A multi-algorithmic computing resource allocation and cost optimization system and method
By constructing a dynamic perception and intelligent decision-making layer, and combining digital twin models and reinforcement learning algorithms, the problems of rigid resource allocation and multi-objective optimization in existing technologies are solved, achieving efficient resource utilization and cost optimization, and improving the overall performance and economic benefits of the system.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- INSPUR SOFTWARE TECH CO LTD
- Filing Date
- 2026-01-15
- Publication Date
- 2026-05-29
AI Technical Summary
Existing technologies suffer from rigid static strategies in resource allocation and management, are unable to cope with heterogeneous computing power, and cannot comprehensively optimize multiple objectives such as performance, cost, and energy consumption, resulting in unreasonable resource allocation and high costs.
By constructing a dynamic perception layer and an intelligent decision-making layer, and combining a digital twin model and a multi-objective optimization decision-making model with reinforcement learning algorithms, we can achieve real-time data acquisition and intelligent resource allocation, generate optimal decisions, and execute them.
It achieves efficient resource utilization and cost optimization, improves resource utilization and return on investment, reduces operation and maintenance complexity, and achieves global optimal multi-objective optimization.
Smart Images
Figure CN122111593A_ABST
Abstract
Description
Technical Field
[0001] This invention discloses a multi-source computing power resource allocation and cost optimization system and method, which relates to the fields of cloud computing and computing power scheduling technology. Background Technology
[0002] With the deepening of digital transformation, enterprises' computing needs are experiencing explosive growth, and the types of computing workloads are becoming increasingly complex, ranging from existing data processing to AI model training, scientific computing, and real-time rendering. At the same time, computing infrastructure is also becoming more diversified, encompassing local private clouds and multiple public clouds such as AWS, Azure, and Alibaba Cloud. These computing resources differ significantly in architecture, type, performance, and cost.
[0003] Currently, mainstream resource allocation and management technologies have the following limitations: 1. Static Allocation Strategy: Most systems use static strategies or automated tools based on simple rules, such as Kubernetes' default scheduler, for resource allocation. This approach cannot detect real-time performance fluctuations and market price changes of underlying resources, leading to rigid resource allocation and making it difficult to achieve global optimization.
[0004] 2. Inability to handle heterogeneity: Existing solutions do not adequately support heterogeneous computing power such as CPU, GPU, NPU, etc., lack unified scheduling capabilities, and are unable to intelligently match specific types of tasks to the most suitable dedicated hardware.
[0005] 3. Limited decision-making dimensions: It is impossible to simultaneously weigh multiple objectives such as performance, cost, and energy consumption within the same decision-making framework, thus failing to maximize overall benefits. Summary of the Invention
[0006] This invention addresses the problems of existing technologies by providing a system and method for allocating and optimizing the cost of diverse computing resources. It can dynamically sense and intelligently allocate computing resources to achieve efficient utilization of diverse computing resources and refined cost control.
[0007] The specific solution proposed in this invention is as follows: This invention also provides a method for allocating and optimizing the cost of diverse computing resources, including: Step 1: Construct a dynamic perception layer to collect real-time data from the task and resource ends. After data fusion, obtain the performance indicators, price, and static attributes corresponding to each computing resource node. Retrieve custom annotations for computation tasks and extract task feature information, including task computation characteristics, resource preferences, elasticity and fault tolerance requirements, performance and service level agreements. Collect resource metrics data for each computing resource node, including CPU, GPU, memory, disk, and network resource metrics data. The API is periodically called to obtain resource prices based on resource metric data and update historical price fluctuations and interruption rates. Step 2: Construct an intelligent decision-making layer. Utilize a digital twin model to simulate the real environment. When making decisions for computational tasks, employ a multi-objective optimization decision-making model to generate the optimal decision and corresponding instructions. The engine retrieves snapshots of the current task and resource status, generates one or more candidate strategies, and applies them to the digital twin model. The comprehensive reward function R of the multi-objective optimization decision model is constructed based on the task's problem. R=W1*(-Cost)+W2*(-ExecutionTime)+W3(-SLA_Penalty), Cost represents the financial cost incurred in executing the task, ExecutionTime represents the task completion time, SLA_Penalty represents the service level agreement penalty, and w1, w2, and w3 are configurable weight coefficients set according to business preferences; the decision generation is guided by maximizing the comprehensive reward function R. Based on reinforcement learning algorithms, simulations are performed on digital twin models to generate an optimal resource allocation scheme that maximizes the achievement of optimization goals while meeting task performance requirements. Step 3: Build the execution layer to execute the instructions of the intelligent decision-making layer and provide feedback on the execution results.
[0008] Furthermore, the computational features extracted in step 1 of the aforementioned multi-dimensional computing power resource allocation and cost optimization method include matching computing power type (task-type), GPU architecture (gpu-type), and CPU architecture (cpu-type). The extracted resilience and fault tolerance requirements include whether interruption is allowed and the maximum number of restarts (max-restart-count). The extracted performance and service level agreement includes the expected duration and the maximum acceptable latency.
[0009] Furthermore, in step 1 of the aforementioned method for allocating and optimizing the cost of multi-source computing power, when obtaining resource prices and updating historical price fluctuations and interruption rates, the real-time status of the local resource pool and the remote cloud resource pool is maintained, and tasks are prioritized for scheduling to available local resources; when local resources are insufficient or the cost is uneconomical, remote cloud resources are dynamically applied for, and the optimal cloud service provider and instance type are selected based on real-time price and performance data.
[0010] Furthermore, in step 2 of the aforementioned method for allocating and optimizing the cost of multi-dimensional computing resources, the decision result of the optimal decision generated by the digital twin model is rendered into a standard Kubernetes Job YAML or other resource definition file, wherein the scheduling constraints specifically include container images, resource requests and limits, node selectors, and tolerances, to ensure that computing tasks can be accurately deployed to the target computing resource nodes.
[0011] Furthermore, step 3 of the aforementioned method for allocating and optimizing the cost of diverse computing resources specifically includes: Upon receiving instructions from the intelligent decision-making layer, the computing tasks are deployed to the target computing resource nodes using extended container orchestration tools. At the same time, the elastic controller listens for instructions from the intelligent decision-making layer to achieve scaling up and down of the computing tasks. After execution, the time and cost consumed are recorded and used as new data points for feedback. The predicted value is compared with the actual value. If the deviation exceeds the threshold, the parameters of the digital twin model and the reinforcement learning model are adjusted to form an optimization closed loop.
[0012] This invention also provides a multi-dimensional computing resource allocation and cost optimization system, including a task awareness module, a resource awareness module, a market awareness module, a digital twin module, a multi-objective optimization decision-making module, a strategy generator module, and an execution module. The system constructs a dynamic perception layer, which collects real-time data from the task end and the resource end. After data fusion, it obtains the performance indicators, prices, and static attributes corresponding to each computing resource node. The task awareness module acquires custom annotations for computation tasks and extracts task feature information, including the task's computational characteristics, resource preferences, elasticity and fault tolerance requirements, performance and service level agreements. The resource awareness module collects resource indicator data for each computing resource node, including CPU, GPU, memory, disk, and network resource indicator data. The market perception module periodically calls the API to obtain resource prices based on resource indicator data and updates historical price fluctuations and interruption rates. The system constructs an intelligent decision-making layer. The digital twin module uses a digital twin model to simulate the real environment. When making decisions for computational tasks, the multi-objective optimization decision-making module uses a multi-objective optimization decision-making model to generate the optimal decision and corresponding instructions. The digital twin module retrieves snapshots of the current task and resource status through the engine, generates one or more candidate strategies, and applies them to the digital twin model. The multi-objective optimization decision module constructs a comprehensive reward function R for the multi-objective optimization decision model based on the task's problem. R=W1*(-Cost)+W2*(-ExecutionTime)+W3(-SLA_Penalty), Cost represents the financial cost incurred in executing the task, ExecutionTime represents the task completion time, SLA_Penalty represents the service level agreement penalty, and w1, w2, and w3 are configurable weight coefficients set according to business preferences; the decision generation is guided by maximizing the comprehensive reward function R. The multi-objective optimization decision module is based on reinforcement learning algorithms and performs simulations on a digital twin model to generate an optimal resource allocation scheme that maximizes the optimization objectives while meeting task performance requirements. The system constructs an execution layer, where the execution module executes the instructions of the intelligent decision-making layer and returns the execution results.
[0013] Furthermore, the computational features extracted by the resource awareness module of the aforementioned multi-dimensional computing power resource allocation and cost optimization system include matching computing power type (task-type), GPU architecture (gpu-type), and CPU architecture (cpu-type). The extracted resilience and fault tolerance requirements include whether interruption is allowed and the maximum number of restarts (max-restart-count). The extracted performance and service level agreement includes the expected duration and the maximum acceptable latency.
[0014] Furthermore, when the market perception module of the aforementioned multi-dimensional computing power resource allocation and cost optimization system obtains resource prices and updates historical price fluctuations and interruption rates, it maintains the real-time status of the local resource pool and the remote cloud resource pool, and prioritizes scheduling tasks to available local resources; when local resources are insufficient or the cost is uneconomical, it dynamically applies for remote cloud resources and selects the optimal cloud service provider and instance type based on real-time price and performance data.
[0015] Furthermore, the strategy generator module of the multi-dimensional computing resource allocation and cost optimization system renders the decision results of the optimal decision generated by the digital twin model into a standard Kubernetes Job YAML or other resource definition file, wherein the scheduling constraints specifically include container images, resource requests and limits, node selectors, and tolerances, to ensure that computing tasks can be accurately deployed to target computing resource nodes.
[0016] Furthermore, after receiving instructions from the intelligent decision layer, the execution module of the multi-dimensional computing power resource allocation and cost optimization system deploys computing tasks to target computing power resource nodes through extended container orchestration tools, and at the same time uses an elastic controller to listen for instructions from the intelligent decision layer to realize the scaling up and down of computing tasks. After execution, the time and cost consumed are recorded and used as new data points for feedback. The predicted value is compared with the actual value. If the deviation exceeds the threshold, the parameters of the digital twin model and the reinforcement learning model are adjusted to form an optimization closed loop.
[0017] The advantages of this invention are: 1. Dynamic and precise perception: It breaks through the limitations of static configuration and provides accurate and comprehensive input for intelligent decision-making through real-time and multi-dimensional data collection.
[0018] 2. Intelligent multi-objective optimization: By introducing digital twins and AI decision-making models, it can simultaneously optimize multiple conflicting objectives such as performance, cost, and energy efficiency to achieve global optimization rather than local optimization.
[0019] 3. Significant cost optimization: By intelligently matching tasks with the most economical computing resources and dynamically utilizing low-cost options such as auction instances, the total cost of ownership (TCO) is effectively reduced.
[0020] 4. Improved resource utilization: It avoids resource idleness and misallocation, enabling expensive dedicated computing power to be used more efficiently, thus improving the return on investment of the entire infrastructure.
[0021] 5. Automation and Adaptability: The system achieves full-process automation from perception and decision-making to execution. It can adapt to changes in workload and resource environment, greatly reducing the complexity of operation and maintenance. Attached Figure Description
[0022] Figure 1 This is a schematic diagram of the overall architecture of the present invention.
[0023] Figure 2 This is a flowchart of the workflow of the present invention. Detailed Implementation
[0024] The present invention will be further described below with reference to the accompanying drawings and specific embodiments, so that those skilled in the art can better understand and implement the present invention. However, the embodiments described are not intended to limit the present invention. Example
[0025] This invention also provides a method for allocating and optimizing the cost of diverse computing resources, including: Step 1: Construct a dynamic perception layer to collect real-time data from the task and resource ends. After data fusion, obtain the performance indicators, price, and static attributes corresponding to each computing resource node. Retrieve custom annotations for computing tasks and extract task feature information, including task computing characteristics, resource preferences, elasticity and fault tolerance requirements, performance and service level agreements. The computing characteristics include matching computing power type (task-type), GPU architecture (gpu-type), and CPU architecture (cpu-type). The extracted resilience and fault tolerance requirements include whether interruption is allowed and the maximum number of restarts (max-restart-count). The extracted performance and service level agreement includes the expected duration and the maximum acceptable latency.
[0026] Collect resource metrics data for each computing resource node, including CPU, GPU, memory, disk, and network resource metrics data.
[0027] Regularly call the API to obtain resource prices based on resource metric data and update historical price fluctuations and interruption rates. Maintain the real-time status of the local resource pool and the remote cloud resource pool, and prioritize scheduling tasks to available local resources. When local resources are insufficient or the cost is uneconomical, dynamically apply for remote cloud resources and select the optimal cloud service provider and instance type based on real-time price and performance data.
[0028] Step 2: Construct an intelligent decision-making layer. Utilize a digital twin model to simulate the real environment. When making decisions for computational tasks, employ a multi-objective optimization decision-making model to generate the optimal decision and corresponding instructions. The engine retrieves snapshots of the current task and resource status, generates one or more candidate strategies, and applies them to the digital twin model. The comprehensive reward function R of the multi-objective optimization decision model is constructed based on the task's problem. R=W1*(-Cost)+W2*(-ExecutionTime)+W3(-SLA_Penalty), Cost represents the financial cost of task execution, ExecutionTime represents the task completion time, SLA_Penalty represents the service level agreement penalty, and w1, w2, and w3 are configurable weight coefficients set according to business preferences; the decision generation is guided by maximizing the comprehensive reward function R.
[0029] Based on reinforcement learning algorithms, simulations are performed on a digital twin model to generate an optimal resource allocation scheme that maximizes the optimization objective while meeting task performance requirements. The optimal decision is obtained through reinforcement learning algorithms, employing Actor-Critic methods such as Proximal Policy Optimization (PPO) or Deep Deterministic Policy Gradient (DDPG). The Actor network takes the current state S as input and outputs a probability distribution representing the likelihood of choosing each action A. The Critic network evaluates the value of taking action A in state S, i.e., the expected cumulative reward. The model is trained using historical data or online simulations to make decisions with optimal long-term returns in complex environments.
[0030] The decision results generated by the digital twin model are rendered into standard Kubernetes Job YAML or other resource definition files, where scheduling constraints specifically include container images, resource requests and limits, node selectors, and tolerances, to ensure that computing tasks can be accurately deployed to target computing resource nodes.
[0031] Step 3: Construct the execution layer to execute the instructions of the intelligent decision-making layer and provide feedback on the execution results. Specifically, this includes: Upon receiving instructions from the intelligent decision-making layer, the computing tasks are deployed to the target computing resource nodes using extended container orchestration tools. At the same time, the elastic controller listens for instructions from the intelligent decision-making layer to achieve scaling up and down of the computing tasks. After execution, the time and cost consumed are recorded and used as new data points for feedback. The predicted value is compared with the actual value. If the deviation exceeds the threshold, the parameters of the digital twin model and the reinforcement learning model are adjusted to form an optimization closed loop. Example
[0032] This invention also provides a multi-dimensional computing resource allocation and cost optimization system, including a task awareness module, a resource awareness module, a market awareness module, a digital twin module, a multi-objective optimization decision-making module, a strategy generator module, and an execution module. The system constructs a dynamic perception layer, which collects real-time data from the task end and the resource end. After data fusion, it obtains the performance indicators, prices, and static attributes corresponding to each computing resource node. The task awareness module acquires custom annotations for computation tasks and extracts task feature information, including the task's computational characteristics, resource preferences, elasticity and fault tolerance requirements, performance and service level agreements. The resource awareness module collects resource indicator data for each computing resource node, including CPU, GPU, memory, disk, and network resource indicator data. The market perception module periodically calls the API to obtain resource prices based on resource indicator data and updates historical price fluctuations and interruption rates. The system constructs an intelligent decision-making layer. The digital twin module uses a digital twin model to simulate the real environment. When making decisions for computational tasks, the multi-objective optimization decision-making module uses a multi-objective optimization decision-making model to generate the optimal decision and corresponding instructions. The digital twin module retrieves snapshots of the current task and resource status through the engine, generates one or more candidate strategies, and applies them to the digital twin model. The multi-objective optimization decision module constructs a comprehensive reward function R for the multi-objective optimization decision model based on the task's problem. R=W1*(-Cost)+W2*(-ExecutionTime)+W3(-SLA_Penalty), Cost represents the financial cost incurred in executing the task, ExecutionTime represents the task completion time, SLA_Penalty represents the service level agreement penalty, and w1, w2, and w3 are configurable weight coefficients set according to business preferences; the decision generation is guided by maximizing the comprehensive reward function R. The multi-objective optimization decision module is based on reinforcement learning algorithms and performs simulations on a digital twin model to generate an optimal resource allocation scheme that maximizes the optimization objectives while meeting task performance requirements. The system constructs an execution layer, where the execution module executes the instructions of the intelligent decision-making layer and returns the execution results.
[0033] The information interaction and execution process between the modules in the above system are based on the same concept as the method embodiment of the present invention, and the specific details can be found in the description in the method embodiment of the present invention, and will not be repeated here.
[0034] Similarly, the advantages of the system of the present invention are: 1. Dynamic and precise perception: It breaks through the limitations of static configuration and provides accurate and comprehensive input for intelligent decision-making through real-time and multi-dimensional data collection.
[0035] 2. Intelligent multi-objective optimization: By introducing digital twins and AI decision-making models, it can simultaneously optimize multiple conflicting objectives such as performance, cost, and energy efficiency to achieve global optimization rather than local optimization.
[0036] 3. Significant cost optimization: By intelligently matching tasks with the most economical computing resources and dynamically utilizing low-cost options such as auction instances, the total cost of ownership (TCO) is effectively reduced.
[0037] 4. Improved resource utilization: It avoids resource idleness and misallocation, enabling expensive dedicated computing power to be used more efficiently, thus improving the return on investment of the entire infrastructure.
[0038] 5. Automation and Adaptability: The system achieves full-process automation from perception and decision-making to execution. It can adapt to changes in workload and resource environment, greatly reducing the complexity of operation and maintenance.
[0039] It should be noted that not all steps and modules in the above processes and system structures are mandatory; some steps or modules can be omitted as needed. The execution order of each step is not fixed and can be adjusted as required. The system structures described in the above embodiments can be physical or logical structures. That is, some modules may be implemented by the same physical entity, or some modules may be implemented by multiple physical entities, or they may be jointly implemented by certain components in multiple independent devices.
[0040] The above-described embodiments are merely preferred embodiments provided to fully illustrate the present invention, and the scope of protection of the present invention is not limited thereto. Equivalent substitutions or modifications made by those skilled in the art based on the present invention are all within the scope of protection of the present invention. The scope of protection of the present invention is defined by the claims.
Claims
1. A method for multi-element computing power resource allocation and cost optimization, characterized in that include: Step 1: Construct a dynamic perception layer to collect real-time data from the task and resource ends. After data fusion, obtain the performance indicators, price, and static attributes corresponding to each computing resource node. Retrieve custom annotations for computation tasks and extract task feature information, including task computation characteristics, resource preferences, elasticity and fault tolerance requirements, performance and service level agreements. Collect resource metrics data for each computing resource node, including CPU, GPU, memory, disk, and network resource metrics data. The API is periodically called to obtain resource prices based on resource metric data and update historical price fluctuations and interruption rates. Step 2: Construct an intelligent decision-making layer. Utilize a digital twin model to simulate the real environment. When making decisions for computational tasks, employ a multi-objective optimization decision-making model to generate the optimal decision and corresponding instructions. The engine retrieves snapshots of the current task and resource status, generates one or more candidate strategies, and applies them to the digital twin model. The comprehensive reward function R of the multi-objective optimization decision model is constructed based on the task's problem. R=W1*(-Cost)+W2*(-ExecutionTime)+W3(-SLA_Penalty), Cost represents the financial cost incurred in executing the task, ExecutionTime represents the task completion time, SLA_Penalty represents the service level agreement penalty, and w1, w2, and w3 are configurable weight coefficients set according to business preferences; the decision generation is guided by maximizing the comprehensive reward function R. Based on reinforcement learning algorithms, simulations are performed on digital twin models to generate an optimal resource allocation scheme that maximizes the achievement of optimization goals while meeting task performance requirements. Step 3: Build the execution layer to execute the instructions of the intelligent decision-making layer and provide feedback on the execution results.
2. The method for multi-source computing power allocation and cost optimization according to claim 1, characterized in that: The computational features extracted in step 1 include matching computational power type (task-type), GPU architecture (gpu-type), and CPU architecture (cpu-type). The extracted resilience and fault tolerance requirements include whether interruption is allowed and the maximum number of restarts (max-restart-count). The extracted performance and service level agreement includes the expected duration and the maximum acceptable latency.
3. The method for allocating and optimizing the cost of multi-source computing power as described in claim 1 is characterized in that, in step 1, when obtaining resource prices and updating historical price fluctuations and interruption rates, the real-time status of the local resource pool and the remote cloud resource pool is maintained, and tasks are prioritized for scheduling to available local resources; when local resources are insufficient or the cost is uneconomical, remote cloud resources are dynamically applied for, and the optimal cloud service provider and instance type are selected based on real-time price and performance data.
4. The method for allocating and optimizing the cost of multi-dimensional computing resources according to claim 1, characterized in that the decision result of the optimal decision generated by the digital twin model in step 2 is rendered into a standard Kubernetes Job YAML or other resource definition file, wherein the scheduling constraints specifically include container images, resource requests and limits, node selectors, and tolerances, to ensure that computing tasks can be accurately deployed to the target computing resource nodes.
5. The method for multi-source computing power allocation and cost optimization according to claim 1, characterized in that: Step 3 specifically includes: Upon receiving instructions from the intelligent decision-making layer, the computing tasks are deployed to the target computing resource nodes using extended container orchestration tools. At the same time, the elastic controller listens for instructions from the intelligent decision-making layer to achieve scaling up and down of the computing tasks. After execution, the time and cost consumed are recorded and used as new data points for feedback. The predicted value is compared with the actual value. If the deviation exceeds the threshold, the parameters of the digital twin model and the reinforcement learning model are adjusted to form an optimization closed loop.
6. A multi-source computing power allocation and cost optimization system, characterized by: It includes a task awareness module, a resource awareness module, a market awareness module, a digital twin module, a multi-objective optimization decision-making module, a strategy generator module, and an execution module. The system constructs a dynamic perception layer, which collects real-time data from the task end and the resource end. After data fusion, it obtains the performance indicators, prices, and static attributes corresponding to each computing resource node. The task awareness module acquires custom annotations for computation tasks and extracts task feature information, including the task's computational characteristics, resource preferences, elasticity and fault tolerance requirements, performance and service level agreements. The resource awareness module collects resource indicator data for each computing resource node, including CPU, GPU, memory, disk, and network resource indicator data. The market perception module periodically calls the API to obtain resource prices based on resource indicator data and updates historical price fluctuations and interruption rates. The system constructs an intelligent decision-making layer. The digital twin module uses a digital twin model to simulate the real environment. When making decisions for computational tasks, the multi-objective optimization decision-making module uses a multi-objective optimization decision-making model to generate the optimal decision and corresponding instructions. The digital twin module retrieves snapshots of the current task and resource status through the engine, generates one or more candidate strategies, and applies them to the digital twin model. The multi-objective optimization decision module constructs a comprehensive reward function R for the multi-objective optimization decision model based on the task's problem. R=W1*(-Cost)+W2*(-ExecutionTime)+W3(-SLA_Penalty), Cost represents the financial cost incurred in executing the task, ExecutionTime represents the task completion time, SLA_Penalty represents the service level agreement penalty, and w1, w2, and w3 are configurable weight coefficients set according to business preferences; the decision generation is guided by maximizing the comprehensive reward function R. The multi-objective optimization decision module is based on reinforcement learning algorithms and performs simulations on a digital twin model to generate an optimal resource allocation scheme that maximizes the optimization objectives while meeting task performance requirements. The system constructs an execution layer, where the execution module executes the instructions of the intelligent decision-making layer and returns the execution results.
7. A multi-source computing power resource allocation and cost optimization system according to claim 6, characterized in that: The computational features extracted by the resource awareness module include matching computational power type (task-type), GPU architecture (gpu-type), and CPU architecture (cpu-type). The extracted resilience and fault tolerance requirements include whether interruption is allowed and the maximum number of restarts (max-restart-count). The extracted performance and service level agreement includes the expected duration and the maximum acceptable latency.
8. A multi-source computing power resource allocation and cost optimization system according to claim 6, characterized in that: When the market perception module obtains resource prices and updates historical price fluctuations and interruption rates, it maintains the real-time status of the local resource pool and the remote cloud resource pool, and prioritizes scheduling tasks to available local resources. When local resources are insufficient or the cost is uneconomical, it dynamically requests remote cloud resources and selects the optimal cloud service provider and instance type based on real-time price and performance data.
9. A multi-source computing power resource allocation and cost optimization system according to claim 6, characterized in that: The strategy generator module renders the optimal decision generated by the digital twin model into a standard Kubernetes Job YAML or other resource definition file. The scheduling constraints specifically include container images, resource requests and limits, node selectors, and tolerances, ensuring that computing tasks can be accurately deployed to the target computing resource nodes.
10. A multi-source computing power resource allocation and cost optimization system according to claim 6, characterized in that: After receiving instructions from the intelligent decision layer, the execution module deploys computing tasks to the target computing resource nodes using extended container orchestration tools, and simultaneously uses an elastic controller to listen for instructions from the intelligent decision layer to achieve scaling up and down of computing tasks. After execution, the time and cost consumed are recorded and used as new data points for feedback. The predicted value is compared with the actual value. If the deviation exceeds the threshold, the parameters of the digital twin model and the reinforcement learning model are adjusted to form an optimization closed loop.