Deep model resource fair scheduling method oriented to multi-task service quality guarantee in cloud edge collaborative environment

Abstract

The invention discloses a deep model resource fair scheduling method for multi-task service quality assurance in a cloud-edge collaborative environment. The method comprises the following steps: S1, a task request sensing module; s2, an online model evaluation module; s3, a model fair selection module; s4, an isomorphic model migration module; s5, a heterogeneous model collaborative scheduling module; and S6, a service quality fairness guarantee module. The method has high engineering practical value, and can be applied to various real-time reasoning scenes such as intelligent medical treatment, intelligent security and protection, industrial internet and the like.

Description

Technical Field

[0001] This invention belongs to the field of cloud computing and edge intelligence technology, specifically relating to a deep learning resource management method for multi-task inference services in a cloud-edge collaborative environment, and in particular a fair scheduling method for deep model resources for quality assurance of multi-task services. Background Technology

[0002] With the rapid proliferation of IoT devices, a large number of deep learning tasks, such as video analytics, image recognition, and behavior monitoring, are generated in real time at the edge and require efficient processing via a cloud-edge collaborative architecture. Edge computing offers advantages such as low latency and low bandwidth consumption, while the cloud boasts powerful computing and storage capabilities. Therefore, collaborating between cloud and edge nodes to complete deep learning inference tasks has become the mainstream technology approach in applications such as smart cities, intelligent security, and smart healthcare.

[0003] In cloud-edge collaborative environments, there are significant differences in computing power, resource consumption, and network bandwidth among different edge nodes. Furthermore, various deep neural network (DNN) models exhibit variations in inference accuracy, computational complexity, and latency costs, making the selection and scheduling of deep models in multi-task scenarios increasingly complex. This is particularly true in scenarios such as video surveillance, autonomous driving, and industrial quality inspection, where tasks demand high inference response time and accuracy. Inappropriate model selection can easily lead to insufficient inference accuracy, excessive response latency, or task backlog, thereby impacting overall system performance.

[0004] Existing cloud-edge collaborative inference methods primarily focus on improving overall throughput or average system accuracy, often employing a best-effort strategy to allocate model resources to tasks. However, when multiple tasks run continuously and simultaneously, this strategy leads to resource skew, where high-quality inputs or high-priority tasks continuously consume high-performance models, while low-quality or weaker tasks struggle to obtain sufficient resources, thus failing to meet their minimum inference accuracy or response requirements. This unfair resource allocation affects the overall availability of tasks and is detrimental to practical applications requiring long-term stable operation.

[0005] Furthermore, the lack of an effective coordination mechanism between edge nodes is a major problem in the current system. When some nodes are under high load, have insufficient model instances, or experience abnormal latency, the system cannot migrate tasks to idle nodes in a timely manner, leading to increased inference latency, resource waste, and a higher task failure rate. At the same time, existing methods often focus on local node model selection and lack joint scheduling strategies for heterogeneous models across nodes, making it difficult to fully utilize all the deep learning model resources within the system.

[0006] In summary, the current cloud-edge collaborative inference system still has the following prominent problems: (1) Lack of fairness in model selection: weak tasks cannot obtain deep models that meet the minimum accuracy requirements for a long time, resulting in a decline in service quality; (2) Insufficient resource utilization: lack of load balancing mechanism for homogeneous models, some nodes are overloaded while other nodes are idle; (3) Insufficient cross-node collaboration: the cross-node calling capability of heterogeneous deep models is limited, resulting in the overall inference capability of the system not being effectively utilized; (4) Difficulty in maintaining stable service quality: in scenarios with dynamic changes in multiple tasks, it is impossible to guarantee that each task meets the minimum inference accuracy and response requirements for a long time.

[0007] Therefore, there is an urgent need to propose a deep model resource scheduling method for cloud-edge collaborative environments that can balance task fairness and system resource utilization, so as to improve the service quality assurance capability between tasks and achieve long-term stable operation of the system. Summary of the Invention

[0008] This invention addresses the technical limitations of existing cloud-edge collaborative inference services, such as unfair model selection, insufficient resource utilization, and difficulty in consistently guaranteeing task service quality. It provides a deep model resource fairness scheduling method for multi-task service quality assurance in a cloud-edge collaborative environment. Compared to existing cloud-edge inference scheduling strategies, this method achieves fair allocation of model resources among tasks, improves cross-node collaboration capabilities, significantly enhances overall inference accuracy, service response speed, and resource utilization, thereby strengthening the system's long-term stable service capabilities in complex business scenarios.

[0009] The technical solution of this invention is:

[0010] A deep model-based fair scheduling method for service quality assurance in a cloud-edge collaborative environment, comprising the following steps:

[0011] S1, Task Request Awareness Module: Real-time collection of task request stream, minimum inference accuracy requirements of tasks, resource usage of each node, running status of model instances and network link status, and construction of task status table and node resource table to provide real-time input for subsequent model selection and scheduling decisions;

[0012] S2, Online Model Evaluation Module: Based on the structural complexity of deep models, historical inference performance records, and the current node running status, it dynamically estimates the inference accuracy, execution latency, and computational cost of each deep model on different tasks, and generates a task-model performance evaluation matrix to reflect the suitability of different models for different tasks.

[0013] S3, Model Fair Selection Module: Based on the minimum inference accuracy threshold of the task and the performance evaluation results, select model instances that can meet the service quality requirements for each task; when the local node cannot meet the task accuracy or resource requirements, trigger the cross-node model selection process to ensure the continuous quality of service for the task.

[0014] S4, Homogeneous Model Migration Module: When some nodes experience situations such as long task queues, resource shortages, or insufficient model instances, the module evaluates the inference benefits and network costs of task migration and migrates tasks to other nodes with homogeneous model instances to achieve load balancing and reduce inference waiting time.

[0015] S5, Heterogeneous Model Collaborative Scheduling Module: For heterogeneous models with different structures and capabilities deployed in the cloud-edge system, it selects the optimal model instance for inference among multiple nodes by comprehensively considering factors such as inference accuracy, execution latency, and network transmission overhead, thereby achieving global collaborative optimization of heterogeneous resources.

[0016] S6, Service Quality Fairness Assurance Module: Constructs task fairness indicators based on task inference results, resource allocation ratios, and migration behavior; when a task fails to meet the minimum accuracy requirements or resource allocation is unbalanced, it improves the overall service quality fairness and operational stability of the system by dynamically adjusting model selection weights, scheduling strategies, and inter-node collaboration methods.

[0017] Furthermore, in the task request perception module described in step S1, the task and node status data are updated periodically by continuously monitoring the CPU / GPU utilization, memory capacity, number of model instances, and network latency of the nodes to support real-time scheduling decisions.

[0018] Furthermore, in the online model evaluation module described in step S2, a performance prediction function based on historical inference samples is constructed to achieve continuous learning and dynamic updating of model inference performance, thereby improving the accuracy and adaptability of model performance evaluation.

[0019] Furthermore, in the model fair selection module described in step S3, the model selection process is comprehensively decided based on factors such as task accuracy requirements, task priority, node load status, and model selection weights, so that the model allocation process has transparency and stability.

[0020] Furthermore, in the isomorphic model migration module described in step S4, by setting a task migration threshold, the migration operation is ensured to be carried out under the premise of improving inference performance and node load level, thus guaranteeing the rationality of the migration behavior.

[0021] Furthermore, in the heterogeneous model collaborative scheduling module described in step S5, the system can flexibly select high-precision or high-performance models under different task scenarios through cross-node deep model capability analysis and task matching mechanism, thereby improving the overall reasoning ability and operating efficiency.

[0022] Furthermore, in the service quality fairness assurance module described in step S6, a task fairness index is constructed by statistically analyzing the actual inference accuracy, resource allocation ratio, and migration call count of each task. When it is detected that some tasks have not met the minimum accuracy requirements or the resource allocation is unbalanced, the system automatically adjusts the model selection weight, node load factor, and migration triggering conditions to enhance the overall fairness among tasks and maintain the long-term stable operation of the system.

[0023] Compared with existing technologies, the deep model resource fair scheduling method for multi-task service quality assurance in a cloud-edge collaborative environment provided by this invention has the following advantages:

[0024] The model resource fair scheduling mechanism proposed in this invention, through the establishment of task accuracy constraints and fairness indicators, makes the allocation of model resources among tasks more reasonable, solving the problem of uneven service quality among tasks in existing methods. The homogeneous model migration mechanism proposed in this invention enables nodes to share load, effectively reducing task backlog and improving the overall utilization efficiency of edge resources. The heterogeneous model cross-node collaborative scheduling strategy proposed in this invention enables the system to achieve optimal model selection under heterogeneous resource conditions, enhancing the system's adaptability to complex task scenarios. Through online model evaluation, this invention can reflect model performance changes in real time, improving the accuracy of model selection and scheduling decisions. This invention has high engineering practical value and can be applied to various real-time inference scenarios such as smart healthcare, intelligent security, and the industrial internet. Attached Figure Description

[0025] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0026] Figure 1 This is a schematic diagram of the multi-task deep model selection mechanism in the cloud-edge collaborative environment of the present invention. It shows the overall structural relationship of the task after reaching the edge node, based on the accuracy and latency requirements and the real-time resource status of the node, to complete the candidate model screening and decision matching among the various model instances available to the node. It reflects the interaction logic between task input, model instance distribution and node load information, and is used to illustrate the task perception, model selection and local decision-making mechanism of the present invention.

[0027] Figure 2 This is the overall flowchart of the multi-node deep model resource fair scheduling of the present invention. It presents the sequential relationship of key steps such as task and node status acquisition, online model performance evaluation, local optimization, load balancing between homogeneous models, collaborative scheduling between heterogeneous models, and service quality fairness adjustment. It reflects the overall scheduling process of the present invention in a multi-node environment to achieve accuracy assurance, latency control and resource balance through "online selection - local optimization - cross-node collaboration".

[0028] Specific implementation methods

[0029] To enable those skilled in the art to better understand the technical solutions in the embodiments of the present invention, and to make the above-mentioned objectives, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be further described below.

[0030] The endpoints and any values ​​of the ranges disclosed herein are not limited to the precise ranges or values, and these ranges or values ​​should be understood to include values ​​close to these ranges or values. For numerical ranges, the endpoint values ​​of the various ranges, the endpoint values ​​of the various ranges and individual point values, and individual point values ​​can be combined with each other to obtain one or more new numerical ranges, which should be considered as specifically disclosed herein.

[0031] A deep model-based fair scheduling method for multi-task service quality assurance in a cloud-edge collaborative environment includes the following steps:

[0032] S1, Task Request Awareness Module: Real-time collection of task request stream, minimum inference accuracy requirements of tasks, resource usage of each node, running status of model instances and network link status, and construction of task status table and node resource table to provide real-time input for subsequent model selection and scheduling decisions;

[0033] S2, Online Model Evaluation Module: Based on the structural complexity of deep models, historical inference performance records, and the current node running status, it dynamically estimates the inference accuracy, execution latency, and computational cost of each deep model on different tasks, and generates a task-model performance evaluation matrix to reflect the suitability of different models for different tasks.

[0034] S3, Model Fair Selection Module: Based on the minimum inference accuracy threshold of the task and the performance evaluation results, select model instances that can meet the service quality requirements for each task; when the local node cannot meet the task accuracy or resource requirements, trigger the cross-node model selection process to ensure the continuous quality of service for the task.

[0035] S4, Homogeneous Model Migration Module: When some nodes experience situations such as long task queues, resource shortages, or insufficient model instances, the module evaluates the inference benefits and network costs of task migration and migrates tasks to other nodes with homogeneous model instances to achieve load balancing and reduce inference waiting time.

[0036] S5, Heterogeneous Model Collaborative Scheduling Module: For heterogeneous models with different structures and capabilities deployed in the cloud-edge system, it selects the optimal model instance for inference among multiple nodes by comprehensively considering factors such as inference accuracy, execution latency, and network transmission overhead, thereby achieving global collaborative optimization of heterogeneous resources.

[0037] S6, Service Quality Fairness Assurance Module: Constructs task fairness indicators based on task inference results, resource allocation ratios, and migration behavior; when a task fails to meet the minimum accuracy requirements or resource allocation is unbalanced, it improves the overall service quality fairness and operational stability of the system by dynamically adjusting model selection weights, scheduling strategies, and inter-node collaboration methods.

[0038] Preferably, in the task request perception module described in step S1, the task and node status data are updated periodically by continuously monitoring the CPU / GPU utilization, memory capacity, number of model instances, and network latency of the nodes to support real-time scheduling decisions.

[0039] Preferably, in the online model evaluation module described in step S2, a performance prediction function based on historical inference samples is constructed to achieve continuous learning and dynamic updating of model inference performance, thereby improving the accuracy and adaptability of model performance evaluation.

[0040] Preferably, in the model fair selection module described in step S3, the model selection process is comprehensively decided based on factors such as task accuracy requirements, task priority, node load status, and model selection weight, so that the model allocation process has transparency and stability.

[0041] Preferably, in the isomorphic model migration module described in step S4, by setting a task migration threshold, the migration operation is ensured to be carried out under the premise of improving inference performance and node load level, thus ensuring the rationality of the migration behavior.

[0042] Preferably, in the heterogeneous model collaborative scheduling module described in step S5, the system can flexibly select high-precision or high-performance models under different task scenarios through cross-node deep model capability analysis and task matching mechanism, thereby improving the overall reasoning ability and operating efficiency.

[0043] Preferably, in the service quality fairness assurance module described in step S6, a task fairness index is constructed by statistically analyzing the actual inference accuracy, resource allocation ratio, and migration call count of each task. When it is detected that some tasks have not met the minimum accuracy requirements or the resource allocation is unbalanced, the system automatically adjusts the model selection weight, node load factor, and migration triggering conditions to enhance the overall fairness among tasks and maintain the long-term stable operation of the system. Specific Implementation Example 1

[0045] Please refer to Figure 1 and Figure 2 This embodiment provides a multi-task deep model resource scheduling mechanism in a cloud-edge collaborative environment. This method achieves dynamic mapping between tasks, models, and nodes through functional modules such as task and node state awareness, model performance evaluation, node internal optimization, homogeneous collaborative scheduling, and heterogeneous collaborative scheduling, ensuring overall service quality.

[0046] I. System Structure Description

[0047] like Figure 1 As shown, this system mainly includes a task input layer, a node resource management layer, a deep model instance layer, and a scheduling decision layer. After multiple tasks arrive, they are characterized based on their accuracy requirements, latency constraints, and data types. Edge nodes report their current computing power, memory, bandwidth, and model deployment status to the scheduling layer through a periodic status reporting mechanism. The system constructs a task-model-node selection relationship based on task and node information for subsequent decision-making.

[0048] The system architecture includes the following key features: (1) Node heterogeneity: Different nodes have different computing power and model deployment. (2) Task differentiation: Tasks have their own minimum accuracy requirements and maximum allowable latency. (3) State awareness: The system is aware of resource status and task rate in real time. (4) Hierarchical decision-making: This invention adopts a two-layer scheduling structure of intra-node selection and inter-node cooperation.

[0049] II. Scheduling Process Description

[0050] like Figure 2 As shown, the scheduling process of the present invention includes the following steps:

[0051] S1: Task and node status awareness, real-time updates on task requirements and resource usage.

[0052] S2: Model performance estimation, which uses accuracy prediction, cost estimation and other methods to obtain the feasibility of the model at different nodes.

[0053] S3: Node-specific optimization: Select the local optimal model instance based on task characteristics and node resources.

[0054] S4: Homogeneous collaborative scheduling, when multiple nodes deploy the same model, load balancing is achieved through task migration.

[0055] S5: Heterogeneous collaborative scheduling improves the overall availability of the system by migrating across models when node models are inconsistent.

[0056] S6: Fairness adjustment, ensuring service fairness through mechanisms such as dynamic penalties and resource weight adjustments.

[0057] The above steps constitute the entire process mechanism from task input to final inference execution. Specific Implementation Example 2

[0059] This embodiment further illustrates the mathematical modeling method, scheduling optimization strategy, and experimental verification process of the present invention.

[0060] I. Task and Resource Modeling

[0061] 1. Parameter Definition

[0062] Let I be the set of tasks, J be the set of nodes, and K be the set of models. The arrival rate of task i at node j is... The set of available models for node j is { , , …}.

[0063] 2. Task Accuracy Model

[0064] The accuracy of model k in processing task i is denoted as . The cumulative inference accuracy is:

[0065] (1)

[0066] Where K represents the set of depth model types available in the system; This represents the level of unit inference accuracy that deep model k can achieve when processing task i; The decision variable is selected for the model to indicate whether task i selects model k at node j to perform inference. It takes a non-zero value when task i selects model k, and a zero value otherwise.

[0067] The minimum accuracy constraint for the task is:

[0068] (2)

[0069] in, The minimum unit inference accuracy requirement for task i is given in advance by the task's quality of service requirements; This represents the arrival rate of inference requests for task i on node j, or the task size coefficient, used to characterize the task load intensity.

[0070] 3. Node Resource Model

[0071] The remaining available computing resources for node j in the current scheduling period are:

[0072] (3)

[0073] in, This represents the total resource budget for node j, including computing power, video memory, or equivalent resource capacity. This represents the unit resource cost required to deploy one instance of model k; This indicates the number of model k instances deployed on node j.

[0074] 4. Optimization Objectives

[0075] This invention employs a joint objective function that combines task accuracy improvement and resource overhead control:

[0076] (4)

[0077] Where I represents the set of tasks in the system; This represents the cumulative inference accuracy achieved by task i at node j; This represents the adjustment function of the remaining resources of node j on the overall revenue, used to characterize the weight of the impact of the remaining computing resources of a node on scheduling revenue.

[0078] The global optimization form is:

[0079] (5)

[0080] Where J represents the set of nodes participating in the coordinated scheduling in the system; denoted as the comprehensive benefit function of node j; x represents the set of all task-model selection decision variables; y represents the set of all node-model instance deployment variables; Formula (5) is used to describe the optimization objective of maximizing the overall service quality and resource utilization efficiency of the system under various constraints.

[0081] II. In-node model selection strategy

[0082] 1. Logarithmic transformation simplifies the objective.

[0083] The additable optimization objective is obtained through logarithmic transformation:

[0084] (6)

[0085] Where I represents the set of tasks to be scheduled in the system. j represents the index of the computing node participating in collaborative reasoning, including cloud nodes and edge nodes; k represents the index of the available deep model instance on the node; The decision variable represents whether task i selects the k-th deep model instance at node j; This indicates that model instance k represents the inference accuracy or quality of service benefit provided by task i; This represents the minimum precision or quality of service requirement threshold for task i. This represents the constraint penalty coefficient when task i is executed on node j, used to adjust the penalty intensity on the objective function when the model accuracy is lower than the task requirements; This represents the total computing resource capacity of node j; This represents the unit resource consumption of model instance k; The variable representing the deployment decision for node j is whether the k-th deep model instance is enabled. This represents the total amount of resources consumed by the currently deployed or activated model instances at node j. This represents the adjustment value of the remaining resources of node j in the current model deployment state to the overall scheduling benefit of the system. When the node has more remaining resources, this function takes a larger value to encourage tasks to continue to be allocated to this node. It is a logarithmic function used to balance the growth rates of different benefit items, preventing a single benefit item from dominating the optimization process, thereby improving the stability and fairness of the scheduling process.

[0086] 2. Online learning model

[0087] Implemented using an improved ε-greedy strategy:

[0088] (1) Real-time updates of model accuracy and latency overhead;

[0089] (2) Balancing exploration and utilization;

[0090] (3) Maintain convergence stability under load fluctuations.

[0091] III. Homogeneous Load Balancing Strategy

[0092] When multiple nodes use the same model type, this invention employs a water level balancing task migration algorithm, defining the current task load deviation of node i as:

[0093] (7)

[0094] Where TaskNum[i] represents the number of tasks currently carried by node i; Cap[i] represents the theoretical processing capacity or capacity limit of node i; AveUtil represents the current average resource utilization level of the system; this deviation value is used to guide the task migration decision between isomorphic models, so that the system load gradually tends to a balanced state.

[0095] By migrating tasks, the node load is brought closer to a balanced state, thereby: (1) avoiding node overload; (2) improving the success rate; and (3) enhancing the overall resource utilization rate.

[0096] IV. Heterogeneous Collaborative Migration Mechanism

[0097] When the model deployments between nodes are inconsistent, the present invention performs the following steps:

[0098] (1) Determine the source node and the migrated tasks;

[0099] (2) Determine whether the target node resources can accommodate the target model;

[0100] (3) Conduct migration simulations to assess changes in system benefits;

[0101] (4) If the profit increases, the migration will be executed; otherwise, the original state will be maintained.

[0102] This mechanism is suitable for complex collaborative reasoning scenarios involving multiple tasks, multiple models, and multiple nodes.

[0103] V. Verification of Experimental Results

[0104] The performance of the proposed method is evaluated on a typical multi-task dataset, including 5 task classes, 4 nodes, and 3 model types. Comparisons are made with traditional methods and local methods.

[0105] Table 1 shows the comparison of different scheduling methods in terms of average accuracy, resource overhead, fairness, and task success rate. The results show that the method of the present invention significantly improves the overall fairness and success rate of the system while ensuring accuracy, thus verifying the effectiveness of the proposed scheduling mechanism in a multi-task cloud-edge collaborative environment.

[0106] Table 1 System Performance Comparison

[0107] method AA (↑) AC (↓) FI (↑) succRatio (↑) Maximum Profit Method 0.7588 <![CDATA[ 0.0972 ]]> 0.4572 0.9206 Minimum fair allocation method 0.7596 0.1046 0.5181 0.8748 Single-node optimization method 0.7715 0.1033 0.5531 0.9141 Method of the present invention <![CDATA[ 0.7933 ]]> 0.1083 <![CDATA[ 0.7178 ]]> <![CDATA[ 0.9772 ]]>

[0108] Table 2 shows the system performance changes when different functional mechanisms are not involved in scheduling decisions. The results indicate that online selection, homogeneous collaboration, and heterogeneous collaboration are all key scheduling mechanisms; their absence leads to varying degrees of decline in core indicators such as average accuracy, fairness, and request success rate. Compared with the complete scheme, the method of this invention achieves higher accuracy, better resource utilization, and significantly improved scheduling fairness through the synergistic effect of the three mechanisms, while also resulting in more stable overall system operation. Among these, the heterogeneous collaboration mechanism makes the most significant contribution to resource efficiency and performance gains, further validating the effectiveness of the scheduling system of this invention.

[0109] Table 2 Ablation Experiment

[0110] method AA (↑) AC (↓) FI (↑) succRatio (↑) No online selection 0.8042 0.1056 0.6345 0.9117 Non-isomorphic collaboration <![CDATA[ 0.8085 ]]> <![CDATA[ 0.1004 ]]> 0.6676 0.7945 Heterogeneous collaboration 0.7848 0.1258 0.5059 0.9327 Method of the present invention 0.7933 0.1083 <![CDATA[ 0.7178 ]]> <![CDATA[ 0.9772 ]]>

[0111] Experiments have verified that the multi-layer collaborative mechanism of this invention plays a key role in performance improvement.

[0112] VI. Summary of Implementation Methods

[0113] In summary, this invention focuses on cloud-edge collaborative multi-task inference scenarios. Through a series of collaborative mechanisms, including task state awareness, precision-cost joint modeling, intra-node model selection, online learning optimization, homogeneous load balancing, and heterogeneous collaborative scheduling, it achieves fair scheduling of deep model resources under conditions of precision lower bound, resource budget, and latency constraints. The joint objective modeling method, decomposable online solution strategy, and two-level collaborative scheduling system proposed in this invention have been experimentally verified to significantly outperform comparative methods in core indicators such as average precision, fairness, and request success rate, demonstrating the stability and effectiveness of the method in multi-node, heterogeneous model environments.

[0114] The embodiments of the present invention have been described in detail above, but the present invention is not limited to the described embodiments. For those skilled in the art, various changes, modifications, substitutions, and variations made to these embodiments without departing from the principles and spirit of the present invention still fall within the protection scope of the present invention.

Claims

1. A deep model-based fair scheduling method for multi-task service quality assurance in a cloud-edge collaborative environment, characterized in that, The method includes the following steps: S1, Task Request Awareness Module: Real-time collection of task request stream, minimum inference accuracy requirements of tasks, resource usage of each node, running status of model instances and network link status, and construction of task status table and node resource table to provide real-time input for subsequent model selection and scheduling decisions; S2, Online Model Evaluation Module: Based on the structural complexity of deep models, historical inference performance records, and the current node running status, it dynamically estimates the inference accuracy, execution latency, and computational cost of each deep model on different tasks, and generates a task-model performance evaluation matrix to reflect the suitability of different models for different tasks. S3, Model Fair Selection Module: Based on the minimum inference accuracy threshold of the task and the performance evaluation results, select model instances that can meet the service quality requirements for each task; when the local node cannot meet the task accuracy or resource requirements, trigger the cross-node model selection process to ensure the continuous quality of service for the task. S4, Homogeneous Model Migration Module: When some nodes experience situations such as long task queues, resource shortages, or insufficient model instances, the module evaluates the inference benefits and network costs of task migration and migrates tasks to other nodes with homogeneous model instances to achieve load balancing and reduce inference waiting time. S5, Heterogeneous Model Collaborative Scheduling Module: For heterogeneous models with different structures and capabilities deployed in the cloud-edge system, it selects the optimal model instance for inference among multiple nodes by comprehensively considering factors such as inference accuracy, execution latency, and network transmission overhead, thereby achieving global collaborative optimization of heterogeneous resources. S6, Service Quality Fairness Assurance Module: Constructs task fairness indicators based on task inference results, resource allocation ratios, and migration behavior; when a task fails to meet the minimum accuracy requirements or resource allocation is unbalanced, it improves the overall service quality fairness and operational stability of the system by dynamically adjusting model selection weights, scheduling strategies, and inter-node collaboration methods.

2. The deep model resource fair scheduling method for multi-task service quality assurance in a cloud-edge collaborative environment according to claim 1, characterized in that, In the task request awareness module described in step S1, the CPU / GPU utilization, memory capacity, number of model instances, and network latency of the nodes are continuously monitored, and task and node status data are updated periodically to support real-time scheduling decisions.

3. The deep model resource fair scheduling method for multi-task service quality assurance in a cloud-edge collaborative environment according to claim 2, characterized in that, In the online model evaluation module described in step S2, a performance prediction function based on historical inference samples is constructed to achieve continuous learning and dynamic updating of model inference performance, thereby improving the accuracy and adaptability of model performance evaluation.

4. The deep model resource fair scheduling method for multi-task service quality assurance in a cloud-edge collaborative environment as described in claim 3, characterized in that, In the model fair selection module described in step S3, the model selection process is comprehensively decided based on factors such as task accuracy requirements, task priority, node load status, and model selection weight, so that the model allocation process has transparency and stability.

5. The deep model resource fair scheduling method for multi-task service quality assurance in a cloud-edge collaborative environment according to claim 4, characterized in that, In the isomorphic model migration module described in step S4, by setting a task migration threshold, the migration operation is ensured to be carried out under the premise of improving inference performance and node load level, thus guaranteeing the rationality of the migration behavior.

6. The deep model resource fair scheduling method for multi-task service quality assurance in a cloud-edge collaborative environment according to claim 5, characterized in that, In the heterogeneous model collaborative scheduling module described in step S5, the system can flexibly select high-precision or high-performance models under different task scenarios through cross-node deep model capability analysis and task matching mechanism, thereby improving the overall reasoning ability and operating efficiency.

7. The deep model resource fair scheduling method for multi-task service quality assurance in a cloud-edge collaborative environment as described in claim 6, characterized in that, In the service quality fairness assurance module described in step S6, a task fairness index is constructed by statistically analyzing the actual inference accuracy, resource allocation ratio, and migration call count of each task. When it is detected that some tasks have not met the minimum accuracy requirements or the resource allocation is unbalanced, the system automatically adjusts the model selection weight, node load factor, and migration triggering conditions to enhance the overall fairness among tasks and maintain the long-term stable operation of the system.

Images