Artificial intelligence computing resource collaborative scheduling method and system oriented to edge nodes

Abstract

The invention relates to the technical field of data analysis, and particularly discloses an edge node-oriented artificial intelligence computing resource collaborative scheduling method and system, and the method comprises the steps: obtaining a target execution task, allocating tasks based on edge node resource collaborative scheduling capability, and analyzing scheduling process parameters to adjust allocation; monitoring task execution of the edge nodes, and evaluating the quality index set to optimize execution parameters; collecting and integrating execution results, and judging whether to give an early warning or not after comparison; the core represents the edge node scheduling capability by calculating a collaborative stability factor, dynamically adjusts task distribution, and optimizes abnormal edge node parameters based on a task execution quality evaluation index; dynamically adapting a scheduling strategy according to the dynamic nature of an edge environment; high-quality edge nodes are preferentially used, and resource overload is avoided; self-adaptive optimization is realized in combination with an artificial intelligence model, the overall performance and stability of the system are improved, and the method is suitable for complex edge scenes.

Description

Technical Field

[0001] This invention relates to the field of data analysis technology, specifically to a collaborative scheduling method and system for artificial intelligence computing resources oriented towards edge nodes. Background Technology

[0002] With the widespread adoption of edge computing, the computing tasks undertaken by edge nodes such as smart terminals and IoT sensors are becoming increasingly complex. Edge nodes face resource constraints, network instability, and highly heterogeneous task requirements, making efficient resource scheduling crucial for improving system performance. Existing technologies employ two main approaches: firstly, they achieve rapid resource allocation through predefined rules and utilize weighted summation algorithms or Pareto front analysis to seek a dynamic balance between task execution latency, system energy consumption, and resource allocation fairness; secondly, they collect historical node operation data to construct reinforcement learning models for intelligent decision-making on resource scheduling strategies; and thirdly, existing technologies combine containerized deployment technology with preemptive scheduling mechanisms to enhance the adaptability and resource utilization efficiency of edge nodes.

[0003] For example, Chinese invention patent CN116909742A discloses a multi-agent collaborative computing resource scheduling method, apparatus, and system, which includes configuring a task request queue; acquiring the historical running state sequence of the edge cluster, inputting it into a multi-layer superimposed LSTM neural network, and outputting a system state matrix with temporal characteristics; inputting the system state matrix into the Actor-Critic network configured in the edge cluster for multi-agent reinforcement learning, calculating the state value of the edge cluster, and selecting a suitable node from the corresponding edge cluster to handle the next task in the task request queue, thus completing resource scheduling; and then updating the Actor-Critic network parameters based on the task reward by calculating the loss function and gradient.

[0004] For example, Chinese invention patent CN112579283B discloses a resource scheduling method and system for edge nodes of the power Internet of Things (IoT). This method obtains the power of each application container in all edge nodes of the power IoT system and the electrical energy value of the power IoT system; determines the current operating mode based on the power of each application container; determines the power consumption value of the power IoT system based on the power of each application container under the current operating mode; and adjusts the current operating mode according to the relationship between the power consumption value and the electrical energy value until the power consumption value and the electrical energy value meet preset requirements. Through long-term accumulation and optimization of data records, it achieves edge computing and edge collaboration under low power consumption constraints.

[0005] The aforementioned technologies have at least the following technical problems: On the one hand, existing technologies focus on resource scheduling under a single constraint, which limits their ability to optimize multi-dimensional resource constraints; although historical time-series features are extracted using LSTM, the model input is mainly historical data; on the other hand, existing technologies exhibit significant heterogeneity in the computing resources and energy supply of edge nodes, and the task allocation strategies of existing methods are easily affected by short-term resource fluctuations of edge nodes, resulting in low resource utilization of some edge nodes and overload of others, making it difficult to achieve optimal resource allocation. Summary of the Invention

[0006] To address the limited resource scheduling and optimization capabilities of existing technologies, this invention provides a method and system for collaborative scheduling of artificial intelligence computing resources for edge nodes. The technical solution is as follows:

[0007] On the one hand, a collaborative scheduling method for AI computing resources for edge nodes is provided, including: Step 1, obtaining target execution tasks from the task request queue, and allocating the target execution tasks to each edge node based on the resource collaborative scheduling capabilities of AI computing; Step 2, processing the target execution tasks through each edge node, monitoring the task execution process of each edge node, collecting and evaluating the task execution quality index set of each edge node, and determining whether to optimize the task execution parameters of each edge node; Step 3, collecting the execution results of each edge node, integrating the execution results marked as target execution tasks, and comparing the execution results to determine whether to issue an early warning for the target execution tasks.

[0008] On the other hand, an AI computing resource collaborative scheduling system for edge nodes is provided, including: a target execution task acquisition module, used to acquire target execution tasks from the task request queue, and allocate target execution tasks to each edge node based on the resource collaborative scheduling capabilities of AI computing; a task execution parameter optimization module, used to process target execution tasks through each edge node, monitor the task execution process of each edge node, collect and evaluate the task execution quality index set of each edge node, and thus determine whether to optimize the task execution parameters of each edge node; and an edge node execution result collection module, used to collect the execution results of each edge node, integrate the execution results marked as target execution tasks, and compare the execution results to determine whether to issue an early warning for the target execution task.

[0009] The beneficial effects of the technical solutions provided in the embodiments of the present invention include at least the following:

[0010] (1) This invention provides a collaborative scheduling method and system for artificial intelligence computing resources for edge nodes, which closely matches the "dynamic" characteristics of the edge environment. By analyzing the collaborative stability factor of each target edge node, comparing the total number of pre-allocated edge nodes for the target execution task with the total number of currently schedulable edge nodes, and monitoring the task execution quality evaluation index of each edge node, the method dynamically adjusts the target execution task edge node allocation strategy and edge node resource configuration to ensure that the resource scheduling strategy always matches the current scenario requirements. It is suitable for edge scenarios with large fluctuations in task requirements, strong heterogeneity of node resources, and high requirements for latency and reliability. Its advantages are not only reflected in the efficient execution of a single task, but also in the continuous optimization of the overall performance and stability of the edge node cluster through long-term quality monitoring and strategy iteration.

[0011] (2) This invention analyzes the collaborative stability factor of each target edge node and, based on the average collaborative stability factor of the currently schedulable nodes, obtains the pre-allocated task amount. If the pre-allocated amount exceeds the total number of available nodes, the edge nodes are optimized by intelligently adjusting the granularity of the sharding and the number of CPU cores through "batch processing + intelligent adjustment", and the batch processing size is automatically reduced to avoid task accumulation due to insufficient nodes. This dynamic adjustment mechanism avoids resource overload and ensures that tasks are executed efficiently within the node's capacity.

[0012] (3) This invention analyzes task allocation, segmentation adjustment and failure reasons based on artificial intelligence model, and analyzes the execution completion rate of target execution tasks in combination with artificial intelligence model. It achieves automatic early warning by comparing the execution completion rate with the threshold value, replacing manual judgment, improving the efficiency and accuracy of result evaluation, accurately identifying task execution risks, automatically isolating abnormal nodes when the execution completion rate is lower than the threshold value, preventing the spread of local faults, supporting cross-node task dependency management, and significantly improving the system's adaptability to complex edge scenarios. Attached Figure Description

[0013] 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.

[0014] Figure 1 This is a schematic diagram of the steps of the collaborative scheduling method for artificial intelligence computing resources oriented towards edge nodes provided in an embodiment of the present invention.

[0015] Figure 2 This is a schematic diagram of the structure of the AI ​​computing resource collaborative scheduling system for edge nodes provided in an embodiment of the present invention;

[0016] Figure 3This is a schematic diagram of the task allocation and adjustment process provided in an embodiment of the present invention;

[0017] Figure 4 This is a schematic diagram of the task execution parameter optimization process provided in an embodiment of the present invention;

[0018] Figure 5 This is a schematic diagram of the execution result evaluation and early warning process provided in the embodiments of the present invention. Detailed Implementation

[0019] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0020] Reference Figure 1 As shown, the first aspect of the present invention provides a method for collaborative scheduling of artificial intelligence computing resources for edge nodes, comprising: Step 1, obtaining a target execution task from a task request queue, and allocating the target execution task to each edge node based on the resource collaborative scheduling capability of each edge node using artificial intelligence computing; Step 2, processing the target execution task through each edge node, monitoring the task execution process of each edge node, collecting and evaluating the task execution quality index set of each edge node, and determining whether to optimize the task execution parameters of each edge node; Step 3, collecting the execution results of each edge node, integrating the execution results marked as the target execution task, and comparing the execution results to determine whether to issue an early warning for the target execution task.

[0021] Specifically, the resource coordination and scheduling capabilities of each edge node are calculated based on artificial intelligence. The specific analysis process is as follows: the artificial intelligence model analyzes historical data and collects the status information of each edge node. The status information includes: the historical task throughput entropy value of each edge node, the scheduling initialization latency of each edge node, and the energy consumption-performance balance value of each edge node.

[0022] The historical task throughput entropy value represents the fluctuation in the number of tasks completed by edge nodes per unit time during historical task processing. Its stability is quantified by information entropy. The task throughput data of edge nodes within a fixed time window in the past is recorded, and the throughput is discretized into k intervals. The probability p of the occurrence of each interval is calculated, and the historical task throughput entropy value is obtained by using the Shannon entropy formula. The scheduling initialization latency refers to the time interval from when a task is issued from the scheduling center to when the edge node completes the initialization of the local scheduler. The scheduling system records the "timestamp of task allocation instruction issuance" and the "timestamp of edge node returning initialization completion confirmation". The initialization latency of a single scheduling is the initialization completion time minus the task allocation time. The average of the last M scheduling timestamps is taken as the scheduling initialization latency. The energy consumption-performance balance value measures the comprehensive balance between energy consumption and performance of edge nodes in task processing. It is defined as "effective performance output per unit of energy consumption". The real-time energy consumption and task performance of edge nodes are monitored synchronously, and the ratio of performance to energy consumption is calculated. Specifically, performance is divided by energy consumption to obtain the energy consumption-performance balance value.

[0023] Based on the weighted ratio coefficients stored in the database, the ratios of the historical task throughput entropy value of each edge node to its corresponding threshold value, the ratio of the scheduling initialization latency of each edge node to its corresponding threshold value, and the ratio of the energy consumption-performance balance value of each edge node to its corresponding threshold value are quantified. In this way, the degree of influence of each ratio on the collaborative stability factor of each edge node is evaluated. By summarizing the degree of influence of each edge node, the collaborative stability factor of each edge node is obtained.

[0024] The collaborative stability factor of each edge node is used to digitally characterize the comprehensive ability of each edge node to coordinate tasks. Its specific expression is as follows:

[0025]

[0026] In the formula, CN i CH is the cooperative stability factor of the i-th edge node. i Let CS be the historical task throughput entropy value of the i-th edge node. i For the scheduling initialization delay of the i-th edge node, CE iLet be the energy consumption-performance balance value of the i-th edge node, C_CH be the predefined historical task throughput entropy value in the database, C_CS be the predefined scheduling initialization delay in the database, C_CE be the predefined energy consumption-performance balance value in the database, ca1 be the predefined weighted ratio coefficient of the historical task throughput entropy value in the database, ca2 be the predefined weighted ratio coefficient of the scheduling initialization delay in the database, ca3 be the predefined weighted ratio coefficient of the energy consumption-performance balance value in the database, i be the node number of each edge node, i = 1, 2, 3, ..., n, and n be the total number of edge nodes.

[0027] The historical task throughput entropy value is defined as the minimum acceptable entropy value standard for edge nodes, used to measure the uncertainty boundary of task processing; the scheduling initialization latency is defined as the minimum latency requirement for task response of edge nodes, serving as a baseline for task startup efficiency; the energy consumption-performance balance value is defined as the maximum allowable value for the overall performance of edge nodes, reflecting the optimal trade-off boundary between energy saving and performance of edge nodes.

[0028] Furthermore, the higher the historical task throughput entropy value, the worse the stability of edge node task processing; the lower the historical task throughput entropy value, the more stable the throughput, and the greater the predictability of the edge node resource scheduling and coordination process. The smaller the scheduling initialization latency, the faster the edge node responds to scheduling commands, and the stronger the coherence of the edge node resource scheduling and coordination process; the larger the scheduling initialization latency, the higher the risk of task startup delay for edge nodes, and the easier it is for tasks to be interrupted during edge node initialization. A higher energy consumption-performance balance value indicates that the node achieves more efficient energy consumption control while meeting performance requirements, and the edge node resource scheduling and coordination process is more efficient. The stronger the sustainability of the process, the lower the energy consumption-performance balance value, and the more prominent the contradiction between energy consumption and performance. If the entropy value of historical task throughput increases, it will cause nodes to frequently adjust resources, increase the resource coordination complexity during scheduling initialization, and thus increase the scheduling initialization delay. Excessive scheduling initialization delay may cause nodes to be in a "semi-active" state for a long time (such as some hardware not being released), increase standby power consumption, and destroy the optimal ratio of energy consumption and performance. The entropy value of historical task throughput and the scheduling initialization delay are negatively correlated with the collaborative stability factor of edge nodes, while the energy consumption-performance balance value is positively correlated with the collaborative stability factor of edge nodes.

[0029] Historical task throughput entropy weighting coefficient: This quantifies the impact of the ratio between the historical task throughput entropy value and its threshold value on the collaborative stability factor of each edge node. The value range of this coefficient is 0 to 1. Scheduling initialization delay weighting coefficient: This quantifies the impact of the ratio between the scheduling initialization delay and its threshold value on the collaborative stability factor of each edge node. The value range of this coefficient is 0 to 1. Energy consumption-performance balance weighting coefficient: This quantifies the energy consumption-performance balance of each edge node. The influence of the proportional relationship between the value and its defined value on the collaborative stability factor of each edge node is considered. The weighted ratio coefficient of the energy consumption-performance balance value ranges from 0 to 1. The above three types of weighted ratio coefficients are all pre-stored in the database with mapping relationship tables to realize the weighted configuration of the three types of parameters. When the historical task throughput entropy value, scheduling initialization latency and energy consumption-performance balance value of a specific edge node are input, the corresponding weighted coefficient can be directly retrieved based on the stored mapping relationship table to dynamically quantify the collaborative stability factor of the edge node, realize the standardized integration of multi-dimensional performance indicators, and ensure the accuracy and scalability of collaborative control of the edge computing cluster.

[0030] Specifically, the target execution tasks are allocated to various edge nodes. The allocation process is as follows: Target execution tasks are retrieved from the task request queue (the highest priority task in the queue); feature extraction is performed on the target execution tasks, analyzing their features to form a task feature vector; based on this vector, the total resource requirement of the target execution task is calculated using a resource demand model in artificial intelligence; key task features are extracted using a convolutional neural network; a gradient boosting tree algorithm is used to output a predicted resource demand; and the total number of currently schedulable edge nodes (i.e., the number of nodes that meet basic operating conditions and are not overloaded) is counted to obtain the required resources for the target execution task. The total amount of resources and the total number of currently schedulable edge nodes are considered. Each currently schedulable edge node is marked as a target edge node. The collaborative stability factor of each schedulable edge node and the total number of target edge nodes are obtained. Based on the collaborative stability factor of each schedulable edge node, the total amount of resources required for the target execution task is allocated. The total number of pre-allocated edge nodes for the target execution task is obtained through analysis using an artificial intelligence resource allocation model. This represents the total number of pre-allocated edge nodes required for the total amount of resources needed for the target execution task. Based on the total number of pre-allocated edge nodes for the target execution task, a comparison is made with the total number of currently schedulable edge nodes. The total number of currently schedulable edge nodes is equal to the total number of target edge nodes.

[0031] The above-mentioned allocation of the total resources required for the target execution task refers to: obtaining the total remaining resources of each currently schedulable edge node, multiplying the total remaining resources of each currently schedulable edge node by the corresponding collaborative stability factor of that edge node to calculate the schedulable resource amount of each currently schedulable edge node, sorting the schedulable resource amounts of each currently schedulable edge node from largest to smallest, and allocating the schedulable resource amounts of the currently schedulable edge nodes to the resources required for the target execution task from top to bottom. If the currently schedulable edge nodes can allocate the total resources required for the target execution task, it indicates that the target execution task does not need to be processed in batches; if the currently schedulable edge nodes cannot allocate the total resources required for the target execution task, it indicates that the target execution task needs to be processed in batches.

[0032] The specific analysis process is as follows: If the total number of pre-allocated edge nodes for the target task is greater than the total number of currently schedulable edge nodes, it indicates a mismatch between the total number of pre-allocated edge nodes and the current resource supply capacity of the edge nodes. In this case, the target task is processed in batches. Batch processing involves dividing the target task into batches with defined batch numbers to reduce the total number of pre-allocated edge nodes. The batch number is increased by determining whether the total number of pre-allocated edge nodes is greater than the total number of currently schedulable edge nodes. This process continues until the total number of pre-allocated edge nodes is less than or equal to the total number of currently schedulable edge nodes. Finally, the batch number is determined, thus updating the required resources for the target task. The total number of pre-allocated edge nodes for the target task is compared with the total number of edge nodes pre-allocated for the target task. The total resource required for the target task is obtained by dividing the total resource required for the target task by the number of batches processed in batches. The total resource required for the first batch of target tasks is then updated to the total resource required for the target task. A second comparison is made with the total number of pre-allocated edge nodes for the target task. If the total number of pre-allocated edge nodes for the target task is less than or equal to the total number of currently schedulable edge nodes, it indicates that the total number of pre-allocated edge nodes for the target task matches the current edge node resource supply capacity. In this case, a second comparison is made with the total number of pre-allocated edge nodes for the target task to further determine whether the total number of pre-allocated edge nodes matches the edge node allocation quantity matched for this task type.

[0033] Furthermore, a second comparison is performed on the total number of pre-allocated edge nodes for the target execution task. The specific comparison process is as follows: Based on an artificial intelligence model, the task type of the target execution task is analyzed. Metadata and contextual information of the target execution task are fused, and fused features are extracted. A CNN-LSTM (Convolutional Neural Network-Long Short-Term Memory Network) hybrid model is used as input for supervised learning to derive the task type of the target execution task. Based on the task type, the number of delimited edge nodes to be allocated is matched from the database. The database stores the number of delimited edge nodes corresponding to the task type of the target execution task. The number of delimited edge nodes is the maximum number of edge nodes that can be allocated based on the task type of the target execution task. If the total number of pre-allocated edge nodes for the target execution task is less than or equal to the number of delimited edge nodes, it indicates that the total number of pre-allocated edge nodes for the target execution task matches the number of edge nodes matched for that task type. Therefore, based on the target execution task... The target execution task is executed with a total number of pre-allocated edge nodes, and each pre-allocated edge node corresponding to the target execution task is marked as an allocated edge node. If the total number of pre-allocated edge nodes for the target execution task is greater than the defined edge node allocation number, it indicates that the total number of pre-allocated edge nodes for the target execution task does not meet the edge node allocation number matched for this task type. In this case, the task sharding granularity is intelligently adjusted based on the collaborative stability factor of each pre-allocated edge node, and the collaborative stability factor of each pre-allocated edge node and the total number of pre-allocated edge nodes for the target execution task are updated. The updated collaborative stability factor of each pre-allocated edge node is marked as the first collaborative stability factor of each pre-allocated edge node, representing the collaborative stability factor of each pre-allocated edge node after the first intelligent adjustment. The total number of pre-allocated edge nodes for the target execution task is marked as the first pre-allocated edge node total number for the target execution task, representing the total number of pre-allocated edge nodes after the first intelligent adjustment.

[0034] The aforementioned intelligent adjustment of task sharding granularity refers to the following: if edge node resources are sufficient, the sharding granularity is reduced by increasing the single-processing task volume and computational complexity of the edge node; if the edge node is close to its load limit, the sharding granularity is increased by reducing the single-processing task volume of the edge node, ensuring load balancing of the edge nodes. The database stores a mapping table between the deviation value of the collaborative stability factor and the change in the single-processing task volume of the edge node. By inputting the deviation value between the collaborative stability factor of each pre-allocated edge node and the average collaborative stability factor of each pre-allocated edge node into the database, the increase or decrease in the single-processing task volume of the edge node can be obtained. The collaborative stability factors of each pre-allocated edge node are summed and divided by the total number of pre-allocated edge nodes to obtain the average collaborative stability factor of each pre-allocated edge node. This increases the collaborative stability factor of each pre-allocated edge node, improves the resource allocation of the edge nodes, and reduces the total number of pre-allocated edge nodes for the target execution task.

[0035] After intelligently adjusting the task granularity, if the total number of pre-allocated edge nodes for the target execution task is less than or equal to the defined edge node allocation number, it indicates that the total number of pre-allocated edge nodes for the target execution task after intelligent adjustment meets the edge node allocation number matched for this task type. In this case, the target execution task is executed based on the total number of pre-allocated edge nodes for the target execution task, and each pre-allocated edge node corresponding to the execution of the target execution task is marked as an allocated edge node. If the total number of pre-allocated edge nodes for the target execution task is greater than the defined edge node allocation number, it indicates that the total number of pre-allocated edge nodes for the target execution task after intelligent adjustment meets the defined edge node allocation number. If the total number of pre-allocated edge nodes still does not meet the required number of edge nodes for the task type, then a secondary intelligent adjustment is made to the number of CPU cores of each pre-allocated edge node, the collaborative stability factor of each pre-allocated edge node and the total number of pre-allocated edge nodes for the target task are updated, and the updated collaborative stability factor of each pre-allocated edge node is marked as the secondary collaborative stability factor of each pre-allocated edge node, representing the collaborative stability factor of each pre-allocated edge node after the secondary intelligent adjustment, and the total number of pre-allocated edge nodes for the target task is marked as the total number of secondary pre-allocated edge nodes for the target task, representing the total number of pre-allocated edge nodes after the secondary intelligent adjustment.

[0036] The above-mentioned intelligent adjustment of the number of CPU cores for each pre-allocated edge node is specifically performed as follows: The difference between the total number of pre-allocated edge nodes for the target task and the number of demarcated edge nodes is multiplied by a CPU core adjustment coefficient to obtain the adjusted CPU core count. A mapping table of CPU core adjustment coefficients corresponding to the difference between the total number of pre-allocated edge nodes for the target task and the number of demarcated edge nodes is stored in the database. The CPU core adjustment coefficient is obtained by retrieving the difference between the total number of pre-allocated edge nodes for the target task and the number of demarcated edge nodes from the database. The adjusted CPU core count is then added to the number of pre-allocated edge nodes for each pre-allocated edge node. The original number of CPU cores at each edge node is used to determine the number of CPU cores for each pre-allocated edge node after intelligent adjustment. If the current number of CPU cores at an edge node is below the maximum number of CPU cores and the task has high real-time requirements, the number of CPU cores is increased to improve parallelism. If the current number of CPU cores at an edge node is high and the task is energy-sensitive, the number of CPU cores is reduced to lower power consumption. This increases and adjusts the collaborative stability factor of each pre-allocated edge node, thereby reducing the total number of pre-allocated edge nodes for the target task. Under the premise of ensuring task execution quality, a smooth and predictable CPU core adjustment path is prioritized, ultimately achieving coordinated optimization of the stability and energy efficiency of edge node resource scheduling.

[0037] If the total number of pre-allocated edge nodes for the target task is less than or equal to the defined edge node allocation number, it indicates that the total number of pre-allocated edge nodes for the target task after intelligent adjustment meets the edge node allocation number matched for this task type. In this case, the target task is executed based on the total number of pre-allocated edge nodes for the target task, and each pre-allocated edge node corresponding to the executed target task is marked as an allocated edge node. If the total number of pre-allocated edge nodes for the target task is still greater than the defined edge node allocation number, it indicates that the total number of pre-allocated edge nodes for the target task after intelligent adjustment still does not meet the edge node allocation number matched for this task type. In this case, an early warning is issued for the task allocation of the edge nodes. The specific warning content is: directly connect to the maintenance work order system, report the specific allocation situation of the target task, trigger cloud resource pre-request, and distribute the task load.

[0038] Figure 3This is a schematic diagram of the task allocation and adjustment process provided in this embodiment of the invention. Through multiple rounds of comparison and adjustment, the reasonable allocation of tasks is ensured. First, the target task to be executed is obtained from the task request queue, features are extracted and the total amount of resources required is determined, the collaborative stability factor of each schedulable edge node is calculated to allocate resources, and the total number of pre-allocated edge nodes is obtained. Then, the number of pre-allocated nodes is compared with the number of currently schedulable nodes. If the former is larger, it is processed in batches to adapt to the resources. Then, the number of edge nodes allocated is defined based on the task type, and the number of pre-allocated nodes is compared again. If it still exceeds the defined number, the task sharding granularity is adjusted first. If it is ineffective, the number of CPU cores is adjusted. Finally, the task allocation is completed or an allocation warning is triggered, and the task execution stage begins.

[0039] This invention analyzes the collaborative stability factor of each target edge node and, based on the average collaborative stability factor of currently schedulable nodes, determines the pre-allocated task volume. If the pre-allocated volume exceeds the total number of available nodes, it optimizes the edge nodes by intelligently adjusting the granularity of the sharding and the number of CPU cores through "batch processing + intelligent adjustment," automatically reducing the batch processing size and avoiding task backlog due to insufficient nodes. This dynamic adjustment mechanism avoids resource overload and ensures that tasks are executed efficiently within the node's capacity.

[0040] Specifically, the task execution quality index set of each edge node is evaluated. The specific analysis process is as follows: obtain the collaborative stability factor of each assigned edge node; the task execution quality index set of the edge node includes the resource fragmentation index of each assigned edge node, the output data integrity assessment value of each assigned edge node, and the error type distribution entropy of each assigned edge node.

[0041] The resource fragmentation index measures the degree of resource dispersion on allocation edge nodes. Resource fragmentation occurs when small, unused blocks of idle resources on allocation edge nodes are frequently allocated and released, preventing large tasks from effectively utilizing them. The index is calculated by using CPU core usage time periods and memory block allocation / release records to obtain the number of idle resource blocks, average block size, and total resource quantity. The product of the number of idle resource blocks and the average block size is divided by the total resource quantity to arrive at the resource fragmentation index. The output data integrity assessment value evaluates the integrity and accuracy of the output data after task execution on allocation edge nodes, reflecting the reliability of task execution. It is determined through a preset verification mechanism using hash verification, comparing task input and output data to obtain the result. The hash consistency score is calculated as follows: if the input and output hash values ​​match exactly, the hash consistency score is 1; if they do not match exactly, the hash consistency score is 0. The error code frequency in the task execution results is statistically analyzed, and the weighted sum of the hash consistency score and the error code frequency is used as the output data integrity assessment value. The error type distribution entropy, based on information entropy theory, reflects the dispersion of error types generated by the allocation edge nodes during task execution. Error logs from the allocation edge nodes are collected, error types are classified and statistically analyzed, and error types are categorized (e.g., defining error type j). The probability of occurrence for each type of error is calculated, and then the error type distribution entropy value is calculated using the Shannon entropy formula.

[0042] By using pre-set weighting coefficients in the database, the proportional relationships between the three types of parameters of each allocation edge node and their corresponding threshold values, as well as the influence of the collaborative stability factor of each allocation edge node on the task execution quality assessment index of each allocation edge node, are quantified. These include the proportional relationship between the resource fragmentation index and the threshold resource fragmentation index of each allocation edge node, the proportional relationship between the output data integrity assessment value and the threshold output data integrity assessment value of each allocation edge node, the proportional relationship between the error type distribution entropy and the threshold error type distribution entropy of each allocation edge node, and the collaborative stability factor of each allocation edge node. The above three proportional relationships and the influence of the collaborative stability factor of each allocation edge node are weighted accordingly and summarized to generate the task execution quality assessment index of the allocation edge node. Based on the task execution quality assessment index of each allocation edge node, it is determined whether the task execution parameters of each allocation edge node should be optimized.

[0043] The task execution quality evaluation index for each assigned edge node is used to characterize the quality of task execution for each assigned edge node. The specific expression is as follows:

[0044]

[0045] In the formula, QI a QR is the quality of execution evaluation index for the task assigned to the a-th edge node. a Let QH be the resource fragmentation index for the a-th allocated edge node. a QP is the output data integrity assessment value for the a-th assigned edge node. a For the error type distribution entropy of the a-th edge node, CNN a Let qa1 be the weighted coefficient of the resource fragmentation index in the database, qa2 be the weighted coefficient of the output data integrity assessment value in the database, qa3 be the weighted coefficient of the error type distribution entropy in the database, qa4 be the weighted coefficient of the collaborative stability factor of the allocated edge node in the database, and qa4 be the node number of each allocated edge node, where a = 1, 2, 3, ..., m, and m is the total number of allocated edge nodes.

[0046] Among them, the resource fragmentation index is defined as the highest tolerable standard for the degree of resource fragmentation of the allocation edge nodes, serving as the upper limit for resource block division in the allocation decision of the allocation edge nodes. The resource fragmentation index of the allocation edge nodes cannot exceed the defined value. The output data integrity assessment value is defined as the minimum guarantee level for the output data integrity of the allocation edge nodes. The output data integrity assessment value of all allocation edge nodes cannot be lower than the defined value. The error type distribution entropy is defined as the maximum allowable value of the error type distribution entropy of the allocation edge nodes. The error type distribution entropy value of the allocation edge nodes cannot exceed the defined value.

[0047] The higher the resource fragmentation index, the higher the degree of fragmentation, and the lower the actual utilization rate of resources allocated to edge nodes. This may lead to a decrease in the scheduling efficiency of target execution tasks or contention for critical resources, failure of output data integrity verification, and the expansion of error types from single resource bottlenecks to multi-dimensional anomalies. The more error types allocated to edge nodes for task execution, the greater the probability of data integrity verification failure, the lower the output data integrity assessment value, and the more diverse errors require more resources for investigation and repair, indirectly exacerbating resource fragmentation. The resource fragmentation index and error type distribution entropy are negatively correlated with the task execution quality assessment index of allocated edge nodes, while the output data integrity assessment value is positively correlated with the task execution quality assessment index of allocated edge nodes.

[0048] Resource fragmentation index weighting coefficient: Used to quantify the impact of the proportional relationship between the resource fragmentation index of each allocation edge node and its threshold value on the task execution quality assessment index of each allocation edge node; Output data integrity assessment value weighting coefficient: Used to quantify the impact of the proportional relationship between the output data integrity assessment value of each allocation edge node and its threshold value on the task execution quality assessment index of each allocation edge node; Error type distribution entropy weighting coefficient: Used to quantify the impact of the proportional relationship between the error type distribution entropy of each allocation edge node and its threshold value on the task execution quality assessment index of each allocation edge node; The proportional weighting coefficient of the collaborative stability factor of the edge node is used to quantify the influence of the collaborative stability factor of each allocated edge node on the task execution quality assessment index of each allocated edge node. The values ​​of the four types of proportional weighting coefficients are all in the range of 0 to 1. The database pre-stores the mapping relationship between the resource fragmentation index, the output data integrity assessment value, the error type distribution entropy, and the collaborative stability factor of each allocated edge node and its corresponding proportional weighting coefficient. Specifically, when the resource fragmentation index, the output data integrity assessment value, the error type distribution entropy, and the collaborative stability factor of each allocated edge node are input, the database can directly retrieve and output the corresponding proportional weighting coefficient.

[0049] Specifically, determining whether to optimize the task execution parameters of each edge node involves the following process: After each edge node executes the target task, its task execution quality assessment index is compared with a task execution quality threshold. The task execution quality threshold is the minimum value of the task execution quality assessment index of each edge node. This comparison is made by acquiring the task execution quality assessment index at different monitoring periods during the task execution process and comparing it with the task execution quality threshold to optimize the target task execution process. If the task execution quality assessment index of an edge node is greater than the task execution quality threshold, it indicates that the edge node... If the task execution of a point-to-point assigned edge node meets the task execution requirements, the task execution results of that edge node are collected to prepare for the next step of evaluating the completion rate of the target task, and it is determined that the task execution parameters of that edge node should not be optimized. If the task execution quality evaluation index of an edge node is less than or equal to the task execution quality threshold, it means that the edge node has not met the task execution requirements, and the edge node is marked as an abnormal edge node. The task execution quality evaluation index of the abnormal edge node is obtained, and the task execution parameters of the abnormal edge node are optimized. This indicates that the edge node has encountered an anomaly during task execution. The task execution parameters of the abnormal edge node are optimized to determine whether a task execution warning is needed for the abnormal edge node.

[0050] The specific execution process of each assigned edge node executing the target task is as follows: trigger the edge node to execute the task instruction, verify the legality of the task and parse the task parameters, evaluate the CPU, memory, bandwidth and other resources of the edge node, allocate resources within the edge node according to priority and process the task, high-priority tasks can preempt resources within the edge node, call local computing power to execute the edge node's task, perform data processing, algorithm inference and device control, track the task progress in real time during task processing, perform local storage and trigger task exception response for abnormalities such as edge node resource exhaustion or timeout, upload to the cloud or cache offline data, and release the computing, storage and network resources within the edge node after the task is completed.

[0051] Furthermore, the task execution parameters of the abnormal edge node are optimized. The specific optimization process is as follows: Based on the task type of the target task and the task execution quality evaluation index of the abnormal edge node, the number of CPU cores of the abnormal edge node is intelligently adjusted once; after the first intelligent adjustment, the task execution quality evaluation index of the abnormal edge node is updated and marked as the first task execution quality evaluation index of the abnormal edge node; if the first task execution quality evaluation index of the abnormal edge node is greater than or equal to the task execution quality threshold, the abnormal mark of the edge node is removed, indicating that after the first intelligent adjustment, the execution of the assigned task by the edge node meets the task execution requirements, and the task execution results of the edge node are collected to prepare for the next step of evaluating the execution completion rate of the target task; if the first task execution quality evaluation index of the abnormal edge node is less than the task execution quality threshold, the number of CPU cores of the abnormal edge node is intelligently adjusted a second time based on the task type of the target task and the first task execution quality evaluation index of the abnormal edge node, and the CPU frequency of the abnormal edge node is also intelligently adjusted.

[0052] The above-mentioned intelligent adjustment of the number of CPU cores of each pre-allocated edge node is specifically performed as follows: The difference between the task execution quality assessment index and the task execution quality threshold of the abnormal edge node is multiplied by the updated CPU core adjustment ratio coefficient to obtain the first adjustment value of the CPU core count for that abnormal edge node. This first adjustment value is then added to the original CPU core count of the abnormal edge node to obtain the CPU core count of each abnormal edge node after the first intelligent adjustment. A mapping table of CPU core adjustment ratio coefficients corresponding to the difference between the task execution quality assessment index and the task execution quality threshold is stored in the database. The CPU core adjustment ratio coefficient can be obtained by retrieving the difference between the task execution quality assessment index and the task execution quality threshold from the database. If the current number of CPU cores of the abnormal edge node is below the maximum number of CPU cores and the task has high real-time requirements, the number of CPU cores is increased to improve parallelism. If the current number of CPU cores of the abnormal edge node is high and the task is energy-sensitive, the number of CPU cores is reduced to lower power consumption. This improves the task execution quality of the edge node, thereby increasing the task execution quality assessment index of the abnormal edge node.

[0053] The aforementioned intelligent adjustment of the number of CPU cores and the intelligent adjustment of the CPU frequency of the abnormal edge node are analyzed as follows: If the task has high real-time requirements and the number of CPU cores of the abnormal edge node has not reached the maximum limit, and the difference between the task execution quality threshold and the first task execution quality evaluation index of the abnormal edge node is greater than the upper limit of the defined difference range, then the increase in the number of CPU cores is further increased to improve parallel processing capability; if the task is energy-sensitive and the current number of cores is redundant, and the difference between the task execution quality threshold and the first task execution quality evaluation index of the abnormal edge node is less than the lower limit of the defined difference range, then the reduction in the number of CPU cores is adjusted to reduce energy consumption while avoiding excessive performance sacrifice; after adjustment, the current number of CPU cores is fixed as the basis for subsequent frequency adjustment.

[0054] A secondary intelligent adjustment is performed on the number of CPU cores of the abnormal edge node. The specific adjustment process is as follows: the difference between the primary task execution quality assessment index and the task execution quality threshold of the abnormal edge node is multiplied by the updated CPU core number adjustment ratio coefficient to obtain the secondary adjustment value of the CPU core number of the abnormal edge node. The secondary adjustment value of the CPU core number of the abnormal edge node is added to the number of CPU cores of the abnormal edge node before the secondary intelligent adjustment to obtain the number of CPU cores of each abnormal edge node after the secondary intelligent adjustment. The database stores a mapping table of CPU core number adjustment ratio coefficients corresponding to the difference between the primary task execution quality assessment index and the task execution quality threshold. The CPU core number adjustment ratio coefficient can be obtained by retrieving the difference between the primary task execution quality assessment index and the task execution quality threshold in the database.

[0055] The CPU frequency of the abnormal edge node is intelligently adjusted. The specific adjustment process is as follows: Based on the analysis of the artificial intelligence model, the CPU frequency adjustment ratio coefficient is obtained. The difference between the task execution quality evaluation index and the task execution quality threshold of the abnormal edge node is multiplied by the CPU frequency adjustment ratio coefficient to obtain the CPU frequency increase adjustment value of the abnormal edge node. The CPU frequency increase adjustment value of the abnormal edge node is added to the CPU frequency of the abnormal edge node before adjustment to obtain the CPU frequency of each abnormal edge node after intelligent adjustment, thereby realizing the increase adjustment of the CPU frequency of each abnormal edge node. The database stores a mapping table of CPU frequency adjustment ratio coefficients corresponding to the difference between the task execution quality evaluation index and the task execution quality threshold. The CPU frequency adjustment ratio coefficient can be obtained by retrieving the difference between the task execution quality evaluation index and the task execution quality threshold in the database.

[0056] The reasons for setting the order of CPU core count and CPU frequency adjustment are as follows: The number of CPU cores determines the parallel processing capability of a node, directly affecting the task sharding strategy and the collaborative stability factor. Therefore, CPU core count adjustment has a higher priority than CPU frequency. Increasing or decreasing the number of CPU cores can quickly change the load on edge nodes, while CPU frequency adjustment requires fine-tuning performance and power consumption under a fixed number of CPU cores. After the number of CPU cores is determined, CPU frequency adjustment is used to fine-tune single-core performance and solve remaining performance bottlenecks. Before the target task is executed, if the total number of pre-allocated edge nodes for the target task is greater than the defined number of edge nodes to be allocated, the number of CPU cores of each pre-allocated edge node is adjusted intelligently in a second round. Allocating CPU cores before task execution can avoid fragmentation of edge node resources, thereby reducing the initial scheduling complexity. During task execution, intelligent adjustment of the number of CPU cores dynamically compensates for the uncertainties in the actual operation of edge nodes, avoiding edge node overload or idleness. By increasing the number of CPU cores to improve edge node performance or decreasing the number of CPU cores to reduce load pressure, the task execution quality evaluation index of edge nodes is increased. Dynamic optimization is achieved through closed-loop feedback, further optimizing the intelligent adjustment of the number of CPU cores.

[0057] After intelligent adjustment is completed, the task execution quality assessment index of the abnormal edge node is updated and marked as the secondary task execution quality assessment index of the abnormal edge node. If the secondary task execution quality assessment index of the abnormal edge node is greater than or equal to the task execution quality threshold, the abnormal mark of the edge node is removed, indicating that the execution of the assigned tasks by the edge node has met the task execution requirements after secondary intelligent adjustment. The task execution results of the edge node are collected to prepare for the next step of evaluating the completion rate of the target execution tasks. If the secondary task execution quality assessment index of the abnormal edge node is still less than the task execution quality threshold, it means that the execution of the assigned tasks by the edge node still does not meet the task execution requirements after secondary intelligent adjustment. An abnormal warning is issued for the abnormal edge node, and the execution tasks of the abnormal edge node are placed in the warning task set. This indicates that the task execution requirements cannot be met by optimizing the resource allocation of the abnormal edge node. Therefore, the abnormal state of the node is marked, and the unqualified tasks executed by it are included in the warning task set to provide a basis for subsequent special processing such as task reallocation, node fault investigation, and resource expansion.

[0058] Obtain the CPU frequency and number of CPU cores of each edge node. Mark the edge nodes whose CPU frequency is greater than the defined CPU frequency or whose number of CPU cores is greater than the defined number of CPU cores as edge nodes that need to be warned and issue a task execution warning. This indicates that the adjustment of the CPU frequency or number of CPU cores is ineffective. The defined CPU frequency is the maximum value within the allowed range of CPU frequency, and the defined number of CPU cores is the maximum value within the allowed range of CPU cores.

[0059] Specifically, the task execution early warning process is as follows: incomplete tasks are automatically migrated to the backup node group, new task assignments on abnormal nodes are suspended to prevent the spread of abnormalities, a unique early warning number is generated, the task number is associated with the edge node number, and early warning information is sent to the management terminal. The early warning information includes edge node status and error logs.

[0060] Figure 4 This is a schematic diagram of the task execution parameter optimization process provided in this embodiment of the invention. It focuses on the parameter optimization of abnormal nodes in task execution. When each node processes a task, it monitors the execution process and collects parameters such as resource fragmentation index to calculate the task execution quality evaluation index. The task execution quality evaluation index is compared with the execution quality evaluation index threshold. If the threshold is met, the result is collected directly; if the threshold is not met, it is marked as an abnormal node. The number of CPU cores is adjusted and the index is updated once. If the threshold is still not met, the number of CPU cores and CPU frequency are adjusted a second time. The execution results of nodes that meet the threshold are collected. Nodes that do not meet the threshold trigger an alert and the task is placed in the alert task set, and then enters the execution result evaluation process.

[0061] This invention analyzes task allocation, partitioning adjustment, and failure causes based on an artificial intelligence model. It combines the analysis of the target task completion rate with the artificial intelligence model to achieve automated early warning by comparing the completion rate with a threshold value, replacing manual judgment, improving the efficiency and accuracy of result evaluation, accurately identifying task execution risks, automatically isolating abnormal nodes when the completion rate is lower than the threshold value to prevent the spread of local faults, supporting cross-node task dependency management, and significantly improving the system's adaptability to complex edge scenarios.

[0062] Furthermore, the specific analysis process for the execution results of the target task is as follows: At the task collection time point, data parameters of the execution results of each edge node are collected. The data parameters of the execution results of each edge node include the basic task identification information, core task execution results, and resource consumption details of each edge node. The data parameters of the execution results of each edge node are integrated and analyzed with an artificial intelligence model to obtain the execution completion rate of the target task. Using a regression model based on historical task data, combined with the data parameters of the execution results of each edge node, the total number of successful task executions of edge nodes in the feature vector is extracted through the regression model. The total number of successful task executions of edge nodes is divided by the total number of task executions of edge nodes to obtain the execution completion rate of the target task, which represents the completion status of the target task.

[0063] Specifically, the process for determining whether to issue an early warning for the target task is as follows: First, obtain the completion rate of the defined task and compare it with the completion rate of the target task. The defined task completion rate represents the minimum acceptable completion rate for the target task. If the target task completion rate is greater than the defined task completion rate, the system reports successful execution of the target task and retrieves the next target task from the task request queue. Second, if the target task completion rate is less than or equal to the defined task completion rate, the system reports a failure, places the target task in the early warning task set, and retrieves the next target task from the task request queue. Third, if the target task completion rate is less than or equal to the defined task completion rate, the system reports a failure, places the target task in the early warning task set, and retrieves the next target task from the task request queue. Fourth, if the target task completion rate is less than or equal to the defined task completion rate, the system reports a failure to the scheduling center, adds the task to the early warning task set for subsequent special processing or rescheduling, and retrieves the next target task from the task request queue.

[0064] Figure 5 This is a schematic diagram of the execution result evaluation and early warning process provided in this embodiment of the invention. The core is to evaluate the overall execution effect of the task. First, the task identifiers, core results and resource consumption details of each node are integrated. The execution completion rate of the target task is calculated by combining the artificial intelligence model and then compared with the defined task execution completion rate. If the execution completion rate of the target task is greater than the defined completion value, the execution is reported as successful and the next task is obtained. If the execution completion rate of the target task is less than or equal to the defined completion value, the execution is reported as failed. The task is placed in the early warning task set and the next task is obtained, forming a closed loop of task processing and completing the result evaluation and connection of the entire scheduling process.

[0065] This invention provides a collaborative scheduling method and system for AI computing resources for edge nodes, closely aligning with the "dynamic" characteristics of edge environments. By analyzing the collaborative stability factors of each target edge node, comparing the total number of pre-allocated edge nodes for the target task with the total number of currently schedulable edge nodes, and monitoring the task execution quality evaluation index of each edge node, the invention dynamically adjusts the edge node allocation strategy and resource configuration for the target task, ensuring that the resource scheduling strategy always matches the current scenario requirements. It is suitable for edge scenarios with large fluctuations in task demands, strong heterogeneity of node resources, and high requirements for latency and reliability. Its advantages are not only reflected in the efficient execution of a single task, but also in the continuous optimization of the overall performance and stability of the edge node cluster through long-term quality monitoring and strategy iteration.

[0066] Reference Figure 2As shown, the second aspect of the present invention provides an artificial intelligence computing resource collaborative scheduling system for edge nodes, including: a module for acquiring target execution tasks, a module for optimizing task execution parameters, a module for collecting execution results from each edge node, and a database.

[0067] The module for obtaining the target execution task is connected to the module for optimizing the task execution parameters. The module for optimizing the task execution parameters is connected to the module for collecting the execution results of each edge node. All three modules are connected to the database.

[0068] The database is used to store the parameters designed in the edge node resource scheduling system.

[0069] The above description is merely an example and illustration of the structure of the present invention. Those skilled in the art can make various modifications or additions to the specific embodiments described, or use similar methods to replace them, as long as they do not deviate from the structure of the invention or exceed the scope defined by the present invention, they should all fall within the protection scope of the present invention.

Claims

1. A collaborative scheduling method for artificial intelligence computing resources for edge nodes, characterized in that, The method includes: Step 1: Obtain the target execution task from the task request queue, and allocate the target execution task to each edge node based on the resource coordination and scheduling capabilities of each edge node using artificial intelligence calculation. Step 2: Process the target execution task through each edge node, monitor the task execution process of each edge node, collect and evaluate the task execution quality index set of each edge node, and thus determine whether the task execution parameters of each edge node need to be optimized. Step 3: Collect the execution results of each edge node, integrate the execution results marked as the target execution task, and compare the execution results to determine whether to issue an early warning for the target execution task.

2. The method for collaborative scheduling of artificial intelligence computing resources for edge nodes according to claim 1, characterized in that: The specific analysis process for the resource collaborative scheduling capability of each edge node based on artificial intelligence is as follows: Collect the status information of each edge node, including: the historical task throughput entropy value of each edge node, the scheduling initialization latency of each edge node, and the energy consumption-performance balance value of each edge node. Based on the weighted ratio coefficients stored in the database, the ratios of the historical task throughput entropy value and its corresponding threshold value, the scheduling initialization delay and its corresponding threshold value, and the energy consumption-performance balance value and its corresponding threshold value of each edge node are quantified. In this way, the degree of influence of each ratio on the collaborative stability factor of each edge node is evaluated. By summarizing the degree of influence of each edge node, the collaborative stability factor of each edge node is obtained. The collaborative stability factor of each edge node is used to digitally characterize the resource collaborative scheduling capability of each edge node.

3. The method for collaborative scheduling of artificial intelligence computing resources for edge nodes according to claim 1, characterized in that: The specific allocation process for distributing the target execution task to each edge node is as follows: Retrieve the target execution task from the task request queue, extract features from the target execution task, and obtain the total amount of resources required for the target execution task and the total number of currently schedulable edge nodes; Obtain the collaborative stability factor of each schedulable edge node and the total number of target edge nodes, and allocate the total amount of resources required for the target execution task based on the collaborative stability factor of each schedulable edge node to obtain the total number of pre-allocated edge nodes for the target execution task. The total number of pre-allocated edge nodes based on the target task is compared and analyzed with the total number of currently schedulable edge nodes. The specific analysis process is as follows: If the total number of pre-allocated edge nodes for the target task is greater than the total number of currently schedulable edge nodes, the target task will be processed in batches until the total number of pre-allocated edge nodes for the target task is less than or equal to the total number of currently schedulable edge nodes. Then, the total amount of resources required for the target task and the total number of pre-allocated edge nodes for the target task will be updated, and a second comparison will be made on the total number of pre-allocated edge nodes for the target task. If the total number of pre-allocated edge nodes for the target task is less than or equal to the total number of currently schedulable edge nodes, then a second comparison is performed on the total number of pre-allocated edge nodes for the target task.

4. The collaborative scheduling method for artificial intelligence computing resources oriented towards edge nodes according to claim 3, characterized in that: The total number of pre-allocated edge nodes for the target task is compared a second time. The specific comparison process is as follows: Based on the analysis of the task type of the target execution task using an artificial intelligence model, the number of edge nodes to be allocated is determined by matching the task type of the target execution task from the database. If the total number of pre-allocated edge nodes for the target execution task is less than or equal to the number of edge nodes allocated, then the target execution task is executed based on the total number of pre-allocated edge nodes for the target execution task, and each pre-allocated edge node corresponding to the execution of the target execution task is marked as an allocated edge node. If the total number of pre-allocated edge nodes for the target task is greater than the number of edge nodes allocated, the task sharding granularity is intelligently adjusted based on the collaborative stability factor of each pre-allocated edge node, and the collaborative stability factor of each pre-allocated edge node and the total number of pre-allocated edge nodes for the target task are updated. The updated collaborative stability factor of each pre-allocated edge node is marked as the first collaborative stability factor of each pre-allocated edge node, and the total number of pre-allocated edge nodes for the target task is marked as the total number of first pre-allocated edge nodes for the target task. After intelligently adjusting the task granularity, if the total number of pre-allocated edge nodes in the target execution task is less than or equal to the number of edge nodes allocated, the target execution task is executed based on the total number of pre-allocated edge nodes in the target execution task, and each pre-allocated edge node corresponding to the execution of the target execution task is marked as an allocated edge node. If the total number of pre-allocated edge nodes for the target task exceeds the number of edge nodes allocated, then the number of CPU cores for each pre-allocated edge node is adjusted in a secondary intelligent manner, the collaborative stability factor of each pre-allocated edge node and the total number of pre-allocated edge nodes for the target task are updated, and the updated collaborative stability factor of each pre-allocated edge node is marked as the secondary collaborative stability factor of each pre-allocated edge node, and the total number of pre-allocated edge nodes for the target task is marked as the total number of secondary pre-allocated edge nodes for the target task. If the total number of secondary pre-allocated edge nodes for the target execution task is less than or equal to the number of edge nodes allocated, then the target execution task is executed based on the total number of secondary pre-allocated edge nodes for the target execution task, and each secondary pre-allocated edge node corresponding to the execution of the target execution task is marked as an allocated edge node. If the total number of pre-allocated edge nodes for the target task is still greater than the number of edge nodes to be allocated, an early warning will be issued for the task allocation of edge nodes.

5. The method for collaborative scheduling of artificial intelligence computing resources for edge nodes according to claim 1, characterized in that: The specific analysis process for evaluating the task execution quality index set of each edge node is as follows: Obtain the collaborative stability factor of each assigned edge node; The task execution quality index set for edge nodes includes the resource fragmentation index of each allocated edge node, the output data integrity assessment value of each allocated edge node, and the error type distribution entropy of each allocated edge node. By using pre-set proportional weighting coefficients in the database, the proportional relationships between the three types of parameters of each allocation edge node and their corresponding boundary values, as well as the influence of the collaborative stability factor of each allocation edge node on the task execution quality assessment index of each allocation edge node, are quantified. Specifically, this includes the proportional relationship between the resource fragmentation index and the boundary resource fragmentation index of each allocation edge node, the proportional relationship between the output data integrity assessment value and the boundary output data integrity assessment value of each allocation edge node, the proportional relationship between the error type distribution entropy and the boundary error type distribution entropy of each allocation edge node, and the collaborative stability factor of each allocation edge node. The above three proportional relationships and the influence of the collaborative stability factor of each allocation edge node are proportionally weighted and summarized to generate the task execution quality assessment index of the allocation edge node. Based on the task execution quality assessment index of each allocation edge node, it is determined whether the task execution parameters of each allocation edge node should be optimized. The task execution quality evaluation index of each assigned edge node is used to characterize the quality of each assigned edge node during task execution.

6. The method for collaborative scheduling of artificial intelligence computing resources for edge nodes according to claim 5, characterized in that: The determination of whether to optimize the task execution parameters of each edge node specifically involves determining whether to optimize the task execution parameters of each assigned edge node. The specific determination process is as follows: After each assigned edge node performs the target task, the task execution quality evaluation index of each assigned edge node is compared with the task execution quality threshold. If the task execution quality evaluation index of a certain allocation edge node is greater than the task execution quality threshold, the task execution results of that allocation edge node will be collected, and it will be determined that the task execution parameters of that allocation edge node will not be optimized. If the task execution quality evaluation index of a certain allocation edge node is less than or equal to the task execution quality threshold, then the allocation edge node is marked as an abnormal edge node. Obtain the task execution quality assessment index of the abnormal edge node and optimize the task execution parameters of the abnormal edge node.

7. The method for collaborative scheduling of artificial intelligence computing resources for edge nodes according to claim 6, characterized in that: The optimization of the task execution parameters for the abnormal edge node is specifically as follows: Based on the task type of the target task and the task execution quality evaluation index of the abnormal edge node, the number of CPU cores of the abnormal edge node is intelligently adjusted. After one intelligent adjustment, update the task execution quality assessment index of the abnormal edge node and mark it as the one-time task execution quality assessment index of the abnormal edge node; If the task execution quality assessment index of the abnormal edge node is greater than or equal to the task execution quality threshold, the abnormal label of the edge node is removed, and the execution result of the target execution task of the edge node is collected. If the primary task execution quality assessment index of the abnormal edge node is less than the task execution quality threshold, then the number of CPU cores of the abnormal edge node is intelligently adjusted again based on the task type of the target task and the primary task execution quality assessment index of the abnormal edge node, and the CPU frequency of the abnormal edge node is intelligently adjusted. After the intelligent adjustment is completed, the task execution quality assessment index of the abnormal edge node is updated and marked as the secondary task execution quality assessment index of the abnormal edge node. If the secondary task execution quality evaluation index of the abnormal edge node is greater than or equal to the task execution quality threshold, then the abnormal label of the edge node is removed, and the execution result of the target execution task of the edge node is collected. If the secondary task execution quality evaluation index of the abnormal edge node is still less than the task execution quality threshold, then an abnormal warning will be issued for the abnormal edge node, and the execution task of the abnormal edge node will be placed in the warning task set. Obtain the CPU frequency and number of CPU cores of each edge node, mark the edge nodes with CPU frequency greater than the defined CPU frequency or number of CPU cores as edge nodes requiring warning, and issue task execution warnings accordingly.

8. The method for collaborative scheduling of artificial intelligence computing resources for edge nodes according to claim 1, characterized in that: The specific analysis process of the execution result of the target task is as follows: At the task collection time point, data parameters of the execution results of each edge node are collected. The data parameters of the execution results of each edge node include the basic task identification information, core task execution results and resource consumption details of each edge node. The data parameters of the execution results of each edge node are integrated and analyzed with artificial intelligence models to obtain the execution completion rate of the target task.

9. The method for collaborative scheduling of artificial intelligence computing resources for edge nodes according to claim 1, characterized in that: The specific determination process for deciding whether to issue a warning to the target task execution is as follows: Obtain the completion rate of the defined task and compare it with the completion rate of the target task; If the target task completion rate is greater than the defined task completion rate, then the target task is reported as successfully executed, and the next target task is retrieved from the task request queue. If the target task completion rate is less than or equal to the defined task completion rate, the target task will be reported as having failed, the target task will be placed in the warning task set, and the next target task will be retrieved from the task request queue.

10. An AI computing resource collaborative scheduling system for edge nodes, wherein the AI ​​computing resource collaborative scheduling system for edge nodes is used to implement the AI ​​computing resource collaborative scheduling method for edge nodes as described in any one of claims 1-9, characterized in that, The system includes: The target execution task acquisition module is used to acquire target execution tasks from the task request queue and, based on artificial intelligence computing, allocate the target execution tasks to each edge node in a coordinated scheduling manner. The task execution parameter optimization module is used to process target execution tasks through each edge node, monitor the task execution process of each edge node, collect and evaluate the task execution quality index set of each edge node, and thus determine whether to optimize the task execution parameters of each edge node. The module for collecting execution results from each edge node is used to collect the execution results of each edge node, integrate the execution results marked as target execution tasks, and compare the execution results to determine whether to issue an early warning for the target execution task.

Images