Cloud edge collaborative algorithm arrangement method and device, equipment and storage medium
Through structured modeling and dynamic resource prediction of cloud edge multi-layer computing resources, combined with deep reinforcement learning and heuristic optimization, the adaptive orchestration of cloud edge collaborative algorithms is realized, and the low resource utilization rate and the The problem of system load imbalance significantly improves the utilization rate of computing resources and task processing efficiency.
Patent Information
- Application Number
- CN202411193082.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-28
- Publication Date
- 2024-11-22
- Estimated Expiration
- Not applicable · inactive patent
AI Technical Summary
The existing cloud-edge collaborative algorithms are difficult to effectively respond to dynamically changing computing environments and diversified computing needs, resulting in low resource utilization and unbalanced system load, which may cause system bottlenecks and performance degradation.
By structuring multi-layer computing resources at the edge of the cloud, a multi-level resource pool and algorithm library are built, and resource use prediction and task decomposition can be used to achieve dynamic matching of algorithms and subtasks, and optimize tasks based on load evolution and delay model. Mapping relationships, combining deep reinforcement learning and heuristic optimization to generate adaptive algorithm orchestration strategies, and dynamically adjust and iteratively optimize through real-time execution results.
It significantly improves the utilization rate of multi-level computing resources at the edge of the cloud, effectively balances the system load, reduces task processing delays, and perceives and adapts to complex and changeable computing environments in real time.
Smart Images

Figure CN119003181A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of cloud-edge collaboration technology, and in particular to a cloud-edge collaboration algorithm orchestration method, device, equipment and storage medium. Background Art
[0002] With the rapid development of cloud computing and edge computing, the traditional centralized computing model can no longer meet the increasingly complex application requirements. As an emerging computing model, cloud-edge collaborative computing combines the powerful processing capabilities of cloud computing with the low latency advantage of edge computing to provide an efficient and flexible solution for data processing and application services.
[0003] The key to cloud-edge collaborative computing is how to effectively distribute and schedule computing tasks between the cloud and the edge. Existing cloud-edge collaborative algorithms mainly rely on predefined rules or static algorithm orchestration strategies, which are difficult to cope with dynamically changing computing environments and diverse computing needs. These static orchestration strategies usually cannot fully utilize the multi-level computing resources of the cloud and the edge, resulting in low resource utilization, unbalanced system load, and may even cause system bottlenecks and performance degradation. Summary of the invention
[0004] The main purpose of the present invention is to solve the technical problem that the existing data organization method is difficult to effectively support fast query and complex data analysis.
[0005] A first aspect of the present invention provides a cloud-edge collaborative algorithm orchestration method, which is applied to an animal trap device, and the cloud-edge collaborative algorithm orchestration method includes: Perform structured modeling on the multi-layer computing resources at the cloud edge to obtain a multi-layer resource pool, and perform classification and mapping on the algorithms corresponding to the multi-layer resource pool to obtain a multi-layer algorithm library; Performing prediction processing on the resource usage data in the multi-level resource pool to obtain a resource usage prediction result, and selecting an applicable algorithm from the multi-level algorithm library according to the prediction result to obtain a candidate algorithm set; Decomposing the computing task according to the resource usage prediction result and the candidate algorithm set to obtain a subtask set, and performing algorithm matching on the subtask set to obtain a task-algorithm mapping relationship; According to the task-algorithm mapping relationship and the system load condition, a load evolution model and a delay model are constructed, and the task-algorithm mapping relationship is optimized according to the load evolution model and the delay model to obtain an initial algorithm scheduling strategy; Performing deep reinforcement learning training and heuristic optimization processing on the initial algorithm orchestration strategy to obtain an adaptive algorithm orchestration optimization strategy; The algorithm is scheduled and executed on the cloud edge according to the adaptive algorithm orchestration optimization strategy, and the adaptive algorithm orchestration optimization strategy is dynamically adjusted and iteratively optimized according to the execution result.
[0006] Optionally, in a first implementation of the first aspect of the present invention, the structured modeling processing is performed on the cloud-edge multi-layer computing resources to obtain a multi-layer resource pool, and the algorithms corresponding to the multi-layer resource pool are classified and mapped to obtain a multi-layer algorithm library including: Perform K-means clustering on cloud computing resources, edge computing resources, and terminal computing resources to obtain multiple computing resource areas; Perform red-black tree construction processing on cloud computing resources, edge computing resources, and terminal computing resources according to the computing resource area to obtain a small resource tree; Performing B+ tree construction processing on the small resource tree to obtain a large resource tree, and combining the small resource tree with the large resource tree to obtain a multi-level resource pool; Classify cloud computing algorithms, edge computing algorithms, and terminal computing algorithms to obtain an algorithm category set; According to the resource types in the multi-level resource pool and the algorithm category set, mapping and associating the cloud computing algorithm, the edge computing algorithm and the terminal computing algorithm are processed to obtain an algorithm-resource mapping table; According to the algorithm-resource mapping table, the cloud computing algorithm, the edge computing algorithm and the terminal computing algorithm are hierarchically organized and processed to obtain a multi-layer algorithm library.
[0007] Optionally, in a second implementation of the first aspect of the present invention, the resource usage data in the multi-level resource pool is predicted and processed to obtain a resource usage prediction result, and an applicable algorithm is selected from the multi-level algorithm library according to the prediction result, and the candidate algorithm set obtained includes: Performing time series reconstruction processing on the historical resource usage data in the multi-level resource pool to obtain a multi-dimensional resource usage matrix, constructing a long short-term memory network model based on the multi-dimensional resource usage matrix, and training the long short-term memory network model to obtain a resource usage prediction model; Inputting the current resource status of the multi-level resource pool into the resource usage prediction model, performing forward propagation calculation processing to obtain a resource usage prediction result, and screening the algorithms in the multi-level algorithm library according to the resource usage prediction result to obtain a preliminary candidate algorithm set; Quantifying resource requirements of algorithms in the preliminary candidate algorithm set to obtain algorithm resource requirement vectors; A similarity calculation process is performed based on the algorithm resource demand vector and the resource usage prediction result to obtain an algorithm-resource matching matrix, and a threshold filtering process is performed on the algorithm-resource matching matrix to obtain a final candidate algorithm set.
[0008] Optionally, in a third implementation of the first aspect of the present invention, the computing task is decomposed according to the resource usage prediction result and the candidate algorithm set to obtain a subtask set, and the subtask set is subjected to algorithm matching processing to obtain a task-algorithm mapping relationship, including: Performing directed acyclic graph modeling processing on the computing tasks to obtain a task dependency graph, and performing node weight allocation processing on the task dependency graph according to the resource usage prediction result to obtain a weighted task dependency graph; Performing modular decomposition processing on the weighted task dependency graph to obtain a preliminary subtask set, and performing granularity adjustment processing on the preliminary subtask set according to the candidate algorithm set to obtain a final subtask set; Performing resource requirement analysis on each subtask in the final subtask set to obtain a subtask resource requirement vector, and performing similarity calculation based on the subtask resource requirement vector and the resource consumption characteristics of the candidate algorithm set to obtain a subtask-algorithm similarity matrix; The subtask-algorithm similarity matrix is matched and optimized to obtain a preliminary task-algorithm mapping relationship, and the scheduling order is optimized based on the preliminary task-algorithm mapping relationship and the weighted task dependency graph to obtain a task-algorithm mapping relationship.
[0009] Optionally, in a fourth implementation of the first aspect of the present invention, constructing a load evolution model and a delay model according to the task-algorithm mapping relationship and the system load condition, and optimizing the task-algorithm mapping relationship according to the load evolution model and the delay model to obtain an initial algorithm orchestration strategy includes: Performing time series analysis on the system load status to obtain a load change trend function, and generating a load evolution model according to the load change trend function and the task-algorithm mapping relationship; Performing graph theory modeling on the cloud-edge network topology structure to obtain a network topology graph, and performing communication delay calculation processing based on the network topology graph and the task-algorithm mapping relationship to obtain a communication delay matrix; Performing queue theory analysis on the load evolution model to obtain a calculation delay function, and combining the communication delay matrix and the calculation delay function to obtain an end-to-end delay model; A multi-objective optimization problem is constructed according to the load evolution model and the end-to-end delay model, the multi-objective optimization problem is solved to obtain a Pareto optimal solution set, and a decision analysis is performed on the Pareto optimal solution set to obtain an initial algorithm orchestration strategy.
[0010] Optionally, in a fifth implementation of the first aspect of the present invention, performing deep reinforcement learning training and heuristic optimization processing on the initial algorithm orchestration strategy to obtain an adaptive algorithm orchestration optimization strategy includes: Constructing a Markov decision process model according to the initial algorithm arrangement strategy, performing state space and action space definition processing on the Markov decision process model to obtain a reinforcement learning environment; Performing reward function design processing on the reinforcement learning environment to obtain a performance evaluation index, and constructing a dual deep Q network model according to the reinforcement learning environment and the performance evaluation index; Performing parameter initialization processing on the dual deep Q network model to obtain an initial policy network and a target network, and performing alternating training processing on the initial policy network and the target network to obtain an optimized deep Q network model; Performing strategy extraction processing on the initial algorithm orchestration strategy according to the optimized deep Q network model to obtain a deep reinforcement learning optimization strategy, and encoding processing on the deep reinforcement learning optimization strategy to obtain a strategy chromosome population; Performing genetic algorithm iteration processing on the strategy chromosome population to obtain an optimized strategy chromosome, and performing decoding processing on the optimized strategy chromosome to obtain a heuristic optimization strategy; The deep reinforcement learning optimization strategy and the heuristic optimization strategy are integrated to obtain an adaptive algorithm orchestration optimization strategy.
[0011] Optionally, in a sixth implementation of the first aspect of the present invention, the algorithm scheduling execution on the cloud edge according to the adaptive algorithm orchestration optimization strategy, and dynamically adjusting and iteratively optimizing the adaptive algorithm orchestration optimization strategy according to the execution result includes: The adaptive algorithm arrangement optimization strategy is divided into time windows to obtain a rolling optimization time series, and the adaptive algorithm arrangement optimization strategy is executed in sections according to the rolling optimization time series to obtain a phased execution result; Perform multi-dimensional performance index extraction processing on the phased execution results to obtain an execution effect evaluation matrix, and perform deviation analysis processing on the adaptive algorithm orchestration optimization strategy according to the execution effect evaluation matrix to obtain a strategy adjustment direction vector; The strategy adjustment direction vector is subjected to gradient descent optimization processing to obtain a strategy update increment, and the strategy update increment is fused with the adaptive algorithm orchestration optimization strategy to obtain and execute an updated adaptive algorithm orchestration optimization strategy.
[0012] A second aspect of the present invention provides a cloud-edge collaborative algorithm orchestration device, the cloud-edge collaborative algorithm orchestration device comprising: A modeling and mapping module is used to perform structured modeling processing on the multi-layer computing resources of the cloud edge to obtain a multi-layer resource pool, and to classify and map the algorithms corresponding to the multi-layer resource pool to obtain a multi-layer algorithm library; An algorithm screening module is used to perform prediction processing on the resource usage data in the multi-level resource pool to obtain a resource usage prediction result, and select an applicable algorithm from the multi-level algorithm library according to the prediction result to obtain a candidate algorithm set; A task decomposition module is used to decompose the computing task according to the resource usage prediction result and the candidate algorithm set to obtain a subtask set, and perform algorithm matching on the subtask set to obtain a task-algorithm mapping relationship; An initial scheduling module, used to construct a load evolution model and a delay model according to the task-algorithm mapping relationship and the system load status, and optimize the task-algorithm mapping relationship according to the load evolution model and the delay model to obtain an initial algorithm scheduling strategy; An adaptive orchestration module, used to perform deep reinforcement learning training and heuristic optimization processing on the initial algorithm orchestration strategy to obtain an adaptive algorithm orchestration optimization strategy; The scheduling module is used to schedule and execute the algorithm on the cloud edge according to the adaptive algorithm orchestration optimization strategy, and dynamically adjust and iteratively optimize the adaptive algorithm orchestration optimization strategy according to the execution results.
[0013] The third aspect of the present invention provides a cloud-edge collaborative algorithm orchestration device, comprising: a memory and at least one processor, the memory storing instructions, the memory and the at least one processor being interconnected via lines; the at least one processor calls the instructions in the memory so that the cloud-edge collaborative algorithm orchestration device executes the steps of the above-mentioned cloud-edge collaborative algorithm orchestration method.
[0014] A fourth aspect of the present invention provides a computer-readable storage medium, which stores instructions that, when executed on a computer, enable the computer to execute the steps of the above-mentioned cloud-edge collaborative algorithm orchestration method.
[0015] The above-mentioned cloud-edge collaborative algorithm orchestration method, device, equipment and storage medium construct a multi-level resource pool and algorithm library by structured modeling of multi-layer computing resources on the cloud-edge side. Then, dynamic matching of algorithms and subtasks is achieved through resource usage prediction and task decomposition. Next, based on the load evolution and delay model, the task-algorithm mapping relationship is optimized to form an initial orchestration strategy. Finally, by combining deep reinforcement learning and heuristic optimization, an adaptive algorithm orchestration optimization strategy is generated, and dynamic adjustment and iterative optimization are performed through real-time execution results. This method can perceive and adapt to complex and changeable computing environments in real time by combining technologies such as multi-level resource modeling, dynamic resource prediction, adaptive task decomposition and deep reinforcement learning, significantly improving the utilization rate of multi-level computing resources on the cloud-edge side, effectively balancing the system load and reducing task processing delays.
[0016] Other features and advantages of the present invention will be described in the following description, and partly become apparent from the description, or understood by practicing the present invention. The purpose and other advantages of the present invention are realized and obtained by the structures particularly pointed out in the description, claims and drawings.
[0017] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, preferred embodiments are given below and described in detail with reference to the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS
[0018] Figure 1 This is a schematic diagram of a first embodiment of a cloud-edge collaborative algorithm orchestration method in an embodiment of the present invention; Figure 2 A schematic diagram of an embodiment of a cloud-edge collaborative algorithm orchestration device in an embodiment of the present invention; Figure 3 A schematic diagram of an embodiment of a cloud-edge collaborative algorithm orchestration device in an embodiment of the present invention. DETAILED DESCRIPTION
[0019] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution of the present invention will be clearly and completely described below in conjunction with the accompanying drawings. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.
[0020] The terms "including" and "having" and any variations thereof mentioned in the embodiments of the present invention are intended to cover non-exclusive inclusions. For example, a process, method, system, product or device end including a series of steps or units is not limited to the listed steps or units, but may optionally include other steps or units that are not listed, or may optionally include other steps or units that are inherent to these processes, methods, products or device ends.
[0021] To facilitate understanding of this embodiment, a cloud-edge collaborative algorithm arrangement method disclosed in an embodiment of the present invention is first introduced in detail, and the cloud-edge collaborative algorithm arrangement method is applied to an animal trap device. Figure 1 As shown, the method comprises the following steps: 101. Perform structured modeling processing on the multi-layer computing resources of the cloud edge to obtain a multi-layer resource pool, and perform classification mapping processing on the algorithms corresponding to the multi-layer resource pool to obtain a multi-layer algorithm library; In one embodiment of the present invention, the structured modeling processing is performed on the cloud-edge multi-layer computing resources to obtain a multi-level resource pool, and the algorithms corresponding to the multi-level resource pool are classified and mapped to obtain a multi-level algorithm library, including: K-means clustering processing is performed on cloud computing resources, edge computing resources and terminal computing resources to obtain multiple computing resource areas; red-black tree construction processing is performed on cloud computing resources, edge computing resources and terminal computing resources according to the computing resource areas to obtain a small resource tree; B+ tree construction processing is performed on the small resource tree to obtain a large resource tree, and the small resource tree and the large resource tree are combined to obtain a multi-level resource pool; cloud computing algorithms, edge computing algorithms and terminal computing algorithms are classified to obtain an algorithm category set; according to the resource type in the multi-level resource pool and the algorithm category set, cloud computing algorithms, edge computing algorithms and terminal computing algorithms are mapped and associated to obtain an algorithm-resource mapping table; cloud computing algorithms, edge computing algorithms and terminal computing algorithms are hierarchically organized according to the algorithm-resource mapping table to obtain a multi-level algorithm library.
[0022] Specifically, the K-means clustering algorithm is first used to cluster cloud computing resources, edge computing resources, and terminal computing resources to obtain multiple computing resource areas. For example, in a smart city scenario, cloud-edge collaborative algorithm orchestration may cluster data centers, edge servers, and IoT devices distributed in different geographical locations into several resource areas based on factors such as their computing power, storage capacity, and network connection characteristics. This clustering not only takes into account the physical distribution of resources, but also the performance characteristics of resources, laying the foundation for subsequent resource management. Next, based on these computing resource areas, the system performs red-black tree construction processing on cloud computing resources, edge computing resources, and terminal computing resources to obtain a small resource tree. The red-black tree is a self-balancing binary search tree that can maintain the balance of the tree while ensuring that the time complexity of search, insertion, and deletion operations is O(log n). In practical applications, for example, for a specific edge computing area, the small resource tree can efficiently organize and manage various computing nodes in the area, so that the system can quickly locate and schedule resources suitable for specific tasks. In order to better manage these small resource trees, the system further performs B+ tree construction processing on them to obtain a large resource tree. B+ tree is a multi-way search tree, which is particularly suitable for the storage and retrieval of large-scale data. In the cloud-edge collaboration scenario, the large resource tree can effectively organize and index resource clusters distributed in different geographical locations. For example, when processing large-scale data analysis tasks across regions, the large resource tree can help the system quickly locate the most suitable resource combination, including high-performance computing resources in the cloud and low-latency processing capabilities at the edge. By combining the small resource tree and the large resource tree, the system finally obtains a multi-level resource pool. This resource pool not only contains the computing resource information at each level, but also retains the hierarchical relationship and organizational structure between resources. This multi-level resource organization method enables the system to flexibly schedule and allocate resources at different levels and granularities according to the characteristics and requirements of the task. At the same time, the system also needs to classify cloud computing algorithms, edge computing algorithms, and terminal computing algorithms to obtain a set of algorithm categories. This process may involve analysis and classification of multiple dimensions such as the algorithm's functions, performance characteristics, and resource requirements. For example, some machine learning algorithms may be classified as cloud algorithms that require high computing power, while some real-time data processing algorithms may be classified as algorithms suitable for running on edge nodes. After obtaining the multi-level resource pool and algorithm category set, the system performs a key mapping association process and generates an algorithm-resource mapping table. This mapping table establishes the correspondence between algorithm types and resource characteristics, providing an important reference for subsequent task scheduling.For example, for a deep learning task that requires a lot of GPU computing, the mapping table may prioritize it to cloud computing resources with high-performance GPU clusters; while for a real-time image processing task that requires low-latency response, the mapping table may associate it with the edge computing node closest to the data source. Finally, based on this algorithm-resource mapping table, the system hierarchically organizes and processes cloud computing algorithms, edge computing algorithms, and terminal computing algorithms, and finally builds a multi-layer algorithm library. This multi-layer algorithm library not only contains detailed information about various algorithms, but also reflects the adaptation relationship between algorithms and resources. In practical applications, when the system receives a complex task, it can quickly select the most appropriate algorithm combination from this multi-layer algorithm library based on the characteristics of the task and the current resource status. For example, for an intelligent transportation system task involving large-scale data processing and real-time decision-making, the system may choose algorithms that perform data analysis and model training in the cloud, algorithms that perform real-time data filtering and preliminary processing at edge nodes, and algorithms that perform data collection and simple preprocessing at terminal devices, so as to achieve efficient execution of the entire task.
[0023] 102. Performing prediction processing on the resource usage data in the multi-level resource pool to obtain a resource usage prediction result, and selecting an applicable algorithm from the multi-level algorithm library according to the prediction result to obtain a candidate algorithm set; In one embodiment of the present invention, the resource usage data in the multi-level resource pool is predicted and processed to obtain a resource usage prediction result, and an applicable algorithm is selected from the multi-layer algorithm library according to the prediction result to obtain a candidate algorithm set, including: performing time series reconstruction processing on the historical resource usage data in the multi-level resource pool to obtain a multi-dimensional resource usage matrix, and constructing a long short-term memory network model according to the multi-dimensional resource usage matrix, and training the long short-term memory network model to obtain a resource usage prediction model; inputting the current resource status of the multi-level resource pool into the resource usage prediction model, performing forward propagation calculation processing to obtain a resource usage prediction result, and screening the algorithms in the multi-layer algorithm library according to the resource usage prediction result to obtain a preliminary candidate algorithm set; performing resource demand quantification processing on the algorithms in the preliminary candidate algorithm set to obtain an algorithm resource demand vector; performing similarity calculation processing based on the algorithm resource demand vector and the resource usage prediction result to obtain an algorithm-resource matching matrix, and performing threshold filtering processing on the algorithm-resource matching matrix to obtain a final candidate algorithm set.
[0024] Specifically, we first start by reconstructing the time series of historical resource usage data in the multi-level resource pool, with the goal of obtaining a multi-dimensional resource usage matrix. This matrix not only contains the time dimension of resource usage, but also covers the usage of different types of resources. For example, in a smart city's traffic management system, this matrix may contain data in multiple dimensions such as CPU usage, memory usage, and network bandwidth utilization at various time points in the past few months. Next, based on this multi-dimensional resource usage matrix, the system constructs a long short-term memory (LSTM) network model. The LSTM network is a special recurrent neural network that can effectively capture long-term and short-term temporal dependencies. This feature of LSTM is particularly important in resource usage prediction, because resource usage patterns usually have both short-term fluctuations and long-term trends. For example, in the above traffic management system, the LSTM model can simultaneously learn the resource usage peak of the daily rush hour and the long-term resource demand growth trend brought about by urban development. Training the LSTM model is the next key step. During the training process, the model continuously adjusts its internal parameters to minimize the difference between the predicted value and the actual resource usage value. After training, the cloud-edge collaborative algorithm orchestration obtains a model that can accurately predict future resource usage. The predictive ability of this model is crucial for subsequent algorithm selection and resource allocation. With a trained resource usage prediction model, the system can use the current resource status of the multi-level resource pool as input and obtain the resource usage prediction result through the forward propagation calculation of the model. This prediction result may tell the cloud-edge collaborative algorithm orchestration that in a certain period of time in the future, the computing resources of some edge nodes may be in short supply, while the storage resources in the cloud may be in surplus. Based on these prediction results, the system will screen the algorithms in the multi-layer algorithm library to obtain a preliminary set of candidate algorithms. This screening process will consider the degree of match between the algorithm's resource requirements and the predicted resource availability. For example, if the prediction shows that the CPU usage of a certain edge node will reach a peak within the next hour, the system may tend to select some algorithms with lower computational complexity, or consider transferring computationally intensive tasks to the cloud with more abundant resources. In order to more accurately evaluate the degree of match between algorithms and resources, the system will quantify the resource requirements of each algorithm in the preliminary candidate algorithm set to obtain the algorithm resource requirement vector. This vector may contain the specific requirements of the algorithm in terms of CPU, memory, network bandwidth, etc. For example, a complex machine learning algorithm may require a lot of CPU and memory resources, while a data compression algorithm may rely more on I / O performance. Subsequently, the system will perform similarity calculations based on these algorithm resource requirement vectors and the previously obtained resource usage prediction results to generate an algorithm-resource matching matrix. Each element in this matrix represents the degree of matching between a specific algorithm and a specific resource node.The method for calculating similarity may use common metrics such as cosine similarity or Euclidean distance. For example, if the resource demand vector of an algorithm is highly similar to the predicted resource state vector of an edge node in the next hour, then the match between them will be high. Finally, the system will threshold filter this algorithm-resource matching matrix to obtain the final set of candidate algorithms. This process will set a matching threshold, and only algorithm-resource combinations with a matching degree exceeding this threshold will be retained in the final candidate set. The purpose of this is to ensure that the selected algorithm can not only operate under the current resource conditions, but also maintain good performance under the predicted future resource conditions.
[0025] 103. Decomposing the computing task according to the resource usage prediction result and the candidate algorithm set to obtain a subtask set, and performing algorithm matching on the subtask set to obtain a task-algorithm mapping relationship; In one embodiment of the present invention, the decomposition processing of the computing task according to the resource usage prediction result and the candidate algorithm set to obtain a subtask set, and the algorithm matching processing of the subtask set to obtain a task-algorithm mapping relationship includes: performing directed acyclic graph modeling processing on the computing task to obtain a task dependency graph, and performing node weight assignment processing on the task dependency graph according to the resource usage prediction result to obtain a weighted task dependency graph; modularizing the weighted task dependency graph to obtain a preliminary subtask set, and performing granularity adjustment processing on the preliminary subtask set according to the candidate algorithm set to obtain a final subtask set; performing resource demand analysis processing on each subtask in the final subtask set to obtain a subtask resource demand vector, and performing similarity calculation processing based on the subtask resource demand vector and the resource consumption characteristics of the candidate algorithm set to obtain a subtask-algorithm similarity matrix; matching and optimizing processing on the subtask-algorithm similarity matrix to obtain a preliminary task-algorithm mapping relationship, and performing scheduling order optimization processing based on the preliminary task-algorithm mapping relationship and the weighted task dependency graph to obtain a task-algorithm mapping relationship.
[0026] Specifically, we first start by modeling the computing task as a directed acyclic graph (DAG) to obtain a task dependency graph. For example, in a complex smart city data analysis task, the DAG may include multiple interdependent steps such as data collection, preprocessing, feature extraction, model training, and result visualization. This graphical representation not only clearly shows the internal structure of the task, but also provides a basis for subsequent task decomposition and parallel processing. Next, the system assigns node weights to the task dependency graph based on the previously obtained resource usage prediction results to obtain a weighted task dependency graph. The purpose of this step is to associate the predicted resource availability with the various components of the task. For example, if the prediction shows that the computing resources of the edge nodes are relatively sufficient within a certain period of time, some computing-intensive task nodes may be assigned higher weights, indicating that they are more suitable for execution at the edge. Subsequently, the system performs modular decomposition on this weighted task dependency graph to obtain a preliminary set of subtasks. The goal of modular decomposition is to decompose the original task into a series of relatively independent but interrelated subtasks. In this process, the system considers factors such as the dependency between task nodes, data flow, and resource requirements. For example, in the above-mentioned smart city data analysis task, data collection and preprocessing may be decomposed into one subtask, while model training may be decomposed into another subtask. After obtaining the preliminary subtask set, the system will adjust the granularity of these subtasks according to the candidate algorithm set to obtain the final subtask set. The purpose of this step is to ensure that the granularity of the subtask matches the available algorithms and resources. For example, if the candidate algorithm set contains some efficient parallel processing algorithms, the system may further subdivide some large subtasks to fully utilize the advantages of these algorithms. Next, the system performs resource requirement analysis on each subtask in the final subtask set to obtain a subtask resource requirement vector. This vector may include the specific requirements of the subtask in terms of CPU, memory, network bandwidth, etc. For example, a subtask involving large-scale data processing may require a large amount of memory and I / O resources, while a subtask of complex calculations may rely more on CPU performance. Then, the system performs similarity calculation based on these subtask resource requirement vectors and the resource consumption characteristics of the candidate algorithm set to obtain a subtask-algorithm similarity matrix. Each element in this matrix represents the degree of match between a specific subtask and a specific algorithm. The method for calculating similarity may use common metrics such as cosine similarity or Euclidean distance. For example, if the resource requirement vector of a subtask is highly similar to the resource consumption characteristics of an algorithm, then the match between them will be very high. Next, the system performs matching optimization on the subtask-algorithm similarity matrix to obtain a preliminary task-algorithm mapping relationship. This process may use optimization methods such as the Hungarian algorithm to find the globally optimal matching solution between subtasks and algorithms.For example, in a complex image processing task, the system may match the edge detection subtask to an efficient edge computing algorithm, and the image classification subtask to a deep learning algorithm running in the cloud. Finally, the system performs scheduling order optimization based on this preliminary task-algorithm mapping relationship and the previously obtained weighted task dependency graph to obtain the final task-algorithm mapping relationship. The purpose of this step is to maximize the possibility of parallel processing while ensuring that the task dependencies are met, thereby improving the overall processing efficiency. For example, the system may schedule some independent subtasks to execute in parallel, while executing some subtasks with strict dependencies in a specific order. Through this series of carefully designed steps, the cloud-edge collaborative system can effectively decompose complex computing tasks into subtasks suitable for distributed processing and find the most suitable execution algorithm and resources for each subtask. This approach not only takes into account the internal structure and dependencies of the task, but also makes full use of the resource usage prediction results and the characteristics of the available algorithms, so that it can better adapt to the dynamically changing computing environment. For example, in a large-scale intelligent traffic management system, a complex traffic flow optimization task may be decomposed into multiple subtasks, including real-time data collection, traffic flow prediction, signal light control, etc. The system may assign subtasks of data collection and preliminary processing to edge nodes to take advantage of the low latency of edge computing; assign complex traffic flow prediction model training tasks to the cloud to take advantage of the powerful computing power of cloud computing; and assign subtasks that require real-time response, such as traffic light control, to local traffic controllers. This fine-grained task decomposition and precise algorithm matching can not only make full use of the distributed resources of the cloud and edge, but also significantly improve the response speed and processing efficiency of the entire system, providing strong technical support for intelligent traffic management.
[0027] 104. Construct a load evolution model and a delay model according to the task-algorithm mapping relationship and the system load status, and optimize the task-algorithm mapping relationship according to the load evolution model and the delay model to obtain an initial algorithm scheduling strategy; In one embodiment of the present invention, a load evolution model and a delay model are constructed according to the task-algorithm mapping relationship and the system load condition, and the task-algorithm mapping relationship is optimized according to the load evolution model and the delay model to obtain an initial algorithm orchestration strategy, including: performing time series analysis on the system load condition to obtain a load change trend function, and generating a load evolution model according to the load change trend function and the task-algorithm mapping relationship; performing graph theory modeling on the cloud-edge network topology structure to obtain a network topology diagram, and performing communication delay calculation processing according to the network topology diagram and the task-algorithm mapping relationship to obtain a communication delay matrix; performing queue theory analysis processing on the load evolution model to obtain a calculation delay function, and combining the communication delay matrix and the calculation delay function to obtain an end-to-end delay model; constructing a multi-objective optimization problem according to the load evolution model and the end-to-end delay model, solving the multi-objective optimization problem to obtain a Pareto optimal solution set, and performing decision analysis processing on the Pareto optimal solution set to obtain an initial algorithm orchestration strategy.
[0028] Specifically, we first start with a time series analysis of the system load status, with the goal of obtaining a trend function that can describe the dynamic changes in the system load. For example, in a smart city traffic management system, this load change trend function may reflect the characteristics of a sharp increase in load during peak hours in the morning and evening on weekdays and a relatively flat load on weekends. Next, the system generates a load evolution model based on this load change trend function and the previously obtained task-algorithm mapping relationship. This model not only takes into account the load change trend of the system itself, but also takes into account the impact of task execution on the system load. For example, if a large-scale data analysis task is mapped to a specific edge node, the load of this node may increase significantly during task execution. The load evolution model can predict this dynamic change and provide an important basis for subsequent optimization decisions. At the same time, the system performs graph modeling on the cloud-edge network topology to obtain a network topology diagram. This diagram may include the connection relationship and network characteristics between cloud data centers, edge servers, and terminal devices. Based on this network topology diagram and the task-algorithm mapping relationship, the system performs communication delay calculation processing to obtain a communication delay matrix. This matrix describes the time required to transmit data between different nodes. For example, in a distributed machine learning task, each step from collecting data from the terminal device to the edge node for preliminary processing and then to the cloud for model training will generate corresponding communication delays. Subsequently, the system performs queue theory analysis on the load evolution model to obtain the computational delay function. This function describes the processing time of tasks on each computing node under different load conditions. For example, when the load of an edge node is close to saturation, newly arrived tasks may need to wait in the queue for a long time before being processed. By combining the communication delay matrix and the computational delay function, the system obtains an end-to-end delay model. This model comprehensively describes the delay of the entire process from submission to completion of the task, including data transmission time and actual calculation time. For example, in a real-time video analysis system, the end-to-end delay model can predict the time required for the entire process from capturing the image from the camera, transmitting it through the network, performing preliminary processing at the edge node, and then performing in-depth analysis in the cloud. With the load evolution model and the end-to-end delay model, the system can construct a multi-objective optimization problem. This optimization problem usually considers multiple objectives, such as minimizing overall delay, balancing system load, maximizing resource utilization, etc. For example, in a production line monitoring system of a smart factory, the optimization objectives may include ensuring the real-time performance of key monitoring tasks while maximizing the overall data processing throughput. By solving this multi-objective optimization problem, the system obtains a Pareto optimal solution set. Each solution in the Pareto optimal solution set represents a different trade-off between multiple objectives.For example, one solution may reduce latency to a great extent, but may result in a high load on some nodes; while another solution may achieve better load balance, but slightly increase the overall latency. Finally, the system performs decision analysis on this Pareto optimal solution set to obtain an initial algorithm orchestration strategy. This process may take into account some additional factors, such as the system's quality of service requirements, energy efficiency, etc. For example, in an IoT application that needs to process a large amount of real-time data, the system may choose a solution that strikes a good balance between latency and load balancing, while ensuring that the load on key nodes does not exceed a preset threshold. Through this series of carefully designed steps, the cloud-edge collaborative system is able to find an initial, approximately optimal algorithm orchestration strategy for complex distributed tasks based on a comprehensive load and latency model. This approach not only takes into account the dynamic characteristics of the system and the network topology, but also seeks a balance between multiple performance objectives, so that it can better adapt to complex and changing computing environments.
[0029] 105. Perform deep reinforcement learning training and heuristic optimization processing on the initial algorithm orchestration strategy to obtain an adaptive algorithm orchestration optimization strategy; In one embodiment of the present invention, the initial algorithm orchestration strategy is subjected to deep reinforcement learning training and heuristic optimization processing to obtain an adaptive algorithm orchestration optimization strategy, including: constructing a Markov decision process model according to the initial algorithm orchestration strategy, defining the state space and action space of the Markov decision process model to obtain a reinforcement learning environment; designing a reward function for the reinforcement learning environment to obtain a performance evaluation index, and constructing a dual deep Q network model according to the reinforcement learning environment and the performance evaluation index; initializing the parameters of the dual deep Q network model to obtain an initial strategy network and a target network, and The initial strategy network and the target network are alternately trained to obtain an optimized deep Q network model; the initial algorithm orchestration strategy is subjected to strategy extraction processing according to the optimized deep Q network model to obtain a deep reinforcement learning optimization strategy, and the deep reinforcement learning optimization strategy is encoded to obtain a strategy chromosome population; the strategy chromosome population is subjected to genetic algorithm iteration processing to obtain an optimized strategy chromosome, and the optimized strategy chromosome is decoded to obtain a heuristic optimization strategy; the deep reinforcement learning optimization strategy and the heuristic optimization strategy are integrated to obtain an adaptive algorithm orchestration optimization strategy.
[0030] Specifically, we first start by building a Markov decision process model based on the initial algorithm orchestration strategy, which can capture the randomness of system state transitions and the long-term impact of decisions. For example, in an intelligent traffic management system, the Markov decision process can describe the dynamic changes in traffic flow and system load under different traffic control strategies. Next, the system defines the state space and action space of this Markov decision process model to obtain a complete reinforcement learning environment. The state space may include information such as the current system load, task queue length, network status, etc., while the action space may include task scheduling decisions, resource allocation strategies, etc. For example, in a distributed computing task, the state may include the current computing load and network bandwidth usage of each node, while the action may include the decision to assign a specific task to a specific node. Subsequently, the system designs a reward function for this reinforcement learning environment to obtain performance evaluation indicators. The design of the reward function directly affects the optimization direction of the reinforcement learning algorithm, and usually considers multiple factors, such as task completion time, resource utilization, load balancing, etc. For example, in a real-time data processing system, the reward function may seek a balance between reducing processing delays and maintaining system stability. Based on this reinforcement learning environment and performance evaluation indicators, the system built a double deep Q network (Double DQN) model. Double DQN is an advanced variant of deep reinforcement learning. It uses two networks (one for action selection and the other for value evaluation) to reduce the overestimation problem of Q values, thereby improving the stability and efficiency of learning. Then, the system initializes the parameters of this double deep Q network model to obtain the initial policy network and the target network. Initialization may use some pre-trained weights or random initialization methods. Then, the system alternately trains the two networks. During this process, the policy network is continuously updated to optimize the decision, while the target network is relatively fixed to provide a stable learning goal. Through a lot of interaction and learning, the system finally obtains an optimized deep Q network model. Based on this optimized model, the system extracts the policy of the initial algorithm orchestration strategy to obtain a deep reinforcement learning optimization strategy. This strategy can make intelligent scheduling decisions based on the current system state. For example, in a complex cloud-edge collaborative computing environment, this strategy may dynamically decide to allocate specific types of computing tasks to the cloud, edge nodes, or local devices based on real-time load conditions and network status. To further optimize this strategy, the system encodes the deep reinforcement learning optimization strategy and converts it into a set of chromosomes that can be operated by the genetic algorithm. These chromosomes constitute the initial strategy chromosome population. Then, the system iterates the strategy chromosome population with the genetic algorithm. In this process, the system simulates the biological evolution process, continuously generates new strategy variants and evaluates their performance through selection, crossover and mutation operations.After multiple rounds of iterations, the system obtains a set of optimized policy chromosomes. Next, the system decodes these optimized policy chromosomes and converts them back into executable scheduling policies to obtain heuristic optimization policies. This heuristic optimization policy may perform well in certain specific scenarios, such as when dealing with highly structured or periodic tasks. Finally, the system integrates the deep reinforcement learning optimization policy and the heuristic optimization policy to obtain the final adaptive algorithm orchestration optimization policy. This integrated strategy combines the adaptability of deep reinforcement learning and the efficiency of heuristic methods, and can make intelligent scheduling decisions in various complex scenarios.
[0031] 106. Schedule and execute the algorithm on the cloud edge according to the adaptive algorithm orchestration optimization strategy, and dynamically adjust and iteratively optimize the adaptive algorithm orchestration optimization strategy according to the execution result.
[0032] In one embodiment of the present invention, the algorithm scheduling and execution on the cloud edge according to the adaptive algorithm orchestration optimization strategy, and the dynamic adjustment and iterative optimization processing of the adaptive algorithm orchestration optimization strategy according to the execution result include: dividing the adaptive algorithm orchestration optimization strategy into time windows to obtain a rolling optimization time series, and performing segmented execution processing on the adaptive algorithm orchestration optimization strategy according to the rolling optimization time series to obtain a phased execution result; performing multi-dimensional performance indicator extraction processing on the phased execution result to obtain an execution effect evaluation matrix, and performing deviation analysis processing on the adaptive algorithm orchestration optimization strategy according to the execution effect evaluation matrix to obtain a strategy adjustment direction vector; performing gradient descent optimization processing on the strategy adjustment direction vector to obtain a strategy update increment, and fusing the strategy update increment with the adaptive algorithm orchestration optimization strategy to obtain and execute the updated adaptive algorithm orchestration optimization strategy.
[0033] Specifically, we first start by dividing the adaptive algorithm orchestration optimization strategy into time windows, with the goal of obtaining a rolling optimization time series. For example, in a smart city's traffic management system, the cloud-edge collaborative algorithm orchestration may divide a day into multiple time windows, such as morning peak, off-peak, evening peak, etc., each of which may last 2-3 hours. This division allows the system to adopt different optimization strategies in different time periods to adapt to the dynamic changes in traffic flow. Next, the system will segment the adaptive algorithm orchestration optimization strategy according to this rolling optimization time series to obtain phased execution results. In the example of the above traffic management system, this means that the system will execute a specific optimization strategy in each time window. For example, during the morning peak period, the system may prioritize the tasks of real-time traffic flow analysis and signal light control, while during the off-peak period, more resources may be allocated to tasks such as long-term traffic pattern analysis or road condition prediction. Subsequently, the system will extract multi-dimensional performance indicators from these phased execution results to obtain an execution effect evaluation matrix. This matrix may contain multiple performance indicators, such as task completion time, resource utilization, load balancing, energy efficiency, etc. For example, in a traffic management system, cloud-edge collaborative algorithm orchestration may focus on indicators such as average vehicle travel time, congestion level at intersections, and traffic accident response time. Based on this execution effect evaluation matrix, the system will perform deviation analysis on the adaptive algorithm orchestration optimization strategy to obtain a strategy adjustment direction vector. This process involves comparing the difference between the actual execution effect and the expected target and determining the direction that needs to be adjusted. For example, if the system finds that the task processing delay within a certain time window is significantly higher than expected, the adjustment direction may tend to increase the computing resource allocation during that period or improve the task scheduling algorithm. Next, the system will perform gradient descent optimization on this strategy adjustment direction vector to obtain a strategy update increment. Gradient descent is a commonly used optimization algorithm that can find the direction that minimizes the objective function in a high-dimensional space. In the scenario of cloud-edge collaborative algorithm orchestration, the objective function may be a weighted sum that comprehensively considers multiple performance indicators. Through gradient descent, the system can find an optimal strategy adjustment plan to maximize the overall performance. Finally, the system merges this strategy update increment with the original adaptive algorithm orchestration optimization strategy to obtain and execute the updated adaptive algorithm orchestration optimization strategy. This fusion process needs to be handled carefully to ensure that the new strategy can transition smoothly without causing sudden fluctuations in system performance. For example, in a traffic management system, if the new strategy recommends a significant increase in the green time of a signal light on a certain road section, the system may choose to adjust gradually rather than make a drastic change all at once to avoid sudden negative impacts on surrounding roads. Through this continuous dynamic adjustment and iterative optimization process, the cloud-edge collaborative system can continuously adapt to changes in the environment and improve its responsiveness to complex and changing computing needs.For example, in a large smart city management platform, this approach can help the system effectively respond to various routine and emergency situations. In daily operations, the system can dynamically adjust the resource allocation strategy according to the characteristics of different time periods, such as increasing the allocation of computing resources to the office area during office hours on weekdays, while paying more attention to the needs of the entertainment area on weekends. When faced with emergencies, such as large-scale events or natural disasters, the system can quickly adjust its strategy to allocate more resources to key tasks. For example, if the city is hosting a large-scale sports event, the system may prioritize tasks related to traffic control, crowd monitoring, and security during the event, while temporarily reducing the priority of some non-emergency tasks. By monitoring the execution effect in real time and constantly adjusting the strategy, the system can ensure that it remains efficient during special periods. In addition, this approach can also help the system gradually optimize its long-term performance. By analyzing the execution effect evaluation matrix over a long period of time, the system can identify certain recurring performance bottlenecks or resource waste points, so as to formulate more targeted optimization strategies. For example, if the system finds that some edge nodes are always in a low-load state during a specific period of time, it may adjust the resource allocation strategy to use part of the computing power of these nodes to process some non-real-time but computationally intensive background tasks, such as data analysis or model training, thereby improving overall resource utilization.
[0034] In this embodiment, a multi-level resource pool and algorithm library are constructed by structured modeling of multi-layer computing resources on the cloud edge. Then, dynamic matching of algorithms and subtasks is achieved through resource usage prediction and task decomposition. Next, based on the load evolution and delay model, the task-algorithm mapping relationship is optimized to form an initial orchestration strategy. Finally, deep reinforcement learning and heuristic optimization are combined to generate an adaptive algorithm orchestration optimization strategy, and dynamic adjustment and iterative optimization are performed through real-time execution results. This method can perceive and adapt to complex and changeable computing environments in real time by combining technologies such as multi-level resource modeling, dynamic resource prediction, adaptive task decomposition, and deep reinforcement learning, which significantly improves the utilization rate of multi-level computing resources on the cloud edge, effectively balances the system load, and reduces task processing delays.
[0035] The above describes the cloud-edge collaborative algorithm arrangement method in the embodiment of the present invention. The following describes the cloud-edge collaborative algorithm arrangement device in the embodiment of the present invention. Figure 2 , an embodiment of a cloud-edge collaborative algorithm orchestration device in an embodiment of the present invention includes: The modeling and mapping module 201 is used to perform structured modeling processing on the multi-layer computing resources of the cloud edge to obtain a multi-layer resource pool, and classify and map the algorithms corresponding to the multi-layer resource pool to obtain a multi-layer algorithm library; The algorithm screening module 202 is used to perform prediction processing on the resource usage data in the multi-level resource pool to obtain a resource usage prediction result, and select an applicable algorithm from the multi-level algorithm library according to the prediction result to obtain a candidate algorithm set; The task decomposition module 203 is used to decompose the computing task according to the resource usage prediction result and the candidate algorithm set to obtain a subtask set, and perform algorithm matching on the subtask set to obtain a task-algorithm mapping relationship; An initial scheduling module 204 is used to construct a load evolution model and a delay model according to the task-algorithm mapping relationship and the system load status, and optimize the task-algorithm mapping relationship according to the load evolution model and the delay model to obtain an initial algorithm scheduling strategy; The adaptive orchestration module 205 is used to perform deep reinforcement learning training and heuristic optimization processing on the initial algorithm orchestration strategy to obtain an adaptive algorithm orchestration optimization strategy; The scheduling module 206 is used to schedule and execute the algorithm on the cloud edge according to the adaptive algorithm orchestration optimization strategy, and dynamically adjust and iteratively optimize the adaptive algorithm orchestration optimization strategy according to the execution result.
[0036] In an embodiment of the present invention, the cloud-edge collaborative algorithm orchestration device runs the above-mentioned cloud-edge collaborative algorithm orchestration method, and the cloud-edge collaborative algorithm orchestration device constructs a multi-level resource pool and algorithm library by structured modeling of multi-layer computing resources on the cloud-edge side. Then, dynamic matching of algorithms and subtasks is achieved through resource usage prediction and task decomposition. Next, based on the load evolution and delay model, the task-algorithm mapping relationship is optimized to form an initial orchestration strategy. Finally, combining deep reinforcement learning and heuristic optimization, an adaptive algorithm orchestration optimization strategy is generated, and dynamic adjustment and iterative optimization are performed through real-time execution results. This method can perceive and adapt to complex and changeable computing environments in real time by combining technologies such as multi-level resource modeling, dynamic resource prediction, adaptive task decomposition, and deep reinforcement learning, which significantly improves the utilization rate of multi-level computing resources on the cloud-edge side, effectively balances the system load, and reduces task processing delays. .
[0037] above Figure 2 The cloud-edge collaborative algorithm orchestration device in the embodiment of the present invention is described in detail from the perspective of modular functional entities. The cloud-edge collaborative algorithm orchestration device in the embodiment of the present invention is described in detail from the perspective of hardware processing.
[0038] Figure 3It is a structural schematic diagram of a cloud-edge collaborative algorithm orchestration device provided by an embodiment of the present invention. The cloud-edge collaborative algorithm orchestration device 300 may have relatively large differences due to different configurations or performances, and may include one or more processors (central processing units, CPU) 310 (for example, one or more processors) and a memory 320, and one or more storage media 330 (for example, one or more massive storage device terminals) storing application programs 333 or data 332. Among them, the memory 320 and the storage medium 330 can be short-term storage or persistent storage. The program stored in the storage medium 330 may include one or more modules (not shown in the figure), and each module may include a series of instruction operations in the cloud-edge collaborative algorithm orchestration device 300. Furthermore, the processor 310 may be configured to communicate with the storage medium 330, and execute a series of instruction operations in the storage medium 330 on the cloud-edge collaborative algorithm orchestration device 300 to implement the steps of the above-mentioned cloud-edge collaborative algorithm orchestration method.
[0039] The cloud-edge collaborative algorithm orchestration device 300 may also include one or more power supplies 340, one or more wired or wireless network interfaces 350, one or more input and output interfaces 360, and / or one or more operating systems 331, such as Windows Serve, Mac OS X, Unix, Linux, FreeBSD, etc. Those skilled in the art will appreciate that Figure 3 The cloud-edge collaborative algorithm orchestration device structure shown does not constitute a limitation on the cloud-edge collaborative algorithm orchestration device provided by the present invention, and may include more or fewer components than shown in the figure, or a combination of certain components, or a different arrangement of components.
[0040] The present invention also provides a computer-readable storage medium, which may be a non-volatile computer-readable storage medium or a volatile computer-readable storage medium. Instructions are stored in the computer-readable storage medium. When the instructions are executed on a computer, the computer executes the steps of the cloud-edge collaborative algorithm orchestration method.
[0041] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working process of the above-described system, device, or unit can refer to the corresponding process in the aforementioned method embodiment and will not be repeated here.
[0042] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention is essentially or the part that contributes to the prior art or the whole or part of the technical solution can be embodied in the form of a software product. The computer software product is stored in a storage medium, including several instructions to enable a computer device (which can be a personal computer, server, or network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM), random access memory (RAM), disk or optical disk and other media that can store program code.
[0043] As described above, the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that the technical solutions described in the aforementioned embodiments may still be modified, or some of the technical features thereof may be replaced by equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.
Claims
1. A cloud-edge collaborative algorithm orchestration method, characterized in that: The cloud-edge collaborative algorithm orchestration method includes: Perform structured modeling on the multi-layer computing resources at the cloud edge to obtain a multi-layer resource pool, and perform classification and mapping on the algorithms corresponding to the multi-layer resource pool to obtain a multi-layer algorithm library; Performing prediction processing on the resource usage data in the multi-level resource pool to obtain a resource usage prediction result, and selecting an applicable algorithm from the multi-level algorithm library according to the prediction result to obtain a candidate algorithm set; Decomposing the computing task according to the resource usage prediction result and the candidate algorithm set to obtain a subtask set, and performing algorithm matching on the subtask set to obtain a task-algorithm mapping relationship; According to the task-algorithm mapping relationship and the system load condition, a load evolution model and a delay model are constructed, and the task-algorithm mapping relationship is optimized according to the load evolution model and the delay model to obtain an initial algorithm scheduling strategy; Performing deep reinforcement learning training and heuristic optimization processing on the initial algorithm orchestration strategy to obtain an adaptive algorithm orchestration optimization strategy; The algorithm is scheduled and executed on the cloud edge according to the adaptive algorithm orchestration optimization strategy, and the adaptive algorithm orchestration optimization strategy is dynamically adjusted and iteratively optimized according to the execution result.
2. The cloud-edge collaborative algorithm orchestration method according to claim 1, characterized in that: The structured modeling process is performed on the cloud-edge multi-layer computing resources to obtain a multi-layer resource pool, and the algorithms corresponding to the multi-layer resource pool are classified and mapped to obtain a multi-layer algorithm library including: Perform K-means clustering on cloud computing resources, edge computing resources, and terminal computing resources to obtain multiple computing resource areas; Perform red-black tree construction processing on cloud computing resources, edge computing resources, and terminal computing resources according to the computing resource area to obtain a small resource tree; Performing B+ tree construction processing on the small resource tree to obtain a large resource tree, and combining the small resource tree with the large resource tree to obtain a multi-level resource pool; Classify cloud computing algorithms, edge computing algorithms, and terminal computing algorithms to obtain an algorithm category set; According to the resource types in the multi-level resource pool and the algorithm category set, mapping and associating the cloud computing algorithm, the edge computing algorithm and the terminal computing algorithm are processed to obtain an algorithm-resource mapping table; According to the algorithm-resource mapping table, the cloud computing algorithm, the edge computing algorithm and the terminal computing algorithm are hierarchically organized and processed to obtain a multi-layer algorithm library.
3. The cloud-edge collaborative algorithm orchestration method according to claim 1, characterized in that: The resource usage data in the multi-level resource pool is predicted and processed to obtain a resource usage prediction result, and a suitable algorithm is selected from the multi-level algorithm library according to the prediction result to obtain a candidate algorithm set including: Performing time series reconstruction processing on the historical resource usage data in the multi-level resource pool to obtain a multi-dimensional resource usage matrix, constructing a long short-term memory network model based on the multi-dimensional resource usage matrix, and training the long short-term memory network model to obtain a resource usage prediction model; Inputting the current resource status of the multi-level resource pool into the resource usage prediction model, performing forward propagation calculation processing to obtain a resource usage prediction result, and screening the algorithms in the multi-level algorithm library according to the resource usage prediction result to obtain a preliminary candidate algorithm set; Quantifying resource requirements of algorithms in the preliminary candidate algorithm set to obtain algorithm resource requirement vectors; A similarity calculation process is performed based on the algorithm resource demand vector and the resource usage prediction result to obtain an algorithm-resource matching matrix, and a threshold filtering process is performed on the algorithm-resource matching matrix to obtain a final candidate algorithm set.
4. The cloud-edge collaborative algorithm orchestration method according to claim 1, characterized in that: The step of decomposing the computing task according to the resource usage prediction result and the candidate algorithm set to obtain a subtask set, and performing algorithm matching on the subtask set to obtain a task-algorithm mapping relationship includes: Performing directed acyclic graph modeling processing on the computing tasks to obtain a task dependency graph, and performing node weight allocation processing on the task dependency graph according to the resource usage prediction result to obtain a weighted task dependency graph; Performing modular decomposition processing on the weighted task dependency graph to obtain a preliminary subtask set, and performing granularity adjustment processing on the preliminary subtask set according to the candidate algorithm set to obtain a final subtask set; Performing resource requirement analysis on each subtask in the final subtask set to obtain a subtask resource requirement vector, and performing similarity calculation based on the subtask resource requirement vector and the resource consumption characteristics of the candidate algorithm set to obtain a subtask-algorithm similarity matrix; The subtask-algorithm similarity matrix is matched and optimized to obtain a preliminary task-algorithm mapping relationship, and the scheduling order is optimized based on the preliminary task-algorithm mapping relationship and the weighted task dependency graph to obtain a task-algorithm mapping relationship.
5. The cloud-edge collaborative algorithm orchestration method according to claim 1, characterized in that: The step of constructing a load evolution model and a delay model according to the task-algorithm mapping relationship and the system load condition, and optimizing the task-algorithm mapping relationship according to the load evolution model and the delay model to obtain an initial algorithm scheduling strategy includes: Performing time series analysis on the system load status to obtain a load change trend function, and generating a load evolution model according to the load change trend function and the task-algorithm mapping relationship; Performing graph theory modeling on the cloud-edge network topology structure to obtain a network topology graph, and performing communication delay calculation processing based on the network topology graph and the task-algorithm mapping relationship to obtain a communication delay matrix; Performing queue theory analysis on the load evolution model to obtain a calculation delay function, and combining the communication delay matrix and the calculation delay function to obtain an end-to-end delay model; A multi-objective optimization problem is constructed according to the load evolution model and the end-to-end delay model, the multi-objective optimization problem is solved to obtain a Pareto optimal solution set, and a decision analysis is performed on the Pareto optimal solution set to obtain an initial algorithm orchestration strategy.
6. The cloud-edge collaborative algorithm orchestration method according to claim 1, characterized in that: The performing of deep reinforcement learning training and heuristic optimization processing on the initial algorithm orchestration strategy to obtain an adaptive algorithm orchestration optimization strategy includes: Constructing a Markov decision process model according to the initial algorithm arrangement strategy, performing state space and action space definition processing on the Markov decision process model to obtain a reinforcement learning environment; Performing reward function design processing on the reinforcement learning environment to obtain a performance evaluation index, and constructing a dual deep Q network model according to the reinforcement learning environment and the performance evaluation index; Performing parameter initialization processing on the dual deep Q network model to obtain an initial policy network and a target network, and performing alternating training processing on the initial policy network and the target network to obtain an optimized deep Q network model; Performing strategy extraction processing on the initial algorithm orchestration strategy according to the optimized deep Q network model to obtain a deep reinforcement learning optimization strategy, and encoding processing on the deep reinforcement learning optimization strategy to obtain a strategy chromosome population; Performing genetic algorithm iteration processing on the strategy chromosome population to obtain an optimized strategy chromosome, and performing decoding processing on the optimized strategy chromosome to obtain a heuristic optimization strategy; The deep reinforcement learning optimization strategy and the heuristic optimization strategy are integrated to obtain an adaptive algorithm orchestration optimization strategy.
7. The cloud-edge collaborative algorithm orchestration method according to claim 1, characterized in that: The algorithm scheduling and executing on the cloud edge according to the adaptive algorithm orchestration optimization strategy, and dynamically adjusting and iteratively optimizing the adaptive algorithm orchestration optimization strategy according to the execution result includes: The adaptive algorithm arrangement optimization strategy is divided into time windows to obtain a rolling optimization time series, and the adaptive algorithm arrangement optimization strategy is executed in sections according to the rolling optimization time series to obtain a phased execution result; Perform multi-dimensional performance index extraction processing on the phased execution results to obtain an execution effect evaluation matrix, and perform deviation analysis processing on the adaptive algorithm orchestration optimization strategy according to the execution effect evaluation matrix to obtain a strategy adjustment direction vector; The strategy adjustment direction vector is subjected to gradient descent optimization processing to obtain a strategy update increment, and the strategy update increment is fused with the adaptive algorithm orchestration optimization strategy to obtain and execute an updated adaptive algorithm orchestration optimization strategy.
8. A cloud-edge collaborative algorithm orchestration device, characterized in that: The cloud-edge collaborative algorithm orchestration device includes: A modeling and mapping module is used to perform structured modeling processing on the multi-layer computing resources of the cloud edge to obtain a multi-layer resource pool, and to classify and map the algorithms corresponding to the multi-layer resource pool to obtain a multi-layer algorithm library; An algorithm screening module is used to perform prediction processing on the resource usage data in the multi-level resource pool to obtain a resource usage prediction result, and select an applicable algorithm from the multi-level algorithm library according to the prediction result to obtain a candidate algorithm set; A task decomposition module is used to decompose the computing task according to the resource usage prediction result and the candidate algorithm set to obtain a subtask set, and perform algorithm matching on the subtask set to obtain a task-algorithm mapping relationship; An initial scheduling module, used to construct a load evolution model and a delay model according to the task-algorithm mapping relationship and the system load status, and optimize the task-algorithm mapping relationship according to the load evolution model and the delay model to obtain an initial algorithm scheduling strategy; An adaptive orchestration module, used to perform deep reinforcement learning training and heuristic optimization processing on the initial algorithm orchestration strategy to obtain an adaptive algorithm orchestration optimization strategy; The scheduling module is used to schedule and execute the algorithm on the cloud edge according to the adaptive algorithm orchestration optimization strategy, and dynamically adjust and iteratively optimize the adaptive algorithm orchestration optimization strategy according to the execution results.
9. A cloud-edge collaborative algorithm orchestration device, characterized in that: The cloud-edge collaborative algorithm orchestration device comprises: a memory and at least one processor, wherein the memory stores instructions; The at least one processor calls the instructions in the memory so that the cloud-edge collaborative algorithm orchestration device performs the steps of the cloud-edge collaborative algorithm orchestration method as described in any one of claims 1-7.
10. A computer-readable storage medium having instructions stored thereon, characterized in that: When the instructions are executed by the processor, the steps of the cloud-edge collaborative algorithm orchestration method as described in any one of claims 1-7 are implemented.
Citation Information
Cited By
Task migration method and device and storage medium
CN119271421A
Resource transfer method, system and device based on cloud edge collaboration and storage medium
CN119440856A
Intelligent fault early warning system of ship environment monitoring equipment
CN119596805A
Heterogeneous end-side cloud cooperative computing method and system based on elastic coupling mechanism
CN119652891A
Intelligent traffic signal edge computing resource dynamic allocation method and system
CN119902902A
