Edge cloud integrated scheduling system based on cooperation of global optimizer and local optimizer

Abstract

The invention relates to an edge cloud integrated scheduling system based on cooperation of a global optimizer and a local optimizer, and the system comprises a prediction module which is used for predicting the load trend, the network state and the service request amount of each edge computing node based on the current monitoring data and the historical monitoring data of the equipment of each edge computing node; the global optimizer is used for performing optimization based on the prediction result of the prediction module by taking the system total cost minimization as a target to obtain a global task scheduling scheme, and sending the global task scheduling scheme to each edge computing node; each local optimizer is used for carrying out validity verification and local adjustment on the global task scheduling scheme according to own real-time state data, and sending a task migration request to the global optimizer when the global task scheduling scheme is not feasible; and the global optimizer is also used for regenerating the scheduling scheme based on the new system state. The scheduling scheme can be dynamically adjusted according to the system state and demand, and the total cost of the system is effectively reduced.

Description

Technical Field

[0001] This invention relates to the field of task scheduling technology, and in particular to an edge-cloud integrated scheduling system based on the collaboration of a global optimizer and a local optimizer. Background Technology

[0002] With the development of IoT technology, the number of IoT devices continues to grow. Traditional IoT architectures primarily rely on cloud computing centers for data processing and decision support. This model has greatly promoted data integration and analysis capabilities in the IoT field, making centralized processing of large-scale data possible and providing strong support for the digital transformation of numerous industries. In industrial production, cloud computing centers can perform real-time analysis of large amounts of data on the production line, helping companies optimize production processes and improve production efficiency. In the field of intelligent transportation, cloud computing centers can integrate traffic flow data to achieve intelligent control of traffic signals and alleviate urban traffic congestion.

[0003] When facing the challenges of IoT data processing and decision support, existing technologies typically employ the following approaches. One approach relies entirely on cloud computing centers, uploading all data generated by IoT devices to these centers for centralized processing and decision-making. This method fully leverages the powerful computing resources and storage capabilities of cloud computing centers to implement complex data analysis and decision-making algorithms. Another approach utilizes edge computing technology, processing data at the network edge to reduce the amount of data transmitted to the cloud and improve the real-time performance of data processing. Edge computing enables preliminary processing and analysis of data on the device side, uploading only necessary data to the cloud, thus reducing network bandwidth pressure. Furthermore, some solutions attempt to simply combine edge computing and cloud computing, but these often only involve simple data transmission and distribution between the two, lacking an effective collaborative mechanism.

[0004] However, existing technologies have significant drawbacks. Traditional reliance on cloud computing centers performs poorly in scenarios with high real-time requirements, unstable networks, or sensitive data, because data transmission to the cloud and back to the device incurs substantial latency, failing to meet real-time response needs. Furthermore, current edge computing and cloud computing are often deployed independently, lacking unified scheduling and intelligent collaboration mechanisms, resulting in complex system architectures, low resource utilization, and high maintenance costs, making it difficult to meet dynamically changing business demands. Summary of the Invention

[0005] The technical problem to be solved by the present invention is to provide an edge-cloud integrated scheduling system based on the collaboration of global optimizer and local optimizer, aiming to solve at least one of the above-mentioned technical problems.

[0006] The technical solution of the present invention to solve the above-mentioned technical problems is as follows: This application provides an edge-cloud integrated scheduling system based on the collaboration of a global optimizer and a local optimizer, employing the following technical solution: An edge-cloud integrated scheduling system based on the collaboration of a global optimizer and a local optimizer includes a cloud platform and various edge computing nodes. The cloud platform is deployed with a global optimizer and a prediction module, and each edge computing node is deployed with a local optimizer. The prediction module is used to predict the load trend, network status and service request volume of each edge computing node based on the current monitoring data and historical monitoring data of each edge computing node's own device. The current monitoring data includes resource status data, network status data, task status data and service priority. The global optimizer is used to optimize based on the prediction results of the prediction module with the goal of minimizing the total system cost, to obtain a global task scheduling scheme, and to send the global task scheduling scheme to each edge computing node. The total cost includes the total system energy consumption, average task latency, and bandwidth cost. Each of the local optimizers is used to verify the effectiveness of the global task scheduling scheme and make local adjustments based on its own real-time status data. When the global task scheduling scheme is not feasible or the local node resources are insufficient, a task migration request is sent to the global optimizer. The global optimizer is also used to regenerate the scheduling scheme based on the latest system state in response to the task migration request.

[0007] The beneficial effects of this invention are as follows: By constructing a two-layer scheduling architecture with a global optimizer and a local optimizer working together, dynamic balance and precise control of multi-objective optimization in the edge-cloud system are achieved. The prediction module uses current and historical monitoring data of edge computing nodes to make predictions, enabling advance understanding of the load, network status, and service request volume of edge computing nodes. The global optimizer optimizes based on the prediction results with the goal of minimizing the total system cost, resulting in a global task scheduling scheme that balances total energy consumption, average task latency, and bandwidth costs. The local optimizer verifies the effectiveness of the global scheme and makes local adjustments, requesting task migration when the scheme is infeasible or resources are insufficient. The global optimizer responds to migration requests and regenerates the scheduling scheme, thereby achieving dynamic scheduling integrating edge and cloud, improving the overall response speed and processing efficiency of the platform, reducing cloud resource pressure, lowering bandwidth costs, enhancing system stability, improving system resource utilization, and supporting local data processing to protect data privacy.

[0008] Based on the above technical solution, the present invention can be further improved as follows.

[0009] Furthermore, the prediction module, when predicting the load trend, network status, and service request volume of each edge computing node based on the current and historical monitoring data of each edge computing node's own device, specifically uses the following: The current and historical monitoring data of each edge computing node's own device are preprocessed to obtain preprocessed current and historical monitoring data. The preprocessing includes cleaning, normalization and input sequence construction. The current monitoring data and historical monitoring data are input into a pre-trained time series prediction model to obtain the load trend, network status and service request volume of each edge computing node. The time series prediction model is a long short-term memory network model.

[0010] The beneficial effects of adopting the above-mentioned further solutions are as follows: cleaning, normalizing, and constructing input sequences for monitoring data can improve data quality and availability; the Long Short-Term Memory (LSTM) network model can capture the time-series characteristics of the data, accurately predict the load trends, network status, and service request volume of edge computing nodes, and provide a reliable basis for the global optimizer to generate task scheduling schemes. This helps the system achieve optimized scheduling with the goal of minimizing the total system cost, reducing total system energy consumption, average task latency, and bandwidth costs. Furthermore, the global optimizer, used to optimize based on the prediction results of the prediction module with the goal of minimizing the total system cost, specifically performs the following when obtaining a global task scheduling scheme: Construct an objective function that minimizes the total system cost; Based on the system deployment scenario information, the prediction results of the prediction module, and the current monitoring data of each edge computing node's own device, the weight coefficients in the objective function are dynamically calculated. The weight coefficients include the weight of the total system energy consumption, the weight of the average task latency, and the weight of the bandwidth cost. The system size is determined based on the number of tasks to be scheduled and the number of edge computing nodes, and the target optimization algorithm is determined based on the system size. Based on the objective optimization algorithm, objective function, preset constraints, and prediction results of the prediction module, a constrained optimization problem is constructed and solved to generate a global task scheduling scheme. The preset constraints include resource constraints, SLA constraints, data location constraints, and migration cost limits. If a solution is not found within the preset maximum time limit during the solution process, a rollback strategy is activated. The rollback strategy includes: prioritizing the allocation of tasks that meet the Service Level Agreement (SLA) delay constraints, then allocating high-priority tasks, and finally allocating low-priority tasks.

[0011] The beneficial effects of adopting the above-mentioned further solutions are as follows: By constructing an objective function that minimizes the total system cost, a clear optimization direction can be provided for task scheduling; based on the system deployment scenario, prediction results, and current monitoring data, the weight coefficients can be dynamically calculated, and the weights of total system energy consumption, average task latency, and bandwidth costs can be flexibly adjusted according to the actual situation, making the scheduling scheme more in line with actual needs; determining the objective optimization algorithm according to the system scale allows for the selection of the most suitable algorithm to solve the constrained optimization problem, improving the efficiency and accuracy of the solution; constructing and solving the constrained optimization problem to generate a global task scheduling scheme can minimize the total system cost under conditions such as resource constraints, SLA constraints, data location constraints, and migration cost limitations; enabling a rollback strategy can prioritize the allocation of tasks that meet SLA latency constraints and high-priority tasks when a solution is not found within a preset time, ensuring the stable operation of the system.

[0012] Furthermore, the global optimizer, when determining the target optimization algorithm based on the system size, is specifically used for: When the size of the edge computing node is less than the first threshold, the target optimization algorithm is determined to be the Mixed Integer Linear Programming (MILP) algorithm. When the size of the edge computing node is not less than the first threshold, the target optimization algorithm is determined to be either a genetic algorithm or a simulated annealing algorithm.

[0013] The beneficial effects of adopting the above-mentioned further scheme are as follows: the target optimization algorithm is determined according to the scale of the edge computing nodes. When the scale of the edge computing nodes is less than the first threshold, the mixed integer linear programming (MILP) algorithm is selected, which can obtain the optimal solution and is suitable for small-scale scenarios. When the scale of the edge computing nodes is not less than the first threshold, the genetic algorithm or simulated annealing algorithm is selected, which can meet the real-time requirements in large-scale scenarios, thereby improving the accuracy and efficiency of global optimizer scheduling and achieving a more reasonable task allocation with the goal of minimizing the total system cost.

[0014] Furthermore, the global optimizer, when dynamically calculating the weight coefficients in the objective function based on the system deployment scenario information, the prediction results of the prediction module, and the current monitoring data of each edge computing node's own device, is specifically used for: Based on the system deployment scenario information, the first weights for total energy consumption, average task latency, and bandwidth cost are determined respectively. Based on the prediction results and the current monitoring data of each edge computing node's own devices, the first weight of total energy consumption, the first weight of average task latency, and the first weight of bandwidth cost are dynamically adjusted to obtain the second weight of total energy consumption, the second weight of average task latency, and the second weight of bandwidth cost. Based on the normalized adaptive formula, the three indicators of total energy consumption, average task latency and bandwidth cost are normalized respectively to obtain the third weight of total energy consumption, the third weight of average task latency and the third weight of bandwidth cost. Based on each second weight and each third weight, determine the weights for total energy consumption, average task latency, and bandwidth cost.

[0015] The beneficial effects of adopting the above-mentioned further solution are as follows: by combining the initial weights with the system deployment scenario information, then dynamically adjusting the weights based on the prediction results and current monitoring data, and finally performing normalization processing to determine the final weights, the global optimizer can more accurately balance total energy consumption, average task latency and bandwidth costs in different scenarios, generate a more reasonable global task scheduling scheme with the goal of minimizing the total system cost, improve system resource utilization, reduce system costs, and meet business needs in different scenarios.

[0016] Furthermore, the global optimizer, when determining the first weight of total energy consumption, the first weight of average task latency, and the first weight of bandwidth cost based on the system deployment scenario information, specifically uses the following: In scenarios with low latency requirements, the first weight for average task latency is greater than the first weight for total energy consumption and the first weight for bandwidth cost. In energy-saving scenarios, the first weight of total energy consumption is set to be greater than the first weight of average task latency and the first weight of bandwidth cost. In scenarios where bandwidth is expensive, the first weight for bandwidth cost is greater than the first weight for total energy consumption and the first weight for average task latency.

[0017] The beneficial effects of adopting the above-mentioned further solutions are: They achieve a precise match between optimization goals and business needs, effectively solving the pain point of traditional scheduling systems where optimization goals are singular and unable to adapt to diverse scenario requirements. This enables the system to automatically ensure real-time performance in low-latency scenarios, prioritize energy efficiency optimization in energy-saving scenarios, and intelligently control communication costs in bandwidth-intensive scenarios.

[0018] Furthermore, the global optimizer, when dynamically adjusting the first weight of total energy consumption, the first weight of average task latency, and the first weight of bandwidth cost based on the prediction results and the current monitoring data of each edge computing node's own devices, is specifically used for: When the average task latency is detected to exceed the preset SLA threshold, the first weight of the average task latency is increased by a preset ratio; When the total energy consumption of the system is detected to be close to the preset upper limit threshold, the first weight of the total energy consumption is increased by a preset ratio. When the prediction module predicts that network congestion will occur in the future, increasing bandwidth costs takes precedence.

[0019] The beneficial effects of adopting the above-mentioned further solutions are as follows: When the average task latency exceeds the preset SLA threshold, the first weight of the average task latency is increased by a preset ratio, which makes the global optimizer pay more attention to task latency in subsequent optimizations, so as to reduce the average task latency and meet the service level agreement requirements; When the total system energy consumption is detected to be close to the preset energy consumption upper limit threshold, the first weight of the total energy consumption is increased by a preset ratio, which can prompt the global optimizer to pay more attention to reducing the total system energy consumption when scheduling tasks and avoid excessive energy consumption; When the prediction module predicts that the network will be congested in the future, the first weight of bandwidth cost is temporarily increased, which reduces the amount of data transmission between the edge and the cloud when the global optimizer generates the scheduling scheme, reduces bandwidth cost, and thus achieves the goal of minimizing the total system cost.

[0020] Furthermore, when each of the local optimizers verifies the effectiveness of the global task scheduling scheme based on its own real-time state data, it specifically performs the following: For any of the local optimizers, based on the current task queue length and bandwidth status of its own device, the latency value of each task executed locally is predicted; For any of the local optimizers, obtain the current resource utilization and link bandwidth of its own device; For any of the local optimizers, the global task scheduling scheme is determined to be invalid when the deviation between the predicted latency value and the global scheme preset value is greater than a set deviation threshold, or the resource utilization is greater than a set safety threshold, or the link bandwidth does not meet the set bandwidth requirements.

[0021] The beneficial effects of adopting the above-mentioned further scheme are: the local optimizer predicts task latency based on task queue length and bandwidth status, obtains resource utilization and link bandwidth, and judges whether the global task scheduling scheme has failed based on deviation threshold, safety threshold and bandwidth requirements. It can promptly detect the infeasibility of the global scheme locally, avoid executing unreasonable scheduling schemes, improve the accuracy and reliability of task scheduling, and ensure the stable operation of the system.

[0022] Furthermore, when each of the local optimizers makes local adjustments to the global task scheduling scheme based on its own real-time state data, it is specifically used for: Based on its own real-time status data and system deployment scenario information, the system selects an adjustment mode, which includes feedback rejection mode, local degradation mode and hybrid adjustment mode. Based on the adjustment mode, the global task scheduling scheme is locally adjusted to obtain the adjustment scheme; Among them, the feedback rejection mode represents the processing method of sending a solution failure notification to the global optimizer and waiting for an updated solution; Local degradation mode represents a processing approach that generates alternatives locally and executes them immediately; The hybrid adjustment mode represents the approach of using different processing strategies according to task priority.

[0023] The beneficial effects of adopting the above-mentioned further solutions are as follows: the local optimizer can select an appropriate adjustment mode from three modes—feedback rejection, local degradation, and hybrid adjustment—to locally adjust the global task scheduling scheme based on real-time status data and system deployment scenarios. The feedback rejection mode can ensure global consistency of the system and is suitable for businesses with extremely high requirements for global consistency; the local degradation mode can generate alternative solutions locally and execute them immediately, which is suitable for scenarios with high latency requirements and where waiting for the cloud is not feasible; the hybrid adjustment mode adopts different processing strategies according to task priority, which can meet the different needs of tasks with different priorities, enabling the system to respond more flexibly and efficiently to adjustments to the global task scheduling scheme in different scenarios, thereby improving the system's adaptability and reliability.

[0024] Furthermore, when each of the local optimizers sends a task migration request to the global optimizer when the global task scheduling scheme is infeasible or the local node's resources are insufficient, it is specifically used for: When the computing or storage resources of the edge computing node are lower than the preset task requirement threshold, a task migration request is sent to the global optimizer to trigger cloud migration. When the latency tolerance of a task exceeds a preset task latency threshold, a task migration request is sent to the global optimizer, triggering cloud migration. When a task involves multi-node collaboration and the communication overhead between edges exceeds a preset overhead limit threshold, a task migration request is sent to the global optimizer to trigger cloud migration. When a task contains data that sets a security level, a task migration request is sent to the global optimizer, triggering a cloud migration.

[0025] The beneficial effects of adopting the above-mentioned further solutions are as follows: Triggering cloud migration when edge computing nodes lack sufficient computing or storage resources avoids impacting task execution due to insufficient local resources; triggering cloud migration when the task latency tolerance exceeds a preset threshold fully utilizes cloud resources to process latency-insensitive tasks; triggering cloud migration when a task involves multi-node collaboration and inter-edge communication overhead is excessive reduces communication costs; triggering cloud migration when a task contains data with a set security level enables centralized data management and ensures data security. Simultaneously, these migration strategies allow the system to dynamically adjust task deployment based on different task characteristics and node states, improving system flexibility, stability, and resource utilization, and minimizing total system cost. Attached Figure Description

[0026] Figure 1A schematic diagram of the structure of an edge-cloud integrated scheduling system based on the collaboration of a global optimizer and a local optimizer, provided as an embodiment of the present invention; Figure 2 This is a schematic diagram of another edge-cloud integrated scheduling system based on the collaboration of a global optimizer and a local optimizer, provided as an embodiment of the present invention. Detailed Implementation

[0027] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0028] Furthermore, the term "and / or" in this article is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, or B existing alone. Additionally, the character " / " in this article, unless otherwise specified, generally indicates that the preceding and following related objects have an "or" relationship.

[0029] like Figure 1 and Figure 2 As shown, this application provides an edge-cloud integrated scheduling system based on the collaboration of a global optimizer and a local optimizer. It mainly includes a cloud platform and various edge computing nodes. A global optimizer and a prediction module are deployed on the cloud platform, and a local optimizer is deployed on each edge computing node. The prediction module is used to predict the load trend, network status and service request volume of each edge computing node based on the current monitoring data and historical monitoring data of each edge computing node's own device. The current monitoring data includes resource status data, network status data, task status data and service priority. A global optimizer is used to optimize based on the prediction results of the prediction module with the goal of minimizing the total system cost, to obtain a global task scheduling scheme, and to send the global task scheduling scheme to each of the edge computing nodes. The total cost includes the total system energy consumption, average task latency, and bandwidth cost. Each local optimizer is used to verify the effectiveness of the global task scheduling scheme and make local adjustments based on its own real-time status data. When the global task scheduling scheme is not feasible or the local node resources are insufficient, a task migration request is sent to the global optimizer. The global optimizer is also used to regenerate the scheduling scheme based on the latest system state in response to the task migration request.

[0030] In this embodiment, each edge computing node acquires the current monitoring data of its own device in real time and periodically reports the current monitoring data to the cloud platform.

[0031] Business priorities can be determined through rule configuration and dynamic adjustment. Different types of business priorities are predefined during system deployment. These priorities are typically determined by industry standards and business department requirements. For example, priorities can be divided into four levels: highest priority P1, high priority P2, medium priority P3, and low priority P4, as detailed in Table 1. Table 1: The system can automatically raise or lower task priorities based on runtime status. For example, when a device enters a dangerous operating condition, the priority of its related tasks is raised from P3 to P1; when there is network congestion or excessive energy consumption, low-priority tasks (such as batch log uploads) are postponed.

[0032] In this embodiment of the application, the prediction module, when predicting the load trend, network status, and service request volume of each edge computing node based on the current and historical monitoring data of each edge computing node's own device, is specifically used for: The current and historical monitoring data of each edge computing node's own device are preprocessed to obtain preprocessed current and historical monitoring data. The preprocessing includes cleaning, normalization and input sequence construction. The current monitoring data and historical monitoring data are input into a pre-trained time series prediction model to obtain the load trend, network status and service request volume of each edge computing node. The time series prediction model is a long short-term memory network model.

[0033] In the above embodiments, the prediction includes network-level prediction, which predicts the system's bandwidth utilization, latency trend, and packet loss rate; computing-level prediction, which predicts the CPU / GPU utilization, memory usage, and task queue length of edge computing nodes; and business-level prediction, which predicts the number of task requests and business peak fluctuations.

[0034] Suppose that at time t, the system collects the following data: Xt=[CPU_usaget,Mem_usaget,BW_upt,BW_downt,Latencyt,Loss_ratet,Req_numt,Alert_numt,Hour_t,Day_t], is a 10-dimensional feature vector that is input into an LSTM unit. The sequence length can be the data points from the most recent 10 minutes or 1 hour.

[0035] The prediction module, through the collaborative work of the data acquisition and prediction computation components, can accurately predict the future state of each edge computing node, providing a reliable basis for the global optimizer. The data acquisition component ensures the comprehensiveness and accuracy of the data, while the prediction computation component uses advanced models to analyze and predict the data, enabling the system to prepare for resource allocation and scheduling in advance.

[0036] In this embodiment of the application, the global optimizer, when optimizing based on the prediction results of the prediction module with the goal of minimizing the total system cost, to obtain a global task scheduling scheme, specifically performs the following: Construct an objective function that minimizes the total system cost; Based on the system deployment scenario information, the prediction results of the prediction module, and the current monitoring data of each edge computing node's own device, the weight coefficients in the objective function are dynamically calculated. The weight coefficients include the weight of the total system energy consumption, the weight of the average task latency, and the weight of the bandwidth cost. The system size is determined based on the number of tasks to be scheduled and the number of edge computing nodes, and the target optimization algorithm is determined based on the system size. Based on the objective optimization algorithm, objective function, preset constraints, and prediction results of the prediction module, a constrained optimization problem is constructed and solved to generate a global task scheduling scheme. The preset constraints include resource constraints, SLA constraints, data location constraints, and migration cost limits. If a solution is not found within the preset maximum time limit during the solution process, a rollback strategy is activated. The rollback strategy includes: prioritizing the allocation of tasks that meet the Service Level Agreement (SLA) delay constraints, then allocating high-priority tasks, and finally allocating low-priority tasks.

[0037] In the above implementation, the global optimizer takes minimizing the total system cost as its objective function and considers the following factors: Total energy consumption (edge ​​device energy consumption + cloud computing energy consumption); Average task latency (including transmission latency and computation latency); Bandwidth costs (the amount of data transferred across domains).

[0038] The objective function building component can quantify and combine various cost factors according to different cost calculation models to form a comprehensive objective function.

[0039] Objective function: ; Where: E: System energy consumption (including edge computing nodes and cloud server power consumption); D: Task latency (average response time or 95th percentile latency); B: Bandwidth cost (data transmission volume between edge and cloud * unit bandwidth price); α, β, γ: Weighting coefficients.

[0040] Optimization goal: minCost.

[0041] Under constraints (such as Service Level Agreement (SLA), node computing power limit, and maximum tolerable latency), the task allocation scheme is output through optimization algorithms (such as linear programming, genetic algorithms, or reinforcement learning scheduling strategies).

[0042] Furthermore, the global optimizer, when determining the target optimization algorithm based on the system size, is specifically used for: When the size of the edge computing node is less than the first threshold, the target optimization algorithm is determined to be the Mixed Integer Linear Programming (MILP) algorithm. When the size of the edge computing node is not less than the first threshold, the target optimization algorithm is determined to be either a genetic algorithm or a simulated annealing algorithm.

[0043] In the above implementation, the algorithm selection component determines the system size based on the number of tasks to be scheduled and the number of edge computing nodes, and determines the target optimization algorithm based on the system size. When the number of edge computing nodes is less than a first threshold, the target optimization algorithm is determined to be the Mixed Integer Linear Programming (MILP) algorithm, which can obtain the optimal solution; when the number of edge computing nodes is not less than the first threshold, the target optimization algorithm is determined to be either a genetic algorithm or a simulated annealing algorithm, which can quickly obtain approximate optimal solutions in large-scale problems.

[0044] The solution generation component constructs and solves a constrained optimization problem based on the objective optimization algorithm, objective function, preset constraints, and prediction results from the prediction module, generating a global task scheduling scheme. Preset constraints include resource constraints to ensure that the computing resources of edge computing nodes do not exceed hardware limits; SLA constraints to ensure that the latency of certain high-priority tasks does not exceed thresholds; data location constraints to limit the transmission range of sensitive data; and migration cost constraints to avoid the high costs associated with frequent migrations. During the solution process, if a solution is not found within the preset maximum time limit, the solution generation component will activate a fallback strategy, prioritizing tasks that meet the SLA latency constraints, then allocating high-priority tasks, and finally low-priority tasks.

[0045] Furthermore, the global optimizer, when dynamically calculating the weight coefficients in the objective function based on the system deployment scenario information, the prediction results of the prediction module, and the current monitoring data of each edge computing node's own device, is specifically used for: Based on the system deployment scenario information, the first weights for total energy consumption, average task latency, and bandwidth cost are determined respectively. Based on the prediction results and the current monitoring data of each edge computing node's own devices, the first weight of total energy consumption, the first weight of average task latency, and the first weight of bandwidth cost are dynamically adjusted to obtain the second weight of total energy consumption, the second weight of average task latency, and the second weight of bandwidth cost. Based on the normalized adaptive formula, the three indicators of total energy consumption, average task latency and bandwidth cost are normalized respectively to obtain the third weight of total energy consumption, the third weight of average task latency and the third weight of bandwidth cost. Based on each second weight and each third weight, determine the weights for total energy consumption, average task latency, and bandwidth cost.

[0046] Furthermore, the global optimizer, when determining the first weight of total energy consumption, the first weight of average task latency, and the first weight of bandwidth cost based on the system deployment scenario information, specifically uses the following: In scenarios with low latency requirements, the first weight for average task latency is greater than the first weight for total energy consumption and the first weight for bandwidth cost. In energy-saving scenarios, the first weight of total energy consumption is set to be greater than the first weight of average task latency and the first weight of bandwidth cost. In scenarios where bandwidth is expensive, the first weight for bandwidth cost is greater than the first weight for total energy consumption and the first weight for average task latency.

[0047] Furthermore, the global optimizer, when dynamically adjusting the first weight of total energy consumption, the first weight of average task latency, and the first weight of bandwidth cost based on the prediction results and the current monitoring data of each edge computing node's own devices, is specifically used for: When the average task latency is detected to exceed the preset SLA threshold, the first weight of the average task latency is increased by a preset ratio; When the total energy consumption of the system is detected to be close to the preset upper limit threshold, the first weight of the total energy consumption is increased by a preset ratio. When the prediction module predicts that network congestion will occur in the future, increasing bandwidth costs takes precedence.

[0048] In the above implementation, the weight coefficients in the objective function are dynamically calculated based on the system deployment scenario information, the prediction results of the prediction module, and the current monitoring data of each edge computing node's own devices. In low-latency scenarios, such as industrial control scenarios, the weight of average task latency is set greater than the weight of total energy consumption and bandwidth cost; in energy-saving scenarios, such as battery-powered edge device scenarios, the weight of total energy consumption is set greater than the weight of average task latency and bandwidth cost; in bandwidth-intensive scenarios, such as satellite link scenarios, the weight of bandwidth cost is set greater than the weight of total energy consumption and average task latency.

[0049] The weight calculation component also dynamically adjusts the weights based on real-time monitoring and prediction results. When the average task latency exceeds the preset SLA threshold, the weight of the average task latency is increased by a preset ratio. When the total system energy consumption is close to the preset energy consumption upper limit threshold, the weight of the total energy consumption is increased by a preset ratio. When the prediction module predicts that the network will experience congestion in the future, the weight of the bandwidth cost is temporarily increased.

[0050] Meanwhile, in order to prevent an imbalance caused by a certain indicator value being too large or too small, the indicators will be normalized to obtain the final weight.

[0051] In the above implementation, the constraints are as follows: Resource constraints: Established for various resources; SLA constraints: For critical task t, ; Data location constraints: If the data is sensitive or requires physical proximity to the device, then Only specific sets of nodes are allowed; Migration cost limits: Limit the data limit or number of migrations per round to avoid high jitter.

[0052] To avoid weight imbalance caused by differences in the units of E, D, and B, each item is first normalized to its estimated maximum value. The normalization formula is as follows: Reweighted combination.

[0053] In this embodiment of the application, each of the local optimizers, when verifying the effectiveness of the global task scheduling scheme based on its own real-time state data, is specifically used for: For any of the local optimizers, based on the current task queue length and bandwidth status of its own device, the latency value of each task executed locally is predicted; For any of the local optimizers, obtain the current resource utilization and link bandwidth of its own device; For any of the local optimizers, the global task scheduling scheme is determined to be invalid when the deviation between the predicted latency value and the global scheme preset value is greater than a set deviation threshold, or the resource utilization is greater than a set safety threshold, or the link bandwidth does not meet the set bandwidth requirements.

[0054] In this embodiment, the migration is divided into two steps: pre-copying (warm-up) and switching. For large tasks that need to be migrated (such as model weights or large files), most of the data is copied asynchronously first (to reduce switching downtime), and then an atomic switch is performed at an appropriate time.

[0055] The edge computing nodes then continuously monitor execution latency, resource consumption, and anomalies. If an SLA violation or node failure occurs, an emergency re-optimization is triggered, and high-priority tasks are migrated first. If it is network jitter, the weights are adjusted according to the prediction module to avoid congested paths.

[0056] When the local optimizer verifies that the global task scheduling scheme is not feasible or that local resources are insufficient, it triggers a migration request to the global optimizer, which then regenerates the scheduling scheme based on the latest system state. Each edge computing node feeds back the actual result data of task execution to the cloud platform, updates the model parameters of the prediction module, and adjusts the objective function weights of the global optimizer, forming a closed-loop optimization.

[0057] Edge computing nodes report actual execution data (real latency, real energy consumption, real cross-domain traffic) to the cloud; enabling the cloud platform to update prediction models such as LSTM based on this data (through online or periodic fine-tuning); and to adjust weights α, β, and γ based on the actual execution data. For example, PID control, gradient descent with a learning rate, or reinforcement learning strategies can be used to optimize the weights to minimize the long-term average cost. .

[0058] In this embodiment of the application, each of the local optimizers, when making local adjustments to the global task scheduling scheme based on its own real-time state data, is specifically used for: Based on its own real-time status data and system deployment scenario information, the system selects an adjustment mode, which includes feedback rejection mode, local degradation mode and hybrid adjustment mode. Based on the adjustment mode, the global task scheduling scheme is locally adjusted to obtain the adjustment scheme; Among them, the feedback rejection mode represents the processing method of sending a solution failure notification to the global optimizer and waiting for an updated solution; Local degradation mode represents a processing approach that generates alternatives locally and executes them immediately; The hybrid adjustment mode represents the approach of using different processing strategies according to task priority.

[0059] In the above implementation, for example, in the global scheme, task A is assigned to node B with an expected delay of 2 seconds; the actual delay of node B, monitored locally by the edge computing node, is 4.5 seconds, exceeding the SLA constraint. Then, the local optimizer immediately executes a degradation scheme, keeping task A to be executed locally (delay of 2.3 seconds, satisfying the SLA); at the same time, it reports the deviation information to the cloud platform; the cloud-based global optimizer discovers that node B is overloaded based on the feedback, reduces its task allocation ratio in the next round of optimization, and generates a new global task scheduling scheme, migrating the subsequent subtasks of task A to node C to avoid further default.

[0060] In this embodiment of the application, when the local optimizer sends a task migration request to the global optimizer when the global task scheduling scheme is infeasible or the local node resources are insufficient, it is specifically used for: When the computing or storage resources of the edge computing node are lower than the preset task requirement threshold, a task migration request is sent to the global optimizer to trigger cloud migration. When the latency tolerance of a task exceeds a preset task latency threshold, a task migration request is sent to the global optimizer, triggering cloud migration. When a task involves multi-node collaboration and the communication overhead between edges exceeds a preset overhead limit threshold, a task migration request is sent to the global optimizer to trigger cloud migration. When a task contains data that sets a security level, a task migration request is sent to the global optimizer, triggering a cloud migration.

[0061] In this embodiment, tasks are further categorized into multiple types. Based on at least one of the following criteria: latency sensitivity, resource requirements, data locality, security level, task persistence, parallelizability, and bandwidth requirements, input tasks are classified into strongly constrained tasks, weakly constrained tasks, and unconstrained tasks. When latency sensitivity is below a first threshold, the task is classified as a strongly constrained task; when latency sensitivity is above the first threshold and resource requirements are above a second threshold, the task is classified as a weakly constrained task; when latency sensitivity is above a third threshold and task persistence exceeds a preset duration, the task is classified as an unconstrained task.

[0062] The global optimizer receives task classification results and generates global scheduling schemes for weakly constrained and unconstrained tasks. Weakly constrained tasks are allowed to migrate between edge nodes, while unconstrained tasks are allowed to be freely scheduled between the edge and the cloud. Local optimizers are deployed on each edge node to directly execute strongly constrained tasks and perform local resource allocation and execution monitoring for weakly constrained and unconstrained tasks distributed to their respective nodes. Strongly constrained tasks are bound to the edge node that generated them for local execution and do not participate in global scheduling. Weakly constrained tasks migrate between edge nodes for execution according to the global optimizer's scheduling scheme. Unconstrained tasks are allocated by the global optimizer to the cloud or edge nodes with sufficient resources for execution.

[0063] This system achieves dynamic balancing and precise control of multi-objective optimization in the edge-cloud system by constructing a two-layer scheduling architecture that coordinates a global optimizer and local optimizers. The prediction module uses current and historical monitoring data of edge computing nodes to predict their load, network status, and service request volume in advance. The global optimizer optimizes based on the prediction results, aiming to minimize the total system cost, resulting in a global task scheduling scheme that balances total energy consumption, average task latency, and bandwidth costs. The local optimizer verifies the effectiveness of the global scheme and makes local adjustments, requesting task migration when the scheme is infeasible or resources are insufficient. The global optimizer responds to migration requests and regenerates the scheduling scheme, thereby achieving integrated edge-cloud dynamic scheduling, improving the overall platform response speed and processing efficiency, reducing cloud resource pressure, lowering bandwidth costs, enhancing system stability, improving system resource utilization, and supporting local data processing to protect data privacy.

[0064] The terms “comprising,” “including,” or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0065] The above description is merely a preferred embodiment of this application and an explanation of the technical principles employed. Those skilled in the art should understand that the scope of this application is not limited to technical solutions formed by specific combinations of the above-described technical features, but should also cover other technical solutions formed by arbitrary combinations of the above-described technical features or their equivalents without departing from the foregoing application concept. For example, technical solutions formed by substituting the above features with (but not limited to) technical features with similar functions claimed in this application.

Claims

1. A cloud-edge integrated scheduling system based on the collaboration of a global optimizer and a local optimizer, characterized in that, This includes a cloud platform and various edge computing nodes. The cloud platform is equipped with a global optimizer and a prediction module, while each edge computing node is equipped with a local optimizer. The prediction module is used to predict the load trend, network status and service request volume of each edge computing node based on the current monitoring data and historical monitoring data of each edge computing node's own device. The current monitoring data includes resource status data, network status data, task status data and service priority. The global optimizer is used to optimize based on the prediction results of the prediction module with the goal of minimizing the total system cost, to obtain a global task scheduling scheme, and to send the global task scheduling scheme to each edge computing node. The total cost includes the total system energy consumption, average task latency, and bandwidth cost. Each of the local optimizers is used to verify the effectiveness of the global task scheduling scheme and make local adjustments based on its own real-time status data. When the global task scheduling scheme is not feasible or the local node resources are insufficient, a task migration request is sent to the global optimizer. The global optimizer is also used to regenerate the scheduling scheme based on the latest system state in response to the task migration request.

2. The edge-cloud integrated scheduling system based on the collaboration of a global optimizer and a local optimizer as described in claim 1, characterized in that, The prediction module, used to predict the load trend, network status, and service request volume of each edge computing node based on the current and historical monitoring data of each edge computing node's own devices, specifically includes: The current and historical monitoring data of each edge computing node's own device are preprocessed to obtain preprocessed current and historical monitoring data. The preprocessing includes cleaning, normalization and input sequence construction. The current monitoring data and historical monitoring data are input into a pre-trained time series prediction model to obtain the load trend, network status and service request volume of each edge computing node. The time series prediction model is a long short-term memory network model.

3. The edge-cloud integrated scheduling system based on the collaboration of a global optimizer and a local optimizer as described in claim 1, characterized in that, The global optimizer, used to optimize based on the prediction results of the prediction module with the goal of minimizing the total system cost, specifically performs the following when obtaining a global task scheduling scheme: Construct an objective function that minimizes the total system cost; Based on the system deployment scenario information, the prediction results of the prediction module, and the current monitoring data of each edge computing node's own device, the weight coefficients in the objective function are dynamically calculated. The weight coefficients include the weight of the total system energy consumption, the weight of the average task latency, and the weight of the bandwidth cost. The system size is determined based on the number of tasks to be scheduled and the number of edge computing nodes, and the target optimization algorithm is determined based on the system size. Based on the objective optimization algorithm, objective function, preset constraints, and prediction results of the prediction module, a constrained optimization problem is constructed and solved to generate a global task scheduling scheme. The preset constraints include resource constraints, SLA constraints, data location constraints, and migration cost limits. If a solution is not found within the preset maximum time limit during the solution process, a rollback strategy is activated. The rollback strategy includes: prioritizing the allocation of tasks that meet the Service Level Agreement (SLA) delay constraints, then allocating high-priority tasks, and finally allocating low-priority tasks.

4. The edge-cloud integrated scheduling system based on the collaboration of a global optimizer and a local optimizer as described in claim 3, characterized in that, The global optimizer, when determining the target optimization algorithm based on the system size, is specifically used for: When the size of the edge computing node is less than the first threshold, the target optimization algorithm is determined to be the Mixed Integer Linear Programming (MILP) algorithm. When the size of the edge computing node is not less than the first threshold, the target optimization algorithm is determined to be either a genetic algorithm or a simulated annealing algorithm.

5. The edge-cloud integrated scheduling system based on the collaboration of a global optimizer and a local optimizer as described in claim 3, characterized in that, The global optimizer, used to dynamically calculate the weight coefficients in the objective function based on the system deployment scenario information, the prediction results of the prediction module, and the current monitoring data of each edge computing node's own devices, is specifically used for: Based on the system deployment scenario information, the first weights for total energy consumption, average task latency, and bandwidth cost are determined respectively. Based on the prediction results and the current monitoring data of each edge computing node's own devices, the first weight of total energy consumption, the first weight of average task latency, and the first weight of bandwidth cost are dynamically adjusted to obtain the second weight of total energy consumption, the second weight of average task latency, and the second weight of bandwidth cost. Based on the normalized adaptive formula, the three indicators of total energy consumption, average task latency and bandwidth cost are normalized respectively to obtain the third weight of total energy consumption, the third weight of average task latency and the third weight of bandwidth cost. Based on each second weight and each third weight, determine the weights for total energy consumption, average task latency, and bandwidth cost.

6. The edge-cloud integrated scheduling system based on the collaboration of a global optimizer and a local optimizer as described in claim 5, characterized in that, The global optimizer, when determining the first weight of total energy consumption, the first weight of average task latency, and the first weight of bandwidth cost based on the system deployment scenario information, is specifically used for: In scenarios with low latency requirements, the first weight for average task latency is greater than the first weight for total energy consumption and the first weight for bandwidth cost. In energy-saving scenarios, the first weight of total energy consumption is set to be greater than the first weight of average task latency and the first weight of bandwidth cost. In scenarios where bandwidth is expensive, the first weight for bandwidth cost is greater than the first weight for total energy consumption and the first weight for average task latency.

7. The edge-cloud integrated scheduling system based on the collaboration of a global optimizer and a local optimizer as described in claim 5, characterized in that, The global optimizer, used to dynamically adjust the first weight of total energy consumption, the first weight of average task latency, and the first weight of bandwidth cost based on the prediction results and the current monitoring data of each edge computing node's own devices, is specifically used for: When the average task latency is detected to exceed the preset SLA threshold, the first weight of the average task latency is increased by a preset ratio; When the total energy consumption of the system is detected to be close to the preset upper limit threshold, the first weight of the total energy consumption is increased by a preset ratio. When the prediction module predicts that network congestion will occur in the future, increasing bandwidth costs takes precedence.

8. The edge-cloud integrated scheduling system based on the collaboration of a global optimizer and a local optimizer as described in claim 1, characterized in that, When each of the local optimizers verifies the effectiveness of the global task scheduling scheme based on its own real-time state data, it is specifically used for: For any of the local optimizers, based on the current task queue length and bandwidth status of its own device, the latency value of each task executed locally is predicted; For any of the local optimizers, obtain the current resource utilization and link bandwidth of its own device; For any of the local optimizers, the global task scheduling scheme is determined to be invalid when the deviation between the predicted latency value and the global scheme preset value is greater than a set deviation threshold, or the resource utilization is greater than a set safety threshold, or the link bandwidth does not meet the set bandwidth requirements.

9. The edge-cloud integrated scheduling system based on the collaboration of a global optimizer and a local optimizer as described in claim 8, characterized in that, When each of the local optimizers makes local adjustments to the global task scheduling scheme based on its own real-time state data, it is specifically used for: Based on its own real-time status data and system deployment scenario information, the system selects an adjustment mode, which includes feedback rejection mode, local degradation mode and hybrid adjustment mode. Based on the adjustment mode, the global task scheduling scheme is locally adjusted to obtain the adjustment scheme; Among them, the feedback rejection mode represents the processing method of sending a solution failure notification to the global optimizer and waiting for an updated solution; Local degradation mode represents a processing approach that generates alternatives locally and executes them immediately; The hybrid adjustment mode represents the approach of using different processing strategies according to task priority.

10. The edge-cloud integrated scheduling system based on the collaboration of a global optimizer and a local optimizer as described in claim 8, characterized in that, When the local optimizer sends a task migration request to the global optimizer when the global task scheduling scheme is infeasible or the local node's resources are insufficient, it is specifically used for: When the computing or storage resources of the edge computing node are lower than the preset task requirement threshold, a task migration request is sent to the global optimizer to trigger cloud migration. When the latency tolerance of a task exceeds a preset task latency threshold, a task migration request is sent to the global optimizer, triggering cloud migration. When a task involves multi-node collaboration and the communication overhead between edges exceeds a preset overhead limit threshold, a task migration request is sent to the global optimizer to trigger cloud migration. When a task contains data that sets a security level, a task migration request is sent to the global optimizer, triggering a cloud migration.

Images