An intelligent resource scheduling method and system based on container behavior fingerprint recognition

Abstract

The present invention discloses an intelligent resource scheduling method and system based on container behavior fingerprint recognition. The method includes: analyzing the resource usage signals during container runtime to construct fingerprints that can reflect the container behavior characteristics; constructing a container symbiotic relationship map based on the real-time runtime data metrics of containers to perform intelligent pairing of different types of containers; establishing a multi-objective programming model to obtain a resource scheduling strategy; judging the resource requirements of containers at different time points based on the fingerprints, classifying high-demand and low-demand containers, and establishing a cross-container resource lending contract through a distributed negotiation and resource credit mechanism to maintain system resource balance; constructing a reinforcement learning optimizer with the container fingerprints, pairing results, and lending contracts as the state space and resource allocation and lending behaviors as the action space, and defining a comprehensive reward function to achieve the autonomous evolution of the scheduling strategy. The present invention realizes predictive resource allocation and improves resource utilization rate.

Description

Technical Field

[0001] The present invention relates to the field of power dispatching automation systems, and particularly to an intelligent resource scheduling method and system based on container behavior fingerprint recognition. Background Art

[0002] Container technology has become a core tool for modern application development and deployment, especially in cloud environments. Compared with traditional virtual machines, containers provide a lighter and more efficient solution, enabling fine-grained resource management and rapid startup, and significantly enhancing application portability and deployment flexibility.

[0003] With the popularization of microservices architecture and the rise of cloud-native applications, container technology has been widely used in key fields such as finance, healthcare, telecommunications, and power. In the power industry's smart grid and dispatching automation systems, container technology has significantly improved the system's flexibility and scalability. Currently, the power system faces challenges such as large-scale access of renewable energy, growth of electric vehicle loads, and rapid development of distributed energy, which pose higher requirements for data processing systems. However, the expansion of application scale and the increase in complexity also pose higher requirements for container system management and intelligent expansion.

[0004] In a multi-container deployment and dynamically changing workload environment, elastic expansion has become a key capability. Existing expansion methods mainly include rule-based expansion and threshold-based automatic expansion. These methods perform well in certain scenarios but often struggle to meet the following requirements:

[0005] 1. Handling unpredictable workload peaks: Traditional methods usually rely on historical data and predefined rules and are difficult to adapt to sudden and irregular load changes. In the power system, this is manifested as difficulty in coping with intermittent fluctuations in distributed photovoltaic power generation, concentrated growth of electric vehicle charging loads, etc., which may lead to system response delays or service quality degradation during peak loads.

[0006] 2. Ensuring real-time data processing: In applications with high real-time requirements such as power dispatching, the system needs to be able to quickly respond to load changes to ensure the timeliness of data processing. The reaction speed and accuracy of existing methods are often insufficient to meet the requirements of key functions such as power grid real-time monitoring, fault diagnosis, and load forecasting.

[0007] 3. Maintaining system reliability under various conditions: Changes in environmental conditions (such as network fluctuations, hardware failures, etc.) may affect the performance of the container system. For critical infrastructure such as the power system, even under extreme weather or partial equipment failures, it must maintain high-reliability operation, and traditional expansion methods are difficult to maintain system stability and reliability under complex conditions.

[0008] 4. Adapt to diverse application requirements: Different types of applications in the power system may have different resource demand patterns, such as SCADA data acquisition, Energy Management System (EMS), Distribution Management System (DMS), and Transmission Management System (TMS). A single scaling strategy is difficult to meet the unique needs of various applications.

[0009] 5. Handle multi-dimensional resource constraints: Container scaling not only involves CPU and memory but also may need to consider resource limitations in multiple dimensions such as network bandwidth and storage I / O. When dealing with the massive, diverse, and heterogeneous data in the power system (such as PMU phasor measurement unit data, SCADA real-time data, electric vehicle charging data, distributed energy generation data, etc.), traditional methods often struggle to handle the multi-dimensional resource balance.

[0010] These challenges lead to over-allocation of resources, performance degradation, and cost increase. In severe cases, it may affect the stable operation of grid dispatching and even trigger cascading failures. Machine learning and artificial intelligence technologies provide new ideas for solving these problems. Multimodal learning is expected to achieve more intelligent and accurate container scaling decisions by integrating multi-source heterogeneous data, which is of great significance for dealing with new business entities such as microgrids and resource aggregators and various types of power data.

[0011] However, applying multimodal learning to container scaling still faces challenges such as how to effectively integrate different power data, how to achieve millisecond-level decision response, and ensure model interpretability.

[0012] Developing an intelligent container scaling system that comprehensively considers multiple factors and adapts to complex environments is of great significance for improving the performance of power regulation systems and resource utilization efficiency. This not only enhances the system operation effect but also contributes to the intelligent and green development of the power grid. Containerized services enable flexible access and real-time data processing for multiple business forms and will become a key technical path for the digital transformation of the power system. Summary of the Invention

[0013] Object of the Invention: To solve the problems in the above-mentioned prior art, the present invention proposes an intelligent resource scheduling method and system based on container behavior fingerprint recognition, which solves the core pain points such as resource contention, scaling delay, and resource fragmentation in a multi-container complex environment.

[0014] To achieve the above object of the invention, the technical solution of the present invention is as follows:

[0015] In the first aspect, an intelligent resource scheduling method based on container behavior fingerprint recognition includes:

[0016] Analyze the resource usage signals during the running process of containers, and combine information entropy calculation to construct fingerprints that can reflect the behavioral characteristics of containers changing over time, so as to quantify the resource consumption patterns of containers in different time periods, possible burst behaviors, and the periodic characteristics of workloads;

[0017] Based on the real-time running data metrics of containers, establish a symbiotic advantage scoring function considering temporal correlation, resource complementarity, and load transfer efficiency. Take containers as nodes, the relationships between containers as edges, and the symbiotic advantage score as the edge weight to construct a container symbiotic relationship graph, and perform intelligent pairing of different types of containers according to the symbiotic relationship graph;

[0018] Continuously obtain the real-time behavior data and historical fingerprints of containers, calculate the deviation between the two, calculate the resource competition relationship based on the real-time behavior fingerprints of containers, and calculate the saturation of various resources. Establish a multi-objective programming model with the goal of minimizing the container behavior deviation and resource competition relationship and maximizing the resource saturation or average resource utilization rate to obtain a resource scheduling strategy;

[0019] Based on the container behavior fingerprints, judge the resource requirements of different containers at different time points, divide the containers with low resource requirements and the containers with high resource requirements, and adopt a distributed negotiation framework. Through a resource credit mechanism, establish a cross-container resource lending contract between the containers with low resource requirements and the containers with high resource requirements to maintain the balance of system resources;

[0020] Take the container behavior fingerprint features, the container pairing information extracted from the dynamic symbiotic relationship graph, and the status of the resource lending contract as the input state space, and the control operations for resource allocation and lending behaviors as the action space to establish a deep reinforcement learning optimizer. The optimizer takes the comprehensive reward function as the optimization goal, and realizes the autonomous evolution of the scheduling strategy by rewarding or punishing the resource scheduling strategy.

[0021] Furthermore, the method for constructing container behavior fingerprints specifically includes:

[0022] Dynamically collect the resource consumption time series data, burst behavior characteristics, resource saturation point parameters, and workload cycle maps of containers through the entire life cycle;

[0023] Use wavelet transform to perform time-frequency decomposition on the original data, identify periodic patterns through fast Fourier transform, establish a time-frequency joint distribution matrix, and calculate quantile features , approximate entropy and mutual information Three types of indicators to form a set of characteristic indicators ;

[0024] Based on the set of characteristic indicators , the clustering algorithm is used to classify containers, and a dynamic evolution model is established. Continuously update the container behavior fingerprint, where represents the behavior fingerprint of the container at the current moment, and the behavior fingerprint uses the statistical features in it as input, comprehensively reflecting the time-domain, frequency-domain, and cross-resource association behavior characteristics of the container, and is used for subsequent anomaly detection and resource scheduling optimization; is the behavior fingerprint of the previous moment; represents the change amount of the newly extracted behavior features in the current cycle; is the smoothing coefficient, .

[0025] Furthermore, the specific calculation method of the symbiotic advantage scoring function is as follows:

[0026] ,

[0027] where represents the symbiotic advantage score between container and container ; represents the temporal correlation between container and , and is calculated by the Pearson correlation coefficient based on the time series behavior fingerprints of containers and ; represents the resource complementarity between container and , and is calculated based on the resource demand feature quantity; represents the load transfer efficiency between container and , and is obtained by calculating the average transmission success rate of unit load flowing from container i to container j based on the container load transfer relationship; α, β, and γ are the weight coefficients of the corresponding three indicators.

[0028] Furthermore, the cross-container resource lending contract adopts a multi-agent bidding game framework to realize cross-container resource flow by establishing a resource lending contract mechanism. According to the resource lending contract, the system dynamically evaluates the resource supply and demand relationship and lending feasibility based on the resource price changes output by the real-time dynamic pricing model. When it is judged that lending is feasible according to the resource surplus, price, or load conditions, the container is allowed to initiate a resource lending request, and the lending operation is preferentially approved according to the container priority; after the resource is lent out, the system continuously monitors the actual usage status and load evolution trend of the resource through the smart contract monitor. When it is detected that the resource price, borrowing time, or load conditions reach the recovery conditions, the system automatically triggers the resource recovery process, requiring the borrowing container to return the resource, or calculating and executing the penalty function according to the default situation.

[0029] Furthermore, the comprehensive reward function used by the reinforcement learning agent is obtained by weighted summation of the service level agreement compliance, resource efficiency metrics, and system stability scores. The service level agreement compliance is determined based on the proportion of service level agreement violation events in the total task volume. The resource efficiency metrics are obtained based on the statistical characteristics of resource utilization. The system stability score is obtained based on the normalized relative deviation of the system's key performance indicators before and after configuration changes.

[0030] Furthermore, during the autonomous evolution of the resource scheduling policy, historical experience is reused during policy migration by identifying the similarity of container behavior fingerprints; and the performance of the scheduling policy is regularly evaluated, and the policy parameters are continuously adjusted and optimized according to the actual operation results.

[0031] In a second aspect, an intelligent resource scheduling system based on container behavior fingerprint recognition includes:

[0032] A container behavior fingerprint construction module, which is used to analyze the resource usage signals during the operation of the container, and combine information entropy calculation to construct a fingerprint that can reflect the behavioral characteristics of the container changing over time, so as to quantify the resource consumption patterns, possible burst behaviors, and periodic characteristics of the workload of the container in different time periods;

[0033] A symbiotic relationship graph construction module, which is used to establish a symbiotic advantage scoring function considering temporal correlation, resource complementarity, and load transfer efficiency based on the real-time operation data metrics of the container. Taking the container as a node, the relationship between containers as an edge, and the symbiotic advantage score as the edge weight, construct a container symbiotic relationship graph, and perform intelligent pairing of different types of containers according to the symbiotic relationship graph;

[0034] A resource allocation module, which is used to continuously obtain the real-time behavior data and historical fingerprints of the container, calculate the deviation between the two, calculate the resource competition relationship according to the real-time behavior fingerprint of the container, and calculate the saturation of various resources, and establish a multi-objective programming model with the goal of minimizing the container behavior deviation and resource competition relationship and maximizing the resource saturation or average resource utilization rate, and obtain a resource scheduling policy;

[0035] A cross-container resource lending module, which is used to judge the resource demand situations of different containers at different time points based on the container behavior fingerprint, divide the containers with low resource demand and the containers with high resource demand, adopt a distributed negotiation framework, and establish a cross-container resource lending contract between the containers with low resource demand and the containers with high resource demand through a resource credit mechanism to maintain system resource balance;

[0036] The closed-loop reinforcement learning optimization module uses the container behavior fingerprint features, the container pairing information extracted from the dynamic symbiotic relationship graph, and the resource lending contract status as the input state space, and uses the regulation operations for resource allocation and lending behaviors as the action space to establish a deep reinforcement learning optimizer. The optimizer takes the comprehensive reward function as the optimization goal and realizes the autonomous evolution of the scheduling strategy by rewarding or punishing the resource scheduling strategy.

[0037] In a third aspect, the present invention also provides an electronic device, including: one or more processors; a memory; and one or more programs, wherein the one or more programs are stored in the memory and are configured to be executed by the one or more processors. When the program is executed by the processor, the steps of the intelligent resource scheduling method based on container behavior fingerprint recognition described above are implemented.

[0038] In a fourth aspect, the present invention also provides a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, the steps of the intelligent resource scheduling method based on container behavior fingerprint recognition described above are implemented.

[0039] In a fifth aspect, the present invention also provides an intelligent container resource scheduling device, including:

[0040] A heterogeneous computing unit cluster configured to process container behavior data streams in parallel;

[0041] A distributed storage array configured to store a dynamically updated behavior fingerprint library and a symbiotic relationship graph;

[0042] A programmable logic device configured to implement an adaptive scheduling strategy according to the method described in the first aspect of the present invention.

[0043] Beneficial effects: (1) Through container behavior fingerprint recognition and symbiotic relationship optimization, the present invention changes the traditional passive threshold-triggered expansion mode, realizes predictive resource allocation, can improve resource utilization rate, and at the same time reduces performance fluctuations caused by resource competition. (2) By establishing a cross-container resource lending contract to implement a predictive resource arbitration mechanism, the present invention can reduce the over-configuration cost and avoid resource waste. (3) Through the autonomous learning ability of the intelligent agent, it can adapt to various workload characteristics and continuously optimize the decision-making strategy without manual intervention. This resource scheduling method centered on container behavior provides an excellent solution for large-scale microservice architectures and establishes a new paradigm for container orchestration technology. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] Figure 1 is the overall architecture diagram of the intelligent resource scheduling system based on container behavior fingerprint recognition proposed by the present invention;

[0045] Figure 2 It is an example of a container symbiotic relationship map, showing the complementary, competitive, and transfer relationships between containers;

[0046] Figure 3 It is a flowchart of closed-loop reinforcement learning and continuous optimization, including evaluation, policy generation, pattern recognition, effect verification, and feedback mechanisms. Specific implementation manners

[0047] To make the objectives, technical solutions, and advantages of the present invention clearer, the technical solutions of the present invention will be clearly and completely described below in conjunction with specific embodiments and corresponding drawings. At the same time, it should be understood that these embodiments are only used to illustrate the present invention and not to limit the scope of the present invention. After reading the present invention, those skilled in the art's various equivalent modifications of the present invention all fall within the scope defined by the appended claims of this application.

[0048] Referring to Figure 1 , the intelligent resource scheduling system based on container behavior fingerprints (Container Behavior Profile Scheduling, CBPS) proposed by the present invention adopts a distributed architecture and realizes intelligent resource scheduling through three core mechanisms: behavior fingerprint construction, symbiotic relationship recognition, and predictive resource arbitration.

[0049] S1: Container behavior fingerprint construction

[0050] The CBPS system of the present invention realizes digital modeling of container behavior characteristics through a behavior fingerprint construction engine. This engine continuously captures multi-dimensional runtime characteristics during the entire life cycle of the container, including resource consumption time series data, burst behavior characteristics, resource saturation point parameters, and workload cycle maps. The system records the fluctuation curves of CPU / memory / IO at a sampling frequency of 1 second for subsequent behavior modeling and anomaly detection.

[0051] When the utilization rate of a resource (referring to key resource indicators such as CPU utilization rate, memory utilization rate, and disk I / O usage rate) reaches a preset threshold (usually 85%), the system records the corresponding concurrent pressure characteristics as an indicator of the resource saturation point. The concurrent pressure characteristics refer to a series of performances and characteristics related to system concurrent operations when the system resource utilization rate reaches the preset threshold, such as request response time, throughput change, system load balancing situation, request failure rate, resource waiting queue length, and other performance indicators.

[0052] For the identification of periodic workloads, the system uses Fourier transform technology to extract frequency components. When the energy of a certain frequency exceeds 10% of the maximum frequency energy, the system determines it as a valid period.

[0053] These obtained original monitoring data, namely time-series resource utilization information, are then processed in three levels:

[0054] First, wavelet transform is used for time-frequency decomposition to capture local features through discrete wavelet transform:

[0055] ,

[0056] where represents the coefficient of the signal after continuous wavelet transform, reflecting the local features of the signal at different scales and positions ; is the input signal to be analyzed, referring to the CPU or memory utilization data collected by the system; [[ID=2!]]is the scale factor , used to adjust the stretching degree of the wavelet function. A smaller value is used to capture the high-frequency details in the signal, and a larger value is used to analyze the low-frequency trend; is the position factor, used to determine the position or offset of the wavelet function on the time axis; is the conjugate complex form of the wavelet mother function. Commonly used mother wavelets include Morlet wavelet or Mexican hat wavelet, etc., which have good time-frequency local characteristics and are suitable for dynamic behavior recognition.

[0057] Then, the periodic pattern is identified through the fast Fourier transform, and finally, a time-frequency joint distribution matrix covering multi-dimensional resource metrics such as CPU utilization, memory utilization, and I / O usage is constructed. This matrix comprehensively captures the dynamic characteristics of container behavior in the time domain and frequency domain, supporting the comprehensive characterization of the temporal correlation, resource complementarity, and load transferability during container operation. The matrix adopts a hierarchical structure, with time segments and frequency bandwidths as dimensions, combined with resource types, to achieve multi-scale and multi-granularity modeling of resource dynamic behavior.

[0058] During the feature quantization process, the system calculates three types of metrics, namely quantile features , approximate entropy , and mutual information , to form a feature index set . Based on , the system uses a clustering algorithm for container classification and establishes a dynamic evolution model to continuously update the container behavior fingerprint. Among them, represents the behavior fingerprint of the container at the current moment. The behavior fingerprint takes the statistical features in as input, comprehensively reflecting the temporal, frequency domain, and cross-resource association behavior characteristics of the container, and is used for subsequent anomaly detection and resource scheduling optimization; is the behavior fingerprint of the previous moment; Represents the change in newly extracted behavioral features in the current period; is the smoothing coefficient , which is used to regulate the degree of fusion between historical behavioral features and current change features. When has a larger value, the model relies more on historical information and is suitable for stable environments; when has a smaller value, the model is more sensitive to the latest behavioral changes and is suitable for dynamic change scenarios.

[0059] This dynamic evolution model is based on the feature set with a 5-minute update period, fuses historical and current behavioral data through weighted smoothing, and at the same time uses a mutation detection algorithm to identify abnormal changes. The anomaly detection mechanism is achieved by calculating the deviation degree between real-time behavior and historical fingerprints:

[0060] ,

[0061] where : The weight coefficient of the th feature, reflecting the importance of this feature in the deviation degree calculation.

[0062] : The th behavioral feature value collected at the current moment.

[0063] : The historical reference value or expected value, representing the benchmark of normal behavior.

[0064] When the deviation degree exceeds a predetermined threshold (3 times the standard deviation), the system triggers an alarm and generates a uniquely identified abnormal event record, providing an important basis for subsequent scheduling decisions.

[0065] S2: Construction of the container symbiotic relationship graph

[0066] This system constructs a multi-dimensional container symbiotic relationship graph to achieve resource collaborative optimization. This graph is based on real-time monitoring data streams and historical operation logs, and establishes a container association analysis model that includes three dimensions: temporal correlation, resource complementarity, and load transferability. The system extracts dynamic interaction features between containers through a time window sliding mechanism, calculates the correlation of cross-container resource fluctuations using the covariance matrix, and then identifies potential collaborative or competitive relationships, which are called container symbiotic relationships in this invention.

[0067] This invention uses a symbiotic relationship quantification module to construct a multi-dimensional collaborative scoring model. This model comprehensively considers temporal correlation, resource complementarity, and load transfer efficiency to form a symbiotic advantage scoring function:

[0068] ,

[0069] where Represents the container The symbiotic advantage score between the container and the container The higher the value, the stronger the coupling degree between the two in terms of resource synergy. The definitions of each index are as follows:

[0070] Represents the container And The temporal correlation of, based on the container And The time series behavior fingerprints of are calculated. Specifically, the Pearson Correlation Coefficient is used to measure the consistency of the running trends of the two. The calculation formula is:

[0071] ,

[0072] Where Is the covariance, And Are the respective standard deviations.

[0073] Represents the container And The resource complementarity degree of, which is the complementary score calculated based on the resource demand characteristic quantities (such as CPU and memory demand rates), and is defined as:

[0074] ,,

[0075] Where Represents the overlapping part of the resource demands, And Represent the norms of their respective demand vectors respectively. The higher the complementarity, the larger the score.

[0076] Represents the container And The load transfer efficiency of, based on the container load transfer relationship, calculates the average transmission success rate of unit load flowing from container i to container j, and is defined as:

[0077] ,

[0078] Where Is the number of successfully forwarded loads, Is the total number of requests.

[0079] α, β, and γ are the weight coefficients corresponding to the three indicators, and are dynamically optimized through algorithms such as gradient descent to ensure that the scoring results can accurately reflect the actual synergy effect between containers.

[0080] The graph construction uses a graph data structure , where vertices represent container instances, and edges represent the relationships between containers. The weight matrix W stores the calculated symbiotic advantage score results for each pair of containers . As Figure 2 shown, the dynamic relationship graph generated by the system achieves an intuitive visualization effect. The vertex size is positively correlated with the total resource consumption, the color of the edge distinguishes complementary or competitive relationships, and the transparency reflects the relationship strength.

[0081] First, the system identifies container types based on container behavior fingerprint features. By analyzing key metrics such as CPU utilization and disk I / O read / write rates, containers are classified as compute-intensive or I / O-intensive. After type identification, combined with the weight matrix (i.e., the symbiotic advantage score matrix ), according to the high and low of the complementarity score , heterogeneous (i.e., different types) container pairs with good load transferability are preferentially screened for intelligent pairing. By preferentially scheduling complementary containers to adjacent nodes or shared resource pools, the system improves the overall resource utilization and reduces the performance bottleneck caused by homogeneous competition.

[0082] In the deployment optimization phase, the modularity algorithm is applied to group containers, and the optimal deployment plan is generated by identifying container clusters with high cohesion and low coupling. The system has established a graph version control mechanism. When container behavior fingerprints are updated or resource policies are adjusted, incremental graph reconstruction is triggered to ensure that the graph always reflects the latest system state.

[0083] S3: Predictive Resource Demand Analysis

[0084] The CBPS system realizes intelligent pre-allocation of resources through a predictive resource analysis engine. The engine has established a three-level time-scale prediction system, including short-term, medium-term, and long-term prediction modules.

[0085] 1. Short-term prediction: A time series model is used to process minute-level data. The prediction equation is suitable for capturing instantaneous fluctuations in resource usage. Among them, represents the change in resource usage at time , represents the change at the previous time, and represent the current and lagged random disturbance terms respectively, is the influence coefficient of the historical change trend, is the regression coefficient of the lagged disturbance. This model is suitable for capturing the instantaneous fluctuation characteristics of resource usage and improving the short-term prediction accuracy at the minute level;

[0086] ​​​​2. Medium-term prediction: Using a deep learning network to model the hourly change trend;

[0087] 3. Long-term prediction: Based on the Fourier series expansion of the periodic component, identifying and predicting the periodic patterns at the daily / weekly / monthly levels.

[0088] To enhance the scheduling guidance value of the prediction results, the system performs auxiliary analyses such as resource competition detection, load propagation analysis, and node influence ranking in combination with the container coexistence relationship graph.

[0089] This invention takes predictive resource analysis as the basis for formulating forward-looking scheduling strategies and graph analysis as a means of real-time deconstruction of the current state of the system. The two are combined to construct a "prediction-identification-optimization" resource management closed-loop. Among them:

[0090] The prediction module provides a quantitative expectation of future resource usage;

[0091] The graph module identifies the conflict risk paths and key nodes in the current resource distribution structure.

[0092] For example, the system identifies potential resource conflict risks by calculating the competition index between containers:

[0093] ,

[0094] where represents the resource competition index between container i and container j, and the larger the value, the higher the intensity of their competition for the same type of resources; According to the calculation of the overlap degree of resource requests of container and container on the same type of resources (such as CPU, memory, I / O), the more the overlap, the stronger the competition. The calculation formula can be expressed as:

[0095] ,

[0096] where represents the request volume of container on resource , represents the request volume of container on resource . The global resource complementarity degree is defined as the normalized complementarity degree in the overall resource dimension in this module, and can be defined as 1 minus the resource overlap rate, that is , and the larger the value, the stronger the resource complementarity. After the two are independently calculated based on the container resource usage fingerprint characteristics, they are applied to construct the resource competition detection index ; is a very small constant used to prevent the denominator from being zero to ensure the stability of the model.

[0097] Load propagation analysis is based on a propagation impact diagram derived from the symbiosis graph. This diagram uses containers as nodes, and edge weights no longer use a symbiosis advantage score. Instead, they calculate the efficiency or impact of load transfer from the source container to the target container per unit time based on the actual historical resource request chain, network traffic, and log metrics between containers. The propagation impact diagram is used to assess potential load diffusion paths and their dynamic risks. While constructed in the same manner as the static symbiosis graph, the edge weights focus more on measuring runtime propagation capabilities.

[0098] Node influence ranking is based on a symbiotic relationship graph. The system analyzes the graph's structural characteristics and uses graph algorithms such as degree centrality, betweenness centrality, and load propagation capacity to comprehensively assess the importance of container nodes and identify key nodes. The ranking results are used in the resource anomaly prediction and scheduling decision-making stages. When triggering resource scheduling, load balancing, or disaster recovery optimization, the system prioritizes resource allocation for key nodes based on the priority results, thereby improving overall system stability and risk mitigation.

[0099] The resource bottleneck warning mechanism calculates the saturation of various resources in real time:

[0100] ,

[0101] Among them, represents the normalized saturation of resource r; is the resource usage rate at the current moment; The average usage rate of the resource in the historical time window; The standard deviation of resource utilization. This metric measures whether current resource usage is significantly higher than the historical average. A larger value indicates that the resource is closer to a bottleneck.

[0102] The system establishes a three-level early warning mechanism. In different ranges, the system triggers yellow, orange or red warnings respectively, and reserves appropriate resource buffers in advance based on the prediction results. The abnormal response link defines a weighted abnormality detection function and implements a graded response strategy to ensure that the system can make corresponding adjustments for abnormalities of different degrees. Based on the aforementioned prediction results and graph analysis, the system constructs a multi-objective optimization model in the decision optimization stage, while comprehensively considering the two core goals of minimizing risk and maximizing resource utilization. In the decision optimization stage, the system constructs a multi-objective optimization model, while comprehensively considering the two core goals of minimizing risk and maximizing resource utilization. Under the constraints, the optimal resource allocation plan is obtained by solving the multi-objective optimization problem. Among them, the risk minimization goal can be achieved through the container behavior deviation index (such as deviation ) and resource competition index ( ), etc. for quantification; while the goal of maximizing resource utilization is reflected by relying on resource saturation indicators ( ) or average resource utilization, etc.

[0103] This optimization process can be formally expressed as the following multi-objective programming model:

[0104] ,

[0105] Among them, and are the weight coefficients of the risk objectives, which are flexibly set according to the scenario.

[0106] By performing weighted solution and dynamic balance between the performance and stability objectives, the system can generate the optimal resource allocation strategy for the container cluster, achieving a double improvement in the overall system efficiency and security.

[0107] S4: Predictive Resource Arbitration Mechanism

[0108] The predictive resource arbitration mechanism of the present invention establishes a multi-agent resource negotiation framework based on market game theory. In this framework, containers and nodes are regarded as the buyer and seller in resource trading respectively, because nodes provide resources (such as CPU, memory, storage, network resources), and containers consume resources. A resource trading triple including trading subjects (buyer B, seller S) and time window (T) is defined .

[0109] The system constructs a dynamic pricing model:

[0110] ,

[0111] Among them, represents the dynamic pricing of resource at time ; is the initial benchmark price of the resource; is the actual utilization rate of resource at the current moment; is the critical utilization rate threshold of resource , which is used to measure the degree of system congestion or load approaching the bottleneck; is the adjustment coefficient, which is used to control the sensitivity of price fluctuations with the utilization rate.

[0112] Adjust the resource price according to the real-time resource utilization rate, and achieve supply-demand matching through the double auction mechanism, maximizing the trading benefits under the premise of meeting the total resource constraint.

[0113] Before the system performs resource dynamic pricing and lending transactions, it identifies the current resource demand status based on the container behavior fingerprint analysis model, distinguishing containers with high resource demand from those with low resource demand. Containers with high resource demand participate in the auction as buyers, while containers with low demand provide surplus resources as sellers. The system realizes the matching of resource supply and demand between buyers and sellers through a double auction mechanism, improving the overall resource utilization rate.

[0114] In the present invention, the lending protocol execution unit designs the resource lending process based on the resource lending contract (Contract = (LoanAmount, LoanDuration, PriorityWeight, PenaltyFunction)), clarifying the resource quantity, borrowing duration, priority weight, and default penalty clause of the lending transaction.

[0115] During the lending decision-making process, the system dynamically evaluates the resource supply and demand relationship and lending feasibility based on the resource price changes output by the real-time dynamic pricing model. When the system detects that a node or container has surplus resources, the current resource unit price is lower than the set threshold, and the load change risk is judged to be low within a certain future time window according to the prediction module, it allows the container to initiate a resource lending request and preferentially approves the lending operation according to the container priority.

[0116] After the resources are lent out, the system continuously monitors the actual usage status of the resources and the load evolution trend through the smart contract monitor. When it detects that the resource unit price rises above the warning threshold, the borrowing time approaches or exceeds the agreement, or it predicts an increase in the overall system load risk, the system automatically triggers the resource recovery process, requiring the borrowing container to return the resources, or calculates and executes the penalty function (PenaltyFunction) according to the default situation to ensure the execution consistency of the lending contract and the overall stability of the system resource allocation.

[0117] Through the above dynamic lending and recovery mechanism, the present invention realizes the maximum improvement of the system resource utilization rate, container operation stability, and service quality while ensuring the elastic adjustment of resource supply and demand.

[0118] In terms of multi-objective optimization, the system formally defines the resource utility function:

[0119] ,

[0120] where represents the resource utility score between container and resource , which is used to comprehensively evaluate the cost performance of the scheduling configuration; Indicates the degree of improvement in task performance confidence, reflecting the positive promotion effect of resource allocation on container performance; Is the expected delay change caused by resource scheduling, measuring the potential impact on response latency after resource allocation; Is the expected cost of resource usage; , , Are weight coefficients, respectively controlling the importance of performance, latency, and cost factors in the total utility function.

[0121] Comprehensively consider performance improvement, latency change, and cost factors, and continuously update the weight coefficients through backpropagation learning. The optimization algorithm adopts an improved genetic algorithm framework, designs a dedicated coding method, fitness function, and mutation strategy, and simultaneously establishes a change cost model to evaluate the system impact brought by configuration adjustment.

[0122] To ensure the security of resource reallocation, the system designs a phased migration protocol, including three stages: preheating, double writing, and switching. In the preheating stage, a shadow resource pool is established. In the double writing stage, resource status synchronization is guaranteed. In the switching stage, a blue-green deployment mode is adopted to achieve a smooth transition. At the same time, the system implements a rollback decision model and a distributed transaction coordinator. The former triggers a rollback operation when the service quality deteriorates, and the latter ensures the consistency of cross-node resource operations.

[0123] Through a multi-layer architecture of market bidding, contract constraints, optimization solving, and secure execution, this system realizes the Nash equilibrium state of resource allocation, ensuring the optimal global resource configuration under the premise of meeting the service level agreement.

[0124] S5: Closed-loop Reinforcement Learning and Continuous Optimization

[0125] The closed-loop reinforcement learning module of the present invention establishes a dynamic tuning system based on deep reinforcement learning. Referring to Figure 3 , this system first defines key performance indicators through a multi-dimensional evaluation unit, including SLA compliance, resource efficiency, and system stability, for comprehensively evaluating the current state of the system and the pros and cons of the scheduling strategy.

[0126] Among them, the SLA compliance score ( ) reflects the degree to which the system meets the commitments of the service level agreement (SLA), and is defined as:

[0127] ,

[0128] Among them Is the proportion of SLA violation events detected per unit time in the total task volume, which is collected in real time through an intelligent contract monitor. The resource efficiency indicator ( ) Reflects the utilization of computing resources by the system, defined as:

[0129] ,

[0130] where is the average value of resource utilization, is the standard deviation of resource utilization, is the load fluctuation adjustment coefficient, and the statistic is derived from the time-series sampling and fluctuation analysis results of real-time resource usage characteristics (such as CPU and memory occupancy rates) in the container behavior fingerprint modeling module. This metric encourages the system to reduce the instability caused by load fluctuations while maintaining high utilization.

[0131] System stability score ( ) measures the impact of resource configuration changes on system stability and is defined as:

[0132] ,

[0133] where represents the normalized relative deviation of the system's key performance indicators before and after configuration changes, and the calculation formula is:

[0134] ,

[0135] where includes key performance indicators such as average response latency, throughput, and task completion rate, and represent the average values of performance samples within a certain time window before and after configuration adjustment respectively, and ε is a very small positive number to prevent division by zero anomalies.

[0136] Data for each performance indicator is continuously collected through the real-time monitoring module and combined with the historical performance baseline recorded in the container behavior fingerprint ( ) as a reference standard. After each adjustment of the scheduling policy, the system triggers a performance backtracking detection process to calculate , and then derives , which is used as a component of the comprehensive reward function ([[ID=�0]] ) in the closed-loop reinforcement learning module to provide a quantitative feedback basis for subsequent policy optimization. To comprehensively evaluate the effect of scheduling actions, the system defines a comprehensive reward function R(t) as the optimization objective of deep reinforcement learning:

[0137] ,

[0138] where are the weight coefficients of SLA compliance, resource efficiency, and system stability respectively, which can be adjusted according to different scenario requirements or optimized adaptively through training. The system uses the comprehensive reward function With the goal of dynamically adjusting the scheduling policy parameters, continuously improving the overall performance of the system and the efficiency of resource management.

[0139] The policy optimization adopts deep reinforcement learning technology. The system generates resource allocation decisions by constructing a policy network and evaluates the long-term value in different system states using a value network. The system uses the container behavior fingerprint features, the container pairing information extracted from the dynamic symbiotic relationship graph, and the resource lending contract status as the input state space, specifically including the resource utilization rate, burstiness characteristics, and complementarity scores of each container , the current dynamic pricing of resources and the existing lending protocol parameters. On this basis, the action space is defined as the regulatory operations for resource allocation and lending behaviors, including the application or return of resource lending between containers, the increase or decrease of resource quotas, and the dynamic adjustment of lending contract parameters. To balance training efficiency and decision-making accuracy, the system appropriately discretizes the action space to reduce the computational complexity brought by the continuous action space.

[0140] During the policy training process, the reinforcement learning optimizer uses the comprehensive reward function as the optimization goal, where comprehensively considers the SLA compliance score, resource efficiency metrics, and system stability score. The system gives rewards or punishments according to the execution effects of different policy actions, thereby guiding the resource scheduling policy to continuously adapt and evolve, and continuously optimizing the overall resource utilization rate and system performance stability.

[0141] By identifying the similarity of container behavior patterns:

[0142] ,

[0143] where, and are the behavior feature vectors of two containers, represents the cosine similarity between them, and the closer the value is to 1, the higher the similarity. This metric can be used to reuse historical experience during the policy migration process to accelerate model convergence. The system can reuse historical experience to accelerate the learning process while maintaining the lightweight and high efficiency of the model.

[0144] The system regularly evaluates the performance of the scheduling policy and continuously adjusts and optimizes the policy parameters according to the actual operation results. When it detects that the continuous performance deviation is too large, satisfying:

[0145] ,

[0146] where, is the performance deviation ratio, is the current actual performance metric (such as response time, resource utilization rate), is the system prediction or target value, is the tolerance threshold.

[0147] The system will automatically trigger the retraining process to ensure that the scheduling policy always adapts to the current workload characteristics. Through this closed-loop feedback mechanism, the system continuously improves itself during long-term operation, gradually improving resource utilization efficiency and reducing the service quality violation rate.

[0148] The present invention also provides an intelligent resource scheduling system based on container behavior fingerprint recognition, including:

[0149] A container behavior fingerprint construction module, which is used to analyze the resource usage signals during the operation of the container, and combine information entropy calculation to construct a fingerprint that can reflect the behavioral characteristics of the container changing over time, so as to quantify the resource consumption patterns of the container in different time periods, possible burst behaviors, and the periodic characteristics of the workload;

[0150] A symbiotic relationship graph construction module, which is used to establish a symbiotic advantage scoring function considering temporal correlation, resource complementarity, and load transfer efficiency based on the real-time operation data metrics of the container. Taking the container as a node, the relationship between containers as an edge, and the symbiotic advantage score as the edge weight, construct a container symbiotic relationship graph, and perform intelligent pairing of different types of containers according to the symbiotic relationship graph;

[0151] A resource allocation module, which is used to continuously obtain the real-time behavior data and historical fingerprints of the container, calculate the deviation between the two, calculate the resource competition relationship according to the real-time behavior fingerprint of the container, and calculate the saturation of various resources, and establish a multi-objective programming model with the goal of minimizing the container behavior deviation and resource competition relationship and maximizing the resource saturation or average resource utilization rate, and obtain a resource scheduling strategy;

[0152] A cross-container resource lending module, which is used to judge the resource demand situations of different containers at different time points according to the container behavior fingerprint, divide the containers with low resource demand and the containers with high resource demand, and adopt a distributed negotiation framework to establish a cross-container resource lending contract between the containers with low resource demand and the containers with high resource demand through a resource credit mechanism to maintain the balance of system resources;

[0153] A closed-loop reinforcement learning optimization module, which is used to use the container behavior fingerprint features, the container pairing information extracted from the dynamic symbiotic relationship graph, and the status of the resource lending contract as the input state space, and the regulation operations for resource allocation and lending behaviors as the action space, establish a deep reinforcement learning optimizer, and the optimizer takes the comprehensive reward function as the optimization goal, and realizes the autonomous evolution of the scheduling strategy by rewarding or punishing the resource scheduling strategy.

[0154] The present invention also provides an electronic device, comprising: one or more processors; a memory; and one or more programs, wherein the one or more programs are stored in the memory and configured to be executed by the one or more processors, and when the program is executed by the processor, the steps of the intelligent resource scheduling method based on container behavior fingerprint recognition as described above are implemented.

[0155] The present invention also provides a computer-readable storage medium, on which a computer program is stored, and when the computer program is executed by a processor, the steps of the intelligent resource scheduling method based on container behavior fingerprint recognition as described above are implemented.

[0156] The present invention also provides an intelligent container resource scheduling device, comprising:

[0157] a heterogeneous computing unit cluster configured to process container behavior data streams in parallel;

[0158] a distributed storage array configured to store a dynamically updated behavior fingerprint library and a symbiotic relationship map;

[0159] a programmable logic device configured to implement an adaptive scheduling strategy by the method as described above.

[0160] Through container behavior fingerprint recognition and symbiotic relationship optimization, the present invention changes the traditional passive threshold-triggered expansion mode. The predictive resource arbitration mechanism effectively improves the resource utilization efficiency, significantly reduces the performance fluctuations caused by resource competition and the resource waste caused by over-allocation, and avoids resource waste. Through the autonomous learning ability of the intelligent agent, it can adapt to various workload characteristics and continuously optimize the decision-making strategy without manual intervention. This container behavior-centered resource scheduling method provides an excellent solution for large-scale microservice architectures and establishes a new paradigm for container orchestration technology.

Claims

1. An intelligent resource scheduling method based on container behavior fingerprint recognition, characterized in that, Including: Analyze the resource usage signals during the running process of the container, and combine with information entropy calculation to construct a fingerprint that can reflect the behavioral characteristics of the container changing over time, so as to quantify the resource consumption patterns of the container in different time periods, possible burst behaviors, and the periodic characteristics of the workload; Based on the real-time running data metrics of the container, establish a symbiotic advantage scoring function considering temporal correlation, resource complementarity, and load transfer efficiency. Take the container as a node, the relationship between containers as an edge, and the symbiotic advantage score as the edge weight to construct a container symbiotic relationship graph, and perform intelligent pairing of different types of containers according to the symbiotic relationship graph; Continuously obtain the real-time behavior data and historical fingerprints of the container, calculate the deviation between the two, calculate the resource competition relationship according to the real-time behavior fingerprint of the container, and calculate the saturation of various resources. Establish a multi-objective programming model with the goal of minimizing the container behavior deviation and resource competition relationship, and maximizing the resource saturation or average resource utilization rate to obtain a resource scheduling strategy; Based on the container behavior fingerprint, judge the resource demand situation of different containers at different time points, divide the containers with low resource demand and the containers with high resource demand, and adopt a distributed negotiation framework. Through the resource credit mechanism, establish a cross-container resource lending contract between the containers with low resource demand and the containers with high resource demand to maintain the system resource balance; Take the container behavior fingerprint characteristics, the container pairing information extracted from the dynamic symbiotic relationship graph, and the status of the resource lending contract as the input state space, and take the control operations for resource allocation and lending behaviors as the action space to establish a deep reinforcement learning optimizer. The optimizer takes the comprehensive reward function as the optimization goal, and realizes the autonomous evolution of the scheduling strategy by rewarding or punishing the resource scheduling strategy; 2. The method according to claim 1, characterized in that, The specific method for constructing the container behavior fingerprint includes: Dynamically collect the resource consumption time series data, burst behavior characteristics, resource saturation point parameters, and workload cycle graph of the container through the full life cycle; Perform time-frequency decomposition on the original data using wavelet transform, identify periodic patterns through fast Fourier transform, establish a time-frequency joint distribution matrix, and form a set of characteristic indicators by calculating three types of indicators: quantile feature , approximate entropy , and mutual information ; ; Based on the feature index set , use the clustering algorithm to classify containers and establish a dynamic evolution model Continuously update the container behavior fingerprint, where represents the behavior fingerprint of the container at the current moment, and the behavior fingerprint takes the statistical features in as the input, comprehensively reflects the time-domain, frequency-domain and cross-resource association behavior characteristics of the container, and is used for subsequent anomaly detection and resource scheduling optimization; is the behavior fingerprint of the previous moment; represents the change amount of the newly extracted behavior features in the current cycle; is the smoothing coefficient, .

3. The method according to claim 1, wherein The specific calculation method of the symbiotic advantage scoring function is as follows: ， Among them, represents the symbiotic advantage score between container and container ; represents the temporal correlation between container and , which is calculated by the Pearson correlation coefficient based on the time series behavior fingerprints of container and ; represents the resource complementarity between container and , which is calculated based on the resource demand characteristic quantities; represents the load transfer efficiency between container and , which is obtained by calculating the average transmission success rate of unit load flowing from container i to container j based on the container load transfer relationship; α, β, and γ are the weight coefficients of the corresponding three indicators.

4. The method according to claim 1, characterized in that, The cross-container resource lending contract adopts a multi-agent bidding game framework to realize cross-container resource flow by establishing a resource lending contract mechanism. According to the resource lending contract, the system dynamically evaluates the resource supply-demand relationship and lending feasibility according to the resource price changes output by the real-time dynamic pricing model. When it is judged that the lending is feasible according to the resource balance, price, or load conditions, the container is allowed to initiate a resource lending request, and the lending operation is preferentially approved according to the container priority; after the resource is lent out, the system continuously monitors the actual usage status and load evolution trend of the resource through the intelligent contract monitor. When it is detected that the resource price, borrowing time, or load conditions reach the recovery conditions, the system automatically triggers the resource recovery process, requiring the borrowing container to return the resource, or calculating and executing the penalty function according to the default situation.

5. The method according to claim 1, wherein The comprehensive reward function used by the reinforcement learning agent is obtained by the weighted sum of the service level agreement compliance, resource efficiency metrics, and system stability scores. Among them, the service level agreement compliance is determined according to the proportion of service level agreement violation events in the total task volume, the resource efficiency metrics are obtained based on the statistical characteristics of resource utilization, and the system stability score is obtained according to the normalized relative deviation of the system's key performance indicators before and after configuration changes.

6. The method according to claim 1, characterized in that, During the autonomous evolution of the resource scheduling strategy, the historical experience is reused during the strategy migration by identifying the similarity of container behavior fingerprints; and the performance of the scheduling strategy is regularly evaluated, and the strategy parameters are continuously adjusted and optimized according to the actual operation results.

7. An intelligent resource scheduling system based on container behavior fingerprint recognition, characterized in that, Including: A container behavior fingerprint construction module, which is used to analyze the resource usage signals during the container operation process, and combine the information entropy calculation to construct a fingerprint that can reflect the behavioral characteristics of the container changing over time, so as to quantify the resource consumption patterns of the container in different time periods, possible burst behaviors, and the periodic characteristics of the workload; A symbiotic relationship graph construction module, which is used to establish a symbiotic advantage scoring function considering temporal correlation, resource complementarity, and load transfer efficiency based on the container real-time operation data metrics, construct a container symbiotic relationship graph with containers as nodes, the relationships between containers as edges, and the symbiotic advantage scores as edge weights, and perform intelligent pairing of different types of containers according to the symbiotic relationship graph; A resource allocation module, which is used to continuously obtain the container real-time behavior data and historical fingerprints, calculate the deviation between the two, calculate the resource competition relationship according to the container real-time behavior fingerprint, calculate the saturation of various resources, and establish a multi-objective programming model with the goal of minimizing the container behavior deviation and resource competition relationship and maximizing the resource saturation or average resource utilization rate, and obtain the resource scheduling strategy; A cross-container resource lending module, which is used to judge the resource demand situations of different containers at different time points according to the container behavior fingerprints, divide the containers with low resource demand and the containers with high resource demand, adopt a distributed negotiation framework, and establish a cross-container resource lending contract between the containers with low resource demand and the containers with high resource demand through a resource credit mechanism to maintain the system resource balance; A closed-loop reinforcement learning optimization module, which is used to use the container behavior fingerprint features, the container pairing information extracted from the dynamic symbiotic relationship graph, and the resource lending contract status as the input state space, and the control operations for resource allocation and lending behaviors as the action space, establish a deep reinforcement learning optimizer, and the optimizer takes the comprehensive reward function as the optimization goal, and realizes the autonomous evolution of the scheduling strategy by rewarding or punishing the resource scheduling strategy.

8. An electronic device, characterized in that, Including: One or more processors; A memory; And one or more programs, where the one or more programs are stored in the memory and are configured to be executed by the one or more processors. When the program is executed by the processor, it implements the steps of the intelligent resource scheduling method based on container behavior fingerprint recognition according to any one of claims 1-6.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the intelligent resource scheduling method based on container behavior fingerprint recognition according to any one of claims 1-6.

10. An intelligent container resource scheduling device, characterized in that, Including: A heterogeneous computing unit cluster configured to process container behavior data streams in parallel; A distributed storage array configured to store a dynamically updated behavior fingerprint library and a symbiotic relationship map; A programmable logic device configured to implement an adaptive scheduling strategy by implementing the intelligent resource scheduling method based on container behavior fingerprint recognition according to any one of claims 1-6.

Images