Scheduling optimization system applied to AI edge computing server

Abstract

The invention discloses a scheduling optimization system applied to an AI edge computing server, and relates to the technical field of server resource scheduling, the scheduling optimization system comprises a data collection module, a preprocessing module, a computing analysis module and a scheduling decision module, the data collection module comprises local data and task network data, and the calculation analysis module comprises a resource supply analysis sub-module, a task allocation sub-module and a collaborative scheduling analysis sub-module. According to the invention, through the resource supply analysis sub-module, the task allocation sub-module and the collaborative scheduling sub-module, a dynamic resource elasticity coefficient, a task urgency index and a collaborative scheduling decision value are output respectively, and multi-dimensional factors such as real-time resources, health risks and energy stability are fused, so that a comprehensive evaluation and dynamic priority mechanism is constructed; according to the method, hardware adaptation, energy consumption and transmission delay are optimized, efficient task allocation and low-consumption operation of an edge cluster are achieved, local unbalance and service congestion are avoided, an excellent scheduling optimization effect is achieved, and intelligence and practicability are better.

Description

TECHNICAL FIELD

[0001] The application relates to the technical field of server resource scheduling, in particular to a scheduling optimization system applied to an AI edge computing server. BACKGROUND

[0002] With the development of the integration of 5G communication and artificial intelligence technology, AI edge computing servers have been widely applied to key fields such as industrial internet, automatic driving and intelligent security. In such scenarios, edge nodes such as industrial gateways, vehicle-mounted servers and edge small stations need to process a large amount of AI tasks in real time, such as real-time video stream inference, industrial quality inspection defect identification and automatic driving environment perception. The core demands are low latency, high reliability and efficient use of resources. However, the problem of resource allocation and task scheduling in the edge computing environment is still a great challenge, especially under the background of the increasing demand of AI applications for computing resources.

[0003] The existing scheduling optimization method usually judges the availability of nodes by a single resource index such as CPU and memory, and may not fully consider the node health degree such as temperature and failure rate, the energy stability such as power grid interruption risk and mobile node charging rate, and the distributed collaboration capability such as communication delay and collaboration success rate with adjacent nodes, thereby possibly leading to a one-sided evaluation of the node state, and possibly causing a high-load node to be misjudged as available or a low-load node to be unable to efficiently execute tasks due to health problems.

[0004] Moreover, the existing scheduling optimization method may be poor in dynamic adaptation to task urgency. The existing method may use static priority scheduling and does not fully consider factors such as real-time data transmission cost, dynamic dependency chain and node resource tension, thereby possibly causing an adverse situation of a critical task being overdue due to priority mismatch.

[0005] In addition, the existing scheduling optimization method may only focus on whether the resource quantity meets the requirement when matching nodes and tasks, and does not quantify the adaptation degree of task hardware requirements such as GPU computing power and NPU instruction set to node hardware capability, thereby possibly leading to a situation of low efficiency of AI task execution on edge nodes, and possibly not considering the matching of task energy consumption sensitivity and node real-time power consumption, thereby leading to a situation of excessive total energy consumption of edge clusters and other adverse situations, and resulting in poor overall scheduling optimization effect. SUMMARY

[0006] The application aims to provide a scheduling optimization system applied to an AI edge computing server, which solves the problems proposed in the background.

[0007] To achieve the above-mentioned purpose, the application provides the following technical scheme: a scheduling optimization system applied to an AI edge computing server, comprising:

[0008] The data collection module is configured to collect local data and task network data of the AI edge computing server.

[0009] The local data includes real-time core temperature of the node j, number of power grid interruptions of the node j in the past 30 days, real-time power consumption of the node j, and real-time idle CPU core number of the node j.

[0010] The task network data includes deadline of the task i, input data volume of the task i, transmission delay of the task i to the node j, average transmission bandwidth of the task i, and retransmission rate.

[0011] The preprocessing module is configured to receive the local data and the task network data of the edge node, and perform standardization and normalization processing on the received data, and input the processed data into the computing analysis module.

[0012] The computing analysis module is configured to analyze network congestion degree of the node j by combining retransmission rate and delay fluctuation characteristics based on real-time idle CPU core number of the node j and real-time idle memory of the node j, fuse temperature data, failure rate data, and energy stability data of the node j, and consider cooperative computing capability of the node j and adjacent edge nodes, to output dynamic resource elasticity coefficient of the node j.

[0013] The computing analysis module is configured to output task i emergency index by considering key degree of the task i in the entire service flow based on introduction of the dynamic resource elasticity coefficient of the node j, combination of deadline of the task i and current system time, introduction of input data volume of the task i and average transmission bandwidth of the task i.

[0014] The computing analysis module is configured to multiply the dynamic resource elasticity coefficient of the node j and the task i emergency index, and output cooperative scheduling decision value by combining heterogeneous computing power adaptation degree of the task i and the node j, real-time power consumption of the node j, and transmission delay of the task i to the node j.

[0015] The scheduling decision module is configured to input the dynamic resource elasticity coefficient of the node j, the task i emergency index, and the cooperative scheduling decision value, analyze the input data, select the node to execute the task, and cope with edge environment changes through an adjustment mechanism.

[0016] Optionally, the computing analysis module includes a resource supply analysis submodule, a task allocation submodule, and a cooperative scheduling analysis submodule.

[0017] Optionally, the processing flow of the resource supply analysis submodule is as follows.

[0018] I. By dividing the real-time idle CPU core number of node j by the total CPU core number of node j, the normalization of the idle CPU core number is introduced into the resource supply analysis submodule;

[0019] II. By operating the real-time idle memory of node j with the total memory of node j, the real-time idle memory in node is normalized and introduced into the resource supply analysis submodule;

[0020] III. By analyzing the TCP retransmission rate and delay standard deviation in the past 10s and processing based on the summation function, the network congestion degree of node j is analyzed, and the network congestion degree is normalized by the maximum network congestion threshold of node j;

[0021] IV. By combining the real-time core temperature of node j, the number of task failures of node j in the past 30 days, and the number of power grid interruptions of node j in the past 30 days, the temperature safety factor of node j, the failure prediction factor of node j, and the energy stability index of node j are evaluated to realize the fusion of temperature, failure rate and energy stability of node j, and to comprehensively evaluate the health index of node j;

[0022] V. By combining the number of neighbor nodes of node j, the real-time communication delay between node j and neighbor node k, and the historical cooperation success rate between node j and neighbor node k, the cooperation potential index of node j is analyzed, and the dynamic resource elasticity coefficient of node j is finally output.

[0023] Optionally, the processing flow of the task allocation submodule is:

[0024] I. By introducing the dynamic resource elasticity coefficient of node j into the task allocation submodule in a weighted form, the resource adaptation degree of node is reflected;

[0025] II. By dividing the estimated execution time of task i by the difference between the deadline of task i and the current system time, the criticality of task i is preliminarily analyzed;

[0026] III. By combining the input data volume of task i with the average transmission bandwidth of task i, the criticality of task i is comprehensively analyzed;

[0027] IV. By combining the criticality index of subsequent task m with the length of the dependency chain from task i to m, the key degree of task i in the entire service flow is evaluated, and the dynamic dependency chain coefficient of task i is analyzed, so as to finally output the criticality index of task i.

[0028] Optionally, the processing flow of the cooperative scheduling analysis submodule is:

[0029] I. The urgency of the task is associated with the resource node by multiplying the dynamic resource elasticity coefficient of node j and the task i urgency index;

[0030] II. The resource node is associated with the execution time by dividing the estimated execution time of task i by the dynamic resource elasticity coefficient of node j;

[0031] III. The hardware demand of task i is quantified and the heterogeneous computing power characteristics of node j are matched to evaluate the heterogeneous computing power adaptation degree of task i and node j by combining the demand vector of task i for the e-type hardware with the e-type hardware capability of node j;

[0032] IV. The transmission delay of task i to node j, the energy consumption sensitivity coefficient of task i and the real-time power consumption of node j are introduced into the collaborative scheduling analysis submodule to finally output the collaborative scheduling decision value.

[0033] Optionally, the scheduling decision module specifically comprises:

[0034] First, the dynamic resource elasticity coefficient of node j is stored in the real-time database of the scheduling center;

[0035] Then, based on the task i urgency index, the urgency indexes of all tasks are sorted in descending order to form a task priority queue;

[0036] Then, the collaborative scheduling decision value is stored in the matching matrix of the scheduling center.

[0037] Optionally, the selected node in the scheduling decision module executes the task, specifically:

[0038] First, traverse the task priority queue to select the node j with the highest collaborative scheduling decision value for each task, and the heterogeneous computing power adaptation degree of task i and node j should be greater than or equal to 0.6 to avoid serious hardware mismatch;

[0039] Second, if the dynamic resource elasticity coefficient of the subsequent node j is less than 0.3, it means that the resource is tight, so even if the collaborative scheduling decision value is high, the node with the second highest value is also selected to avoid single-point overload.

[0040] Optionally, the adjustment mechanism in the scheduling decision module specifically comprises:

[0041] First, the matching matrix of the collaborative scheduling decision value is recalculated every 2s, and for the tasks that have been allocated but not executed, such as tasks with an urgency index increase greater than or equal to 20%, the node migration is triggered, and at this time, the node with a high collaborative scheduling decision value is selected;

[0042] Second, if the failure prediction coefficient of node j is less than 0.5, it indicates that the risk of failure is high, at this time, the task of the node is assigned in advance to the dynamic resource elasticity coefficient of node j is greater than or equal to 0.8, and the transmission delay of task i to the neighbor node of node j is less than or equal to 20ms, in order to avoid task interruption.

[0043] Compared with the prior art, the beneficial effects of the present application are as follows:

[0044] Firstly, the present application outputs the dynamic resource elasticity coefficient of node j through the resource supply analysis submodule, this submodule constructs a comprehensive node comprehensive evaluation system by fusing the real-time resource state, health risk, energy stability and distributed collaboration ability of the edge node, this submodule accurately reflects the instantaneous resource characteristics such as CPU idle state, memory available amount, network congestion degree by real-time high-frequency data collection and calculation, avoids the state lag caused by periodic sampling, and this submodule comprehensively introduces and evaluates the temperature safety state, fault history and energy stability, thereby avoiding task execution failure caused by hardware failure or energy interruption in advance, and this submodule identifies the node cluster with resource complementary ability by evaluating the communication delay of the node and the adjacent node and the collaborative execution success rate, avoiding the local imbalance problem of single node resource idling and adjacent node overload.

[0045] Secondly, the present application outputs the task i emergency index through the task allocation submodule, this submodule constructs a dynamically adjusted task priority mechanism by fusing task deadline pressure, data transmission cost, dynamic dependency and node resource state, this submodule calculates the data transmission time according to the transmission bandwidth and data volume of task input data, to ensure that the data is processed in time within the validity period, and dynamically adjusts the priority weight by analyzing the position of the task in the business process, to avoid the whole business chain blockage caused by single task delay, and when the candidate node resource is tight, the task emergency can be automatically improved to occupy the resource in priority, avoiding task timeout due to waiting for resources.

[0046] Thirdly, the present application outputs the cooperative scheduling decision value through the cooperative scheduling analysis submodule, this submodule constructs a multi-objective optimization task allocation mechanism by fusing task emergency, node resource elasticity, heterogeneous hardware adaptability, transmission delay and energy consumption cost, this submodule analyzes the characteristic demand of task on CPU, GPU and NPU and the hardware capability of node, to ensure that the task is allocated to the node with the most matched hardware characteristics, avoiding the loss of execution efficiency caused by hardware incompatibility, and the energy consumption sensitivity of the task and the real-time power consumption of the node are taken into account in the matching decision, to reduce the total energy consumption of the edge cluster while ensuring task execution, and the transmission delay of task data between nodes and the actual time consumption of node executing task are considered comprehensively to ensure the minimum end-to-end delay of task, to achieve excellent scheduling optimization effect and better practicality. Attached Figure Description

[0047] Figure 1 This is a schematic diagram of the structure of the scheduling optimization system applied to AI edge computing servers;

[0048] Figure 2 This is a schematic diagram of the data collection module and the preprocessing module in the scheduling optimization system applied to AI edge computing servers;

[0049] Figure 3 This is a schematic diagram of the data collection module in the scheduling optimization system applied to AI edge computing servers;

[0050] Figure 4 This is a flowchart of the steps in the computation and analysis module of the scheduling optimization system applied to AI edge computing servers;

[0051] Figure 5 This is a schematic diagram of the computational analysis module and the scheduling decision module in the scheduling optimization system applied to AI edge computing servers. Detailed Implementation

[0052] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0053] Please see Figures 1 to 5 This embodiment provides a scheduling optimization system for AI edge computing servers, including:

[0054] Data collection module: used to collect local data and task network data from AI edge computing servers;

[0055] Local data includes: real-time core temperature of node j, number of power grid outages of node j in the past 30 days, real-time power consumption of node j, and real-time number of idle CPU cores of node j.

[0056] Task network data includes: the deadline of task i, the amount of input data for task i, the transmission delay from task i to node j, the average transmission bandwidth and retransmission rate of task i.

[0057] Preprocessing module: Used to receive local data from edge nodes and task network data, standardize and normalize the received data, and input the processed data into the calculation and analysis module;

[0058] The computing analysis module: based on the real-time idle CPU core number of node j and the real-time idle memory introduction of node j, the retransmission rate and the delay fluctuation characteristics are combined, the network congestion degree of node j is analyzed, the temperature data, the failure rate data and the energy stability data of node j are fused, and the cooperative computing capability of node j and the adjacent edge node is considered, to output the dynamic resource elasticity coefficient of node j;

[0059] The computing analysis module: based on the introduction of the dynamic resource elasticity coefficient of node j, the deadline of task i and the current system time are combined, the input data amount of task i and the average transmission bandwidth of task i are introduced, and the key degree of task i in the whole service flow is considered, to output the emergency index of task i;

[0060] The computing analysis module: based on the multiplication of the dynamic resource elasticity coefficient of node j and the emergency index of task i, and combined with the heterogeneous computing power adaptation degree of task i and node j, the real-time power consumption of node j and the transmission delay of task i to node j, to output the cooperative scheduling decision value;

[0061] The scheduling decision module: for inputting the dynamic resource elasticity coefficient of node j, the emergency index of task i and the cooperative scheduling decision value, the scheduling decision module analyzes the input data, selects the node to execute the task, and adjusts the mechanism to respond to the edge environment change;

[0062] The scheduling decision module specifically includes:

[0063] Firstly, the dynamic resource elasticity coefficient of node j is stored in the real-time database of the scheduling center;

[0064] Then, based on the emergency index of task i, the emergency indexes of all tasks are sorted in descending order to form a task priority queue;

[0065] Then, the cooperative scheduling decision value is stored in the matching matrix of the scheduling center;

[0066] The computing analysis module includes: resource supply analysis submodule, task allocation submodule and cooperative scheduling analysis submodule;

[0067] The selection of node to execute task in the scheduling decision module specifically includes:

[0068] Firstly, traverse the task priority queue, select the node j with the highest cooperative scheduling decision value for each task, and the heterogeneous computing power adaptation degree of task i and node j should be greater than or equal to 0.6, to avoid serious hardware mismatch;

[0069] Secondly, if the dynamic resource elasticity coefficient of the subsequent node j is less than 0.3, it means that the resource is tight, so even if the cooperative scheduling decision value is high, the node with the second highest value is also selected to avoid single point overload;

[0070] The adjustment mechanism in the scheduling decision module, specifically:

[0071] First, the matching matrix of the cooperative scheduling decision value is recalculated every 2s. For tasks that have been allocated but not executed, such as tasks with an urgency index increase greater than or equal to 20%, trigger node migration, at this time, select nodes with high cooperative scheduling decision values;

[0072] Second, if the fault prediction coefficient of node j is less than 0.5, indicating a high risk of failure, the task of node j is allocated to a neighbor node with a dynamic resource elasticity coefficient greater than or equal to 0.8, and the transmission delay of task i to node j is less than or equal to 20ms, to avoid task interruption.

[0073] First of all, it needs to be explained that: the system calculation and analysis module calculates the dynamic resource elasticity coefficient DR j (node state evaluation)→ task i urgency index TL i (task priority sorting)→ cooperative scheduling decision value CSi,j (task and node optimal matching), forming a closed-loop scheduling framework of "node perception-task classification-global decision", realizing the leap from single resource scheduling to system-level resource-task-energy efficiency collaborative optimization;

[0074] Through the dynamic resource elasticity coefficient DR j of node j, the cooperative potential and heterogeneous adaptation can improve the CPU / GPU resource utilization of the edge cluster, the dynamic dependency chain and the execution time compression can reduce the average end-to-end delay of AI inference tasks, meet the real-time requirements of the edge, the combination of health index and energy stability can ensure the success rate of tasks when the node fails or the energy fluctuates, the high-frequency state acquisition of the dynamic resource elasticity coefficient DR j of node j, the real-time priority update of the task i urgency index TL i , and the dynamic matching decision of the cooperative scheduling decision value CS i，j , the scheduling system can respond to the resource fluctuations of the edge node and the changes in task demand in real time, and the system can cover fixed and mobile nodes, heterogeneous hardware, AI tasks with different urgency and energy consumption requirements, which is different from the fixed scheduling optimization method in the traditional prior art, providing a practical scheduling optimization solution for AI edge scenarios such as industry, intelligent transportation, and remote medical care, and enabling large model inference, real-time data processing, and other tasks to be executed efficiently in resource-constrained edge environments.

[0075] Referring to Figure 1 , Figure 2 , Figure 3 , Figure 4 and Figure 5 , the processing flow of the resource supply analysis sub-module is as follows:

[0076] S1, by dividing the number of real-time idle CPU cores of node j by the total number of CPU cores of node j, to introduce the normalization of the number of idle CPU cores into the resource supply analysis submodule;

[0077] S2, by operating the real-time idle memory of node j with the total memory of node j, to introduce the normalization of real-time idle memory in the node into the resource supply analysis submodule;

[0078] S3, by analyzing the TCP retransmission rate and delay standard deviation in the past 10s, and processing based on the summation function, to analyze the network congestion degree of node j, and by the maximum network congestion threshold of node j, to normalize the network congestion degree;

[0079] S4, by combining the real-time core temperature of node j, the number of task failures of node j in the past 30 days, and the number of power grid interruptions of node j in the past 30 days, to evaluate the temperature safety factor of node j, the failure prediction factor of node j, and the energy stability index of node j, to realize the fusion of temperature, failure rate and energy stability of node j, and to comprehensively evaluate the health index of node j;

[0080] S5, by combining the number of neighbor nodes of node j, the real-time communication delay between node j and neighbor node k, and the historical cooperation success rate between node j and neighbor node k, to analyze the cooperation potential index of node j, to finally output the dynamic resource elasticity coefficient of node j.

[0081] Referring to Figure 1 , Figure 2 and Figure 3 , the calculation process of the resource supply analysis submodule is:

[0082]

[0083] Among them:

[0084] DR j refers to the dynamic resource elasticity coefficient of node j, which reflects the cooperative computing ability of node j and the adjacent edge node, that is, the efficiency of task splitting and result sharing, and comprehensively evaluates the real-time resource supply ability of node j, the higher the value, the more suitable the node is to receive new tasks;

[0085] DRA j refers to the number of real-time idle CPU cores of node j, the number of idle CPU cores currently allocable by node j, which reflects the real-time availability of computing resources, the unit is core, and the range is 0 j ≤DRB j, the collection method is through the edge agent process deployed in node j, and the idle core number in the operating system file is directly read every 500 ms;

[0086] DRB j Refers to the total CPU core number of node j, that is, the total number of physical CPU cores of node j, and the collection method is to obtain it through BIOS information or lscpu command when the node is initialized, which is a fixed value;

[0087] DRC j Refers to the real-time free memory of node j, that is, the current free memory available for allocation of node j, and is the value after deducting system cache and reserved memory, unit is GB, range is 0≤DRC j ≤DRD j ;

[0088] DRD j Refers to the total memory of node j, that is, the total capacity of physical memory of node j, unit is GB, and the collection method is to obtain it through dmidecode command when the node is initialized;

[0089] DRE j Refers to the network congestion degree of node j, that is, the real-time congestion degree of node j network link, which integrates retransmission rate and delay fluctuation characteristics, range is 0-1, DRE j =0 means no congestion, DRE j =1 means complete congestion, and the collection method is as follows:

[0090] First, the TCP retransmission rate RRT and delay standard deviation DVT in the past 10s are collected through the network monitoring module of node j, and the calculation formula is as follows:

[0091]

[0092] Among them: refers to

[0093] RRT t Refers to the retransmission rate at time t;

[0094] DVT t Refers to the delay standard deviation at time t, and the unit is ms, and is normalized by dividing by 100;

[0095] DRF j Refers to the maximum network congestion threshold of node j, that is, the upper limit value of node j network link congestion judgment, which is used to normalize the network congestion degree DRE j of node j, which is a fixed value, and the DRF j of the embodiment is set to 1;

[0096] DRGj Health index of node j, that is, the comprehensive health score after fusing the temperature, failure rate and energy stability of node j, reflecting the hardware reliability, ranging from 0 to 1, DRG j = 0, indicating complete failure, DRG j = 1, indicating healthy state, and the calculation method is as follows:

[0097] DRG j = 0.4 x DRGA j + 0.3 x DRGB j + 0.3 x DRGC j ;

[0098] Wherein:

[0099] DRGA j Temperature safety factor of node j, and the calculation formula is as follows:

[0100] DRGA j = 1 - (S1 j - S2 j ) / (S3 j - S2 j );

[0101] Further:

[0102] S1 j Real-time core temperature of node j, in ℃, which can be collected by lm-sensors tool;

[0103] S2 j Minimum temperature of node j, which is preset to 25℃ in the embodiment;

[0104] S3 j Maximum temperature of node j, which is preset to 85℃ in the embodiment, and DRGA j = 0 when S1 j exceeds 85℃;

[0105] DRGB j Failure prediction coefficient of node j, and the calculation formula is as follows:

[0106] DRGB j = 1 - (A1 j - A2 j );

[0107] Further:

[0108] A1 j Number of task failures of node j in the past 30 days;

[0109] A2 jTotal task number, can be counted from edge task scheduling log;

[0110] DRGC j Energy stability index of node j, that is, energy sustainability index, used to reflect the long-term stability of node j energy supply. Fixed nodes focus on power grid reliability, and mobile nodes focus on power and charging efficiency;

[0111] Case one: fixed nodes, such as edge servers in computer rooms, the calculation formula is as follows:

[0112] DRGC j = 1-(W1 j / W2);

[0113] Further:

[0114] W1 j Refers to the number of power grid interruptions of node j in the past 30 days, which can be collected through UPS log, unit is times;

[0115] W2 refers to the maximum allowed interruption times, this embodiment is preset to 3 times / month, and DRGC j = 0 when exceeded;

[0116] Case two: mobile nodes, such as unmanned aerial vehicle edge units, the calculation formula is as follows:

[0117] DRGC j = 1-(Q1 j / Q2 j )×(1+0.5×Q3 j );

[0118] Further:

[0119] Q1 j Refers to the remaining battery capacity of node j, that is, the remaining battery energy of mobile edge node j, such as unmanned aerial vehicle edge unit and vehicle-mounted server, which reflects the current available power level of the battery, and the acquisition device is through the built-in battery management system of the mobile node;

[0120] Q2 j Refers to the total capacity of the battery of node j, that is, the rated total capacity of the battery pack equipped in the mobile edge node j, that is, the theoretical maximum energy storage capacity when the battery is fully charged, which is a basic parameter for measuring battery endurance;

[0121] Q3 j Refers to the real-time charging rate, unit is battery capacity / hour, which can be collected through BMS system;

[0122] DRH jSynergy potential index of node j, representing the synergy computing capability of node j and adjacent edge nodes, reflecting the efficiency of distributed task splitting and result sharing, ranging from 0 to 1, DRH j = 0, no synergy capability, DRH j = 1, optimal synergy state, the calculation formula is as follows:

[0123]

[0124] Wherein:

[0125] NN j Number of neighbor nodes of node j, the embodiment presets NN j ≤ 5, which can be automatically discovered through edge network topology;

[0126] B1 j，k Real-time communication delay between node j and neighbor node k, in ms, which can be detected every 2s through ICMP ping;

[0127] B2 j Average delay between node j and all neighbors, in ms, which can be dynamically updated;

[0128] B3 j，k Historical synergy success rate between node j and neighbor node k, representing the proportion of successful execution in the past 100 synergy tasks, which can be counted from the scheduling log;

[0129] DW j Physical and network neighbor node set of node j, that is, other edge nodes directly connected to node j in the edge network topology or having a communication delay lower than a threshold, in the embodiment, the threshold is 50 ms;

[0130] k ∈ DW j Indicates that all neighbor nodes k of node j are traversed, k is the index of the neighbor node, distinguished from j, and the acquisition method of the neighbor node is further described below:

[0131] The acquisition means is: through an edge network topology discovery protocol, which is a self-defined edge node discovery mechanism based on UDP broadcast, the surrounding nodes can be scanned every 10s, and the nodes having a communication delay of ≤ 50 ms are recorded;

[0132] The number limit is: the physical and network neighbor node set DW of node j j Contains at most 5 nodes, so as to avoid the influence of long-distance nodes on synergy efficiency;

[0133] The dynamic updating mechanism is: when the communication delay of the neighbor node of node j exceeds 50 ms or fails, the physical and network neighbor node set DW of node j is automatically updatedj remove and supplement newly found low latency nodes;

[0134] D1, D2, D3, D4 and D5 respectively refer to the weight factors of the number of idle CPU cores, idle memory, network congestion, node health index and collaborative potential index.

[0135] Secondly, it needs to be explained that the dynamic resource elasticity coefficient DR of node j j By fusing the real-time resource status (CPU, memory and network), health (temperature, failure rate and energy stability) and collaborative potential of node j, a dynamic and multi-dimensional node resource evaluation system is constructed. Traditional scheduling only relies on a single resource index such as CPU utilization, while the calculation results of this sub-module solve the core pain points of dynamic changes in node status in edge computing environment, such as mobile node power fluctuation, industrial environment temperature rise and distributed collaborative demand, such as task splitting execution, and provide a more comprehensive node portrait for scheduling decision-making;

[0136] The dynamic resource elasticity coefficient DR of node j j Directly reflects the comprehensive service capability of node j, the higher the value, the more suitable the node is for receiving new tasks. This sub-module excavates the idle resources of adjacent nodes through the collaborative potential index to avoid local overload, avoids node failure risk in advance through the temperature safety factor and energy stability factor in the health index, and adjusts the multi-weight according to the task type to adapt to the edge heterogeneous hardware environment. This sub-module captures the instantaneous state changes of edge nodes by collecting the real-time idle CPU core number of node j and the real-time idle memory of node j, and calculating the collaborative potential index of node j in real time, and the energy stability index of node j can cover both fixed nodes and mobile nodes, adapting to the diversity of node types in edge environment. When scheduling AI reasoning tasks, the GPU-related weight can be increased to ensure efficient execution of AI models on heterogeneous edge nodes;

[0137] Unlike existing technologies, the collaborative potential index DRH of node j in this sub-module quantifies the distributed collaborative capability through neighbor node delay and collaborative success rate, so that edge nodes with limited resources but adjacent nodes idle, such as industrial gateway clusters, have higher scheduling priority. The energy stability index DRGC of node j in this sub-module j Fuses the reliability of fixed node power grid and the charging rate of mobile node, covers the energy characteristics of all scenarios of edge computing, and reduces the task failure rate caused by energy interruption.

[0138] Reference Figure 1 , Figure 2 , Figure 3 , Figure 4 and Figure 5, the processing flow of the task allocation sub-module is:

[0139] S1, by introducing the dynamic resource elasticity coefficient of node j into the task allocation sub-module in a weighted form, to reflect the resource adaptation degree of the node;

[0140] S2, by dividing the estimated execution time of task i by the difference between the deadline of task i and the current system time, to preliminarily analyze the urgency of task i;

[0141] S3, by combining the input data volume of task i with the average transmission bandwidth of task i, to comprehensively analyze the urgency of task i;

[0142] S4, by combining the emergency index of the subsequent task m with the length of the dependency chain from task i to task m, to evaluate the criticality of task i in the entire service flow, and then analyze the dynamic dependency chain coefficient of task i, so as to finally output the emergency index of task i.

[0143] Referring to Figure 1 , Figure 2 and Figure 3 , the calculation process of the task allocation sub-module is:

[0144]

[0145] Among them:

[0146] TL i refers to the emergency index of task i, which is used to comprehensively evaluate the scheduling priority of task i, and the higher the value, the more priority the resource needs to be allocated;

[0147] TLA i refers to the deadline of task i, that is, the latest time when task i must be completed, which can be specified by the user or the upper application, and the unit is timestamp, and the precision is second, and the acquisition method is to explicitly pass in by the client when submitting the task;

[0148] TLB refers to the current system time, that is, the real-time time of the edge scheduling system, which is used to calculate the remaining deadline of the task, and the acquisition method is to directly read the system time of the scheduling server;

[0149] TLC i refers to the estimated execution time of task i, that is, the expected execution time of task i on a resource sufficient node, which is used to normalize the remaining deadline, and the unit is second, and the acquisition method is to fit the model based on the type of task i, such as image classification and input data volume, and to predict the historical execution time;

[0150] TLD iThe input data volume of task i, that is, the total amount of input data that task i needs to transmit, in MB, obtained by reporting from the client when the task is submitted;

[0151] TLE i The average transmission bandwidth of task i, which means the real-time average network bandwidth from the data source to the edge node of task i, used to predict the data transmission time, in MB / s, TLE i > 0, affected by the edge network type, 5G is usually 50-200 MB / s, Wi-Fi6 is 20-80 MB / s, obtained by real-time collection through the traffic monitoring module of the edge gateway, updated every 1s;

[0152] TLG i The dynamic dependency chain coefficient of task i, which means the criticality of task i in the entire business flow, if the delay of task i will cause the blocking of multiple high-urgency tasks, then the dynamic dependency chain coefficient of task i TLG i Significantly improved, ranging from 1 < TLG i ≤ 3, the calculation formula is as follows:

[0153]

[0154] Among them:

[0155] Mq i Refers to the set of all subsequent tasks that directly and indirectly depend on the output results of task i, that is, the downstream task list of task i, ranging from a non-empty set, if task i is the end of the process, then Mq i is an empty set, at this time TLG i = 1;

[0156] PAQ m Refers to the urgency index of the subsequent task m, obtained by real-time calculation, synchronized with the urgency index TL i of the current task i, ranging from 0 < PAQ m ≤ 5, the calculation process is as follows:

[0157] S1, trigger condition: the urgency index TL i of the current task i starts to calculate, the scheduling system automatically traverses all m in the set Mq i which is the set of all subsequent tasks that directly and indirectly depend on the output results of task i, triggers the real-time calculation of the urgency index PAQ m of the subsequent task m, recursively calls the task allocation submodule;

[0158] Data synchronization: the calculation of the urgency index PAQ m of the subsequent task m is synchronized with the urgency index TL iShare the same set of parameters, such as TLA m and TLD m , can be read from task metadata in real time;

[0159] Cache mechanism: to avoid repeated calculation, the urgency index PAQ m of the subsequent task m is cached for 500ms, and the edge scene task characteristics change quickly, and need to be recalculated after the cache timeout;

[0160] Example: if the urgency index PAQ m1 of the subsequent task m of the downstream task m1 is 3.2, and the PAQ m2 of m2 is 2.8, the dynamic dependency chain coefficient TLG i of task i needs to be calculated by accumulating these two values;

[0161] PAB i，m refers to the length of the dependency chain from task i to m, such as i→k→m, then PAB i，m =2, to avoid over-amplification of long chains, that is, the number of tasks contained on the shortest path from task i to downstream task m, including i and m, is used to weaken the excessive influence of long dependency chains on the dynamic dependency chain coefficient TLG i of task i;

[0162] TA, TB, TC and TD respectively refer to task time, input data volume, resource elasticity coefficient and dependency chain weight factor;

[0163] TC×DR j refers to the node resource adaptation term, that is, the resource elasticity coefficient DR j of node j, the higher the value, the more sufficient the resources, the better the health and the stronger the coordination potential, which is positively correlated with the task i urgency TL i , that is, the higher the quality of the node, the higher the scheduling priority of task i.

[0164] And it should be noted that: this sub-module constructs a dynamic priority evaluation system by integrating the deadline pressure of task i, data transmission cost, dependency chain criticality and node resource tension, traditional priority scheduling only relies on static deadline, and the task i urgency index TL i solves the problem of data timeliness decay of edge AI tasks, such as real-time video stream reasoning and business chain dependency blockage, such as industrial quality inspection pipeline tasks, to ensure that critical tasks are executed first;

[0165] The task i urgency index TL i output value directly determines the scheduling priority of task i, the higher the value, the higher the priority of resource allocation, and the dynamic dependency chain coefficient TLG iTo ensure that critical path tasks, such as obstacle recognition in autonomous driving, are prioritized and scheduled, thereby improving the task deadline compliance rate to a certain extent, the data transmission time in this submodule is reduced. To quantify the impact of edge network fluctuations on tasks and prevent data expiration due to transmission delays, this submodule introduces the dynamic resource elasticity coefficient DR of node j. j This is to avoid assigning high-urgency tasks to nodes with insufficient resources, thereby reducing task queuing delays.

[0166] This submodule can adapt to the diversity of task types in edge scenarios through dynamic adjustment of multiple weights. The urgency index TL of task i changes with the dynamic resource elasticity coefficient DR of node j. j It updates the urgency of dependency chain tasks in real time, responding to dynamic changes in task requirements in edge environments, and is also capable of handling AI tasks sensitive to data freshness, such as real-time object detection. Real-time computation ensures that input data completes inference within its validity period. Unlike existing technologies, traditional scheduling may consider task priorities in isolation, and the dynamic dependency chain coefficient (TLG) of task i... i By relying on the urgency of subsequent tasks and the length of the dependency chain, the global criticality of task i in the business flow is quantified. For example, if task i is a preprocessing step for 5 high-urgency AI inference tasks, the dynamic dependency chain coefficient (TLG) of task i is... i This will significantly improve performance and prevent the entire business chain from being blocked due to a single point of delay.

[0167] See Figure 1 , Figure 2 , Figure 3 , Figure 4 and Figure 5 The processing flow of the collaborative scheduling analysis submodule is as follows:

[0168] S1. By multiplying the dynamic resource elasticity coefficient of node j with the urgency index of task i, the urgency of the task is associated with the resource node.

[0169] S2. By dividing the estimated execution time of task i by the dynamic resource elasticity coefficient of node j, the resource node is associated with the execution time.

[0170] S3. By combining the demand vector of task i for the e-th type of hardware with the e-th type of hardware capability of node j, the degree of matching between the hardware requirements of task i and the heterogeneous computing power characteristics of node j is quantified, so as to evaluate the heterogeneous computing power adaptability between task i and node j.

[0171] S4. The transmission delay from task i to node j, the energy sensitivity coefficient of task i, and the real-time power consumption of node j are introduced into the collaborative scheduling analysis submodule to finally output the collaborative scheduling decision value.

[0172] Referring to Figure 1 , Figure 2 and Figure 3 Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 Figure 1 Figure 2 Figure 3 , the calculation process of the cooperative scheduling analysis submodule is as follows:

[0173]

[0174] Wherein:

[0175] CS i，j refers to a distributed cooperative scheduling decision value, used to comprehensively evaluate the global optimality of task i assigned to node j, and a high score indicates a good matching degree;

[0176] CSA i，j refers to the heterogeneous computing power adaptation degree of task i and node j, that is, the matching degree of the hardware requirements of task i and the heterogeneous computing power characteristics of node j, such as GPU architecture and NPU instruction set, ranging from 0 to 1, and CSA i，j = 0, completely mismatched, and CSA i，j = 1, perfect match, and the calculation formula is as follows:

[0177]

[0178] Wherein:

[0179] LA i，e refers to the demand vector of task i for the e-th type of hardware, and the sum of the vectors is 1, e = 1 refers to GPU computing power, e = 2 refers to NPU computing power, and e = 3 refers to CPU computing power, such as an AI inference task, and the demand vector of task i for the e-th type of hardware LA i，e = [0.8, 0.2, 0.1], which is finally normalized to a sum of 1, and can be obtained by inference, based on the type of task i, specified by the user when submitting the task, and then matched with the preset demand template library, as follows:

[0180] image classification task type, that is, GPU-intensive, LA i，e = [0.8, 0.1, 0.1], wherein the GPU demand accounts for 80%;

[0181] speech recognition task type, that is, NPU-optimized, LA i，e = [0.2, 0.7, 0.1], wherein the NPU demand accounts for 70%;

[0182] log analysis task type, that is, CPU-intensive, LA i，e = [0.1, 0.1, 0.8], wherein the CPU demand accounts for 80%;

[0183] LB j，eThis refers to the e-th type of hardware capability of node j, which can be pre-calibrated using benchmarking tools. For example, GPU computing power is tested through FP16 throughput and then normalized to [0,1]. This is the standardized capability value of node j on the three types of hardware, reflecting its efficiency in executing tasks requiring corresponding hardware requirements. This value can be obtained through benchmarking. When node j is initialized, the hardware benchmarking tool is automatically run to obtain this parameter, as shown below:

[0184] GPU computing power, e=1: Run FP16 precision inference test, record throughput per second, normalization formula is: LB j，1 = Actual throughput / Theoretical maximum throughput;

[0185] NPU computing power, e=2: Run INT8 precision inference test, normalization method is the same as above when e=1;

[0186] CPU computing power, e=3: Benchmark tests are run, and the single-core score is taken. The normalization formula is: LB j，3 = Single-core score / Single-core score of mainstream CPUs;

[0187] LB j，f The f-th type of hardware capability of node j, such as GPU, NPU, CPU, and normalized to [0,1], represents the heterogeneous computing power adaptability (CSA) between task i and node j. i，j The denominator of the calculation formula is the sum of squares of the capabilities of each hardware component, LB. j，1 2 +LB j，2 2 +LB j，3 2 Essentially, it's a measure of the energy of node j's hardware resources. The squared magnitude of the hardware capability vector of node j is used to normalize the dot product result of the numerator, ensuring CSA. i， The value of j is constrained to the range of [0,1] to avoid distortion of the matching degree due to excessively large absolute values ​​of node hardware capabilities;

[0188] CSB i，j This refers to the transmission latency from task i to node j, which is the real-time network latency for the input data of task i to be transmitted from the data source to node j, measured in milliseconds and ranging from CSB. i，j >0, edge LAN is usually 5-50ms, and WAN may reach more than 100ms. The acquisition method is to perform real-time detection through the edge SDN controller, such as ICMP timestamp requests based on tunnels, sampling once every 500ms and taking the average of 3 times;

[0189] CSE iCSE i = 0 means insensitive, SE i = 0 means sensitive, the way to get is specified by user when submitting the task, for example, through parameter setting when submitting the task, such as edge node task CSE i = 0.9, grid-powered task CSE i = 0.3, when the user does not specify, the system automatically assigns according to the task type, as follows:

[0190] Real-time interactive task, such as AR rendering: CSE i = 0.2, means delay priority, energy consumption secondary;

[0191] Offline analysis task, such as log mining: CSE i = 0.8, means energy consumption priority, delay tolerant;

[0192] CSF j refers to the real-time power consumption of node j, the total power consumption of node j at present, including CPU, GPU, memory and network module, that is, the total power consumption of node j at present, used to reflect the hardware energy consumption level, unit: watt, range: 0 < CSF j ≤ CSF j，max , CSF j，max is the maximum power consumption of node j, such as edge server usually 300-500W, the way to get is fixed node through smart meter collection, such as IPMI interface reading power module power consumption, mobile node through BMS battery management system collection;

[0193] CSE i × CSF j refers to: traditional scheduling does not consider energy consumption, CSE i × CSF j realizes dynamic binding of task energy consumption demand and node energy consumption state, such as low emergency task CSE i = 0.8, preferentially assigned to node j with real-time power consumption CSF j <100W;

[0194] CA and CB respectively refer to balance coefficient one and balance coefficient two, CA adjusts the weight of transmission delay and execution delay, CB adjusts the influence strength of energy consumption penalty term, the range of CA is [0, 1], the range of CB is [0, 0.5], CA is preset as 0.5 in the embodiment, indicating balancing transmission and execution delay, CA is preset as 0.7 when the task is delay-sensitive, CB is preset as 0.5 when the task is energy-sensitive, and CB is preset as 0.2 by default, the acquisition method of the parameter is that the system dynamically adjusts according to the global optimization target, for example, when the total energy consumption of the edge cluster exceeds the threshold of 10 kW, CB is automatically increased to 0.5;

[0195] TL i ×DR j The product of task urgency and node resource elasticity represents the scheduling goal that high-urgency tasks should be preferentially allocated to high-resource-elasticity nodes, for example, an urgent AI inference task is allocated to a node with idle CPU or GPU;

[0196] TL i It refers to the estimated execution time of task i, the benchmark time consumption on a resource-sufficient node, divided by DR j It represents that the higher the resource elasticity of node j is, the shorter the actual execution time of task i is, and when the resources are sufficient, the CPU / GPU parallel efficiency is improved, and the execution time is compressed.

[0197] It should be noted that the calculation result of the sub-module is the cooperative scheduling decision value CS i，j By fusing task urgency, node resource elasticity, heterogeneous computing power adaptation degree and energy consumption cost and other factors, a task and node global optimal matching system is constructed, and the traditional scheduling may be based on only a single index, such as task priority or node load, while the cooperative scheduling decision value CS i，j of the sub-module solves the situation that task demand and node capability do not match in edge computing, such as the disadvantageous situation of NPU task allocation to CPU node and the high-energy-consumption scheduling situation, such as the problem of allocating low-urgency tasks to high-power GPU nodes, and balances the system-level performance and energy efficiency;

[0198] The cooperative scheduling decision value CS i，jThe final decision of task i assigned to node j is based on the selection of the highest score of task and node pair to perform scheduling, the heterogeneous computing power adaptation degree is used to ensure that the AI task is preferentially assigned to the matching hardware, such as the edge node with NPU, so as to improve the utilization rate of hardware acceleration to a certain extent, the energy consumption sensitivity coefficient and the real-time power consumption of node j represent the energy consumption penalty term, which can be automatically triggered when the edge node power is low or the machine room PUE is out of limit, thereby reducing the total energy consumption to a certain extent, through the joint optimization of transmission delay and dynamic execution time, the end-to-end delay of AI task is reduced, and the task i urgency index TL i and the dynamic resource elasticity coefficient DR j of node j, so as to avoid resource mismatch caused by static rules, such as assigning high urgency tasks to nodes with sufficient resources but unmatched hardware, the heterogeneous computing power adaptation degree of task i and node j is calculated by the dot product operation of hardware demand vector and node capability vector, which adapts to the coexistence of GPU, NPU and CPU heterogeneous hardware in edge computing, and through the calculation of the heterogeneous computing power adaptation degree of task i and node j, for computing-intensive AI tasks such as large model inference, high-power nodes can be preferentially assigned, and through dynamic compression of execution time, to ensure that the inference delay meets the edge real-time requirement;

[0199] The prior art may only focus on resource amount, such as GPU memory, while the heterogeneous computing power adaptation degree CSA i，j of task i and node j in the present sub-module normalizes the dot product of hardware demand and node capability to quantify the characteristic matching degree of task and hardware, avoiding scheduling errors caused by insufficient ability with resources;

[0200] The prior art may not consider energy consumption, CSE i ×CSF j achieves dynamic binding of task energy consumption demand and node energy consumption state, such as low urgency task CSE i =0.8, preferentially assigned to CSF i <100W node, so as to improve the endurance time of the whole system.

[0201] It should be noted that through further operation of the cooperative scheduling decision value CS i，j , the weight factor D5 of the cooperative potential index in the resource supply analysis submodule is further affected, so that the parameter changes according to a certain adjustment amplitude, so as to adjust the influence degree of the cooperative potential index DRH j of node j in the resource supply analysis submodule, so as to optimize the resource supply analysis submodule, the task allocation submodule and the cooperative scheduling analysis submodule, the specific processing process is as follows:

[0202] First, the cooperative scheduling decision value CSi，j Further operation, the calculation formula is as follows:

[0203]

[0204] Case one: if KPL t <KPLA, KPLA is a preset target threshold, and the embodiment is set to 0.85, which indicates that the current coordination potential weight is insufficient, at this time, the weight adjustment amplitude SSA needs to be increased to improve the influence of node coordination on scheduling, and then the calculation formula of the weight factor D5 of the coordination potential index is as follows:

[0205] D5 t+1 =D5 t +SSA;

[0206] Case two: if KPL t ≥KPLA, and KPL t -KPL t-1 <0, which indicates that the performance decreases, and the weight factor D5 of the coordination potential index is too high to ignore the resource locality, and the weight factor D5 of the coordination potential index needs to be reduced, and then the calculation formula of the weight factor D5 of the coordination potential index is as follows:

[0207] D5 t+1 =D5 t -SSA;

[0208] Case three: if KPL t ≥KPLA, and KPL t -KPL t-1 ≥0, which indicates that the performance is stable or increases, and then the calculation formula of the weight factor D5 of the coordination potential index is as follows:

[0209] D5 t+1 =D5 t ;

[0210] Secondly, the iteration needs to set the convergence condition of iteration, and the embodiment is based on three termination conditions to converge when any one of the following three conditions is met, that is, the iteration is terminated.

[0211] Condition one: |D5 t+1 -D5 t |<0.01, which indicates that the weight change is less than 1%;

[0212] Condition two: |CS i，j，t-1 -CS i，j，t |<0.02, which indicates that the fluctuation of the average coordination scheduling decision value is less than 2%;

[0213] Condition three: the maximum number of iterations is 50 times;

[0214] Wherein:

[0215] KPL refers to the system average scheduling decision score, which reflects the task and node matching quality, and PQA is the number of task and node matching pairs;

[0216] SSA refers to the weight adjustment range, and in the embodiment, SSA = 0.02 is set;

[0217] KPL t refers to the system average scheduling decision score after the tth iteration;

[0218] KPL t-1 refers to the system average scheduling decision score after the t-1th iteration;

[0219] KPLA refers to a preset target threshold, and in the embodiment, KPLA = 0.85 is set;

[0220] CS i，j，t-1 refers to the collaborative scheduling decision value after the t-1th iteration;

[0221] CS i，j，t refers to the collaborative scheduling decision value after the tth iteration;

[0222] D5 t+1 refers to the weight factor of the collaborative potential index after the t+1th iteration;

[0223] D5 t refers to the weight factor of the collaborative potential index after the tth iteration, and the weight factor of the collaborative potential index D5 has a boundary constraint, D5 ∈ [0.05, 0.3], the lower limit avoids the collaborative potential being completely ignored, and the upper limit prevents excessive dependence on collaboration from wasting local resources.

[0224] Finally, it needs to be pointed out that: through the iteration of the weight factor D5 of the collaborative potential index, the overall system can automatically adjust the priority of local resources and node collaboration according to the actual scheduling effect;

[0225] For example, in a densely deployed edge cluster, which means that the communication delay between nodes is low, the weight factor D5 of the collaborative potential index after iteration may be increased from 0.15 to 0.25, enhancing the weight of the collaborative potential, and realizing task splitting and load balancing;

[0226] For example, in a dispersedly deployed edge node, which means that the communication delay is high, the weight factor D5 of the collaborative potential index may be reduced to 0.08, to preferentially ensure that the nodes with sufficient local resources are scheduled;

[0227] Different from the prior art, the prior art is fixed weight, and then the fixed weight may not adapt to the dynamic change of the edge network topology, such as the joining and exiting of nodes and link congestion, and then the iteration mechanism can be calibrated in real time to a certain extent;

[0228] The system optimizes the weight factor D5 of the coordination potential index, so that the coordination scheduling decision value CS i，j More accurately reflects the matching degree of tasks and nodes, and the dynamic adjustment of the coordination potential weight can reduce the resource island situation, such as the idle CPU of a node but not being scheduled due to low coordination weight;

[0229] And the weight factor D5 of the coordination potential index in the resource supply analysis submodule directly affects the coordination potential index DRH j And the coordination potential index DRH j Of node j is relatively strong, which can affect the calculation result of the coordination scheduling analysis submodule to a certain extent, and then form a closed-loop feedback and cyclic iteration form between multiple submodules. By adjusting the weight factor D5 of the coordination potential index, the influence of the local resources and distributed coordination ability of the node on the scheduling decision can be balanced, thereby improving the overall task completion rate of the system.

[0230] Although the embodiments of the present application have been shown and described, it can be understood by those skilled in the art that various changes, modifications, replacements and variations can be made to these embodiments without departing from the principles and spirits of the present application, and the scope of the present application is defined by the appended claims and their equivalents.

Claims

1. A scheduling optimization system for AI edge computing servers, characterized in that, include: Data collection module: used to collect local data and task network data from AI edge computing servers; The local data includes: the real-time core temperature of node j, the number of power grid outages of node j in the past 30 days, the real-time power consumption of node j, and the real-time number of idle CPU cores of node j. The task network data includes: the deadline of task i, the amount of input data of task i, the transmission delay from task i to node j, the average transmission bandwidth and retransmission rate of task i. Preprocessing module: Used to receive local data from edge nodes and task network data, standardize and normalize the received data, and input the processed data into the calculation and analysis module; The computational analysis module introduces the real-time idle CPU cores and real-time idle memory of node j, combines the retransmission rate and latency fluctuation characteristics, analyzes the network congestion of node j, integrates the temperature data, failure rate data and energy stability data of node j, and considers the collaborative computing capabilities of node j with neighboring edge nodes to output the dynamic resource elasticity coefficient of node j. The calculation and analysis module introduces the dynamic resource elasticity coefficient of node j, combines the deadline of task i with the current system time, introduces the input data volume of task i and the average transmission bandwidth of task i, and considers the criticality of task i in the whole business flow to output the urgency index of task i. The calculation and analysis module is based on the multiplication of the dynamic resource elasticity coefficient of node j and the urgency index of task i, combined with the heterogeneous computing power adaptability between task i and node j, the real-time power consumption of node j, and the transmission delay from task i to node j, to output the collaborative scheduling decision value. The scheduling decision module is used to input the dynamic resource elasticity coefficient of node j, the urgency index of task i, and the collaborative scheduling decision value. The scheduling decision module analyzes the input data, selects nodes to execute tasks, and responds to changes in the edge environment through adjustment mechanisms.

2. The scheduling optimization system for AI edge computing servers according to claim 1, characterized in that: The computational analysis module includes: a resource supply analysis submodule, a task allocation submodule, and a collaborative scheduling analysis submodule.

3. The scheduling optimization system for AI edge computing servers according to claim 2, characterized in that: The processing flow of the resource supply analysis submodule is as follows: I. By dividing the real-time idle CPU core count of node j by the total CPU core count of node j, the normalization of the idle CPU core count is introduced into the resource supply analysis submodule. II. By calculating the real-time free memory of node j with the total memory of node j, the real-time free memory in the node is normalized and introduced into the resource supply analysis submodule. III. By analyzing the TCP retransmission rate and delay standard deviation over the past 10 seconds and processing them based on the summation function, the network congestion of node j is analyzed, and the network congestion level is normalized by using the maximum network congestion threshold of node j. IV. By combining and analyzing the real-time core temperature of node j, the number of task failures of node j in the past 30 days, and the number of power grid outages of node j in the past 30 days, we can evaluate the temperature safety factor, fault prediction factor, and energy stability index of node j, so as to achieve the integration of node j's temperature, failure rate, and energy stability and comprehensively evaluate the health index of node j. V. By combining the number of neighboring nodes of node j, the real-time communication delay between node j and neighboring node k, and the historical collaboration success rate between node j and neighboring node k, the collaboration potential index of node j is analyzed, and the dynamic resource elasticity coefficient of node j is finally output.

4. The scheduling optimization system for AI edge computing servers according to claim 3, characterized in that: The processing flow of the task allocation submodule is as follows: I. By introducing the dynamic resource elasticity coefficient of node j into the task allocation submodule in a weighted form, the resource adaptability of the node is reflected. II. By dividing the estimated execution time of task i by the difference between the deadline of task i and the current system time, the urgency of task i can be preliminarily analyzed. III. By combining the amount of input data for task i with the average transmission bandwidth of task i, the urgency of task i can be comprehensively analyzed. IV. By combining the urgency index of subsequent task m with the dependency chain length from task i to m, the criticality of task i in the entire business flow is evaluated, and then the dynamic dependency chain coefficient of task i is analyzed, thus finally outputting the urgency index of task i.

5. The scheduling optimization system for AI edge computing servers according to claim 4, characterized in that: The processing flow of the collaborative scheduling analysis submodule is as follows: I. By multiplying the dynamic resource elasticity coefficient of node j with the urgency index of task i, the urgency of the task is associated with the resource node. II. By dividing the estimated execution time of task i by the dynamic resource elasticity coefficient of node j, the resource node is associated with the execution time. III. By combining the demand vector of task i for the e-th type of hardware with the e-th type of hardware capability of node j, the degree of matching between the hardware requirements of task i and the heterogeneous computing power characteristics of node j is quantified, so as to evaluate the heterogeneous computing power adaptability between task i and node j. IV. The transmission delay from task i to node j, the energy sensitivity coefficient of task i, and the real-time power consumption of node j are introduced into the collaborative scheduling analysis submodule to ultimately output the collaborative scheduling decision value.

6. The scheduling optimization system for AI edge computing servers according to claim 1, characterized in that: The scheduling decision module is specifically as follows: First, the dynamic resource elasticity coefficient of node j is stored in the real-time database of the scheduling center; Then, based on the urgency index of task i, the urgency indices of all tasks are sorted in descending order to form a task priority queue. Then, the collaborative scheduling decision value is stored in the matching matrix of the scheduling center.

7. The scheduling optimization system for AI edge computing servers according to claim 1, characterized in that: The selection of nodes to execute tasks in the scheduling decision module specifically involves: First, traverse the task priority queue and select the node j with the highest collaborative scheduling decision value for each task. The heterogeneous computing power compatibility between task i and node j must be greater than or equal to 0.6 to avoid serious hardware mismatch. Secondly, if the dynamic resource elasticity coefficient of subsequent node j is less than 0.3, it indicates that resources are tight. In this case, even if the collaborative scheduling decision value is high, the node with the second highest value will be forcibly selected to avoid single-point overload.

8. The scheduling optimization system for AI edge computing servers according to claim 1, characterized in that: The adjustment mechanism in the scheduling decision module is as follows: First, the matching matrix of the collaborative scheduling decision value is recalculated every 2 seconds. For tasks that have been assigned but not executed, such as urgent tasks i whose urgency index has increased by more than or equal to 20%, node migration is triggered. At this time, other nodes with higher collaborative scheduling decision values ​​are selected. Secondly, if the fault prediction coefficient of node j is less than 0.5, it indicates a high fault risk. In this case, the task of node j should be assigned in advance to a neighboring node of node j with a dynamic resource elasticity coefficient greater than or equal to 0.8 and a transmission delay from task i to node j less than or equal to 20ms, in order to avoid task interruption.

Images