Cluster load balancing processing method based on cloud computing

Abstract

The invention relates to the technical field of cloud computing, and discloses a cluster load balancing processing method based on cloud computing, which comprises the following steps of: S1, initializing a cluster state: determining computing power, power consumption parameters and a task allocation model of each node in a cloud computing cluster; s2, acquiring a node operation state: providing dynamic input data for construction of a load distribution model; s3, constructing a load distribution model: defining a global optimization target; s4, generating a preliminary task allocation scheme: satisfying an optimization target; s5, aiming at the dynamic change of the task arrival rate and the node state; s6, task migration is executed, and the state is updated; and task migration between the nodes is implemented. Through a closed-loop process of node state acquisition, task allocation model construction, dynamic adjustment and feedback optimization, efficient matching of a task request and a computing node is realized. The method can adapt to dynamic changes in a cloud computing environment, the resource utilization rate is remarkably improved, and task processing delay is effectively reduced.

Description

Technical Field

[0001] The present invention relates to the technical field of cloud computing, and in particular to a cluster load balancing processing method based on cloud computing. Background Art

[0002] As a flexible resource sharing and on-demand allocation technology, cloud computing is widely used in high-performance computing, big data processing, distributed systems and other fields. In the cloud computing environment, cluster load balancing is an important technical means to ensure the efficient operation of the system. Its core task is to reasonably allocate user requests or computing tasks among multiple nodes in the cluster, thereby optimizing resource utilization, improving task processing efficiency, and ensuring system stability and reliability.

[0003] However, in practical applications, cluster load balancing still faces many challenges: First, the cloud computing environment is highly dynamic, and the distribution of task requests and node performance have the characteristics of rapid changes in the time dimension. For example, some nodes may be overloaded due to sudden task requests, while other nodes may have idle resources. Existing load balancing methods usually lack the ability to adapt to this dynamic nature, resulting in reduced resource utilization efficiency and longer task response time.

[0004] Secondly, existing technologies often adopt local optimization strategies, only allocating tasks to a single node or local cluster, but failing to balance the load globally. This limitation leads to uneven resource allocation, with some nodes being in a high-load state for a long time while other nodes are idle, which ultimately affects the overall performance of the system.

[0005] In addition, the real-time and accuracy of task allocation are the key to achieving efficient cloud computing load balancing. In existing technologies, load allocation usually relies on static or semi-dynamic scheduling algorithms, which are difficult to respond to changes in node status in real time in complex dynamic environments, resulting in reduced system stability and reliability. Summary of the invention

[0006] In view of the deficiencies in the prior art, the present invention provides a cluster load balancing processing method based on cloud computing, which solves the problems of inefficient resource allocation in a dynamic environment, difficulty in achieving global load balancing goals, and insufficient real-time task allocation and system reliability.

[0007] To achieve the above objectives, the present invention is implemented through the following technical solutions: A cluster load balancing processing method based on cloud computing includes the following steps: S1. Initialize cluster status: determine the computing power, power consumption parameters and task allocation model of each node in the cloud computing cluster, set optimization goals, and provide basic conditions for subsequent steps; Initializing the cluster state includes determining the overall architecture of the cloud computing cluster, including the number of nodes, the computing power of each node (such as CPU, memory and network bandwidth) and power consumption characteristics (including idle power consumption and dynamic power consumption change rate); setting the task allocation model and clarifying the rules for task arrival and processing; defining the optimization goal, combining the task processing efficiency and energy consumption optimization requirements, and setting the weight parameters for global optimization; initializing the node state, including the task queue length, remaining computing power and initial load, to provide basic support for the task allocation strategy in subsequent steps.

[0008] S2, collect node operation status: obtain the task load, task arrival rate and processing capacity of each node in real time, and provide dynamic input data for the construction of load distribution model; Among them, the node operation status is collected: the operation status of each node in the cloud computing cluster is monitored in real time, including task load (the length of the current task queue of the node), task arrival rate (the number of new tasks arriving at the node per unit time) and task completion rate (the task processing speed based on the remaining computing power of the node); at the same time, the historical operation data and task migration records of the node are recorded to provide dynamic input data for the load distribution model, and support the optimization and adjustment of the task allocation strategy in subsequent steps.

[0009] S3. Build a load distribution model: Based on the collected node status data, establish a dynamic optimization model for load distribution that comprehensively considers power consumption and task delay, and clarify the global optimization goal; Among them, a dynamic optimization model is established based on the collected node status data, combined with the node's task load, task arrival rate and task completion capability; this model comprehensively considers node power consumption (including idle power consumption and load power consumption changes) and task delay (determined by the task queue length and remaining processing capacity), takes global load balancing as the core goal, defines the weight parameters of power consumption optimization and task response time optimization, clarifies the optimization objective function, and provides mathematical description and basic model support for solving task allocation strategies.

[0010] S4. Generate a preliminary task allocation plan: solve the optimization model through game theory methods, determine the preliminary task allocation strategy between nodes, and ensure that the task allocation meets the optimization goal in the current state; Among them, based on the constructed optimization model, the game theory method is used to model the task allocation problem between nodes as a non-cooperative game. The power consumption and task delay of the node are used as the utility function, and the preliminary task allocation strategy between nodes is determined by solving the Nash equilibrium method; by analyzing the current state of the node, including task load, task arrival rate and remaining processing capacity, a task allocation plan that meets the current global optimization goal is generated to ensure load balancing and energy consumption optimization under current conditions.

[0011] S5. According to the dynamic changes of task arrival rate and node status, the initial task allocation plan is adjusted by using the optimization control method to optimize the task allocation ratio; Among them, the optimization control method is used to adjust the initially generated task allocation plan in response to the random fluctuations of the task arrival rate and the dynamic changes of the node status; the task allocation ratio is dynamically optimized by real-time analysis of the node's task load, remaining processing capacity and power consumption characteristics, combined with random optimization or prediction models; in this process, the weight coefficients of power consumption and task delay are dynamically adjusted to meet the priority requirements of performance optimization or energy consumption optimization in different scenarios, thereby improving the adaptability of the task allocation plan and the global optimization effect.

[0012] S6. Execute task migration and update status: According to the optimized task allocation plan, implement task migration between nodes, and update the task queue length and processing capacity status of the node in real time to determine whether the global load balancing and energy consumption optimization goals are achieved. If not, return to S2 and repeat the data collection and optimization process.

[0013] Among them, task migration is executed and the status is updated: according to the optimized task allocation plan, tasks are migrated from nodes with higher loads to nodes with lower loads, and the task queue lengths, remaining computing power and power consumption status of all nodes are updated in real time; the execution effect of task migration between nodes is monitored to determine whether the global load balancing and energy consumption optimization goals are achieved; if the preset convergence conditions are not met, return to step S2, re-collect the node operation status and start a new round of optimization and adjustment process to ensure continuous optimization of the task allocation strategy and achieve dynamic adaptability of the system.

[0014] Preferably, in step S2, the collection of the node operation status includes task load, task arrival rate, task completion rate, node remaining computing capacity and dynamic change data of the task queue; the task completion rate is calculated based on the node remaining processing capacity and the current task load.

[0015] Among them, in step S2, the collection of the node operation status includes task load (the length of the node's current task queue), task arrival rate (the number of new tasks entering the node per unit time), task completion rate (the number of tasks that the node can process per unit time), node remaining computing power (dynamically updated based on the node's hardware resources and current task processing status) and dynamic change data of the task queue; among them, the task completion rate is calculated in real time based on the node's remaining processing power and current task load to reflect the node's actual processing performance.

[0016] Preferably, in step S3, the load distribution model is constructed in the following manner: The power consumption of a node is dynamically calculated based on the node’s current task load and idle power consumption; The task delay of a node is dynamically calculated based on the remaining processing capacity of the node and the length of the task queue; Construct an optimization function with the goal of global load balancing, taking into account the dynamic changes of node power consumption and task delay.

[0017] The load distribution model is constructed in the following way: The power consumption of the node is dynamically calculated by combining the current task load and the idle power consumption, where the power consumption increases nonlinearly with the increase of task load; The task delay of a node is dynamically calculated based on the remaining processing capacity of the node and the length of the task queue. The task delay increases significantly as the processing capacity of the node approaches saturation. Construct a global load balancing optimization function, which comprehensively considers the dynamic changes of node power consumption and task delay, and adjusts the weight parameters of power consumption and delay according to the optimization goals set by the system to ensure the optimal trade-off between energy consumption and performance while load balancing.

[0018] Preferably, in step S4, solving the task allocation strategy by game theory method includes: Model the task allocation problem between nodes as a non-cooperative game; A game model is constructed with node power consumption and task delay as utility functions; The variational inequality method is used to solve the Nash equilibrium point of task allocation between nodes; A preliminary task allocation plan is generated based on the Nash equilibrium point to ensure that the task allocation meets the balance between nodes and the global optimal goal.

[0019] The task allocation strategy solved by game theory method includes: The task allocation problem between nodes is modeled as a non-cooperative game. Each node, as an independent participant in the game, seeks to complete the task allocation under the condition of satisfying its own optimal utility. The game model is constructed with the power consumption and task delay of the node as the key parameters of the utility function. The utility function reflects the comprehensive impact of the node task allocation scheme on system resource consumption and performance. The variational inequality method is used to solve the Nash equilibrium point in the game model, which ensures that the task allocation is balanced under the condition that all nodes compete with each other. A preliminary task allocation plan is generated based on the obtained Nash equilibrium point to ensure that the task allocation between nodes can achieve load balancing and meet the goals of optimizing the system's global energy consumption and performance.

[0020] Preferably, in step S5, the optimization control method for adjusting the task allocation scheme includes: Considering the uncertainty of task arrival rate, the preliminary allocation plan is dynamically adjusted through stochastic optimization method; Optimize the ratio of task migration according to the real-time load and processing capacity of the node; Dynamically adjust the weight coefficients of power consumption and latency to meet the priority requirements of power consumption optimization or performance optimization in different scenarios.

[0021] Among them, the optimization control method adjusts the task allocation plan including: In view of the random fluctuation of task arrival rate, combined with the current task load and processing capacity of the node, a random optimization method is used to dynamically adjust the initially generated task allocation scheme to improve the scheme's adaptability to changes in task arrival; According to the real-time load and remaining processing capacity of each node, the proportion of task migration is optimized to make the task distribution more balanced among nodes and reduce the pressure on high-load nodes; Dynamically adjust the weight coefficients of power consumption and delay in the optimization target, and flexibly switch between power consumption priority optimization and performance priority optimization strategies according to the needs of different application scenarios to meet the global optimization needs in different scenarios.

[0022] Preferably, in step S6, the execution of task migration includes: According to the optimized task allocation plan, tasks are migrated from high-load nodes to low-load nodes; Update the node's task queue length and remaining computing power in real time; Record the historical data of node task migration to provide input for subsequent allocation strategy optimization.

[0023] The execution of task migration includes: According to the optimized task allocation plan, nodes with higher current load are identified and tasks are preferentially selected from their task queues and migrated to nodes with lower loads to achieve load balancing among nodes; Update the task queue length and remaining computing capacity of each node in real time after migration to ensure that the node status data accurately reflects the migration results and provide dynamic support for subsequent steps; Record the historical data of task migration, including the amount of migrated tasks, status changes of source nodes and target nodes, etc., to provide a reference for the optimization of subsequent allocation strategies and the improvement of prediction models.

[0024] Preferably, in step S6, the node status after task migration is fed back to step S2, and the execution effect of the task allocation scheme is judged by monitoring the change of the node queue length, and the optimization model is rebuilt.

[0025] Among them, the node status after task migration is fed back to step S2, and the execution effect of the current task allocation plan is judged by monitoring the changes in the node queue length, remaining computing power and task completion rate; if it is found that the task load between nodes is unbalanced or the global energy consumption and delay optimization goals do not meet expectations, return to step S3, rebuild the load distribution optimization model according to the feedback data, and generate a new task allocation strategy to continuously optimize system performance.

[0026] Preferably, the convergence condition of load balancing is: The change in the task allocation ratio between nodes in two consecutive iterations is less than the preset threshold; The optimization targets of global energy consumption and latency both meet the preset requirements.

[0027] The convergence condition of load balancing is: The change in the task allocation ratio between nodes during two consecutive iterations is less than a preset threshold, which is set according to the dynamic changes of the system's tasks and the node performance characteristics; The global energy consumption optimization target and task delay optimization target both meet the preset requirements. The energy consumption optimization target is based on the lowest overall system power consumption, and the task delay optimization target is based on the balance of task response time of all nodes. Convergence judgment is combined with real-time monitoring data to ensure that the task allocation plan is stable and dynamically adaptable while meeting load balancing requirements.

[0028] Preferably, in step S5, the optimization control method dynamically adjusts the task allocation ratio by introducing historical data and prediction models to minimize the impact of task migration on node load and improve the convergence speed of global optimization.

[0029] Among them, the optimization control method is realized by the following methods: Introduce historical node operation data, including task load change trends, task arrival rate, and completion rate history records, and combine them with the current node status to dynamically adjust the task allocation ratio; Use the prediction model to analyze the changing trend of future task arrival rate and adjust the task allocation strategy in advance to ensure a more balanced distribution of task loads on nodes. By optimizing the selection and allocation strategy of migration tasks, the impact of task migration on node load is minimized, avoiding the increase of system overhead caused by frequent migration; Improve the convergence speed of global optimization so that the task allocation scheme can quickly reach a stable state under dynamically changing workload conditions, thereby achieving higher efficiency and system stability.

[0030] Preferably, during the task migration process, the task allocation priority of the node is adjusted in real time through an energy-aware dynamic scheduling mechanism. The priority is comprehensively evaluated based on the real-time power consumption of the node, the length of the task queue and the remaining computing power, and tasks are preferentially allocated to nodes that meet the low power consumption and load balancing conditions.

[0031] During the task migration process, the following optimizations are achieved through the energy-aware dynamic scheduling mechanism: Comprehensively evaluate the real-time status of each node, including real-time power consumption, task queue length, and remaining computing power; Dynamically adjust the task allocation priority of nodes based on the evaluation results, and prioritize tasks to nodes with lower power consumption and nearly balanced loads to reduce overall energy consumption; Combined with the real-time monitoring data of task migration, the migration strategy is dynamically optimized to avoid resource overhead caused by excessive migration, ensuring that the system minimizes global energy consumption while balancing the load; The dynamic scheduling mechanism is used to respond to changes in task load in real time, making the system adaptive and improving the efficiency and stability of task allocation.

[0032] The present invention provides a cluster load balancing processing method based on cloud computing. It has the following beneficial effects: 1. The present invention realizes efficient matching of task requests and computing nodes through a closed-loop process of node status collection, task allocation model construction, dynamic adjustment and feedback optimization. The method can adapt to dynamic changes in the cloud computing environment, significantly improve resource utilization, and effectively reduce task processing delays.

[0033] 2. In the process of load distribution, the present invention adopts an optimization method combining dynamic adjustment and game theory, which can comprehensively consider the power consumption, task delay and network overhead of nodes with the goal of global load balancing. Compared with traditional local optimization strategies, this method can achieve optimal resource allocation in a larger range and improve the overall performance of the system.

[0034] 3. The present invention can quickly respond to node load changes and fault conditions by collecting node status, task distribution and historical data in real time, and dynamically optimize task allocation schemes by adjusting the optimization model through feedback. This mechanism effectively enhances the adaptability and robustness of load balancing, allowing the system to operate efficiently in complex environments.

[0035] 4. The present invention introduces a random optimization and energy-aware scheduling mechanism in task allocation optimization, and implements a fast-converging task migration strategy in combination with node priority evaluation. In large-scale cloud computing clusters, this method can significantly reduce computational complexity and improve task processing efficiency while maintaining low energy consumption. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Figure 1 The figure is a flow chart of the method of the present invention. DETAILED DESCRIPTION

[0037] The following will be combined with the drawings in the specification of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0038] In order to better understand the present invention, the above contents are described in detail below in conjunction with specific embodiments. Please see attached Figure 1 The embodiment of the present invention provides a cluster load balancing processing method based on cloud computing, which includes the following steps: S1, initializing the cluster state: determining the computing power, power consumption parameters and task allocation model of each node in the cloud computing cluster, setting the optimization target, and providing the basic conditions for the subsequent steps; S2, collecting the node operation status: obtaining the task load, task arrival rate and processing capacity of each node in real time, and providing dynamic input data for the construction of the load distribution model; S3, constructing the load distribution model: based on the collected node status data, establishing a load distribution dynamic optimization model that comprehensively considers power consumption and task delay, and clarifying the global optimization target; S4, generating the initial Step 1: Task allocation plan: Solve the optimization model through game theory method, determine the preliminary task allocation strategy between nodes, and ensure that the task allocation meets the optimization goal under the current state; S5. According to the dynamic changes of task arrival rate and node status, use the optimization control method to adjust the preliminary task allocation plan and optimize the task allocation ratio; S6. Execute task migration and update status: According to the optimized task allocation plan, implement task migration between nodes, and update the task queue length and processing capacity status of the node in real time to determine whether the global load balancing and energy consumption optimization goals are achieved. If not, return to S2 and repeat the data collection and optimization process.

[0039] In this embodiment, step S1 aims to provide basic data support for load balancing optimization. By obtaining the distribution information of task requests in the cluster and the performance parameters of computing nodes, the dynamic characteristics of task requirements and node resources are comprehensively constructed. This step is the prerequisite for the subsequent optimization model construction and the implementation of dynamic allocation strategy. Its data accuracy and real-time performance directly affect the optimization effect of task allocation. The core of this step is to collect and establish a distribution model of task requests and a characterization model of node performance. It should be noted that the task demand distribution and node performance parameters not only include their static characteristics (such as task request volume, node computing power, etc.), but also need to dynamically capture their changing trends to provide support for dynamic optimization.

[0040] It should be noted that the task distribution information and node performance data obtained in this step involve multi-dimensional parameters, including but not limited to the computational workload and latency sensitivity of the task requirements, as well as the computing power, bandwidth availability, and current load of the node. By quantifying and modeling these parameters, accurate support can be provided for the definition and optimization of the cost function in subsequent steps.

[0041] Task requests and node performance are described by the following complete joint formula: ;

[0042] Among them, x represents the specific attribute vector of the task, including: x1 the computing requirements of the task (such as CPU time, memory usage), x2 the task type characteristics (such as video processing, transaction requests), x3 the delay sensitivity of the task (such as real-time tasks with low delay requirements), y represents the performance attribute vector of the node, including: y1 the CPU performance of the node (currently available computing resources), y2 the memory availability of the node, y3 the network bandwidth and delay of the node, the joint distribution P(x,y) represents the matching relationship between tasks and nodes in the current cluster state, μ(x) is the distribution density function of task requests, which represents the statistical distribution of task characteristics; ν(y) is the distribution density function of node performance, which represents the real-time status of node resources in the cluster; Acquisition of task distribution information: Specifically, the task distribution density function μ(x) is obtained and described by: ;

[0043] Among them, P(x1) is obtained through the resource requirement declaration submitted by the task, including the computing amount and memory requirement of the task, P(x2) is the task type feature determined by the task identifier and application classification, for example, video processing tasks have high computing requirements, and online transaction requests have high latency sensitivity, P(x3) is the latency sensitivity marked by the task scenario or business requirements, for example, high-priority tasks have low tolerance for latency, and P(x 1， x 2， x3) ​​represents a three-dimensional probability distribution function, which is used to represent three random variables x 1， x 2， Joint probability distribution between x3;

[0044] It should be noted that the task distribution information can also be supplemented by historical predictions. For example, based on historical task requests, the time series prediction model (such as LSTM) is used to estimate the task distribution formula at the next moment as follows: ;

[0045] Among them, μ t+1 (x) Task distribution at the next moment, μ t(x) is the historical task distribution, It represents the change in task distribution over time and can be calculated by training the model.

[0046] Acquisition of node performance data: The node performance distribution density function ν(y) is obtained and described by the following method: ;

[0047] Among them, P(y1) is collected by real-time monitoring of the CPU utilization of the node, such as the resource statistics provided by the node manager (such as Kubernetes), P(y2) is obtained by the node's memory monitoring tool (such as Prometheus), P(y3) is collected by the network measurement tool in real time for bandwidth usage and latency, and P(y 1， y 2， y3) represents a three-dimensional probability distribution function, which is used to represent three random variables y 1， y 2， The joint probability distribution between y3.

[0048] As a possible implementation, the node performance distribution ν(y) can be dynamically updated to reflect changes in the node resource status. For example, when the node CPU utilization approaches the upper limit, the performance distribution will be adjusted to reduce its resource allocation priority.

[0049] Therefore, the joint modeling of the complete formula, the joint description formula of task request and node performance is: ; Where P(x, y) represents the joint probability distribution of random variables x and y.

[0050] This formula represents the matching relationship between task characteristics and node performance under the current cluster state. Specifically, the computing requirements of the task need to match the available CPU resources of the node, and the delay sensitivity of the task needs to adapt to the network delay of the node. It should be noted that the joint distribution not only provides a reference for the initial matching of task allocation, but also provides data support for the cost function definition of the subsequent optimization model.

[0051] In step S2, the collection of node operation status includes task load, task arrival rate, task completion rate, node remaining computing capacity and dynamic change data of task queue; the task completion rate is calculated based on the node remaining processing capacity and current task load.

[0052] In the present invention, the core of step S2 is to establish a cost model for allocating tasks and computing nodes, and to quantitatively describe the network delay, node energy consumption, and system reliability costs that may be generated during the task allocation process through the cost function, so as to provide a clear mathematical expression for load balancing optimization. To support the construction of the optimization model, this step also combines the collection of node operating status to fully quantify the task distribution characteristics and node dynamic capabilities.

[0053] It should be noted that the task allocation cost model not only relies on the static characteristics of tasks and nodes (such as computing requirements and performance parameters), but also needs to be combined with the dynamic changes in the node operating status, including task load, task arrival rate, task completion rate, node remaining computing power and the real-time status of the task queue, so as to more accurately characterize the allocation cost.

[0054] In this embodiment, the collection of node operation status includes the following: Task load: the total amount of tasks currently being processed by the node, which is a direct indicator of the node operation load.

[0055] Task arrival rate: the number of new tasks arriving at a node per unit time, reflecting the task processing pressure of the node. Task completion rate: calculated based on the node's remaining computing power and the current task load, indicating the node's task digestion capacity. Node remaining computing power: the remaining resources that the node can use to process new tasks, including CPU, memory, etc. Task queue dynamic change data: the length and change rate of the node's internal task queue, used to reflect the task queue time and congestion level.

[0056] As a possible implementation method, the task completion rate of a node can be calculated based on the remaining computing capacity of the node and the current task load, as follows: ;

[0057] Among them, R(y) is the current task load, C avail (y) is the remaining computing power, including the currently available CPU and memory resources, and L(y) is the current task load of the node, which indicates the total number of tasks being processed. It should be noted that the remaining computing power and task load of the node can be collected through real-time monitoring tools. For example, the node management module in Kubernetes can report the resource usage of the node in real time. Specifically, the task allocation cost model is constructed by the following formula: ;

[0058] Among them, c(x,y) represents the total cost of assigning task x to node y; α, β, and γ are the weight parameters of network delay, energy consumption, and failure risk, respectively, which are used to adjust the relative importance of the three costs.

[0059] The definitions of the costs of each part are as follows: Network delay Latency (x, y) describes the communication delay of task x from its initiation location to node y, which can be calculated by the following formula: ;

[0060] Where d(x,y) is the network distance or communication path length, b(y) is the network bandwidth availability; Energy consumption Energy(x, y) describes the energy consumed by task x when executing on node y. The formula is: ;

[0061] Where w(x) is the computing demand, p(y) is the unit computing power consumption; FailureRisk(y): the probability of failure that may occur at node y, usually estimated based on historical data and current operating status; As a possible implementation method of the extension, the impact of dynamic changes in the task queue on the allocation cost can be introduced by the following formula: ;

[0062] Where Q(y) is the length or waiting time of the task queue of the node, which is used to describe the queuing delay that the task may suffer, and δ is the weight of the queue impact factor; It should be noted that the dynamic change of the queue length Q(y) can be estimated by the following formula: ;

[0063] Among them, A(t) is the task arrival rate, R(y,t) is the task completion rate; In summary, the present invention establishes a comprehensive task allocation cost model by collecting dynamic data of node operation status, covering factors such as network delay, node energy consumption, failure risk and task queue impact. Those skilled in the art can reproduce the cost model construction process described in this embodiment according to the formula definition and data collection method to provide support for subsequent load balancing optimization; In step S3, the load distribution model is constructed in the following way: the power consumption of the node is dynamically calculated based on the current task load and idle power consumption of the node; the task delay of the node is dynamically calculated based on the remaining processing capacity of the node and the length of the task queue; an optimization function with global load balancing as the goal is constructed, which comprehensively considers the dynamic changes of node power consumption and task delay.

[0064] Step S3 of the present invention aims to construct an optimization model with global load balancing as the goal, which comprehensively considers the utilization efficiency of cluster resources and task response time by dynamically describing the changes in node power consumption and task delay. In this step, the load distribution optimization model adopts the method of constrained optimization to globally match task requirements with node performance, and finally achieves dynamic load balancing.

[0065] It should be noted that the optimization model in this step aims to minimize the global cost of task allocation, involving the dynamic calculation and modeling of two key indicators: node power consumption and task delay, and ensuring the dynamic adaptability of the optimization model through real-time data-driven approach; In this embodiment, the dynamic calculation of node power consumption is as follows: ;

[0066] Among them, P idle (y) is the idle power consumption of node y, which indicates the basic energy consumption of the node when there is no task processing; P load (y) is the load-dependent power consumption of node y, which represents the additional energy consumed by the node when processing the current task load; Task load related power consumption P load (y) is dynamically calculated based on the current task load L(y) of the node and the unit task power consumption P(y), and the formula is: ;

[0067] As a possible implementation method, the dynamic change of node power consumption can be achieved by collecting task load and node power consumption parameters through real-time monitoring tools (such as Prometheus); In this embodiment, the dynamic calculation of node task delay is as follows: The task delay L(y) of a node includes the processing capacity of the node and the dynamic state of the task queue, which can be expressed as: ;

[0068] Where Q(y) is the length of the node's task queue, indicating the total number of tasks to be processed; C avail (y) is the remaining processing capacity of the node, which indicates the computing resources currently available for the node to process tasks; The task queue length Q(y) is calculated based on the dynamic changes of task arrival rate and completion rate, and the formula is: ;

[0069] Among them, A(t) is the task arrival rate of the node, which indicates the number of new tasks added per unit time; R(t) is the task completion rate of the node, which indicates the number of tasks completed per unit time; It should be noted that the dynamic change of task delay Q(y) can reflect the congestion state and processing pressure of the node, and provide a key reference for the optimization goal of global load balancing; In this embodiment, the optimization function with global load balancing as the goal is constructed as follows: The optimization goal is to minimize the global cost function, taking into account the dynamic changes of node power consumption and task delay, which can be specifically expressed as: ;

[0070] Where F is the optimization objective function of global load balancing; α and β are weight parameters used to balance the relative impact of node power consumption and task delay; E(y) is the total power consumption of node y; D(y) is the task delay of node y; Under the constraints, the optimization function needs to satisfy the matching relationship between task requirements and node performance: The total distribution of task requirements must meet: ;

[0071] Among them, f(x,y) represents the allocation amount of task x to node y; The total load of the node must meet the following requirements: ;

[0072] Among them, C total (y) represents the total computing capacity of node y.

[0073] As a possible implementation method of expansion, the optimization of the global cost function can be decomposed in combination with the distributed computing model, for example, by adopting a hierarchical optimization method: in the local cluster, the task allocation between nodes is prioritized to be balanced; in the global scope, the task allocation across clusters is adjusted according to the load conditions of different clusters. Through hierarchical optimization, the computational complexity of the model can be further reduced and the applicability of the model can be improved.

[0074] In summary, the present invention constructs an optimization model with global load balancing as the goal through dynamic modeling of node power consumption and task delay, and describes the optimization objective function and its constraints through clear mathematical formulas. Those skilled in the art can reproduce the load distribution optimization model described in this embodiment based on the formula content and dynamic collection method, and provide theoretical support and practical path for task allocation in cloud computing environment.

[0075] In step S4, solving the task allocation strategy through game theory methods includes: modeling the task allocation problem between nodes as a non-cooperative game; constructing a game model with node power consumption and task delay as utility functions; using variational inequality methods to solve the Nash equilibrium point of task allocation between nodes; generating a preliminary task allocation plan based on the Nash equilibrium point to ensure that the task allocation meets the balance between nodes and the global optimal goal.

[0076] In the present invention, the core of step S4 is to solve the task allocation strategy through game theory methods to achieve dynamic adaptation between task requests and node performance. By modeling the task allocation problem between nodes as a non-cooperative game, building a game model with node power consumption and task delay as utility functions, and solving the Nash equilibrium point using the variational inequality method, a preliminary task allocation plan is finally generated to ensure that the allocation satisfies the load balance between nodes while taking into account the global optimal goal.

[0077] It should be noted that the game theory method used in this step can dynamically adapt to changes in task requirements and node performance. By combining local optimization with global balance, the task allocation scheme can still maintain efficiency and robustness in a complex dynamic environment.

[0078] In this embodiment, the task allocation problem is modeled as a non-cooperative game: In the non-cooperative game, each node y∈N is regarded as a player, and its goal is to optimize its utility function by adjusting its own task allocation. The task allocation f(x,y) represents the resource requirement of task x to node y; The utility function of a node consists of two parts: power consumption and task delay, specifically: ;

[0079] E(y) is the power consumption of node y (defined in step S3), D(y) is the task delay of node y (defined in step S3), α and β are the weights of power consumption and delay; In the non-cooperative game, the goal of each node is to maximize its utility function E(y) by adjusting the task allocation f(x, y), that is:

[0080] It should be noted that the characteristics of non-cooperative games allow each node to optimize its own goals independently while being influenced by the strategies of other nodes, thus forming a dynamic equilibrium state; In this embodiment, the variational inequality method is used to solve the Nash equilibrium point: In game theory, Nash equilibrium is defined as a task allocation scheme that prevents any node from improving its own utility by unilaterally changing its allocation strategy. Mathematically, the solution to Nash equilibrium can be achieved through the variational inequality method.

[0081] The formal description of variational inequalities is as follows: ;

[0082] Among them, f is the Nash equilibrium point of task allocation; is the gradient of the utility function; F is the feasible solution space for task allocation, which is constrained by task requirements and node performance.

[0083] In this embodiment, a preliminary task allocation plan is generated based on the Nash equilibrium point: The Nash equilibrium point obtained by using the variational inequality method Represents the preliminary task allocation strategy for each node. The specific task allocation plan is generated through the following steps: according to The task allocation ratio in , assigning task x to the corresponding node y satisfies the matching relationship between task demand distribution μ(x) and node performance distribution ν(y); Preliminary allocation plan It can be further adjusted to meet the global optimization goal, such as optimizing load balancing by recalculating task delay and power consumption weight parameters α and β.

[0084] It should be noted that after the preliminary task allocation plan is generated, it can be used as the input of the global optimization model to further adjust the task allocation ratio between nodes in combination with the global goal.

[0085] Extended technical content: As a possible extended implementation method, the generation of preliminary task allocation plans can be combined with distributed optimization methods to reduce the burden of centralized computing: within each local cluster, the game model is independently run to solve the equilibrium strategy between nodes; on a global scale, the local equilibrium results are integrated to form a task allocation plan across clusters.

[0086] In summary, the present invention solves the task allocation strategy through the game theory method, calculates the Nash equilibrium point using the variational inequality method, and generates a preliminary task allocation plan based on the equilibrium point, which provides an efficient and balanced solution for task allocation in a dynamic environment. Those skilled in the art can realize the technical reproduction of this embodiment according to the above formula and method, combined with actual tasks and node parameters.

[0087] In step S5, the optimization control method adjusts the task allocation plan including: dynamically adjusting the preliminary allocation plan through a random optimization method in combination with the uncertainty of the task arrival rate; optimizing the proportion of task migration according to the real-time load and processing capacity of the node; dynamically adjusting the weight coefficients of power consumption and delay to meet the priority requirements of power consumption optimization or performance optimization in different scenarios.

[0088] The core of step S5 is to adjust the preliminary task allocation plan by optimizing the control method to cope with node performance fluctuations, uncertainty in task arrival rate and system dynamic changes. By combining random optimization methods, node real-time load and processing capacity to migrate tasks, and dynamically adjusting the weight coefficients of power consumption and delay, it ensures that the task allocation plan can maintain global optimization in different scenarios.

[0089] It should be noted that the goal of this step is not only to solve the fault tolerance problem in the event of node failure or performance degradation, but also to further optimize the preliminary task allocation scheme to make it more adaptable and elastic in a dynamic environment, thereby meeting the efficiency and reliability requirements of the cloud computing cluster.

[0090] In this embodiment, the preliminary allocation scheme is dynamically adjusted by a random optimization method: The uncertainty of the task arrival rate A(x) is an important factor affecting the allocation scheme. To address this problem, this embodiment adopts a random optimization method to dynamically adjust the preliminary task allocation scheme according to the probability distribution of the task arrival rate.

[0091] The task arrival rate A(x) can be expressed as a random variable, and its probability distribution P(A(x)) is estimated based on historical data. The optimization goal is to adjust the task allocation f(x,y) to minimize the total cost of the task to the node: ;

[0092] in, represents the expectation of the uncertainty of the task arrival rate; c(x,y) is the allocation cost function, which comprehensively considers the impact of power consumption and delay; f(x,y) is the amount of resources allocated to node y from task x.

[0093] It should be noted that the random optimization method can dynamically update the allocation amount f(x, y) through multiple iterations, thereby better adapting to the volatility of the task arrival rate. In some embodiments, Monte Carlo simulation can be used to evaluate the robustness of the task allocation scheme, and the optimization strategy can be adjusted according to the evaluation results.

[0094] In this embodiment, the ratio of task migration is optimized according to the real-time load and processing capacity of the node: Task migration is a key strategy to deal with node load imbalance. Its goal is to balance the task load between nodes without significantly increasing the migration overhead. Indicates migration from node y to node The task volume is calculated as follows: ;

[0095] Among them, L(y) and Represents node y and node The current task load; is the global average load, which represents the average value of the task load in the cluster.

[0096] It should be noted that the task migration ratio The computation is limited by the processing power of the nodes. When the node When the remaining processing capacity is insufficient to carry out the migration task, the migration ratio needs to be further reduced.

[0097] In a possible implementation, the task migration decision can also be combined with task priority. For example, real-time tasks are migrated first, while computing-intensive tasks can be migrated later to reduce the impact on system performance.

[0098] In this embodiment, the weight coefficients of power consumption and delay are dynamically adjusted to meet different optimization priorities: In order to meet the requirements of different scenarios for power consumption optimization or performance optimization, this embodiment dynamically adjusts the weight coefficients α and β of power consumption and delay in the cost function. The adjustment strategy is as follows: ;

[0099] Among them, α represents the weight of power consumption optimization; β represents the weight of delay optimization.

[0100] Specifically: In power-sensitive scenarios (such as large-scale offline computing tasks), priority is given to reducing power consumption and increasing α; in delay-sensitive scenarios (such as real-time audio and video processing tasks), priority is given to reducing delay and increasing β.

[0101] As a possible implementation method, the weight coefficient can be adjusted based on dynamic monitoring of the runtime environment. For example, by analyzing the real-time requirements of tasks in the system and the energy consumption status of nodes, the values ​​of α and β can be dynamically adjusted.

[0102] Extended technical content: Based on task migration and weight adjustment, this embodiment can also combine hierarchical optimization strategy: in the local cluster, the task migration problem between nodes is solved first to ensure local balance; in the global scope, the optimization goals of power consumption and delay are dynamically adjusted according to the resource utilization of different clusters. It should be noted that the hierarchical optimization strategy can further improve the scalability of the model, making it suitable for the load balancing needs of ultra-large-scale cloud computing clusters.

[0103] In summary, the present invention adjusts the task allocation scheme by a random optimization method in step S5, optimizes the task migration ratio in combination with the real-time load and processing capacity of the node, and dynamically adjusts the weight coefficients of power consumption and delay to meet the optimization requirements of different scenarios. The above technical scheme realizes the global optimization of fault-tolerant processing and task redistribution through a clear mathematical model and a dynamic control strategy. Those skilled in the art can reproduce and apply the technical scheme according to the disclosed formulas and methods in combination with the actual cluster environment.

[0104] In step S6, the execution of task migration includes: migrating tasks from high-load nodes to low-load nodes according to the optimized task allocation plan; updating the task queue length and remaining computing power of the node in real time; and recording the task migration history data of the node to provide input for subsequent allocation strategy optimization.

[0105] The core of step S6 is to perform task migration according to the optimized task allocation plan, dynamically update the task queue and computing power of the node, and record the historical data of task migration to provide input for subsequent allocation strategy optimization. By implementing step S6, the execution process of load balancing can be ensured to be stable and efficient, and task allocation optimization in a dynamic environment can be achieved.

[0106] It should be noted that the execution of task migration involves the transfer of tasks to high-load nodes and the resource scheduling of low-load nodes. It is also necessary to provide support for the dynamic adjustment of the system by real-time monitoring of node status and recording of task migration history data.

[0107] In this embodiment, task migration is performed according to the optimized task allocation scheme: Optimized task allocation plan It specifies the target node y to which each task x should be assigned. For high-load node y h , the migration amount is determined by the following formula: ;

[0108] in, From the high load node y h To the low load node y l The amount of tasks to be migrated; L(y h ) and L(y l ) are the current task loads of high-load nodes and low-load nodes respectively; is the global average load of the cluster.

[0109] As a possible implementation method, task migration can be sorted in combination with task priorities. For example, real-time tasks are migrated first, while computing-intensive tasks are migrated when resources are sufficient.

[0110] It should be noted that the amount of tasks to be migrated cannot exceed the low-load node y l The remaining computing capacity C avail (y l ). When the computing power of low-load nodes is insufficient, the migration amount needs to be further reduced.

[0111] In this embodiment, the task queue length and remaining computing capacity of the node are updated in real time: After task migration, high-load node y h and the task queue length Q(y) and remaining computing capacity C of the low-load node y1 avail (y) needs to be updated in real time. The update formula is as follows: ;

[0112] in, is the migration task amount. h ) is the queue length of the high-load node, Q(y l ) is the queue length of the low-load node; Remaining computing power update: ;

[0113] Among them, C avail (y h ) and C avail (y l ) is y h and l The remaining computing power, that is, the amount of resources (such as CPU, memory, etc.) available on the current node, is used to process new tasks. is the task migration amount, u(y h ) and u(y l ) The unit computing resources required to process a single task, such as the CPU time or memory consumption required to process a unit task.

[0114] It should be noted that the update of the task queue length and the remaining computing power can reflect the dynamic state of the node and provide real-time input for subsequent allocation strategy adjustments.

[0115] In this embodiment, the task migration history data of the node is recorded: Task migration history data includes the number of migration tasks, information about the source and target nodes, and the delays and computing resource consumption involved in the migration process. These data can be recorded in the following structured form: ;

[0116] Among them, x is the task identifier of the migration; y h and l are the source node and the target node respectively; is the migration task amount; Latency and Energy are the delay and energy consumption in the migration process respectively.

[0117] As an exemplary implementation, historical task migration data can be stored in a distributed log system (such as Elasticsearch) for optimization of subsequent task allocation strategies. For example, historical data can be used to train machine learning models to predict future task arrival distribution and node performance status, thereby adjusting the allocation plan in advance.

[0118] Extended technical content: This embodiment can also be optimized in combination with the network topology of task migration. For example, for cross-region task migration, node pairs with lower communication path delay are preferentially selected to reduce network overhead during the migration process.

[0119] It should be noted that the execution of task migration can also support the extension of the distributed scheduling framework, for example, by combining with the scheduler plug-in in Kubernetes to realize the task migration management of large-scale cloud computing clusters.

[0120] In summary, in step S6, the present invention ensures that the optimized allocation scheme can be efficiently executed and provides support for dynamic adjustment of the system by clearly modeling and describing the execution of task migration, including the calculation of task migration amount, real-time update of node status, and recording of historical data. Those skilled in the art can reproduce the technical content of this embodiment according to the above formulas and methods in combination with an actual cluster environment.

[0121] In step S6, the node status after task migration is fed back to step S2. By monitoring the change of the node queue length, the execution effect of the task allocation scheme is judged and the optimization model is rebuilt.

[0122] The node status feedback mechanism after task migration is an important part of ensuring the dynamic adjustment of the task allocation plan. This step monitors the changes in the task queue length and node computing power by feeding back the real-time status of the node to the optimization process to determine the execution effect of the task allocation plan after migration. By monitoring and analyzing the changes in node status, the load balancing optimization model is rebuilt to provide accurate input for subsequent task allocation.

[0123] It should be noted that the core of this feedback mechanism is to combine the changes in task load in a dynamic environment with the reconstruction of the optimization model to ensure that the model can continuously adapt to the real-time state of the system and achieve the goal of global load balancing.

[0124] In this embodiment, the node status after task migration is quantified by monitoring the task queue length and the remaining computing power. Specifically, the change in the node task queue length can be used as a direct indicator of the migration effect. For example, when the queue length of the high-load node is significantly reduced after migration, and the queue length of the low-load node is significantly increased, it means that the task migration scheme alleviates the imbalance of node load in the short term.

[0125] Alternatively, the change in the task queue length can be described by the following formula: ;

[0126] in, is the change in queue length, Q new (y) represents the task queue length of node y after migration; Q prev (y) represents the task queue length of node y before migration; It should be noted that when When the absolute value of is large, it may mean that the task migration scheme fails to achieve the expected effect on the load distribution of some nodes, and the optimization model needs to be further adjusted.

[0127] Specifically, in a possible implementation, the changes in the task queue of the node will be fed back to step S2 in real time to redefine the allocation cost model between tasks and nodes. For example, when the queue length of a node continues to grow and exceeds the threshold, the delay cost of the node will increase dynamically, prompting the optimization model to reduce the proportion of tasks allocated to the node.

[0128] During the reconstruction of the optimization model, the weight β of the delay cost can be dynamically adjusted according to the changes in the task queue: ;

[0129] Among them, β0 is the initial weight; k is the adjustment coefficient, which is used to reflect the impact of queue changes on the delay weight.

[0130] It is understandable that by dynamically adjusting the weight coefficients, the optimization model can more accurately reflect the real-time changes in node status, thereby achieving a more balanced task allocation.

[0131] Exemplarily, the task migration history data can also be used to train a machine learning model to predict the changing trend of the task queue length. In some embodiments, based on the historical task allocation and queue length data, a regression model (such as linear regression or LSTM) is used to predict future node queue changes, so as to consider the potential growth of node load in advance during the model building stage.

[0132] It should be noted that the node queue change trend generated by the prediction model can be directly integrated into the optimization model to further improve the forward-looking nature of the load balancing solution. For example, by analyzing the predicted growth trend of the high-load node queue, the proportion of tasks allocated to these nodes can be limited in the initial allocation.

[0133] In a possible implementation, the evaluation of the migration effect can also be combined with the change in the remaining computing power of the node for auxiliary judgment. For example, when the remaining computing power of a node has not improved significantly and the queue length continues to increase, it may indicate that the migration solution has not achieved the expected effect in alleviating the load of the node. At this time, the optimization model needs to adjust the migration strategy and further verify the adjustment results through a dynamic monitoring mechanism.

[0134] In summary, the present invention ensures that the optimization model can be reconstructed according to the actual migration effect through the node state feedback mechanism after task migration, combined with the change of task queue length and the dynamic update of remaining computing power. By combining real-time monitoring, historical data analysis and prediction models, the dynamic adaptability and global optimality of load balancing can be further improved. Those skilled in the art can realize the complete technical reproduction of task migration state feedback and optimization model reconstruction based on the above disclosure and formula description.

[0135] The convergence conditions for load balancing are: the change in the task allocation ratio between nodes in two consecutive iterations is less than the preset threshold; the optimization targets of global energy consumption and delay both meet the preset requirements.

[0136] The optimization of the load balancing solution needs to gradually approach the optimal solution during the iteration process, and its convergence is the key to judging whether the optimization process is completed. The load balancing optimization convergence conditions proposed in the present invention include: the change in the task allocation ratio between nodes is lower than the preset threshold, and the optimization targets of global energy consumption and delay meet the preset requirements. This convergence condition can ensure that the task allocation solution achieves a balance between performance and resource efficiency, while avoiding unnecessary excessive iterations.

[0137] It should be noted that the setting of convergence conditions plays a constraining role in the optimization calculation, enabling the system to achieve a good balance between high efficiency and computational complexity.

[0138] In this embodiment, the convergence condition of the change of the task allocation ratio between nodes is described as follows: In two consecutive iterations, the change in the task allocation ratio f(x, y) of node y needs to be less than the preset threshold ∈ f ,Specifically, the formula for the change in task allocation ratio is expressed as: ;

[0139] Among them, f (k)(x, y) is the ratio of task x assigned to node y in the kth iteration; f (k-1) (x, y) is the ratio of task x assigned to node y in the k-1th iteration; The convergence conditions are: ;

[0140] Among them, ∈ f It is a preset ratio change threshold, usually a positive number less than 0.01, used to control the change accuracy of the task allocation ratio.

[0141] It should be noted that when the distribution ratio changes of all tasks meet the above conditions, it can be considered that the task distribution ratio between nodes has tended to be stable and the optimization process has reached convergence at the local level.

[0142] In this embodiment, the convergence conditions of the optimization objectives of global energy consumption and delay are described as follows: In this embodiment, the convergence conditions of global energy consumption and delay are the key criteria for determining whether the load balancing optimization process is completed. Specifically, global energy consumption refers to the total power consumption of all nodes under the current task allocation scheme, and global delay refers to the total response time of all tasks under the current scheme. The optimization goal of these two indicators is to gradually reduce energy consumption and delay through an iterative process, and stop the optimization when the preset convergence conditions are met.

[0143] The change in global energy consumption refers to the change in the sum of node power consumption between the current iteration and the previous iteration. If the change is less than the preset threshold, the global energy consumption is considered to have reached the optimized convergence state. Similarly, the change in global delay refers to the change in the sum of task response times in two consecutive iterations of the task allocation scheme. If the change is also less than the preset threshold, the delay optimization is considered to have also reached convergence.

[0144] It should be noted that the preset threshold here is a small value, which is usually used to measure the optimization accuracy of global energy consumption and delay. For example, in a high energy efficiency optimization scenario, the threshold can be set lower to improve energy saving, while in a scenario with high real-time requirements, the delay change threshold may need to be more stringent.

[0145] Specifically, the basis for convergence judgment is: if the change in global energy consumption and the change in global delay are both lower than their respective thresholds, it means that the optimization process has been completed, and the iteration can be stopped and the current task allocation plan can be output as the final optimization result.

[0146] In some cases, in order to improve efficiency, the system can also dynamically adjust the threshold based on task priority. For example, in a scenario with more high-priority tasks, the convergence requirements of latency optimization can be appropriately lowered, while in a scenario with a larger proportion of low-priority tasks, more attention can be paid to the convergence effect of energy consumption optimization.

[0147] It should be noted that the convergence conditions of global energy consumption and latency are applicable to scenarios that balance system performance and resource utilization, and can also dynamically adjust the optimization strategy according to different business needs to adapt to a variety of actual application environments. This flexible setting of convergence conditions can ensure the efficiency and target adaptability of the optimization process.

[0148] The optimization control method dynamically adjusts the task allocation ratio by introducing historical data and prediction models to minimize the impact of task migration on node load and improve the convergence speed of global optimization.

[0149] During the task migration process, the task allocation priority of the node is adjusted in real time through the energy-aware dynamic scheduling mechanism. The priority is comprehensively evaluated based on the real-time power consumption of the node, the length of the task queue and the remaining computing power, and tasks are allocated to nodes that meet the low power consumption and load balancing conditions.

[0150] The present invention combines historical data with prediction models by introducing an optimization control method, dynamically adjusts the task allocation ratio, ensures that the impact on node load during task migration is minimized, and significantly improves the convergence speed of global optimization. This implementation method uses an energy-aware dynamic scheduling mechanism to evaluate the task allocation priority of nodes in real time during task migration to achieve the global optimization goals of low power consumption, high efficiency, and load balancing.

[0151] It should be noted that the focus of this step is to dynamically adjust the task allocation priority during the task migration process by using the comprehensive evaluation of the node's real-time power consumption, task queue length, and remaining computing power, so as to ensure that the task allocation can accurately adapt to the dynamic environment.

[0152] In this embodiment, the combination of historical data and prediction model is used to dynamically adjust the task allocation ratio: As a possible implementation method, historical task allocation data includes the task load, queue length and power consumption records of the node. These data are stored in a distributed database to provide support for the input of subsequent prediction models. For example, based on these historical data, regression models (such as linear regression, decision trees) or time series models (such as LSTM, ARIMA) can be used to predict the future load and power consumption of the node.

[0153] It should be noted that the results of the prediction model are used to adjust the task allocation ratio in advance. For example, when it is predicted that the load of a node may increase significantly in the future period, the proportion of tasks allocated to the node can be reduced to avoid node overload.

[0154] In one possible implementation, the dynamic adjustment of the task allocation ratio can also be optimized in combination with a feedback loop. For example, after each task migration, the actual state of the node is compared with the predicted result, and the parameters of the prediction model are updated according to the comparison result to improve the accuracy of the prediction and the adaptability of the allocation scheme.

[0155] In this embodiment, task migration is achieved through an energy-aware dynamic scheduling mechanism: Specifically, the energy-aware dynamic scheduling mechanism comprehensively considers the real-time power consumption of the node, the length of the task queue, and the remaining computing power, and evaluates the priority of task allocation in real time. The priority evaluation principles are as follows: Node power consumption: Monitor the energy consumption level of nodes in real time and prioritize tasks to nodes with lower power consumption to meet the goal of energy efficiency optimization.

[0156] Task queue length: The length of a node's task queue reflects the task waiting time and the degree of node congestion. Tasks are assigned to nodes with shorter task queues first to improve task response speed.

[0157] Remaining computing power: The remaining computing power of a node indicates the amount of its available resources. Tasks are assigned to nodes with higher remaining computing power first to achieve load balancing.

[0158] In a possible implementation, the task allocation priority P(y) can be calculated by the following formula: ;

[0159] Among them, w1, w2, and w3 are weight parameters used to adjust the relative impact of power consumption, queue length, and remaining computing power on priority.

[0160] It should be noted that, through the above priority calculation formula, the task receiving capacity of each node can be dynamically evaluated, so that high-priority nodes are preferentially selected for allocation during task migration.

[0161] Exemplarily, the application of priority adjustment during task migration is as follows: In high-real-time application scenarios, such as online video streaming processing or real-time transaction request processing, the system may pay more attention to the optimization of task latency. In this case, the weight w2 of the queue length parameter can be appropriately increased to prioritize the allocation of tasks to nodes with shorter task queues, ensuring low-latency processing of tasks.

[0162] In another scenario, such as large-scale offline computing tasks, the system may pay more attention to energy consumption optimization. In this case, the weight w1 of the power consumption parameter can be increased to preferentially assign tasks to low-power nodes to meet the requirement of energy efficiency priority.

[0163] Extended technical content: As an extended implementation method, the dynamic scheduling mechanism can also be combined with a distributed load monitoring framework (such as Kubernetes) to achieve cross-cluster task allocation. For example, by monitoring the real-time status and priority evaluation of nodes across regions, the heterogeneity of resources between regions can be fully utilized during task migration, thereby achieving cross-cluster load balancing.

[0164] It should be noted that the cross-region scheduling in the extended technical content needs to be further combined with the network delay model to avoid the additional overhead caused by cross-cluster migration.

[0165] In summary, the present invention realizes dynamic adjustment of task allocation ratio by combining historical data with prediction model, and evaluates the priority of task migration in real time based on energy consumption-aware scheduling mechanism, so as to ensure that the task allocation scheme can adapt to the dynamic environment and achieve the global optimization goal. Those skilled in the art can reproduce the optimization control method and dynamic task allocation mechanism of the present invention in combination with specific application scenarios according to the above content and description.

[0166] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A cluster load balancing processing method based on cloud computing, characterized in that: The following steps are involved: S1. Initialize cluster status: determine the computing power, power consumption parameters and task allocation model of each node in the cloud computing cluster, set optimization goals, and provide basic conditions for subsequent steps; S2, collect node operation status: obtain the task load, task arrival rate and processing capacity of each node in real time, and provide dynamic input data for the construction of load distribution model; S3. Build a load distribution model: Based on the collected node status data, establish a dynamic optimization model for load distribution that comprehensively considers power consumption and task delay, and clarify the global optimization goal; S4. Generate a preliminary task allocation plan: solve the optimization model through game theory methods, determine the preliminary task allocation strategy between nodes, and ensure that the task allocation meets the optimization goal in the current state; S5. According to the dynamic changes of task arrival rate and node status, the initial task allocation plan is adjusted by using the optimization control method to optimize the task allocation ratio; S6. Execute task migration and update status: According to the optimized task allocation plan, implement task migration between nodes, and update the task queue length and processing capacity status of the node in real time to determine whether the global load balancing and energy consumption optimization goals are achieved. If not, return to S2 and repeat the data collection and optimization process.

2. The cluster load balancing processing method based on cloud computing according to claim 1 is characterized in that: In step S2, the collection of the node operation status includes task load, task arrival rate, task completion rate, node remaining computing capacity and dynamic change data of the task queue; the task completion rate is calculated based on the node remaining processing capacity and the current task load.

3. The cluster load balancing processing method based on cloud computing according to claim 1 is characterized in that: In step S3, the load distribution model is constructed in the following manner: The power consumption of a node is dynamically calculated based on the node’s current task load and idle power consumption; The task delay of a node is dynamically calculated based on the remaining processing capacity of the node and the length of the task queue; Construct an optimization function with the goal of global load balancing, taking into account the dynamic changes of node power consumption and task delay.

4. The cluster load balancing processing method based on cloud computing according to claim 1, characterized in that: In step S4, solving the task allocation strategy by game theory method includes: Model the task allocation problem between nodes as a non-cooperative game; A game model is constructed with node power consumption and task delay as utility functions; The variational inequality method is used to solve the Nash equilibrium point of task allocation between nodes; A preliminary task allocation plan is generated based on the Nash equilibrium point to ensure that the task allocation meets the balance between nodes and the global optimal goal.

5. The cluster load balancing processing method based on cloud computing according to claim 1, characterized in that: In step S5, the optimization control method adjusts the task allocation scheme including: Considering the uncertainty of task arrival rate, the preliminary allocation plan is dynamically adjusted through stochastic optimization method; Optimize the ratio of task migration according to the real-time load and processing capacity of the node; Dynamically adjust the weight coefficients of power consumption and latency to meet the priority requirements of power consumption optimization or performance optimization in different scenarios.

6. The cluster load balancing processing method based on cloud computing according to claim 1, characterized in that: In step S6, the execution of task migration includes: According to the optimized task allocation plan, tasks are migrated from high-load nodes to low-load nodes; Update the node's task queue length and remaining computing power in real time; Record the historical data of node task migration to provide input for subsequent allocation strategy optimization.

7. The cluster load balancing processing method based on cloud computing according to claim 1, characterized in that: In step S6, the node status after task migration is fed back to step S2. By monitoring the change of the node queue length, the execution effect of the task allocation scheme is judged and the optimization model is rebuilt.

8. The cluster load balancing processing method based on cloud computing according to claim 1, characterized in that: The convergence conditions for load balancing are: The change in the task allocation ratio between nodes in two consecutive iterations is less than the preset threshold; The optimization targets of global energy consumption and latency both meet the preset requirements.

9. The cluster load balancing processing method based on cloud computing according to claim 1, characterized in that: In step S5, the optimization control method dynamically adjusts the task allocation ratio by introducing historical data and prediction models to minimize the impact of task migration on node load and improve the convergence speed of global optimization.

10. The cluster load balancing processing method based on cloud computing according to claim 9, characterized in that: During the task migration process, the task allocation priority of the node is adjusted in real time through the energy-aware dynamic scheduling mechanism. The priority is comprehensively evaluated based on the real-time power consumption of the node, the length of the task queue and the remaining computing power, and tasks are allocated to nodes that meet the low power consumption and load balancing conditions.

Images