A distributed management system and method for a cloud container cluster

Abstract

The present invention relates to the field of cloud computing technology, and discloses a distributed management system and method for a cloud container cluster, including: a data collection module, configured to collect the remaining resources of each node in the cloud container cluster and the resource requirements of tenants, and provide support for real-time data update; a state modeling module, configured to establish a state variable model describing the dynamic changes of tenant resources based on the resource data collected by the data collection module, wherein the state variables include the resource occupancy, resource application volume, and resource release rate of tenants; a scheduling optimization module, configured to generate a resource allocation strategy based on the dynamic change characteristics of tenant resources and the multi-tenant competition relationship; through the system architecture of modular design such as data collection, state modeling, and scheduling optimization, the present invention introduces a dynamic collection frequency and modeling of resource dynamic changes, improves the real-time performance of data collection and the accuracy of resource allocation, and solves the problems of low resource utilization rate and slow scheduling response in the prior art.

Description

Technical Field

[0001] The present invention relates to the technical field of cloud computing, and specifically provides a distributed management system and method for a cloud container cluster. Background Art

[0002] The distributed management system of cloud container clusters is widely used in microservice architectures. By automating deployment, scaling, monitoring, and operation and maintenance, it optimizes the management process of containerized applications and is widely used in microservice architectures, DevOps, elastic computing, edge computing, high availability, and hybrid cloud scenarios. Its core functions include container orchestration, resource management, network and services, monitoring logs, security authentication, etc., which can effectively improve the availability and scalability of the system. For example, enterprises can use this system to achieve rapid application delivery, dynamic resource allocation, multi-cloud deployment, and fault recovery to meet high-concurrency or complex distributed requirements. However, it should be noted that the learning curve of the system is relatively high, involving complex fields such as networks, security, and scheduling, requiring professional knowledge support and effective fault troubleshooting capabilities.

[0003] Existing cloud container management systems have many deficiencies in dynamic scheduling and resource optimization. Most systems adopt static scheduling strategies or rule-based resource allocation methods, lacking the ability to dynamically respond to load fluctuations and being difficult to meet the needs of sudden tasks and complex scenarios. Scheduling strategies are usually based on single-node or local optimization, failing to optimize resource allocation from a global perspective, resulting in uneven resource utilization or even serious waste. In addition, the predefined resource quota method is prone to generating idle resources, and the system's ability to recycle and reallocate resources is limited, further exacerbating resource waste.

[0004] In summary, in a multi-tenant environment, the distributed management system of cloud container clusters has insufficient service quality guarantee capabilities. High-priority tasks may be delayed due to insufficient resources, while low-priority tasks are frequently interrupted. The traditional system scheduling strategy lacks intelligent optimization capabilities, cannot learn and optimize based on historical operation data, and is also difficult to balance multi-objective requirements such as task latency, energy efficiency, and resource utilization. Summary of the Invention

[0005] In view of the deficiencies of the prior art, the present invention provides a distributed management system and method for a cloud container cluster. By combining modular design, dynamic mathematical modeling, and intelligent optimization algorithms, it solves the problems of lacking the ability to dynamically respond to load fluctuations and being difficult to meet the needs of sudden tasks and complex scenarios.

[0006] To achieve the above objectives, the present invention is realized through the following technical solutions: A distributed management system for a cloud container cluster, comprising:

[0007] A data collection module, used to collect the remaining resources of each node in the cloud container cluster and the resource requirements of tenants, and provide support for real-time data updates;

[0008] A state modeling module, configured to establish a state variable model describing the dynamic changes of tenant resources based on the resource data collected by the data collection module, wherein the state variables include the resource occupancy, resource application volume, and resource release rate of the tenant;

[0009] A scheduling optimization module, configured to generate a resource allocation strategy through an optimization algorithm based on the dynamic change characteristics of tenant resources and the multi-tenant competition relationship;

[0010] A scheduling execution module, configured to dynamically adjust the allocation of cloud container resources according to the resource allocation strategy generated by the scheduling optimization module, including allocating computing resources and adjusting the container deployment between nodes;

[0011] A feedback optimization module, configured to adjust and optimize the strategy according to the operation status and scheduling results of resource allocation by monitoring the scheduling execution effect in real time.

[0012] Preferably, the state variable model established by the state modeling module describes the tenant resource changes with a dynamic equation. The resource occupancy of the tenant changes over time and is jointly determined by the tenant's resource application volume and release rate. The model includes the correlation between the tenant's resource occupancy and resource request behavior.

[0013] Preferably, the core function of the scheduling optimization module is to generate a resource allocation strategy based on the data provided by the state modeling module and the utility function. The scheduling optimization module aims to maximize the tenant utility and adopts the following optimization function:

[0014] ,

[0015] where, is the cumulative utility of tenant i, represents the revenue of resource allocation, is the resource application cost, dt represents the differential of time, and T is the optimization period.

[0016] Preferably, the resource allocation strategy is restricted by the following constraint conditions:

[0017] The total resource allocation volume of the cluster cannot exceed its total available resources;

[0018] The resource application volume of a single tenant cannot exceed its maximum application capacity;

[0019] Resource allocation needs to meet the tenant service quality requirements and ensure priority allocation in case of resource shortage.

[0020] Preferably, the scheduling optimization module generates a resource allocation strategy through the Hamilton-Jacobi-Bellman equation, and the equation is:

[0021] , where is the value function of the tenant, represents the benefit coefficient of resource usage, represents the cost coefficient of resource application, is the current resource occupancy of tenant i, represents the rate of change of the value function of tenant i over time, is the partial derivative of the value function with respect to the state variable, is the dynamic change of the state variable, is the optimization operation of the resource application strategy, is the benefit item of resource allocation, is the cost item of resource application.

[0022] Preferably, the feedback optimization module updates the resource allocation strategy through the deep Q-network of reinforcement learning, and the update rule is:

[0023] ,

[0024] where is the value function of the current state and action, represents the state of the system at the current moment k, represents the action taken by the system at the current moment k, is the immediate reward, is the discount factor, is the maximum estimated value of the future return.

[0025] Preferably, the scheduling execution module includes the following functions:

[0026] Dynamically adjust the computing resources and memory resource allocation of tenants according to the resource allocation strategy;

[0027] Redeploy containers between nodes to balance the computing load of the cluster;

[0028] Trigger the priority management mechanism when resources are scarce, and give priority to ensuring the resource allocation of high-priority tasks;

[0029] Provide a resource recycling function, and dynamically release redundant resources by identifying idle resources or resources occupied by low-priority tasks to improve the overall resource utilization rate of the cluster.

[0030] Preferably, the data collection module obtains the following information through real-time monitoring:

[0031] The remaining resources of each node;

[0032] The resource usage of each tenant;

[0033] The overall load status and performance metrics of the cluster.

[0034] Preferably, the container network optimization function of the scheduling execution module includes the following steps:

[0035] Dynamically adjust the data path according to the real-time communication requirements of containers and the network topology to optimize network performance;

[0036] Prevent unauthorized access by defining network security rules based on container namespaces and tenant isolation;

[0037] Adopt bandwidth allocation optimization and priority scheduling methods to reduce network latency and congestion;

[0038] Automatically trigger network optimization strategies by real-time monitoring of performance metrics such as latency and packet loss rate of communication between containers.

[0039] Preferably, the present invention also provides a distributed management method for a cloud container cluster, including the following steps:

[0040] Real-time collect the resource usage of each node in the cloud container cluster and the resource demand information of tenants;

[0041] Build a state variable model according to the collected resource data, where the dynamic change of tenant resources depends on resource application behavior and task completion release;

[0042] Generate a resource allocation strategy through the utility function optimization method, which is based on the balance of revenue and cost and is restricted by resource allocation constraints;

[0043] Dynamically adjust the computing resource allocation, deployment location and load balancing of containers according to the generated resource allocation strategy;

[0044] Update the resource allocation strategy according to the real-time monitored scheduling effect through reinforcement learning technology and dynamically adapt to load changes.

[0045] The present invention provides a distributed management system for a cloud container cluster. It has the following beneficial effects:

[0046] 1. Through the modular design of the data acquisition, state modeling, scheduling optimization, scheduling execution and feedback optimization system architecture, the present invention achieves the technical effect of efficiently managing cloud container resources. Compared with the prior art technical solutions lacking clear cooperation between modules, it solves the problems of low resource utilization rate and slow scheduling response speed;

[0047] 2. Through the dynamically adjusted acquisition frequency and integration with Kubernetes Metrics Server or Prometheus API, the present invention achieves the technical effect of flexibly adjusting the data acquisition frequency according to the system load. Compared with the prior art solutions with a fixed acquisition period, it solves the problems of low acquisition efficiency and insufficient real-time performance;

[0048] 3. By introducing a resource dynamic change equation in the state modeling module, the present invention adopts a comprehensive modeling method of task requirements, historical resource usage patterns, and resource release rates, achieving the technical effect of accurately describing the resource changes of tenants. Compared with the models lacking dynamic change descriptions in the prior art, it solves the problem that the resource allocation model cannot adapt to the dynamic environment.

[0049] 4. By using the Hamilton-Jacobi-Bellman equation for optimization and solution in the scheduling optimization module and combining the utility function of revenue and cost, the present invention achieves the technical effect of generating a globally optimal resource allocation strategy. Compared with the technical solutions that only rely on simple heuristic algorithms in the prior art, it solves the problems of limited optimization objectives and insufficient globality.

[0050] 5. By combining with the Kubernetes API in the scheduling execution module, the present invention realizes the technical effects of resource allocation, container migration, and dynamic load balancing. Compared with the solutions with fixed resource allocation and slow load balancing response in the prior art, it solves the problems of low system resource utilization efficiency and inability to guarantee service quality. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] Figure 1 is the system architecture diagram of the present invention.

[0052] Figure 2 is the flowchart of the method steps of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0053] Next, the technical solutions of the present invention will be clearly and completely described in conjunction with the drawings of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0054] Please refer to the attached Figure 1 , the embodiment of the present invention provides a distributed management system for a cloud container cluster, aiming at the resource scheduling problem in the distributed management of the cloud container cluster. It includes five modules: data collection, state modeling, scheduling optimization, scheduling execution, and feedback optimization. These modules are based on data collection, realize the efficient allocation of resources through mathematical models and optimization algorithms, and combine a dynamic feedback mechanism to ensure the real-time optimization of the scheduling strategy.

[0055] Data Collection Module

[0056] The main function of the data collection module is to provide real-time operation data support for the state modeling module. Specifically, it includes:

[0057] The current resource usage of each node, such as CPU, memory, and network bandwidth, etc., ensures the accuracy of the global resource distribution. The resource demand data of each tenant, including the resource types, magnitudes required for tasks, and the expected task completion time. The overall load situation of the cluster, including the average load of nodes and the length of the task queue. As an option, the data collection module realizes seamless integration with the cloud container platform through a standardized interface (Kubernetes Metrics Server). It should be noted that in order to ensure the real-time nature of the data, the collection period is usually set to 5 seconds, and this parameter can be dynamically adjusted according to the actual load situation.

[0058] Specifically, in the case of low load, the collection period can be extended to 10 seconds to reduce system overhead; while in the case of high load, the collection period can be shortened to 2 seconds to provide more accurate operation data in a timely manner.

[0059] It should be noted that the output of the data collection module not only serves the state modeling module, but can also be directly passed to the scheduling optimization module to quickly respond to the resource demands of tenants in special cases.

[0060] Exemplarily, the collected data will be directly passed to the state modeling module to generate a mathematical model describing the dynamic changes of resources, providing accurate basic data support for subsequent scheduling optimization.

[0061] State modeling module

[0062] The state modeling module is used to describe the dynamic process of tenant resource application, occupancy, and release, providing an accurate mathematical basis for scheduling optimization.

[0063] In one implementation, the state modeling module establishes a model based on the following dynamic change equation:

[0064] ,

[0065] where, is the change rate of tenant resource occupancy, is the resource occupancy of the i-th tenant at time t, is the resource amount applied by the i-th tenant at time t, is the resource release rate of the i-th tenant, is the resource amount released by the i-th tenant at time t due to task completion.

[0066] It should be noted that the resource application amount can be expressed by the following formula:

[0067] ,

[0068] where, : The current task requirement of tenant i, which describes the resource requirement at time t. : The historical resource usage pattern of tenant i.

[0069] As an option, the state modeling module can adjust the model parameters for different types of tenant tasks to more accurately describe the resource change behavior. The resource release rate for high-priority real-time video processing tasks is usually set higher so that resources can be quickly released after the task is completed, while the resource release rate for low-priority batch processing tasks can be set lower to reduce frequent resource reallocation operations.

[0070] Modeling of high-priority tasks: For tasks with high real-time requirements, the changing trend of the resource application volume can be described by the following formula:

[0071] .

[0072] Modeling of low-priority tasks: For delayable tasks, a lag factor can be introduced into the calculation formula of their resource application volume;

[0073] ,

[0074] where represents the cumulative delay of tenant i's high-priority tasks at time t, represents the unit resource efficiency of tenant i in executing high-priority tasks, To ensure that the resource application volume is not negative, represents the maximum delay that tenant i can tolerate for low-priority tasks.

[0075] Exemplarily, for a mixed task scenario with complex task types, the state modeling module can build a hierarchical model by combining multi-task characteristics, classify tasks by priority for modeling, and finally generate an overall description of resource changes.

[0076] Scheduling optimization module

[0077] The scheduling optimization module generates an optimal resource allocation strategy through a utility function and constraint conditions. Specifically: Construction of the utility function:

[0078] In a possible implementation, the utility function of the tenant is defined as follows:

[0079] ,

[0080] With the goal of maximizing the total utility of the tenant, while satisfying the following constraint conditions:

[0081] ,

[0082] The total cluster resources cannot exceed its available capacity.

[0083] ,

[0084] Among them, is the cumulative utility of tenant i, represents the benefit of resource allocation, is the resource application cost, dt represents the differential of time, and T is the optimization period. represents the total resource occupancy of all n tenants in the system at time t, and n represents the total number of tenants participating in resource competition in the system. represents the amount of resources allocated to tenant i by the system at time t. represents the total available amount of resources in the system. represents the resource application amount of tenant i. cannot be negative, indicating that the resource application amount of tenant i at time t cannot exceed its maximum application capacity. , represents the maximum amount of resources that tenant i can apply for at a single time point.

[0085] The resource application amount of a single tenant shall not exceed its maximum application capacity.

[0086] The scheduling optimization module uses the Hamilton-Jacobi-Bellman equation to solve the resource allocation strategy:

[0087] , where is the value function of the tenant, represents the benefit coefficient of resource use, represents the cost coefficient of resource application, is the current resource occupancy of tenant i, represents the rate of change of the value function of tenant i over time, is the partial derivative of the value function with respect to the state variable, is the dynamic change of the state variable, is the optimization operation of the resource application strategy, is the benefit term of resource allocation, is the cost term of resource application.

[0088] As an option, the scheduling optimization module can adjust the optimization goal in combination with the resource allocation requirements of different scenarios. For example, in the edge computing scenario, the optimization goal of the resource allocation strategy can be changed from maximizing tenant utility to minimizing task latency. Specifically:

[0089] Latency optimization goal: In a latency-sensitive scenario, the optimization goal can be defined as:

[0090] ,

[0091] Among them, min indicates that the goal of optimization is to minimize the value of the objective function. represents the cumulative execution delay of tenant i's tasks under the current resource allocation strategy. represents the contribution degree of the unit resources allocated to tenant i to the reduction of delay. represents the delay penalty caused by insufficient resource allocation.

[0092] Energy efficiency optimization goal: For energy consumption sensitive scenarios, the scheduling optimization module can add an energy consumption cost item to the utility function:

[0093] ,

[0094] Among them, represents the total utility of tenant i within the time interval. represents the energy consumption cost coefficient. is the energy consumption for task execution. represents the energy consumption cost generated by tenant i using resources at time T.

[0095] Exemplarily, the scheduling optimization module can use a reinforcement learning algorithm to replace the traditional dynamic programming method, so as to achieve more efficient resource allocation optimization in the high-dimensional state space.

[0096] It should be noted that regardless of the optimization goal adopted, the finally generated resource allocation strategy needs to meet the total cluster resource limit and the service quality requirements of tenants.

[0097] Scheduling execution module

[0098] The scheduling execution module allocates resources according to the optimization strategy and implements specific operations. Its functions include:

[0099] Dynamically allocate CPU and memory resources for tenants, and update resource limits in real time through the Kubernetes quota mechanism;

[0100] When the node load is unbalanced, trigger container migration, and schedule tasks from high-load nodes to low-load nodes;

[0101] Give priority to allocating resources for high-priority tasks, and recycle resources for low-priority tasks when resources are insufficient.

[0102] The scheduling execution module is integrated with the Kubernetes API, and realizes the automation of resource allocation and container migration through the scheduler interface.

[0103] As an option, the scheduling and execution module can implement dynamic load balancing among containers by combining service mesh technology. The introduction of service mesh can significantly improve the fault tolerance and communication efficiency of the system.

[0104] Specifically:

[0105] Through the service mesh control plane, the traffic allocation is dynamically adjusted according to the resource usage of each container, avoiding performance degradation caused by a single container being overloaded.

[0106] Using the data plane of the service mesh, the communication data of different tenants is encrypted and isolated to ensure security in a multi-tenant environment.

[0107] Exemplarily, when migrating containers between nodes, the scheduling and execution module can combine the traffic routing ability of the service mesh to achieve seamless migration of tasks and avoid interruptions during the migration process.

[0108] It should be noted that the functional expansion of the scheduling and execution module does not affect the implementation of the core resource allocation strategy, but provides higher reliability and flexibility for the system.

[0109] Feedback and optimization module

[0110] The feedback and optimization module dynamically adjusts the resource allocation strategy through reinforcement learning, and the specific implementation is as follows: The current system state includes the resource allocation amount, task completion time, and overall cluster load.

[0111] Immediate reward Calculated based on resource utilization and task latency, with the goal of maximizing resource utilization efficiency and minimizing task latency.

[0112] Deep Q-network update: Optimize the Q-network parameters through the following update formula:

[0113] ,

[0114] where is the value function of the current state and action, represents the state of the system at the current time step k, represents the action taken by the system at the current time step k, is the discount factor, is the maximum estimated value of the future return, is the immediate reward, taking into account resource utilization, task completion time, and quality of service, and is specifically defined as:

[0115] ,

[0116] where represents the resource utilization efficiency of the system at time step k, Denote the time consumed from the arrival of a task at the system to its completion at time step k. Denote the quality loss caused by the system not meeting the tenant service level agreement at time step k. , , : Weight parameter, used to balance the priorities of different objectives.

[0117] As an option, the feedback optimization module can combine the special requirements of the edge computing scenario to dynamically adjust the resource allocation ratio between the edge node and the central node.

[0118] Specifically:

[0119] Central-edge collaborative optimization: The feedback optimization module can dynamically adjust the resource allocation strategy of the edge node according to the geographical distribution and data processing requirements of the task. For example, for tasks that require local processing, edge node resources are preferentially allocated; for tasks that require cross-node collaboration, central node resources are preferentially allocated.

[0120] Dynamically train the reinforcement learning model: The feedback optimization module can use the distributed reinforcement learning method to introduce the real-time data of the edge node into the training process, thereby improving the generalization ability of the Q network.

[0121] Exemplarily, when processing latency-sensitive tasks, the feedback optimization module can combine the prediction ability of the distributed Q network to adjust the resource allocation strategy in advance to cope with sudden loads.

[0122] It can be understood that the design of the feedback optimization module is not only applicable to the centralized cloud computing scenario, but also can meet the requirements of the distributed and edge computing scenarios.

[0123] It should be noted that the feedback optimization module works in cooperation with the scheduling optimization module to dynamically update the resource allocation strategy to ensure that the system remains in an optimal state under load changes.

[0124] Please refer to the appendix Figure 2 , The present invention also provides a distributed management method for a cloud container cluster, which combines various technical means such as dynamic resource monitoring, mathematical modeling, optimization algorithms, and reinforcement learning to ensure the safe sharing and efficient management of resources in the cloud container cluster. The following will describe in detail each step of the method and its implementation manner. The method can be used to achieve efficient resource utilization and service quality optimization during the dynamic allocation of cloud computing resources and task scheduling.

[0125] S1. Data collection provides real-time monitoring of cloud container cluster resources

[0126] In this step, the data collection process can utilize Kubernetes Metrics Server and Prometheus API to obtain the usage status of node resources in real time, including key metrics such as CPU, memory, and network bandwidth. It should be noted that the real-time nature of data collection is controlled by the collection period, which is set to 5 seconds by default, and the specific interval can be dynamically adjusted according to the system load.

[0127] Specifically, the collection of node resource usage includes the remaining number of CPU cores, available memory capacity, and network bandwidth occupancy. The resource demand information of the tenant includes the resource types, resource requirements of the currently running tasks, as well as the task priority and estimated completion time.

[0128] In some embodiments, the collected tenant demand information is updated in real time to the system's state modeling module to ensure that fluctuations in the tenant's resource requirements can be captured in a timely manner. It should be noted that this process ensures that the input data for scheduling optimization is always consistent with the actual operating state of the system.

[0129] S2. State modeling provides dynamic resource state modeling for cloud containers

[0130] In this step, when constructing the state variable model, the dynamic characteristics of resource application behavior and task completion and release are combined. As a possible implementation, the state variable model describes the mapping relationship between the tenant's resource usage and its resource requirements.

[0131] Specifically: Resource application behavior reflects the tenant's resource usage requirements and is affected by task type, running duration, and priority. For example, tasks with strong real-time requirements will exhibit a higher resource application frequency, while the resource applications of non-real-time tasks are more predictable. Task completion and release describe the process of the tenant releasing the occupied resources after completing the task. As an option, this release process can dynamically adjust the release rate according to the task completion rate. For example, for high-priority tasks, their resource release rate is higher to quickly free up resources for other tasks. It should be noted that during the construction of the state variable model, the tenant's historical resource usage patterns are combined to further improve the prediction accuracy of future resource requirements. Exemplarily, the output of this model is the dynamic change data of resource usage, which is directly used by the subsequent optimization module.

[0132] S3. Generate resource allocation strategies

[0133] In this step, the resource allocation strategy is generated based on the utility function optimization method. The relevant formulas and algorithm logics have been described in detail in the system section and will not be repeated here. The generated resource allocation strategy can be further refined into multi-layer strategies, such as node allocation strategy and container allocation strategy. At the same time, the formulation of the strategy can be based on a multi-objective optimization model, comprehensively considering the resource allocation utility and the load balancing objective. In edge computing or multi-cloud environments, the resource allocation strategy can be extended to cross-region or cross-platform resource coordination to ensure unified scheduling in multiple scenarios.

[0134] S4. Dynamically adjust the computing resource allocation, deployment location, and load balancing of containers

[0135] In this step, the resource allocation and deployment adjustment operations of containers have been described in detail in the system section and will not be repeated here. When adjusting the deployment location, according to the resource sensitivity of the task and the network latency requirements, nodes with the lowest latency can be preferentially selected to deploy compute-intensive tasks, and storage-intensive tasks are preferentially deployed to high-storage nodes. In some embodiments, service mesh technology can be combined to optimize the routing of traffic between containers, thereby reducing the communication overhead of task execution.

[0136] S5. Dynamically update the resource allocation strategy through reinforcement learning technology

[0137] In this step, the dynamic update of the resource allocation strategy through reinforcement learning technology has been described in detail in the system section and will not be repeated here. The reinforcement learning module can be combined with the historical data analysis module to enhance the initial performance of the model through offline learning and adapt to the real-time changes of the environment through online learning. In some embodiments, the input state of reinforcement learning can be extended to a multi-dimensional state space, such as including more dynamic parameters such as network bandwidth, task priority, and container running environment, so as to further improve the intelligent level of scheduling.

[0138] The present invention provides a distributed management method for a cloud container cluster, which realizes efficient resource scheduling management by collecting resource usage and demand information in real time, constructing a dynamic state model, generating a resource allocation strategy based on utility function optimization, dynamically adjusting the resource allocation and deployment location of containers, and updating the scheduling strategy using reinforcement learning technology. Each step cooperates with each other to ensure the real-time and efficiency of resource allocation in a dynamic load change environment, effectively improving resource utilization and service quality, and enhancing the adaptability and optimization performance of the system.

[0139] Although the embodiments of the present invention have been shown and described, it will be understood by those of ordinary skill in the art that various changes, modifications, substitutions, and variations can be made to these embodiments without departing from the principles and spirit of the present invention. The scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A distributed management system for a cloud container cluster, characterized in that: include; The data collection module is used to collect the remaining resources of each node in the cloud container cluster and the resource requirements of tenants, and provide real-time data update support; A state modeling module is used to establish a state variable model describing the dynamic changes of tenant resources based on the resource data collected by the data collection module, wherein the state variables include the tenant's resource occupancy, resource application amount and resource release rate; a scheduling optimization module is used to generate a resource allocation strategy through an optimization algorithm based on the dynamic change characteristics of tenant resources and the multi-tenant competition relationship; A scheduling execution module, used to dynamically adjust the allocation of cloud container resources according to the resource allocation strategy generated by the scheduling optimization module, including allocating computing resources and adjusting container deployment between nodes; Feedback optimization module, which is used to monitor the scheduling execution effect in real time and adjust the optimization strategy according to the operating status of resource allocation and scheduling results; The state modeling module builds a model based on the following dynamic change equations: in, is the change rate of tenant resource usage, r i (t) is the resource usage of the ith tenant at time t, u i (t) is the amount of resources requested by the ith tenant at time t, α i is the resource release rate of the ith tenant, α i r i (t) is the amount of resources released by the ith tenant due to task completion at time t; The scheduling optimization module generates the optimal resource allocation strategy through utility functions and constraints.

2. A distributed management system for a cloud container cluster according to claim 1, characterized in that: The state variable model established by the state modeling module describes the tenant resource changes with a dynamic equation, that is, the tenant's resource occupancy changes over time, which is jointly determined by the tenant's resource application amount and release rate. The model includes the correlation between the tenant's resource occupancy and resource request behavior.

3. A distributed management system for a cloud container cluster according to claim 1, characterized in that: The core function of the scheduling optimization module is to generate resource allocation strategies based on the data and utility function provided by the state modeling module. The scheduling optimization module aims to maximize tenant utility and adopts the following optimization function: Among them, J i is the cumulative utility of tenant i, Q i r i (t) represents the benefit of resource allocation, β i u i (t) 2 is the resource application cost, dt represents the differential of time, and T is the optimization period.

4. A distributed management system for a cloud container cluster according to claim 3, characterized in that: The resource allocation strategy is subject to the following constraints: The total resource allocation of a cluster cannot exceed its total available resources; The amount of resources requested by a single tenant must not exceed its maximum request capacity; Resource allocation must meet tenant service quality requirements and ensure priority allocation when resources are insufficient.

5. A distributed management system for a cloud container cluster according to claim 1, characterized in that: The scheduling optimization module generates a resource allocation strategy through the Hamilton-Jacobi-Bellman equation, which is: Among them, V i is the tenant’s value function, Q i represents the benefit coefficient of resource use, β i represents the cost coefficient of resource application, r i is the current resource usage of tenant i, represents the rate at which the value function of tenant i changes over time, is the partial derivative of the value function with respect to the state variable, is the dynamic change of state variables, For the optimization of resource application strategy, Q i r i is the benefit item of resource allocation, The cost item requested for a resource.

6. A distributed management system for a cloud container cluster according to claim 1, characterized in that: The feedback optimization module updates the resource allocation strategy through the deep Q network of reinforcement learning, and the update rule is: Among them, Q(s k , a k ) is the value function of the current state and action, s k represents the state of the system k at the current time, a k represents the action taken by the system at the current moment, R k is the immediate reward, γ is the discount factor, is the maximum estimate of future returns.

7. A distributed management system for a cloud container cluster according to claim 1, characterized in that: The scheduling execution module includes the following functions: Dynamically adjust tenants' computing and memory resource allocations based on resource allocation policies; Redeploy containers across nodes to balance the computing load of the cluster; When resources are tight, the priority management mechanism is triggered to ensure resource allocation for high-priority tasks first; Provides resource recovery function, which can dynamically release redundant resources to improve the overall resource utilization of the cluster by identifying idle resources or resources occupied by low-priority tasks.

8. A distributed management system for a cloud container cluster according to claim 1, characterized in that: The data acquisition module obtains the following information through real-time monitoring: The remaining amount of resources at each node; Resource usage of each tenant; The overall cluster load status and performance indicators.

9. A distributed management system for a cloud container cluster according to claim 1, characterized in that: The container network optimization function of the scheduling execution module includes the following steps: Dynamically adjust data paths based on real-time container communication requirements and network topology to optimize network performance; Prevent unauthorized access by defining network security rules based on container namespaces and tenant isolation; Use bandwidth allocation optimization and priority scheduling methods to reduce network delays and congestion; By real-time monitoring of the latency and packet loss performance indicators of inter-container communication, network optimization strategies are automatically triggered.

10. A distributed management method for a cloud container cluster, applied to the system according to any one of claims 1 to 9, characterized in that: The following steps are involved: Collect resource usage of each node in the cloud container cluster and tenants' resource demand information in real time; Based on the collected resource data, a state variable model is constructed, in which the dynamic changes of tenant resources depend on resource application behavior and task completion release; Generate a resource allocation strategy through a utility function optimization method, wherein the strategy is based on a benefit and cost balance and is subject to resource allocation constraints; Dynamically adjust the computing resource allocation, deployment location, and load balancing of containers based on the generated resource allocation strategy; Through reinforcement learning technology, the resource allocation strategy is updated according to the scheduling effect of real-time monitoring, and dynamically adapts to load changes.

Images