SLA demand-oriented computing network multi-element fusion arrangement method

Abstract

The invention provides an SLA demand-oriented computing network multi-element fusion arrangement method, which comprises the following steps of: matching a demand timeliness priority queue with a directed acyclic graph of container arrangement, and searching an arrangement scheme with optimal resource utilization efficiency and most balanced load by applying a dynamic planning method to form an initial arrangement decision; in combination with a preset calculation network element arrangement strategy template, dynamic selection and adjustment are carried out according to actual conditions, including selection of an arrangement strategy with a preferential resource utilization rate in a service peak period and selection of an arrangement strategy with a preferential cost in a service valley period; and recording and analyzing various calculation network elements generated in a container arrangement process, including resource occupation, starting time, a dependency relationship and an arrangement decision, and optimizing a resource prediction and scheduling algorithm according to historical data.

Description

Technical Field

[0001] The present invention relates to the field of information technology, and in particular to a computing network multi-factor fusion orchestration method oriented to SLA requirements. Background Art

[0002] In the multi-factor fusion orchestration of computing networks, the introduction of demand timeliness has brought new technical challenges. The time sensitivity and urgency of business needs vary greatly. How to accurately evaluate the priority of different needs and reasonably arrange resource scheduling and task execution order based on priority is a complex decision-making problem. Although container technology can achieve rapid deployment and dynamic scaling, in actual applications, the startup time, resource occupation and dependency of different containers vary greatly. How to perform refined orchestration based on demand timeliness and container characteristics still faces many difficulties. The decision-making of multi-factor orchestration of computing networks involves multiple dimensions such as resource selection, scheduling and allocation. How to balance resource utilization efficiency, load balancing and system stability while meeting demand timeliness requires balancing multiple decision-making goals. In summary, the multi-factor fusion orchestration of computing networks oriented to demand timeliness still needs in-depth research in demand assessment, container orchestration, decision optimization and dynamic adjustment to meet increasingly complex business needs. Summary of the invention

[0003] The present invention provides a computing network multi-factor fusion orchestration method oriented to SLA requirements, which mainly includes:

[0004] According to the time when the requirements are submitted and the expected completion time, the preset requirements priority evaluation rules are used to analyze the time sensitivity and urgency of each requirement, calculate the priority values ​​of different requirements, and sort the requirements in descending order of priority values ​​to form a requirements time priority queue;

[0005] Use heuristic methods to model the startup time, resource usage, and dependencies of containers in the computing network. For each container to be orchestrated, evaluate its expected startup time and resource consumption. Consider the dependency constraints between containers and build a directed acyclic graph of container orchestration.

[0006] Match the demand time priority queue with the directed acyclic graph of container orchestration, and use dynamic programming methods to find the orchestration solution with the best resource utilization efficiency and the most balanced load to form an initial orchestration decision;

[0007] During the initial orchestration decision execution process, the system operation status and resource usage are continuously monitored, potential performance bottlenecks and stability risks are identified through anomaly detection methods, and the orchestration decisions are dynamically adjusted in combination with reinforcement learning methods to optimize resource allocation and task scheduling to ensure the fulfillment of the service level agreement (SLA).

[0008] Combine the pre-set computing network element orchestration strategy templates and dynamically select and adjust according to actual conditions, including selecting an orchestration strategy that prioritizes resource utilization during business peaks and selecting an orchestration strategy that prioritizes costs during business troughs;

[0009] Record and analyze various computing network elements generated in the container orchestration process, including resource usage, startup time, dependencies, and orchestration decisions, and optimize resource prediction and scheduling algorithms based on historical data;

[0010] Based on the historical data of demand timeliness and container orchestration, case reasoning and rule reasoning methods are used to assist in the formulation and optimization of orchestration decisions, and provide users with the basis and adjustment suggestions for orchestration decisions according to the provisions of the service level agreement (SLA).

[0011] The technical solution provided by the embodiment of the present invention may have the following beneficial effects:

[0012] The present invention discloses a multi-factor fusion orchestration method for a computing network oriented to SLA requirements. The method first establishes a demand time priority queue, and constructs an orchestration directed acyclic graph in combination with container startup time, resource occupancy and dependency. The optimal orchestration scheme is found through dynamic programming, and the system status is continuously monitored during execution, and the decision is dynamically adjusted using anomaly detection and reinforcement learning methods. The present invention also dynamically selects orchestration strategies based on business peaks and troughs, and optimizes resource prediction and scheduling algorithms based on historical data. In addition, the present invention uses case reasoning and rule reasoning to assist decision making, and provides orchestration suggestions based on SLA agreements. The method achieves a multi-objective balance of demand timeliness, resource utilization and service quality, and effectively improves the intelligence level and operating efficiency of computing network resource scheduling. BRIEF DESCRIPTION OF THE DRAWINGS

[0013] Figure 1 The present invention is a flow chart of a method for fusion orchestration of multiple elements of a computing network oriented to SLA requirements.

[0014] Figure 2 It is a schematic diagram of a computing network multi-factor fusion orchestration method oriented to SLA requirements of the present invention.

[0015] Figure 3 It is another schematic diagram of a computing network multi-factor fusion orchestration method oriented to SLA requirements of the present invention. DETAILED DESCRIPTION

[0016] In order to enable those skilled in the art to better understand the technical solutions in this specification, the technical solutions in the embodiments of this specification will be clearly and completely described below in conjunction with the drawings in the embodiments of this specification. Obviously, the described embodiments are only part of the embodiments of this specification, not all of the embodiments. Based on the embodiments in this specification, all other embodiments obtained by ordinary technicians in this field without creative work should fall within the scope of protection of this specification.

[0017] like Figure 1-3 In this embodiment, a computing network multi-factor fusion orchestration method oriented to SLA requirements may specifically include:

[0018] Step S101, based on the time when the demand is submitted and the expected completion time, the preset demand priority evaluation rules are used to analyze the time sensitivity and urgency of each demand, calculate the priority values ​​of different demands, sort the demands in descending order of priority values, and form a demand time priority queue.

[0019] The time difference is calculated based on the time when the demand is proposed and the expected completion time, and the time difference is quantitatively scored using the preset evaluation rules to obtain the time sensitivity score of the demand; wherein the preset evaluation rules include a mapping relationship between the time difference interval and the corresponding urgency level. The demand content keywords are obtained, and the demand business importance score is determined by matching the preset business importance vocabulary; the preset business importance vocabulary includes the corresponding relationship between keywords and importance. It is judged whether the time sensitivity score and the business importance score are obtained. If the scores are obtained, the urgency of the demand is calculated using the weighted summation method; wherein the time sensitivity weight is the first weight value of the weighted summation, and the business importance weight is the second weight value of the weighted summation. The urgency of the demand is converted into a priority value through a linear mapping function; the linear mapping function includes the minimum and maximum parameters of all the urgency of the demand. According to the priority value, all the demands are sorted in descending order using a merge sort algorithm to obtain a demand sequence arranged from high to low priority; demand information is extracted one by one from the demand sequence, and a maximum heap data structure is constructed as a priority queue.

[0020] Exemplarily, the time difference is calculated based on the time when the demand is proposed and the expected completion time, and the time difference is quantitatively scored according to the preset evaluation rules to obtain the time sensitivity score of the demand. The preset evaluation rules include a mapping relationship between the time difference interval and the corresponding urgency level, such as 0-1 days for extremely urgent, 1-3 days for urgent, 3-7 days for general, and more than 7 days for ordinary. The demand content keywords are matched with the preset business importance vocabulary to obtain the demand business importance score. The business importance vocabulary contains the correspondence between keywords and importance, such as "core business" has a weight of 5, "conventional business" has a weight of 3, and "auxiliary function" has a weight of 1. Combine the time sensitivity score and the business importance score for weighted summation, the time sensitivity weight is 0.6, and the business importance weight is 0.4, to obtain the urgency of the demand. Use a linear mapping function to convert the urgency into a priority value of 0-100. The mapping function is f(x)=100*(x-min) / (max-min), where x is the urgency, min and max are the minimum and maximum urgency values ​​of all requirements, respectively. For the priority value of each requirement, use the merge sort algorithm to sort all requirements in descending order to obtain a requirement sequence arranged from high to low priority. Extract the requirement information one by one from the sorted requirement sequence, and construct a maximum heap data structure as a priority queue. The top element of the heap is the highest priority requirement, and the bottom element of the heap is the lowest priority requirement. The maximum heap performs floating and sinking operations by comparing the node priority values ​​to maintain the priority order of the requirements.

[0021] In the demand management system, first obtain the time when the demand is proposed and the expected completion time, and calculate the time difference. For example, a demand is proposed on 2023-10-01, and the expected completion time is 2023-10-05. The calculated time difference is 4 days. According to the preset evaluation rules, 4 days belongs to the "general" urgency level, and the corresponding time sensitivity score is 60 points. Next, extract keywords from the demand content. For example, if the "core business" keyword is extracted, the business importance vocabulary is queried, and the corresponding weight is 5, and the business importance score is calculated to be 100 points. The time sensitivity score and the business importance score are weighted and summed, with weights of 0.6 and 0.4 respectively, and the demand urgency is 76 points. Assuming that the lowest urgency of all demands is 30 points and the highest is 90 points, use a linear mapping function to convert 76 points into a priority value, and get 76.67. Merge and sort the priority values ​​of all demands, and build the sorted demand sequence into a maximum heap structure. The root node of the maximum heap stores the highest priority demand, such as the demand with a priority of 76.67. When a new demand is inserted or the demand priority changes, the heap structure is adjusted through the floating and sinking operations of the heap to ensure that the top of the heap is always the current highest priority demand. In this way, the demand management system can obtain the most urgent demand from the top of the heap at any time for processing, realizing dynamic priority management of demands.

[0022] Obtain demand change frequency and fulfillment assessment indicators, and determine demand stability through threshold comparison. If changes are frequent and SLA agreement assessment indicators are not met, re-evaluate the demand and update the demand queue.

[0023] Obtain the historical records of demand changes, calculate the number of demand changes per month based on the historical records, and obtain the demand change frequency value; compare the demand change frequency value with the preset change frequency threshold, if the demand change frequency value is higher than the preset change frequency threshold, it is determined that the changes are frequent, otherwise it is determined that the changes are stable; extract the performance assessment index value from the SLA agreement, the performance assessment index value includes response time, resolution time, and customer satisfaction; use a multi-factor scoring method to quantitatively evaluate the performance assessment index value, set the response time weight, resolution time weight, and customer satisfaction weight, perform weighted calculation, and obtain the performance assessment index value. Assessment score; calculate a comprehensive assessment index based on the result of the demand change frequency determination and the performance assessment score, the comprehensive assessment index being equal to the change frequency multiplied by the weight coefficient plus the performance assessment score multiplied by the weight coefficient; use a rule engine to adjust the demand priority according to different intervals of the comprehensive assessment index, if the comprehensive assessment index is less than a first threshold, lower the demand priority, if the comprehensive assessment index is greater than a second threshold, increase the demand priority; based on the adjusted priority score, use an insertion sort algorithm to update the demand priority queue, insert the demand into the appropriate position, and ensure that the demands in the queue are arranged from high to low priority.

[0024] For example, the demand change history is obtained, the number of demand changes per month is calculated, and the demand change frequency value is obtained. At the same time, the performance assessment index value is extracted from the SLA agreement, including response time, resolution time, customer satisfaction, etc. The calculated demand change frequency value is compared with the preset change frequency threshold of 3 times / month. If it is higher than the threshold, it is determined to be frequent changes, otherwise it is determined to be stable changes. The multi-factor scoring method is used to quantitatively evaluate the performance assessment index, and the response time weight is set to 0.3, the resolution time weight is set to 0.4, and the customer satisfaction weight is set to 0.3. Weighted calculation is performed to obtain the performance assessment score to determine whether the preset SLA agreement assessment standard is met. According to the determination result of the demand change frequency and the performance assessment score, the comprehensive evaluation index is calculated, and the comprehensive evaluation index = 0.5 change frequency + 0.5 performance assessment score. The rule engine is used to adjust the demand priority according to different intervals of the comprehensive evaluation index. If the comprehensive evaluation index is less than 60 points, the demand priority is reduced by 2 levels; if the comprehensive evaluation index is between 60-80 points, the original priority is maintained; if the comprehensive evaluation index is greater than 80 points, the demand priority is increased by 1 level. According to the adjusted priority score, the insertion sort algorithm is used to update the demand priority queue, insert the demand into the appropriate position, and ensure that the demand in the queue is arranged from high to low priority. In the demand management system, the change records of a certain demand in the past three months are first obtained from the database, and the number of changes is counted as 7 times, and the average monthly change frequency is calculated to be 2.33 times. At the same time, the performance assessment indicators of the demand are extracted from the SLA agreement, including the average response time of 4 hours, the average resolution time of 2 days, and the customer satisfaction score of 85 points. The average monthly change frequency of 2.33 times is compared with the preset threshold of 3 times / month, and it is determined that the change is stable. Then, the performance indicators are evaluated using the multi-factor scoring method, setting the response time weight of 0.3, the resolution time weight of 0.4, and the customer satisfaction weight of 0.3, and the performance assessment score is calculated to be 88 points. According to the change frequency judgment result and the performance assessment score, the comprehensive evaluation index is calculated to obtain 0.5100+0.588=94 points. Using the rule engine to judge, 94 points are greater than 80 points, and it is decided to increase the priority of the demand by 1 level. Assume that the original priority of the demand is 3, and after adjustment it becomes 2. Finally, use the insertion sort algorithm to update the demand priority queue, remove the demand from the original position, and then insert it into the position with priority 2, ensuring that all demands in the queue are arranged in order from high to low priority.

[0025] Step S102, use heuristic methods to model the startup time, resource usage and dependency relationships of containers in the computing network, evaluate the expected startup time and resource consumption of each container to be orchestrated, and consider the dependency constraints between containers to construct a directed acyclic graph of container orchestration.

[0026] The operation data collected every hour in the past preset time period of the container to be orchestrated is obtained, and the operation data includes the startup time, CPU usage rate and memory occupancy; based on the operation data, the expected startup time and resource consumption of each container are predicted by using the autoregressive integrated moving average model; for the container to be orchestrated, the dependency relationship between the containers is analyzed by the Apriori algorithm, and a container dependency matrix is ​​established, wherein the matrix element value represents the dependency strength between the containers; if the dependency strength exceeds the preset threshold, it is determined that there is a dependency constraint; based on the expected startup time, resource consumption and dependency constraints, an initial solution for container orchestration is constructed, and a directed graph including node weights and edge weights is generated, wherein the node weight is the weighted sum of the expected startup time and resource consumption, and the edge weight is the dependency strength; the initial directed graph is processed by using a topological sorting algorithm, the edges that form loops are detected and removed, the directions of the edges are adjusted, and the final container orchestration directed acyclic graph is obtained, wherein the path length in the directed acyclic graph represents the container startup sequence and time; the container startup sequence is determined according to the topological order of the directed acyclic graph, and resources are allocated in combination with the node weights to generate the final container orchestration solution.

[0027] Exemplarily, the operation data collected every hour in the past 30 days of the container to be orchestrated is obtained, including the startup time, CPU usage, and memory usage, and the expected startup time and resource consumption of each container are predicted by the autoregressive integrated moving average model. The Apriori algorithm is used to analyze the dependencies between containers and establish a container dependency matrix. The element values ​​in the matrix represent the dependency strength between containers. If the dependency strength exceeds 0.7, it is determined that there is a dependency constraint. According to the predicted startup time, resource consumption, and dependency constraints, the initial solution for container orchestration is constructed, and a directed graph containing node weights and edge weights is generated. The node weight is the weighted sum of the expected startup time and resource consumption, and the edge weight is the dependency strength. The initial directed graph is processed using a topological sorting algorithm, and the edges that form loops are detected and removed, and the direction of the edges is adjusted to obtain the final container orchestration directed acyclic graph. The path length in the graph represents the container startup order and time. The container startup order is determined according to the topological order of the directed acyclic graph, and the node weights are considered for resource allocation to generate the final container orchestration solution. In the container orchestration system, firstly, the hourly operation data of the orchestrated containers A, B, and C in the past 30 days are obtained from the monitoring database, including the startup time, CPU usage, and memory usage. The autoregressive integrated moving average model is used to analyze these data, and the expected startup time of container A is predicted to be 10 seconds, the CPU usage is 30%, and the memory usage is 500MB; the expected startup time of container B is 15 seconds, the CPU usage is 40%, and the memory usage is 700MB; the expected startup time of container C is 8 seconds, the CPU usage is 25%, and the memory usage is 400MB. Then, the Apriori algorithm is used to analyze the dependencies between containers, and the minimum support is set to 0.3 and the minimum confidence is set to 0.7. The container dependency matrix is ​​obtained: the dependency strength between A and B is 0.8, the dependency strength between A and C is 0.5, and the dependency strength between B and C is 0.6. According to the prediction results and dependency relationships, an initial directed graph is constructed, in which the node weight is the normalized weighted sum of the startup time and resource consumption, and the edge weight is the dependency strength. Use the topological sorting algorithm to process the initial directed graph, remove the edge from A to B, adjust the edge direction from C to B, and get the final directed acyclic graph. According to the path in the graph, the container startup order is determined to be C->A->B. Considering the node weight, resources are allocated: C uses 25% CPU and 400MB memory, A uses 30% CPU and 500MB memory, and B uses the remaining resources.

[0028] Step S103, matching the demand time priority queue with the directed acyclic graph of container orchestration, using a dynamic programming method to find an orchestration solution with optimal resource utilization efficiency and the most balanced load, and forming an initial orchestration decision.

[0029] Obtain a demand time priority queue and a container orchestration directed acyclic graph, and construct a matching matrix according to the demand time priority queue and the container orchestration directed acyclic graph, wherein the element value of the matching matrix represents the matching degree between the demand and the container. Use a non-dominated sorting genetic algorithm II to perform multi-objective optimization on the matching matrix, wherein the multi-objective optimization includes maximizing the sum of matching degrees, maximizing resource utilization, and minimizing load variance, and obtain a set of non-dominated solutions through iterative evolution. Select a balanced solution from the non-dominated solution set as the final solution, specifically including: calculating the comprehensive score of each solution, and selecting the solution with the highest score as the optimal matching solution. For the optimal matching solution, calculate the resource allocation ratio of each container, and the calculation method of the resource allocation ratio is: allocate resources such as CPU and memory according to the weighted sum of demand priority and container dependency strength, wherein the weight coefficient is obtained according to historical data statistics. Based on the resource allocation ratio, use the minimum connection number algorithm to allocate nodes to the container, and the minimum connection number algorithm includes: calculating the current number of connections and resource utilization of each node as load indicators, selecting the node with the smallest load indicator as the container deployment target, and forming an initial orchestration decision.

[0030] Exemplarily, the demand time priority queue and the container orchestration directed acyclic graph are obtained, and a matching matrix is ​​constructed. The matrix element value represents the matching degree between the demand and the container. The matching score is obtained by calculating the cosine similarity between the demand priority vector and the container startup order vector. According to the matching matrix, the non-dominated sorting genetic algorithm II is used for multi-objective optimization, and the optimization objectives are set to maximize the sum of matching degrees, maximize resource utilization, and minimize load variance. A set of non-dominated solutions is obtained through iterative evolution. A balanced solution is selected from the non-dominated solution set as the final solution, and the comprehensive score of each solution is calculated. The solution with the highest score is selected as the optimal matching solution. For the selected optimal matching solution, the resource allocation ratio of each container is calculated, and resources such as CPU and memory are allocated according to the weighted sum of demand priority and container dependency strength. The weight coefficient is obtained according to historical data statistics. Based on the resource allocation results, the minimum connection number algorithm is used to allocate nodes to the container, and the current number of connections and resource utilization of each node are calculated as load indicators. The node with the smallest load indicator is selected as the container deployment target to form an initial orchestration decision. In the container orchestration system, first obtain the demand time priority queue [A(90), B(80), C(70)] and the container orchestration directed acyclic graph {1->2->3} to build a 3x3 matching matrix. By calculating the cosine similarity between the demand priority vector [90,80,70] and the container startup order vector [3,2,1], we can get the matching score matrix

[0031] [[0.9,0.8,0.7],[0.8,0.9,0.6],[0.7,0.6,0.9]]. The non-dominated sorting genetic algorithm II was used for multi-objective optimization, with the population size set to 100, the number of iterations to 50, the crossover probability to 0.8, and the mutation probability to 0.1, to obtain a set of non-dominated solutions. From the non-dominated solution set, the solution with the highest comprehensive score [A->1, B->2, C->3] was selected as the optimal matching solution, with a total matching degree of 2.7, a resource utilization rate of 85%, and a load variance of 0.02. According to the optimal matching solution, the container resource allocation ratio was calculated, the demand priority weight was set to 0.6, and the container dependency strength weight was set to 0.4, and the resource allocation ratios of containers 1, 2, and 3 were 40%, 35%, and 25%, respectively. Use the minimum connection number algorithm to allocate nodes for containers. Assume that there are 3 nodes, the current connection numbers are [5, 3, 4], and the resource utilization rate is [70%, 60%, 65%]. After comprehensive consideration, select node 2 as the deployment target of container 1, update the node load index, and repeat the process until all containers are allocated, forming the initial orchestration decision [1->node 2, 2->node 3, 3->node 1].

[0032] If problems such as container startup failure and sudden increase in resource usage occur during the computing network container orchestration process, machine learning methods are used to analyze the container operation status in real time, detect anomalies in time, and trigger predefined processing strategies, including restarting the container and dynamically adjusting resources.

[0033] The container operation log and resource monitoring data are obtained, and key features are extracted from the data every 5 seconds, wherein the key features include startup time, CPU usage, memory usage, network traffic, and disk I / O; a multidimensional feature vector is constructed based on the key features to obtain a representation of the container operation status; an isolation forest algorithm is used to perform real-time anomaly detection on the multidimensional feature vector, and if the anomaly score exceeds a preset threshold or the resource usage rate increases by more than a preset percentage within a predetermined time, it is determined to be an abnormal state; for the abnormal state, the anomaly type, anomaly score, resource usage rate and its growth rate are extracted to generate a decision tree input feature; according to the decision tree input feature, a processing solution is selected from a predefined policy library through a decision tree algorithm, and the predefined policy library includes restarting the container, increasing the resource quota, and migrating the container; the processing solution is executed to obtain the policy execution effect, wherein the policy execution effect includes the stable operation time of the container and the degree of improvement of the resource utilization rate; the policy execution effect is fed back to the machine learning model to optimize the anomaly detection threshold, the policy selection weight, and the processing solution parameters.

[0034] Exemplarily, key features are extracted from the container operation log and resource monitoring data every 5 seconds, including startup time, CPU usage, memory usage, network traffic and disk I / O, to construct a multidimensional feature vector as a representation of the container operation status. The isolationforest algorithm is used to perform real-time anomaly detection on the feature vector, and the anomaly score threshold is set to 0.8. At the same time, the growth rate of resource usage is monitored. If the CPU or memory usage increases by more than 50% within 1 minute, it is determined to be a sudden increase in resource usage. The detected anomaly information is converted into decision tree input features, including anomaly type, anomaly score, resource usage and its growth rate. According to the input features, the decision tree algorithm is used to select a suitable processing solution from the predefined policy library. The policy library contains specific policies such as restarting the container, increasing resource quotas, and migrating containers. The above-mentioned restarting container is applicable to startup failures, and the parameters are: the maximum number of retries is 3 times; increasing resource quotas is applicable to sudden increases in resource usage, and the parameters are: CPU increases by 20%, memory increases by 30%; migrating containers is applicable to excessive host load, and the parameters are: target node selection rules. Execute the selected processing strategy, monitor the execution effect of the strategy, including the stable running time of the container and the improvement of resource utilization, and feed the execution results and effect evaluation data back to the machine learning model for continuous optimization of anomaly detection thresholds, strategy selection weights, and processing scheme parameters. In the container orchestration system, the monitoring program extracts running data from container A every 5 seconds and records it into the feature vector [30s, 85%, 75%, 200Mbps, 50IOPS], which means the startup time is 30 seconds, CPU usage is 85%, memory usage is 75%, network traffic is 200Mbps, and disk I / O is 50IOPS.

[0035] The IsolationForest algorithm analyzes the feature vector and calculates an anomaly score of 0.92, which exceeds the threshold of 0.8. At the same time, it detects that the CPU usage has increased from 60% to 85% in the past minute, with a growth rate of more than 50%, and determines that it is a sudden increase in resource usage. The system converts the anomaly information into a decision tree input [1, 0.92, 85%, 41.67%], indicating that the anomaly type is a sudden increase in resources, an anomaly score of 0.92, a CPU usage of 85%, and a growth rate of 41.67% / minute. Based on these features, the decision tree algorithm selects the "Increase resource quota" policy from the predefined policy library, with the specific parameters of increasing the CPU quota by 20%, from 2 cores to 2.4 cores. The system executes the policy and re-evaluates it 30 minutes later, finding that the CPU usage of container A has dropped to 65%, and the stable running time has been extended from an average of 10 minutes to 40 minutes. The system feeds this result back to the machine learning model, adjusts the CPU usage threshold for anomaly detection from 85% to 90%, and increases the selection weight of the "Increase resource quota" policy in the CPU sudden increase scenario.

[0036] Step S104: During the execution of the initial orchestration decision, the system operation status and resource usage are continuously monitored, potential performance bottlenecks and stability risks are identified through anomaly detection methods, and the orchestration decision is dynamically adjusted in combination with reinforcement learning methods to optimize resource allocation and task scheduling to ensure the fulfillment of the service level agreement (SLA).

[0037] The system operation indicators collected by the distributed monitoring tool are obtained, and the system operation indicators include CPU utilization, memory usage, network throughput and disk I / O; an adaptive sampling interval mechanism is used according to the system operation indicators to dynamically adjust the sampling frequency and store the collected data in a time series database in real time; an isolation forest algorithm is used to detect anomalies in the multidimensional time series data stored in the time series database, and the degree of anomalies is determined by a multi-level threshold mechanism; a state representation of a reinforcement learning environment is generated according to the anomaly detection result, and the state representation includes anomaly indicators, anomaly degree and duration; a reinforcement learning environment is constructed, and if the state representation meets the preset conditions, the system state is used as the observation space, and the resource allocation and task scheduling operations are used as the action space; a reinforcement learning agent is trained by a deep Q network algorithm, and the deep Q network algorithm inputs the current system state and anomaly detection results, and outputs the resource allocation ratio and task priority adjustment value; the scheduling plan is updated according to the resource allocation ratio and the task priority adjustment value, and the corresponding adjustment operation is performed.

[0038] Exemplarily, the system operation indicators, including CPU utilization, memory usage, network throughput and disk I / O, are collected through distributed monitoring tools. The adaptive sampling interval mechanism is used to dynamically adjust the sampling frequency according to the system load, which is initially set to 30 seconds, and the collected data is stored in the time series database in real time. The isolation forest algorithm is used to detect anomalies in the collected multi-dimensional time series data, and a multi-level threshold mechanism is introduced to set different anomaly judgment criteria for CPU, memory, network and disk I / O, which are 90%, 85%, 80% and 75% respectively. At the same time, the trend of historical data is considered to identify potential performance bottlenecks and stability risk points. The anomaly detection results are converted into the state representation of the reinforcement learning environment, and a state vector containing anomaly indicators, anomaly degree and duration is constructed. According to the state vector, a reinforcement learning environment is constructed, with the system state as the observation space, resource allocation and task scheduling operations as the action space, and SLA fulfillment as the reward function, where the SLA fulfillment is calculated by the weighted sum of the response time weight 0.4, the availability weight 0.3 and the throughput weight 0.3. The reinforcement learning agent is trained using the deep Q network algorithm. The network structure includes an input layer with a state vector dimension, two hidden layers with 64 neurons each, and an output layer with an action space dimension. The current system state and anomaly detection results are input, and the optimal resource allocation ratio and task priority adjustment value are output. The orchestration scheme is updated and the corresponding adjustment operation is performed. In the container orchestration system, the distributed monitoring tool initially collects system indicators at intervals of 30 seconds. When the CPU utilization rate is detected to increase to 85%, the sampling interval is automatically adjusted to 10 seconds to increase data accuracy. The collected data shows that the CPU utilization rate of a container is 92% and the memory usage is 88%, which triggers the anomaly detection of the isolation forest algorithm. The algorithm calculates that the anomaly score of the container is 0.78, which exceeds the preset threshold of 0.7, and the system marks it as a performance bottleneck. The anomaly detection result is converted into a state vector [1, 0.92, 0.88, 0, 5], which indicates that the CPU is abnormal, the utilization rate is 92%, the memory usage is 88%, the network is normal, and the duration is 5 minutes. Based on the state vector, the reinforcement learning environment calculates the current SLA fulfillment as 0.82. The deep Q network receives these inputs, processes them through two hidden layers of 64 neurons, and outputs action recommendations: increase the CPU resource allocation of the container by 20%, increase the memory allocation by 15%, and increase its task priority by 2 levels. The system executes these adjustments and re-evaluates 30 minutes later, finding that the CPU utilization has dropped to 78%, the memory usage has dropped to 75%, and the SLA fulfillment has increased to 0.91, confirming the effectiveness of the adjustments.

[0039] Step S105, combining the preset network element orchestration strategy template, and dynamically selecting and adjusting according to the actual situation, including selecting an orchestration strategy that prioritizes resource utilization during business peak periods, and selecting an orchestration strategy that prioritizes cost during business trough periods.

[0040] Obtain business load data, the business load data including request volume, CPU utilization and memory occupancy. Based on the business load data, use the ARIMA model to predict the business load trend within the next hour. Determine whether the business load trend exceeds the preset threshold. If the business load trend exceeds the preset threshold, select a resource utilization priority strategy from the preset computing network element orchestration policy template library, otherwise select a cost priority strategy. Convert the selected strategy into a resource allocation plan, which includes horizontal and vertical expansion of container instances. Use a deep Q network algorithm to dynamically optimize policy parameters. The deep Q network algorithm uses system load, resource utilization and cost indicators as state space, resource allocation adjustment as action space, and the weighted sum of performance and cost as the reward function to obtain optimized policy parameters. According to the optimized policy parameters, adjust the number of container instances, CPU core allocation, memory limit and network bandwidth configuration to achieve adaptive adjustment of resource allocation.

[0041] Exemplarily, obtain the business load data of the current time period, including the request volume, CPU utilization, and memory occupancy rate, and use the ARIMA model to predict the business load trend within the next hour. According to the predicted business load trend and business type, select a suitable strategy from the preset computing network element orchestration strategy template library. If the predicted load exceeds the preset threshold, select the resource utilization priority strategy, otherwise select the cost priority strategy. Convert the selected strategy into a specific resource allocation plan, including horizontal expansion of container instances, such as increasing the number of instances, and vertical expansion, such as increasing the resource quota of a single instance. For the resource utilization priority strategy, increase the number of container instances by 20% and increase the CPU core allocation by 30%; for the cost priority strategy, reduce the number of idle instances by 15% and reduce the resource quota of non-critical businesses by 10%. Apply the resource allocation plan to the current computing network environment, and adjust the number of container instances, CPU core allocation, memory limit, and network bandwidth configuration. The policy execution effect is continuously monitored, and the deep Q network algorithm is used to dynamically optimize the policy parameters. The system load, resource utilization and cost indicators are used as the state space, and the resource allocation adjustment is used as the action space. The weighted sum of performance and cost is used as the reward function. The resource utilization threshold and cost control indicators are adjusted according to the actual operation data to achieve adaptive adjustment of the policy. In the computing network orchestration system, the monitoring program collects business load data every 5 minutes and records it into a time series [70%, 75%, 80%, 82%, 85%], which represents the change in CPU utilization in the past 25 minutes. The ARIMA model analyzes the sequence and predicts that the CPU utilization will reach 90% in the next hour, exceeding the 80% threshold for computing-intensive services. The system selects the "resource utilization priority" policy from the policy template library and converts it into a specific plan: increase the number of container instances from 50 to 60, and increase the CPU core allocation from 8 cores / instance to 10 cores / instance. The resource scheduler executes the plan and completes the deployment of 60 container instances and the reallocation of CPU resources within 5 minutes. The Deep Q Network algorithm continuously monitors the system status, takes [CPU utilization 90%, memory usage 75%, cost increase 15%] as the current state input, and outputs the action recommendation after the neural network calculation: add 5 more container instances, but keep the CPU core allocation unchanged. After the system executes this recommendation, the CPU utilization drops to 75% within 30 minutes, which is below the threshold. The Deep Q Network receives a positive reward and updates the Q value table to increase the probability of this action being selected in similar states.

[0042] Step S106, record and analyze various network computing elements generated in the container orchestration process, including resource occupancy, startup time, dependencies, and orchestration decisions, and optimize resource prediction and scheduling algorithms based on historical data.

[0043] Use a distributed log collection tool to obtain indicator data in the container orchestration process, and the indicator data includes CPU usage, memory usage, network traffic, disk I / O, startup time, and dependency graph. Execute a sliding window algorithm based on the indicator data to obtain statistical features in each time window, and the statistical features include average, maximum, minimum, and standard deviation. Construct a feature vector based on the statistical features, and the feature vector contains statistical features and dependency information. Input the feature vector into the random forest algorithm to obtain a resource prediction model, and the resource prediction model is used to predict resource requirements within the next hour. If the resource prediction model outputs a prediction result, a genetic algorithm is used to optimize the container scheduling strategy, and the genetic algorithm uses a binary string to represent the allocation relationship between containers and nodes. Generate a new scheduling decision based on the genetic algorithm, and apply the new scheduling decision to the container orchestration system.

[0044] Exemplarily, through the distributed log collection tool, various indicator data in the container orchestration process are collected every 5 minutes, including CPU usage, memory usage, network traffic, disk I / O, startup time and dependency graph, and the collected data is stored in the time series database in a predefined format. The sliding window algorithm is used to preprocess the stored historical data, setting the window size to 24 hours and the step length to 1 hour, extracting the statistical features in each time window, including the average, maximum, minimum and standard deviation, and constructing the dependency matrix between containers. The preprocessed data is converted into a feature vector containing statistical features and dependency information as the input of the random forest algorithm. The random forest algorithm is used to build a resource prediction model, setting the number of trees to 100 and the maximum depth to 10. The converted feature vector is used as input to predict resource requirements within the next hour, including CPU, memory and network bandwidth. Based on the prediction results, the genetic algorithm is used to optimize the container scheduling strategy. The population size is set to 100, the number of iterations is set to 50, the encoding method is a binary string to represent the allocation relationship between containers and nodes, the crossover operation uses a single-point crossover, the mutation probability is 0.01, and the optimization goals are to maximize resource utilization and minimize load imbalance. Generate new scheduling decisions and apply them to the container orchestration system. Regularly evaluate the accuracy of the prediction model and scheduling algorithm, adjust the model parameters and optimize the target weights according to the actual operation results. In the container orchestration system, the distributed log collection tool collects data from 100 containers every 5 minutes, recording indicators such as CPU usage, memory usage, network traffic, and disk I / O. Taking container A as an example, the data collected at a certain time point is: CPU usage 75%, memory usage 2GB, network traffic 100Mbps, disk I / O, 200IOPS, and startup time 30 seconds. At the same time, the dependency relationship between container A and B and C is recorded. The sliding window algorithm processes 24 hours of historical data and calculates the average CPU usage of container A to be 70%, the maximum value 85%, the minimum value 50%, and the standard deviation 10%. The constructed dependency matrix shows that the dependency between A and B is 0.8, and the dependency between A and C is 0.5. Converting these data into feature vectors

[0045] [70,85,50,10,0.8,0.5] Input the random forest model. The random forest algorithm uses 100 decision trees with a maximum depth of 10. After training, it predicts the resource requirements of container A in the next hour as follows: CPU utilization rate 80%, memory requirement 2.5GB. The genetic algorithm receives this prediction result, and the initial population contains 100 candidate solutions. Each solution uses a 200-bit binary string to represent the allocation scheme of 100 containers on 2 nodes. After 50 iterations, crossover and mutation operations, the optimal solution is obtained: container A is allocated to node 1, resource utilization is increased to 85%, and load imbalance is reduced to 0.1. The system executes this solution and evaluates the prediction accuracy after 24 hours, reaching 90%. Based on this, the number of trees in the random forest is fine-tuned to 110, and the mutation probability of the genetic algorithm is increased to 0.015 to improve the effect of the next round of optimization.

[0046] Step S107: Based on the demand timeliness and historical data of container orchestration, case reasoning and rule reasoning methods are used to assist in the formulation and optimization of orchestration decisions, and provide users with the basis and adjustment suggestions for orchestration decisions according to the agreement of the service level agreement (SLA).

[0047] The historical data of demand timeliness and container orchestration are obtained from the database, and the historical data is constructed into a case library. The case library adopts a key-value pair storage structure, in which the key includes the demand ID and timestamp, and the value includes the demand description, timeliness requirement, resource allocation, orchestration decision and execution result; according to the cases in the case library, the K nearest neighbor algorithm is used to calculate the similarity between the current demand and the case, and the optimal K value is selected through five-fold cross validation, and the top K cases with the highest similarity are selected as reference cases; for the orchestration decision in the reference case, the forward chain reasoning algorithm is used to evaluate it. The evaluation is based on a predefined rule set and the rule matching degree of each solution is calculated; if the case similarity and rule matching degree are obtained, the Softmax function is used to calculate the weight of each solution. The weight calculation formula is:

[0048] w_i=exp(αs_i+βr_i) / Σexp(αs_j+βr_j), where s_i is the case similarity, r_i is the rule matching degree, and α and β are adjustable parameters; the final orchestration decision is obtained by linearly weighted fusion of the results of case reasoning and rule reasoning, and the orchestration decision includes resource allocation, scheduling strategy and performance optimization measures.

[0049] Exemplarily, historical data of demand timeliness and container orchestration are extracted from the database to build a case library, using a key-value pair storage structure, where the key contains the demand ID and timestamp, and the value contains the demand description, timeliness requirements, resource configuration, orchestration decision and execution result. The K nearest neighbor algorithm is used to calculate the similarity between the current demand and the cases in the case library, and the optimal K value is selected through five-fold cross validation. The top K cases with the highest similarity are selected as references, and the orchestration decisions in these cases are extracted as candidate solutions. According to the predefined rule set, the candidate solutions are evaluated using the forward chain reasoning algorithm. The rules include "if CPU utilization > 80%, increase resource quota by 20%", "if load imbalance > 0.3, trigger load balancing", and "if response time > SLA requirements, the priority is increased by one level", and the rule matching degree of each solution is calculated. Based on case similarity and rule matching, the Softmax function is used to calculate the weight of each solution. The weight calculation formula is w_i=exp(αs_i+βr_i) / Σexp(αs_j+βr_j), where s_i is case similarity, r_i is rule matching, and α and β are adjustable parameters. The results of case reasoning and rule reasoning are linearly weighted fused to obtain the final orchestration decision. At the same time, a decision basis report and adjustment suggestions are generated according to the SLA agreement, including resource allocation, scheduling strategy and performance optimization measures. Data visualization tools are used to generate resource usage trend charts and performance indicator dashboards. Combined with natural language generation technology, quantitative indicators are converted into easy-to-understand text descriptions to form a complete user report. In the container orchestration system, the latest 1,000 historical records are extracted from the database to build a case library. The format of each record is {Requirement ID: 20230615001, Timestamp: 1686816000, Requirement Description: "High Concurrency Web Service", Timeliness Requirements: "Response Time

[0050] <100ms", resource configuration: "8-core CPU, 16GB memory", orchestration decision: "3 container instances, load balancing", execution result: "average response time 85ms"}. For the new demand "e-commerce promotion activities", the K nearest neighbor algorithm determines the optimal K value as 5 through 5-fold cross validation, and calculates the 5 cases with the highest similarity, with similarities of 0.92, 0.89, 0.85, 0.82, and 0.78 respectively. The forward chain reasoning algorithm applies the rule "CPU utilization 87%>80%, increase resource quota by 20%", and obtains a rule matching degree of 0.95. Using Softmax The function calculates the weights, setting α=0.6, β=0.4, and obtains case weights [0.3, 0.25, 0.2, 0.15, 0.1] and rule weight 0.8. After linear weighted fusion, the final decision is "deploy 4 container instances, each with 10 CPU cores and 20GB memory, and enable automatic scaling." The data visualization tool generates a CPU utilization trend chart, showing that the peak interval is between 14:00-16:00. The natural language generation module converts the decision into a report: "It is recommended to increase computing resources by 25% and enable elastic scaling during the period of 14:00-16:00 to cope with the peak of promotions. It is expected that the response time can be controlled within 90ms to meet the SLA requirements.

[0051] It should be noted that the above examples are only some specific embodiments of the present invention. Obviously, the present invention is not limited to the above examples, and there are many variations. All variations that can be directly derived or associated with the content disclosed by a person skilled in the art should be considered as the protection scope of the present invention.

Claims

1. A multi-factor fusion orchestration method for computing networks oriented to SLA requirements, characterized in that: The method comprises: According to the time when the requirements are submitted and the expected completion time, the preset requirements priority evaluation rules are used to analyze the time sensitivity and urgency of each requirement, calculate the priority values ​​of different requirements, and sort the requirements in descending order of priority values ​​to form a requirements time priority queue; Use heuristic methods to model the startup time, resource usage, and dependencies of containers in the computing network. For each container to be orchestrated, evaluate its expected startup time and resource consumption. Consider the dependency constraints between containers and build a directed acyclic graph of container orchestration. Match the demand time priority queue with the directed acyclic graph of container orchestration, and use dynamic programming methods to find the orchestration solution with the best resource utilization efficiency and the most balanced load to form an initial orchestration decision; During the initial orchestration decision execution process, the system operation status and resource usage are continuously monitored, potential performance bottlenecks and stability risks are identified through anomaly detection methods, and the orchestration decisions are dynamically adjusted in combination with reinforcement learning methods to optimize resource allocation and task scheduling to ensure the fulfillment of the service level agreement (SLA). Combine the pre-set computing network element orchestration strategy templates and dynamically select and adjust according to actual conditions, including selecting an orchestration strategy that prioritizes resource utilization during business peaks and selecting an orchestration strategy that prioritizes costs during business troughs; Record and analyze various computing network elements generated in the container orchestration process, including resource usage, startup time, dependencies, and orchestration decisions, and optimize resource prediction and scheduling algorithms based on historical data; Based on the historical data of demand timeliness and container orchestration, case reasoning and rule reasoning methods are used to assist in the formulation and optimization of orchestration decisions, and provide users with the basis and adjustment suggestions for orchestration decisions according to the provisions of the service level agreement (SLA).

2. The method according to claim 1, characterized in that According to the time when the requirements are proposed and the expected completion time, the preset requirements priority evaluation rules are adopted to analyze the time sensitivity and urgency of each requirement, calculate the priority values ​​of different requirements, sort the requirements in descending order of priority values, and form a requirements time priority queue, including: Calculate the time difference between the time when the demand is proposed and the expected completion time, and quantify and score the time difference using preset evaluation rules to obtain the time sensitivity score of the demand; Wherein, the preset evaluation rule includes a mapping relationship between the time difference interval and the corresponding emergency level; Obtain the keywords of the demand content and determine the demand business importance score by matching the preset business importance vocabulary; The preset business importance word library includes the correspondence between keywords and importance levels; Determine whether the time sensitivity score and the business importance score are obtained, and if both scores are obtained, calculate the urgency of the demand using a weighted summation method; The time sensitivity weight is a weighted sum of the first weight value, and the business importance weight is a weighted sum of the second weight value; The urgency of the requirement is converted into a priority value through a linear mapping function; The linear mapping function includes minimum and maximum parameters of all demand urgency levels; According to the priority value, a merge sort algorithm is used to sort all requirements in descending order to obtain a requirement sequence arranged from high to low priority; Extracting demand information one by one from the demand sequence and constructing a maximum heap data structure as a priority queue; It also includes: obtaining the frequency of demand changes and performance assessment indicators, judging the stability of demand through threshold comparison, and if changes are frequent and SLA agreement assessment indicators are not met, re-evaluating the demand and updating the demand queue.

3. The method according to claim 2, characterized in that The acquisition of demand change frequency and contract performance assessment indicators, and judging the stability of demand by threshold comparison, if the changes are frequent and the SLA agreement assessment indicators are not up to standard, the demand is re-evaluated and the demand queue is updated, including: Obtain demand change history records, calculate the number of demand changes per month based on the history records, and obtain a demand change frequency value; The demand change frequency value is compared with a preset change frequency threshold. If the demand change frequency value is higher than the preset change frequency threshold, it is determined that the change is frequent, otherwise it is determined that the change is stable; Extracting performance evaluation index values ​​from the SLA agreement, wherein the performance evaluation index values ​​include response time, resolution time, and customer satisfaction; A multi-factor scoring method is used to quantitatively evaluate the performance assessment index value, and response time weight, resolution time weight, and customer satisfaction weight are set to perform weighted calculations to obtain performance assessment scores; Calculate a comprehensive evaluation index based on the determination result of the demand change frequency and the performance assessment score, wherein the comprehensive evaluation index is equal to the change frequency multiplied by the weight coefficient plus the performance assessment score multiplied by the weight coefficient; Using a rule engine to adjust the demand priority according to different intervals of the comprehensive evaluation index, if the comprehensive evaluation index is less than a first threshold, the demand priority is lowered, and if the comprehensive evaluation index is greater than a second threshold, the demand priority is increased; According to the adjusted priority score, the insertion sort algorithm is used to update the demand priority queue, insert the demand into the appropriate position, and ensure that the demands in the queue are arranged from high to low priority.

4. The method according to claim 1, characterized in that The heuristic method is used to model the startup time, resource usage and dependency of containers in the computing network. For each container to be orchestrated, its expected startup time and resource consumption are evaluated, and the dependency constraints between containers are considered to construct a directed acyclic graph of container orchestration, including: Obtaining the operation data collected every hour in the past preset time period of the container to be orchestrated, wherein the operation data includes the startup time, CPU usage, and memory usage; Based on the operation data, an autoregressive integrated moving average model is used to predict the expected startup time and resource consumption of each container; For the containers to be orchestrated, the dependency relationship between the containers is analyzed by the Apriori algorithm, and a container dependency matrix is ​​established, wherein the matrix element value represents the dependency strength between the containers; If the dependency strength exceeds a preset threshold, it is determined that a dependency constraint exists; According to the expected startup time, resource consumption and dependency constraints, an initial solution for container orchestration is constructed, and a directed graph containing node weights and edge weights is generated. The node weight is the weighted sum of the expected startup time and resource consumption, and the edge weight is the dependency strength. The initial directed graph is processed using a topological sorting algorithm to detect and remove edges that form loops, and the directions of the edges are adjusted to obtain a final container orchestration directed acyclic graph, where the path length in the directed acyclic graph represents the container startup order and time. The container startup order is determined according to the topological order of the directed acyclic graph, and resources are allocated in combination with the node weights to generate a final container orchestration solution.

5. The method according to claim 1, characterized in that: The demand time priority queue is matched with the directed acyclic graph of container orchestration, and a dynamic programming method is used to find an orchestration solution with optimal resource utilization efficiency and the most balanced load to form an initial orchestration decision, including: Obtaining a demand timeliness priority queue and a container orchestration directed acyclic graph, and constructing a matching matrix according to the demand timeliness priority queue and the container orchestration directed acyclic graph, wherein the element value of the matching matrix represents the matching degree between the demand and the container; A non-dominated sorting genetic algorithm II is used to perform multi-objective optimization on the matching matrix, wherein the multi-objective optimization includes maximizing the sum of matching degrees, maximizing resource utilization, and minimizing load variance, and a set of non-dominated solutions is obtained through iterative evolution; Selecting a balanced solution from the non-dominated solution set as the final solution, specifically including: calculating the comprehensive score of each solution and selecting the solution with the highest score as the optimal matching solution; For the optimal matching solution, the resource allocation ratio of each container is calculated. The resource allocation ratio is calculated by: allocating resources such as CPU and memory according to the weighted sum of demand priority and container dependency strength, wherein the weight coefficient is obtained according to historical data statistics; Based on the resource allocation ratio, a minimum connection number algorithm is used to allocate nodes to the container. The minimum connection number algorithm includes: calculating the current number of connections and resource usage rate of each node as a load index, selecting the node with the smallest load index as the container deployment target, and forming an initial orchestration decision; It also includes: If problems such as container startup failure and sudden increase in resource usage occur during the computing network container orchestration process, the container operation status is analyzed in real time through machine learning methods, anomalies are discovered in time, and predefined processing strategies are triggered, including restarting the container and dynamically adjusting resources.

6. The method according to claim 5, characterized in that In the process of container orchestration, if there are problems such as container startup failure and sudden increase in resource usage, the machine learning method is used to analyze the container running status in real time, detect abnormalities in time and trigger predefined processing strategies, including restarting the container and dynamically adjusting resources, including: Obtain container operation logs and resource monitoring data, and extract key features from the data every 5 seconds. The key features include startup time, CPU usage, memory usage, network traffic, and disk I / O. Constructing a multidimensional feature vector according to the key features to obtain a representation of the container operation status; The isolationforest algorithm is used to perform real-time anomaly detection on the multidimensional feature vector. If the anomaly score exceeds a preset threshold or the resource usage rate increases by more than a preset percentage within a predetermined time, it is determined to be an abnormal state; For the abnormal state, extract the abnormal type, abnormal score, resource usage rate and its growth rate, and generate decision tree input features; According to the decision tree input features, a processing solution is selected from a predefined policy library through a decision tree algorithm, wherein the predefined policy library includes restarting a container, increasing a resource quota, and migrating a container; Execute the processing solution to obtain the policy execution effect, wherein the policy execution effect includes the stable operation time of the container and the degree of improvement of resource utilization; The policy execution effect is fed back to the machine learning model to optimize the anomaly detection threshold, policy selection weight and processing solution parameters.

7. The method according to claim 1, characterized in that During the initial orchestration decision execution process, the system operation status and resource usage are continuously monitored, potential performance bottlenecks and stability risks are identified through anomaly detection methods, and the orchestration decision is dynamically adjusted in combination with reinforcement learning methods to optimize resource allocation and task scheduling to ensure the fulfillment of the service level agreement SLA, including: Obtain system operation indicators collected by distributed monitoring tools, wherein the system operation indicators include CPU utilization, memory usage, network throughput, and disk I / O; Adopting an adaptive sampling interval mechanism according to the system operation index, dynamically adjusting the sampling frequency and storing the collected data in a time series database in real time; An isolation forest algorithm is used to perform anomaly detection on the multidimensional time series data stored in the time series database, and the degree of anomaly is determined through a multi-level threshold mechanism; Generate a state representation of the reinforcement learning environment according to the anomaly detection result, wherein the state representation includes an anomaly indicator, an anomaly degree, and a duration; Construct a reinforcement learning environment, and if the state representation satisfies the preset conditions, use the system state as the observation space, and the resource allocation and task scheduling operations as the action space; Training a reinforcement learning agent through a deep Q-network algorithm, which inputs the current system state and anomaly detection results, and outputs resource allocation ratios and task priority adjustment values; The scheduling scheme is updated according to the resource allocation ratio and the task priority adjustment value, and corresponding adjustment operations are performed.

8. The method according to claim 1, characterized in that The aforementioned combination of the preset network element orchestration strategy template and the dynamic selection and adjustment according to the actual situation includes selecting an orchestration strategy that prioritizes resource utilization during business peak periods and selecting an orchestration strategy that prioritizes cost during business trough periods, including: Obtaining business load data, the business load data including request volume, CPU utilization, and memory occupancy; Based on the business load data, an ARIMA model is used to predict the business load trend in the next hour; Determine whether the business load trend exceeds a preset threshold, if the business load trend exceeds the preset threshold, select a resource utilization priority strategy from a preset computing network element orchestration strategy template library, otherwise select a cost priority strategy; Converting the selected strategy into a resource allocation scheme, wherein the resource allocation scheme includes horizontal expansion and vertical expansion of container instances; A deep Q network algorithm is used to dynamically optimize the policy parameters. The deep Q network algorithm uses system load, resource utilization and cost indicators as state space, resource allocation adjustment as action space, and a weighted sum of performance and cost as a reward function to obtain optimized policy parameters. According to the optimized policy parameters, the number of container instances, CPU core allocation, memory limit and network bandwidth configuration are adjusted to achieve adaptive adjustment of resource allocation.

9. The method according to claim 1, characterized in that: The above records and analyzes various computing network elements generated in the container orchestration process, including resource usage, startup time, dependencies, and orchestration decisions, and optimizes resource prediction and scheduling algorithms based on historical data, including: Use distributed log collection tools to obtain indicator data during container orchestration, including CPU usage, memory usage, network traffic, disk I / O, startup time, and dependency graphs; Execute a sliding window algorithm according to the indicator data to obtain statistical features in each time window, wherein the statistical features include an average value, a maximum value, a minimum value, and a standard deviation; Constructing a feature vector through the statistical features, wherein the feature vector includes the statistical features and dependency information; The feature vector is input into a random forest algorithm to obtain a resource prediction model, where the resource prediction model is used to predict resource demand within the next hour; If the resource prediction model outputs a prediction result, a genetic algorithm is used to optimize the container scheduling strategy, wherein the genetic algorithm uses a binary string to represent the allocation relationship between the container and the node; A new scheduling decision is generated according to the genetic algorithm, and the new scheduling decision is applied to the container orchestration system.

10. The method according to claim 1, characterized in that Based on the historical data of demand timeliness and container orchestration, case reasoning and rule reasoning methods are used to assist in the formulation and optimization of orchestration decisions; and according to the agreement of the service level agreement (SLA), the basis and adjustment suggestions for orchestration decisions are provided to users, including: Obtain historical data on demand timeliness and container orchestration from a database, and construct the historical data into a case library. The case library adopts a key-value pair storage structure, where the key includes the demand ID and timestamp, and the value includes the demand description, timeliness requirement, resource allocation, orchestration decision, and execution result; According to the cases in the case library, the K nearest neighbor algorithm is used to calculate the similarity between the current demand and the case, the optimal K value is selected through five-fold cross validation, and the first K cases with the highest similarity are selected as reference cases; For the orchestration decision in the reference case, a forward chaining reasoning algorithm is used to evaluate it, wherein the evaluation is based on a predefined rule set and the rule matching degree of each solution is calculated; If the case similarity and rule matching are obtained, the Softmax function is used to calculate the weight of each solution. The weight calculation formula is w_i=exp(αs_i+βr_i) / Σexp(αs_j+βr_j), where s_i is the case similarity, r_i is the rule matching, and α and β are adjustable parameters; The final orchestration decision is obtained by linearly weighted fusion of the results of case reasoning and rule reasoning, and the orchestration decision includes resource allocation, scheduling strategy and performance optimization measures.

Images