the required hardware performance parameters from the organized data, ensuring that they include: number of CPU cores, CPU main frequency, number of GPU cores, GPU clock frequency, GPU architecture floating-point performance parameters, and use the following formula to calculate the floating-point computing performance (FLOPS) of the CPU: CPUFLOPS=number of CPU cores×CPU main frequency×2, where the CPU main frequency should be in GHz and needs to be converted to Hz for calculation. Use the following formula to calculate the theoretical floating-point operation peak (FLOPS) of the GPU: GPUFLOPS = number of GPU cores × GPU clock frequency × architecture floating-point performance parameters, where the GPU clock frequency should be in GHz and needs to be converted to Hz for calculation. Substitute the extracted GPU core number, clock frequency, and floating-point performance parameters into the formula for calculation to obtain the theoretical floating-point operation peak of the GPU. Comprehensively evaluate the CPU floating-point operation performance and the GPU theoretical floating-point operation peak: comprehensive computing power quantization value = CPUFLOPS + GPUFLOPS, which will reflect the overall computing power of the device. Store all calculation results, including hardware configuration information, CPU floating-point operation performance, GPU theoretical floating-point operation peak, and comprehensive computing power quantization value, in a database or file. Use a structured format (such as CSV, JSON) to facilitate subsequent analysis and access. Generate a detailed performance evaluation report that summarizes the hardware configuration and performance evaluation results of the device. The report should include visual charts to more intuitively display performance data.

[0105] In this embodiment, the edge topology module is used to locate the position of each node based on the comprehensive state characteristics of the edge device, and perform real-time state mapping modeling, so as to build an edge computing node topology model, which is specifically used for:

[0106] Sequentially encode and mark the available edge computing devices to obtain multiple edge computing device encoding nodes;

[0107] Locate multiple edge computing device encoding nodes one by one to generate the location coordinates of each node;

[0108] Perform spatial distribution analysis on the location coordinates of each node to generate spatial distribution data of edge nodes;

[0109] Based on the comprehensive state characteristics of edge devices, the spatial distribution data of edge nodes is modeled in real time to construct an edge computing node topology model.

[0110] In this embodiment, a unique code is assigned to each device, such as "Node_001", "Node_002", etc., to ensure the continuity and uniqueness of the code. These codes are automatically generated using scripts (such as Python) to ensure the consistency and accuracy of the markings. The code information and detailed information of the device (such as location, performance parameters, etc.) are stored in a structured database for subsequent retrieval and analysis. According to the network topology and physical location of the device, a suitable positioning method is selected, such as using GPS, Wi-Fi positioning or network topology mapping, and a location positioning operation is performed on each edge computing device to obtain its geographic coordinates (such as longitude and latitude). If Wi-Fi or network topology is used, the location of the device is estimated by signal strength or network delay. The location coordinates of each coded node are associated with its corresponding coded information to form a data set containing coded and positioning coordinates. The positioning coordinate data of all edge computing devices are collected for spatial distribution analysis to ensure that the data format is unified for subsequent processing. Spatial analysis tools (such as QGIS, ArcGIS or Python's GeoPandas library) are used for spatial distribution analysis. Appropriate analysis methods are selected, such as cluster analysis, heat map generation, etc., to identify the spatial distribution characteristics of the device and analyze the spatial distribution of the device. The edge node topology model is constructed by analyzing the spatial distribution of edge nodes, identifying the areas where devices are concentrated and their distribution patterns, recording the analysis results, including the number of devices, density distribution and other information, forming the spatial distribution data of edge nodes, collecting the comprehensive status feature data of edge devices, including performance indicators, load conditions and operating status, etc., associating the status features with device codes and location data, using modeling tools (such as MATLAB, Simulink or Python's Matplotlib, NetworkX library) to build an edge node topology model. In the model, the status features of the device are mapped to its location coordinates to generate a state mapping diagram. According to the connection relationship and communication path between devices, the topological structure of edge computing nodes is constructed to ensure that the model can reflect the interaction and dependency relationship between devices. The edge node topology model and spatial distribution data are displayed using visualization tools (such as Tableau, Grafana, D3.js), creating a visualization interface to display the device location, status features and heat map of spatial distribution, so that users can understand the distribution and status of nodes, providing interactive functions to allow users to zoom and view specific device status, etc. According to user feedback and visualization results, the effectiveness and accuracy of the model are evaluated, and the location positioning and status mapping model are continuously optimized to ensure that it adapts to the actual situation of the edge computing network.

[0111] In this embodiment, the specific steps of performing instant state mapping modeling on the spatial distribution data of edge nodes based on the comprehensive state characteristics of edge devices to construct the edge computing node topology model are:

[0112] Perform physical link interaction identification on multiple edge computing device encoding nodes and extract the interaction paths between nodes;

[0113] Mining the interactive connection logic of the interactive paths between nodes to obtain the connection logic between nodes;

[0114] Perform dynamic topological association evolution on the connection logic between nodes to generate dynamic topological association features;

[0115] Based on the dynamic interactive topological association characteristics, the spatial distribution data of edge nodes is fitted with topological distribution to build a topological distribution network of edge devices.

[0116] According to the comprehensive state characteristics of edge devices, the edge device topology distribution network is modeled in real time to build an edge computing node topology model.

[0117] In this embodiment, the communication data between edge computing devices is monitored and collected, and network monitoring tools (such as Wireshark and tcpdump) are used to obtain network traffic between devices, and information such as IP addresses, port numbers, and communication timestamps between devices are recorded for subsequent analysis. By analyzing the collected network traffic data, the physical link interaction between devices is identified, and the interaction path between each pair of devices is extracted to form a path list containing device pairs and their interaction frequency, delay, and other information. The extracted interaction paths are analyzed using data mining techniques (such as association rule mining and graph theory analysis) to identify the connection logic between nodes. For example, the Apriori algorithm or the FP-Growth algorithm is used to discover frequent item sets and identify high-frequency interaction device pairs. The mined connection logic is recorded in a database, including node pairs, interaction frequencies, and their logical relationships (such as master-slave relationships, peer-to-peer relationships, etc.). An appropriate dynamic network model (such as the Barabási-Albert model and the Erdős-Rényi model) is selected to simulate the dynamic topological evolution between nodes, considering the impact of device joining, leaving, and changes in interaction frequency on the network topology, and calculating the characteristics of the dynamic topology based on real-time interaction data. , such as node degree, clustering coefficient, connectivity, etc., record these dynamic features for subsequent analysis and fitting, combine the spatial distribution data of edge nodes with dynamic topological features to form a comprehensive data set, ensure the uniform data format, and facilitate subsequent analysis, select a suitable fitting model (such as Gaussian mixture model, K-means clustering) to analyze the topological distribution, use statistical methods (such as least squares method) to fit the topological distribution network of edge devices, and build a topological distribution network of edge devices based on the fitting results to ensure that the network can reflect the physical links and logical relationships between nodes. Collect the comprehensive status characteristics of edge devices, including performance indicators (such as CPU utilization, memory usage), load information, and network latency, etc., associate these status characteristics with the topological distribution network, use modeling tools (such as NetworkX, Gephi, or MATLAB) to build a topological model of edge computing nodes, and map the status characteristics of the devices to the topological structure, generate a visualization model, and display the status and interaction relationship of the nodes. Use visualization tools to display the constructed edge computing node topology model, provide intuitive network structure and device status information, evaluate the accuracy and effectiveness of the model, collect user feedback, and make necessary adjustments and optimizations.

[0118] In this embodiment, the task traffic trend module is used to obtain the AI ​​model task operation log; perform task queue analysis on the AI ​​model task operation log, and then perform model task traffic trend analysis to obtain AI model task traffic trend data, which is specifically used for:

[0119] Get AI model task running log;

[0120] Perform task queue analysis on AI model task running logs to identify real-time processing task data and unprocessed task data;

[0121] Perform task processing efficiency analysis on real-time processing task data and generate a model for real-time task processing efficiency;

[0122] Calculate the task processing throughput of the real-time processing task data to generate the model real-time task throughput;

[0123] Perform unprocessed task time series change analysis on unprocessed task data to generate unprocessed task change data;

[0124] Based on the model's real-time task processing efficiency and real-time task throughput, the model task traffic trend analysis is performed on the unprocessed task change data to obtain the AI ​​model task traffic trend data.

[0125] In this embodiment, by configuring a logging mechanism (such as using Python's logging module or other logging frameworks), it is ensured that task-related log information is recorded when the model is running, and that the log information includes key information such as task ID, status (such as start, completion, failure), processing time, task type, etc. The log data is stored in a centralized log management system (such as ELKStack, Splunk, or a simple database) for subsequent query and analysis. A log parsing tool (such as Logstash, Fluentd, or a custom script) is used to parse the stored logs. During parsing, key fields such as task ID, start time, end time, status, etc. are extracted. According to the task status, real-time processing tasks and unprocessed tasks are identified, and the real-time processing tasks and unprocessed tasks are stored in two lists for subsequent analysis. For each real-time processing task, its processing efficiency is calculated, and processing efficiency = processing time (seconds) / task complexity (such as data volume or computing volume). The processing efficiency of all real-time processing tasks is counted, and statistical data such as average efficiency, maximum value, and minimum value are calculated. These efficiency data are recorded to generate a real-time task processing efficiency report for the model. Throughput is defined as the number of tasks completed per unit time. Throughput = number of completed tasks / time period (seconds), select an appropriate time period (such as every minute, every hour) to calculate the throughput, collect the number of tasks completed in each time period, and calculate the real-time task throughput, collect the timestamp information of unprocessed tasks, record the number of unprocessed tasks at each time point, organize the unprocessed task data into a time series format for time series analysis, use statistical analysis methods (such as moving average, time series decomposition) to analyze the changing trend of the number of unprocessed tasks, generate unprocessed task change data, including trend charts and key indicators (such as growth rate), and perform trend analysis on the unprocessed task change data based on the real-time task processing efficiency and real-time task throughput of the model. Analysis, using regression analysis, time series forecasting (such as ARIMA model) and other methods to predict future task traffic trends, combined with real-time processing efficiency and throughput data, generate AI model task traffic trend data, and display the results in the form of visual charts (such as line charts, bar charts) to facilitate analysis and decision-making. All generated analysis data, efficiency reports and traffic trends are recorded in the database for subsequent query and analysis, and a comprehensive report including real-time task processing efficiency, task throughput, changes in unprocessed tasks and traffic trends is generated. Use visualization tools (such as Tableau, PowerBI) to create charts and provide intuitive analysis results.

[0126] In this embodiment, the rolling prediction module is used to perform multi-time point rolling prediction on the AI ​​model task traffic trend data and perform model computing power load demand analysis, thereby generating multi-time point computing power load demand data, specifically for:

[0127] Define multiple prediction time points, and perform multi-time point rolling predictions on the AI ​​model task traffic trend data based on the multiple prediction time points to generate task traffic trend prediction data at each time point;

[0128] Serialize the task traffic trend forecast data at each time point and build a task trend forecast sequence;

[0129] Perform model computing power load demand analysis on the task trend prediction sequence to generate computing power load demand data at multiple time points.

[0130] In this embodiment, multiple prediction time points are defined according to business needs and historical data analysis. For example, future hourly, daily, or weekly time points are selected for prediction to ensure that the selected time points can cover the required prediction range. It is generally recommended to select short-term and medium-term prediction time points, convert all defined prediction time points into a unified time format (such as ISO8601), ensure consistency in data processing, use the previously generated AI model task traffic trend data, ensure that the data format is suitable for time series analysis, preprocess the data, including removing outliers, filling missing values, and standardizing, and select a suitable time series prediction model, such as: ARIMA (autoregressive integrated moving average model): suitable for data with trends and seasonality, SARIMA (seasonal ARIMA): suitable for data with seasonal fluctuations, LSTM (long short-term memory network): suitable for complex nonlinear time series data, select the most suitable model according to the characteristics of the data, divide the historical task traffic data into a training set and a test set, use the training set to train the model, perform rolling predictions for each prediction time point, generate task traffic trend prediction data for each time point, and generate task traffic trend data for each time point. The forecast data is organized into a sequence to form a time series data structure, ensuring that the data in the sequence is arranged in chronological order for subsequent analysis. The forecast sequence is converted into a format suitable for storage and analysis, such as JSON, CSV, or a database table. Each time point should contain the forecasted traffic value and its corresponding timestamp. Based on the task traffic data, define the calculation formula for computing power load demand: computing power demand = task traffic × computing power required for each task. Determine the computing power value required for each task, which is obtained through historical data analysis or model-based performance evaluation. Use the above formula to calculate the computing power load demand for each time point in the task trend forecast sequence to generate multi-time point computing power load demand data, record the forecasted traffic at each time point and its corresponding computing power demand, store the generated task traffic trend forecast data and computing power load demand data in the database for subsequent query and analysis, use structured data formats (such as tables, time series databases) to store the results, use visualization tools (such as Tableau, PowerBI, or Matplotlib) to display the forecast results and computing power load demand data, and create charts (such as line charts and bar charts) to intuitively display the trend of task traffic and the corresponding changes in computing power demand.

[0131] In this embodiment, the optimal migration time point module is used to perform edge computing simulation and optimal migration time point analysis on the edge computing node topology model based on multi-time point computing load demand data, so as to obtain optimal migration time point data, which is specifically used for:

[0132] Perform edge computing simulation on the edge computing node topology model based on the computing power load demand data at multiple time points to generate edge computing simulation response data;

[0133] The edge computing simulation response data is subjected to node-by-node response delay calculation, and the node calculation response delay of each node is extracted;

[0134] Perform real-time computing performance evaluation on the node computing response delay of each node, using the real-time computing performance value of each node;

[0135] The optimal migration time point analysis is performed based on the real-time computing power performance value of each node to obtain the optimal migration time point data.

[0136] In this embodiment, an edge computing simulation environment is constructed, and a simulation tool (such as NS-3, OMNeT++, or a custom Python simulator) is used to create a topological model of edge computing nodes. Simulation parameters, including network delay, bandwidth, node performance, etc., are set. Multi-point computing load demand data are imported into the simulation environment to simulate node responses under different load conditions, ensuring that the data format is consistent with the requirements of the simulation tool. The simulation is started, and the response of each edge computing node under different loads is observed. The simulated response data is collected, and the computing performance and response results of each node at each load point are recorded. The response time and related performance indicators of each node are extracted from the edge computing simulation response data, ensuring that the data includes information such as the task processing time and network delay of each node. The response data of each node is processed, and the computing response delay of the node is calculated. Response delay = task completion time − task start time. The response delay data of each node is recorded to form a delay data set, and real-time computing power is defined. The evaluation indicators of performance, such as the processing power, throughput and latency of each node, are used to calculate the real-time computing power performance value of each node using the following formula: Real-time computing power performance = number of processed tasks / response delay. According to the response delay data calculated in the previous step, the real-time computing power performance of each node is evaluated, and the real-time computing power performance value of each node is recorded to form a performance data set. The method for analyzing the optimal migration time point is determined, such as using heuristic algorithms, dynamic programming or machine learning models, and setting evaluation criteria, such as the lowest response delay, the highest computing power performance or load balancing. Based on the real-time computing power performance value of each node, the performance changes at each time point are analyzed to identify the best migration time point. For example, by comparing the changes in the performance values ​​of different nodes at each time point, it is determined at which time point the migration task can obtain the best performance. The optimal migration time point of each node and its corresponding performance data are recorded for subsequent reference and analysis to form an optimal migration time point data set, including the node ID, the migration time point and its computing power performance value.

[0137] In this embodiment, the migration decision module is used to make distributed edge migration decisions based on the optimal migration time point data, and to perform dynamic cloud collaborative management to build an edge migration collaborative management model, specifically for:

[0138] Identify current unprocessed tasks based on the AI ​​model task running log and extract all current unprocessed tasks;

[0139] Performing data lossless encapsulation processing on the current unprocessed task to obtain multiple lossless encapsulation task packages;

[0140] Make distributed edge migration decisions for multiple lossless encapsulation task packages based on the optimal migration time point data and build a distributed edge migration strategy;

[0141] According to the distributed edge migration strategy, the edge computing node topology model is dynamically managed in the cloud and an edge migration collaborative management model is constructed.

[0142] In this embodiment, data is extracted from the task running log of the AI ​​model to ensure that the log contains information such as task ID, status (such as unprocessed, in progress, completed), timestamp, etc., and a log analysis tool (such as ELKStack, Splunk) or a custom script (such as Python) is used to quickly filter out currently unprocessed tasks, parse the log data, identify all tasks with a current status of "unprocessed", store the information of the unprocessed tasks (such as task ID, related parameters, timestamp, etc.) in a data structure (such as a list or database table), define the standard for lossless encapsulation, ensure that the task does not lose any key information during the encapsulation process, and encapsulate the task-related Input data, metadata, and execution environment configuration, encapsulate data for each unprocessed task to form multiple lossless encapsulated task packages, use data serialization technology (such as JSON, Protocol Buffers) to package task data and metadata, ensure the integrity and parsability of the encapsulated data, store the lossless encapsulated task packages in distributed storage systems (such as HDFS, Amazon S3) for subsequent access and migration, obtain the optimal migration time point data generated previously, including the performance indicators of each node and the time point information suitable for migration, ensure that these data are combined with the information of the lossless encapsulated task package to make decisions, and make decisions based on the characteristics of the task (such as computational complexity). , data size) and node performance indicators, build a distributed edge migration decision model, use decision trees, heuristic algorithms or optimization models (such as linear programming) to decide which tasks should be migrated to which edge nodes, generate a distributed edge migration strategy, determine the target edge node and its migration time for each lossless encapsulation task package, store the decision results in the strategy table for subsequent execution, and dynamically update the topology model of the edge computing node according to the distributed edge migration strategy to ensure that the topology model reflects the current node status, load and real-time situation of the migration task. Based on the updated topology model, build an edge migration collaborative management model to achieve dynamic coordination between the cloud and edge devices. Adopt microservice architecture or event-driven architecture to ensure that the system can respond to task migration, load changes and node status in real time, implement dynamic cloud collaborative management strategies to ensure the smooth progress of task migration, monitor performance indicators during the migration process, ensure the normal operation of tasks after migration, monitor the task status and performance of edge computing nodes after migration in real time to ensure that tasks are executed as expected, use monitoring tools (such as Prometheus, Grafana) to collect and display key indicators (such as latency, throughput, load, etc.), analyze the effect of task migration based on the monitoring results, identify bottlenecks and problems, and continuously optimize migration strategies and collaborative management models to adapt to the dynamically changing edge computing environment.

[0143] Therefore, the embodiments should be regarded as illustrative and non-restrictive from all points, and the scope of the present invention is limited by the appended claims rather than the above description, and it is therefore intended that all changes falling within the meaning and range of equivalent elements of the application documents are included in the present invention.

[0144] The above is only a specific embodiment of the present invention, so that those skilled in the art can understand or implement the present invention. Various modifications to these embodiments will be apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but should conform to the widest scope consistent with the principles and novel features invented herein.

Claims

1. An edge computing management system that supports dynamic migration of AI models, characterized in that: The edge computing management system supporting dynamic migration of AI models includes a state mining module, an edge topology module, a task traffic trend module, a rolling prediction module, an optimal migration time point module, and a migration decision module: The state mining module is used to identify available edge computing devices; perform multi-frequency information monitoring on available edge computing devices, and perform device comprehensive state mining to generate comprehensive state features of edge devices; The edge topology module is used to locate the position of each node based on the comprehensive state characteristics of the edge device, and perform real-time state mapping modeling, so as to build an edge computing node topology model; The task traffic trend module is used to obtain the AI ​​model task operation log; perform task queue analysis on the AI ​​model task operation log, and then perform model task traffic trend analysis to obtain AI model task traffic trend data; The rolling prediction module is used to perform multi-time point rolling prediction on the AI ​​model task traffic trend data and perform model computing power load demand analysis to generate multi-time point computing power load demand data; The optimal migration time point module is used to perform edge computing simulation and optimal migration time point analysis on the edge computing node topology model based on the multi-time point computing load demand data, so as to obtain the optimal migration time point data; The migration decision module is used to make distributed edge migration decisions based on the optimal migration time point data, and to perform dynamic cloud collaborative management to build an edge migration collaborative management model; The rolling prediction module is used to perform multi-time point rolling prediction on the AI ​​model task traffic trend data and to perform model computing load demand analysis, thereby generating multi-time point computing load demand data, specifically for: Define multiple prediction time points, and perform multi-time point rolling predictions on the AI ​​model task traffic trend data based on the multiple prediction time points to generate task traffic trend prediction data at each time point; Serialize the task traffic trend forecast data at each time point and build a task trend forecast sequence; Perform model computing load demand analysis on the task trend prediction sequence to generate computing load demand data at multiple time points; The optimal migration time point module is used to perform edge computing simulation and optimal migration time point analysis on the edge computing node topology model based on multi-time point computing load demand data, so as to obtain optimal migration time point data, which is specifically used for: Perform edge computing simulation on the edge computing node topology model based on the computing power load demand data at multiple time points to generate edge computing simulation response data; The edge computing simulation response data is subjected to node-by-node response delay calculation, and the node calculation response delay of each node is extracted; Perform real-time computing performance evaluation on the node computing response delay of each node, using the real-time computing performance value of each node; Analyze the optimal migration time point based on the real-time computing power performance value of each node to obtain the optimal migration time point data; The migration decision module is used to make distributed edge migration decisions based on the optimal migration time point data, and to perform dynamic cloud collaborative management and build an edge migration collaborative management model, specifically for: Identify current unprocessed tasks based on the AI ​​model task running log and extract all current unprocessed tasks; Performing data lossless encapsulation processing on the current unprocessed task to obtain multiple lossless encapsulation task packages; Make distributed edge migration decisions for multiple lossless encapsulation task packages based on the optimal migration time point data and build a distributed edge migration strategy; According to the distributed edge migration strategy, the edge computing node topology model is dynamically managed in the cloud and an edge migration collaborative management model is constructed.

2. The edge computing management system supporting dynamic migration of AI models according to claim 1, characterized in that: The state mining module is used to identify available edge computing devices; perform multi-frequency information monitoring on available edge computing devices, and perform device comprehensive state mining to generate edge device comprehensive state features, specifically for: Perform a global scan of the edge computing network available to the AI ​​model and extract available edge computing devices; Perform multi-frequency information monitoring on available edge computing devices and extract edge computing device status monitoring parameters; Calculate the GPU utilization of the edge computing device status monitoring parameters to obtain the device GPU utilization; Perform device load analysis based on device GPU utilization to generate edge device load data; Quantify the edge computing power of available edge computing devices to obtain a comprehensive computing power quantification value; The comprehensive status of the equipment is mined on the real-time computing power quantization value and edge device load data to generate the comprehensive status characteristics of the edge device.

3. The edge computing management system supporting dynamic migration of AI models according to claim 2, characterized in that: The specific steps of quantifying the edge computing power of the available edge computing devices to obtain the comprehensive computing power quantification value are: Extracting hardware configuration information of the available edge computing device; Performing performance parameter identification on the hardware configuration information and extracting hardware performance parameters, wherein the hardware performance parameters include the number of CPU cores, CPU main frequency, the number of GPU cores, clock frequency and architecture floating-point performance parameters; Calculate the CPU floating point performance based on the number of CPU cores and CPU main frequency to generate the CPU floating point performance; Quantify the theoretical floating-point operation peak value of the GPU core number, clock frequency and architecture floating-point performance parameters to obtain the theoretical floating-point operation peak value; Comprehensive computing performance evaluation is performed on the CPU floating-point computing performance and theoretical floating-point computing peak value to obtain the comprehensive computing power quantification value.

4. The edge computing management system supporting dynamic migration of AI models according to claim 1, characterized in that: The edge topology module is used to locate the position of each node based on the comprehensive state characteristics of the edge device, and to perform real-time state mapping modeling, thereby building an edge computing node topology model, specifically for: Sequentially encode and mark the available edge computing devices to obtain multiple edge computing device encoding nodes; Locate multiple edge computing device encoding nodes one by one to generate the location coordinates of each node; Perform spatial distribution analysis on the location coordinates of each node to generate spatial distribution data of edge nodes; Based on the comprehensive state characteristics of edge devices, the spatial distribution data of edge nodes is modeled in real time to construct an edge computing node topology model.

5. The edge computing management system supporting dynamic migration of AI models according to claim 4, characterized in that: The specific steps of performing instant state mapping modeling on the spatial distribution data of edge nodes based on the comprehensive state characteristics of edge devices to construct the edge computing node topology model are as follows: Perform physical link interaction identification on multiple edge computing device encoding nodes and extract the interaction paths between nodes; Mining the interactive connection logic of the interactive paths between nodes to obtain the connection logic between nodes; Perform dynamic topological association evolution on the connection logic between nodes to generate dynamic topological association features; Based on the dynamic interactive topological association characteristics, the spatial distribution data of edge nodes is fitted with topological distribution to build a topological distribution network of edge devices. According to the comprehensive state characteristics of edge devices, the edge device topology distribution network is modeled in real time to build an edge computing node topology model.

6. The edge computing management system supporting dynamic migration of AI models according to claim 1, characterized in that: The task traffic trend module is used to obtain the AI ​​model task operation log; perform task queue analysis on the AI ​​model task operation log, and then perform model task traffic trend analysis to obtain AI model task traffic trend data, which is specifically used for: Get AI model task running log; Perform task queue analysis on AI model task running logs to identify real-time processing task data and unprocessed task data; Perform task processing efficiency analysis on real-time processing task data and generate a model for real-time task processing efficiency; Calculate the task processing throughput of the real-time processing task data to generate the model real-time task throughput; Perform unprocessed task time series change analysis on unprocessed task data to generate unprocessed task change data; Based on the model's real-time task processing efficiency and real-time task throughput, the model task traffic trend analysis is performed on the unprocessed task change data to obtain the AI ​​model task traffic trend data.

Images