A cloud computing resource scheduling supervision processing system

Abstract

The application discloses a cloud computing resource scheduling supervision processing system and belongs to the technical field of data analysis; an attention mechanism LSTM can effectively reduce SLA demand prediction error by focusing on key time step features, and can provide accurate input for subsequent scheduling; a blockchain smart contract can realize the transparency and real-time updating of weight coefficients, and can avoid scheduling bias caused by SLA clause conflicts between tenants; potential conflicts can be identified in advance through the confidence interval definition of an SLA demand matrix; through the application of a constructed detection model, dynamic perception and priority quantization of multi-tenant SLA demand can be realized; through end-to-end learning of a resource allocation intelligent agent trained by a deep deterministic policy gradient algorithm, the conflict between real-time performance and cost can be resolved; a penalty term of a resource waste rate in a reward function can promote the intelligent agent to dynamically adjust the cloud virtual machine scaling strategy.

Description

Technical Field

[0001] This invention relates to the field of data analysis technology, specifically to a cloud computing resource scheduling and monitoring system. Background Technology

[0002] Cloud computing resource scheduling, data processing and analysis refers to the rational allocation and management of computing, storage, network and other resources in a cloud computing environment, as well as the collection, processing and analysis of the generated operational data or user business data, in order to optimize system performance, improve resource utilization, ensure service quality and support intelligent decision-making.

[0003] Cloud computing needs to simultaneously meet the differentiated Service Level Agreements (SLAs) for multiple tenants, including task response time, data processing accuracy, and resource availability. In contrast, SLAs in other domains often focus on a single metric, such as data processing accuracy, which has a low risk of conflict. There are inherent conflicts between different SLA metrics. For example, one tenant may require "data processing latency ≤ 1 second," corresponding to a real-time SLA, while another tenant may require "resource utilization ≥ 80%," corresponding to a cost SLA. If the scheduling algorithm cannot dynamically balance multi-dimensional constraints, it may over-allocate resources to meet real-time requirements, violating the cost SLA; or it may compress computing nodes to save resources, resulting in failure to meet real-time requirements. This leads to scheduling conflicts under multi-dimensional service level agreement constraints. Summary of the Invention

[0004] The purpose of this invention is to provide a cloud computing resource scheduling and monitoring system to solve the technical problem of scheduling conflicts under the constraints of multi-dimensional service level agreements in existing technical solutions.

[0005] The objective of this invention can be achieved through the following technical solutions:

[0006] A cloud computing resource scheduling and monitoring system includes:

[0007] The priority-based data analysis and processing module constructs a multi-tenant SLA dynamic prediction and priority-based model. Through the AI ​​time series prediction module, it makes short-term predictions of the tenant's task response time requirements, data processing accuracy threshold, and resource fluctuation characteristics for the next few minutes, generating a multi-dimensional SLA requirement matrix that includes real-time constraints, accuracy constraints, and availability constraints. Based on blockchain smart contracts, it records the weight coefficients of each tenant's SLA terms, forming an immutable basis for priority-based stratification.

[0008] The dynamic balancing application closed-loop monitoring module designs an edge-cloud collaborative reinforcement learning scheduler based on the output SLA demand matrix and priority weights. It trains the resource allocation agent through a deep deterministic policy gradient algorithm, taking the edge node computing power, cloud remaining resources, and real-time SLA compliance rate as state inputs. It dynamically outputs the edge offloading ratio and the cloud virtual machine elastic scaling policy. The monitoring obtains the positive and negative feedback values ​​corresponding to the dynamic balancing effect executed by the resource allocation agent, and uses the positive and negative feedback values ​​to dynamically trigger the conventional policy iterative optimization mechanism.

[0009] Preferably, historical task data of tenants is collected to construct the corresponding input feature matrix. Where N is the number of tenants; T is the time window length; and F is the feature dimension, which includes task attributes, resource status, and environmental interference.

[0010] Preferably, the AI ​​time series prediction module constructs a prediction model based on an LSTM network with an attention mechanism. ;in, To predict response time; To predict the accuracy of data processing; To predict resource availability.

[0011] Preferably, based on the prediction results output by the prediction model Generate the SLA demand matrix for tenant j. : ; where the superscripts min and max are the lower and upper bounds of the 95% confidence interval for the corresponding predicted value, respectively.

[0012] Preferably, a base weighting coefficient is defined based on the historical SLA compliance rate. Adjusting the base weight coefficients involves the following expression:

[0013] ;in, This represents the base weighting coefficient of tenant j under the SLA metric k; This represents the updated dynamic weight coefficient; k is the index of the indicator. , These are real-time performance metrics, accuracy metrics, and resource availability metrics; This is a correction factor.

[0014] Preferably, when designing an edge-cloud collaborative reinforcement learning scheduler, a state space S and an action space A are defined; the state space S serves as the input feature vector of the resource allocation agent; the action space is... ;in, This represents the edge unloading ratio. ; To scale the number of virtual machines in the cloud. .

[0015] Preferably, a basic reward is constructed based on the hard constraint compliance rate, soft constraint compliance rate, and resource waste rate. The reward coefficient is dynamically adjusted using priority weights in smart contracts to achieve tenant priority differentiation. The relevant expression is as follows: ;in, This is the adjusted reward coefficient; Basic reward coefficient; Priority weights in smart contracts; This is the correction factor for SLA compliance status.

[0016] Preferably, historical data is used to pre-train the resource allocation agent, the policy network is optimized through multiple rounds of iteration, and it is deployed to the edge-cloud collaborative environment. The trained policy is then solidified into a cloud virtual machine elastic scaling policy and integrated with edge nodes and cloud resource managers through API interfaces.

[0017] Preferably, the positive feedback value and negative feedback value of the resource allocation agent are calculated and obtained respectively. If the positive feedback value is greater than 0 and the negative feedback value is less than 1, it is determined that the dynamic balance of the existing scheme is normal and the implementation of the regular strategy iteration is maintained.

[0018] Conversely, if the dynamic balance of the existing scheme is abnormal, the conventional strategy iterative optimization mechanism will be triggered.

[0019] Preferably, the calculation expressions for the positive feedback value and the negative feedback value are as follows:

[0020] ;in, It is a positive feedback value; This is the regulatory value for the edge unloading ratio; This is the standard value for the edge unloading ratio;

[0021] ;in, This represents the negative feedback value; NL represents the total number of resources with a non-zero waste rate. It is the sum of the total number of resources with a non-zero waste rate and the total number of resources with a zero waste rate; The allowable proportion of resource waste, with a value range of (0,1).

[0022] Compared to existing solutions, the beneficial effects achieved by this invention are:

[0023] The attention mechanism LSTM in this invention can effectively reduce SLA demand prediction errors by focusing on key time step features, providing accurate input for subsequent scheduling. The blockchain smart contract enables transparency and real-time updates of weight coefficients, avoiding scheduling bias caused by SLA term conflicts among tenants. By defining the confidence interval of the SLA demand matrix, potential conflicts can be identified in advance. Through the application-built detection model, dynamic perception and priority quantification of multi-tenant SLA demands can be achieved, thereby providing accurate constraint inputs and conflict coordination basis for subsequent resource scheduling algorithms.

[0024] This invention employs an end-to-end learning approach to train resource allocation agents using a deep deterministic policy gradient algorithm, resolving the conflict between real-time performance and cost. While ensuring low latency for high-priority tenants, it effectively improves resource utilization. The immutable weights recorded in smart contracts ensure fairness in rewards and penalties, preventing scheduling bias due to priority disputes among tenants and significantly improving audit traceability efficiency. The penalty term for resource waste rate in the reward function prompts the agent to dynamically adjust cloud virtual machine scaling strategies, effectively reducing peak resource waste rate compared to traditional solutions. By monitoring and evaluating the dynamic balancing effect executed by the resource allocation agent, and dynamically triggering a conventional strategy iteration optimization mechanism using the evaluation results, closed-loop monitoring of the dynamic balancing scheme is achieved, further enhancing the reliability and robustness of the dynamic balancing scheme implementation. Attached Figure Description

[0025] The invention will now be further described with reference to the accompanying drawings.

[0026] Figure 1 This is a flowchart illustrating the operation of a cloud computing resource scheduling and monitoring system according to the present invention. Detailed Implementation

[0027] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0028] like Figure 1 As shown, the present invention is a cloud computing resource scheduling and monitoring system, comprising:

[0029] The priority-based data analysis and processing module constructs a multi-tenant SLA dynamic prediction and priority-based model. Through an AI time-series prediction module, it performs short-term predictions (within minutes) of tenant task response time requirements, data processing accuracy thresholds, and resource fluctuation characteristics, generating a multi-dimensional SLA requirement matrix that includes real-time constraints, accuracy constraints, and availability constraints. Based on a blockchain smart contract, it records the weight coefficients of each tenant's SLA terms, forming an immutable basis for priority layering. Specific steps include:

[0030] Collect historical task data from tenants and construct the corresponding input feature matrix. Where N is the number of tenants; T is the time window length; and F is the feature dimension, which includes task attributes, resource status, and environmental interference.

[0031] Specifically, task attributes include data volume (MB), computational complexity (FLOPS), and historical response time (seconds).

[0032] Resource status, including CPU utilization (%), memory usage (%), and network bandwidth (Mbps);

[0033] Environmental interference, including the frequency of sudden requests (times / minute) and node failure rate (‰).

[0034] The AI ​​time series prediction module builds a prediction model based on an attention-based LSTM network. ;in, The predicted response time is in seconds, corresponding to the real-time SLA; The accuracy of the predicted data processing is expressed in %, corresponding to the accuracy SLA. To predict resource availability, we need to define the corresponding availability SLA; SLA stands for Service Level Agreement.

[0035] The network structure of the prediction model includes an embedding layer, an LSTM layer, an attention layer, and a prediction layer;

[0036] The embedding layer, in particular, embeds discrete features through an embedding matrix. Convert to a low-dimensional vector; discrete features, such as tenant ID;

[0037] The LSTM layer consists of two bidirectional LSTM layers with a hidden layer dimension of 64. The forget gate bias is initialized to 1.0 to mitigate gradient vanishing. The output sequence is... ;

[0038] Attention layer, calculates the attention weights at time step i. The expression involved is:

[0039] Where W is the linear transformation weight matrix; Let be the hidden state vector at time step t; b be the bias vector; tanh() be the hyperbolic tangent activation function; and w be the attention weight vector. This is the inner product operation; exp() is the exponential function. Let be the hidden state vector at time step i; i is the index of the time step, i=1,2,...,T; T is the length of the time window; This represents the summation of the exponential attention scores over all time steps. The formula dynamically calculates the weight of each feature at each time step. Its core function is to allow the model to automatically focus on time step information that is more critical to the current task, such as resource fluctuation characteristics at the moment of sudden request, thereby improving the accuracy of sequence prediction or decision-making.

[0040] The prediction layer outputs a 3D prediction result through a fully connected layer, with the activation functions being: The corresponding ReLU function ensures non-negativity; The corresponding Sigmoid function is mapped to [0,1].

[0041] The training data for the prediction model consists of approximately 1 million samples of tenant task logs from the past 3 months, divided into training and validation sets in an 8:2 ratio. The training and optimization of the prediction model follows existing conventional technical solutions, and the specific implementation steps are not detailed here.

[0042] Based on the prediction results output by the prediction model Generate the SLA demand matrix for tenant j. : The superscripts min and max represent the lower and upper bounds of the 95% confidence interval for the predicted value, respectively, for example, in real-time SLAs. ≤1 second;

[0043] Deploy smart contracts using the Hyperledger Fabric blockchain and define basic weight coefficients. , ;

[0044] The basic weighting coefficient is determined based on the agreement previously signed between the tenant and the cloud service provider, such as for financial tenants. e-commerce tenants ;

[0045] Based on historical SLA compliance rates Adjusting weighting coefficients, historical SLA compliance rate Specifically, corresponding to the past 24 hours, the relevant expression is:

[0046] ;in, This represents the base weighting coefficient of tenant j under the SLA metric k; This represents the updated dynamic weight coefficient; k is the index of the indicator. , These are real-time performance metrics, accuracy metrics, and resource availability metrics; The correction factor has a default value of 0.2.

[0047] If tenant j's real-time SLA compliance rate is less than the compliance threshold, the weight coefficient calculation update will be triggered; the default value of the compliance threshold is 0.95, which can be customized according to the application requirements of the actual application scenario.

[0048] The SLA demand matrix and the weight coefficients of each tenant's SLA terms are written into the blockchain ledger, and the immutability is ensured through Merkle tree hash verification. The on-chain data update frequency is synchronized with the prediction model.

[0049] In this embodiment of the invention, the attention mechanism LSTM, by focusing on key time-step features, such as resource fluctuations during sudden requests, can effectively reduce SLA demand prediction errors and provide accurate input for subsequent scheduling. The blockchain smart contract enables transparency and real-time updates of weight coefficients, avoiding scheduling bias caused by SLA term conflicts among tenants. For example, when the weight of a real-time tenant is dynamically increased, other tenants can trace and adjust the basis through on-chain data. By defining the confidence interval of the SLA demand matrix, such as a delay limit of ≤1 second, potential conflicts can be identified in advance. Through the detection model built by the application, dynamic perception and priority quantification of multi-tenant SLA demands can be achieved, thereby providing accurate constraint inputs and conflict coordination basis for subsequent resource scheduling algorithms.

[0050] The dynamic balancing application closed-loop monitoring module, based on the output SLA demand matrix and priority weights, designs an edge-cloud collaborative reinforcement learning scheduler. It trains a resource allocation agent using a deep deterministic policy gradient algorithm, taking edge node computing power, remaining cloud resources, and real-time SLA compliance rate as state inputs. The module dynamically outputs the edge offloading ratio and the cloud virtual machine elastic scaling strategy. The monitoring module obtains positive and negative feedback values ​​corresponding to the dynamic balancing effect executed by the resource allocation agent, and uses these positive and negative feedback values ​​to dynamically trigger a conventional policy iterative optimization mechanism. Specific steps include:

[0051] When designing a reinforcement learning scheduler for edge-cloud collaboration, a state space S and an action space A are defined; the state space S serves as the input feature vector of the resource allocation agent; specifically, the state space is as follows:

[0052] ;in, These are CPU utilization, memory usage, and bandwidth, all of which are characteristics of edge nodes. These are the remaining number of VMs and resource utilization, both of which are cloud-based characteristics. These are the real-time performance compliance rate, accuracy compliance rate, and availability compliance rate, all of which are real-time SLA metrics. These are predicted response time, predicted accuracy, and predicted availability, all of which are SLA prediction requirements. These are the weights for real-time performance indicators, accuracy indicators, and resource availability indicators, respectively.

[0053] Action space A is specifically as follows ;in, This represents the edge unloading ratio. 0 indicates local processing, and 1 indicates complete uninstallation to the cloud; To scale the number of virtual machines in the cloud. Negative numbers indicate the release of cloud virtual machines, while positive numbers indicate the addition of new cloud virtual machines.

[0054] The edge / cloud status is collected in real time through edge node monitoring tools; edge node monitoring tools, such as Prometheus, have a sampling frequency of 1 second / time.

[0055] Furthermore, cloud virtual machine (VM) scaling refers to a dynamic management mechanism that dynamically adjusts the number of virtual machines or resource configuration based on actual business load to maximize resource utilization and optimize costs. In this embodiment of the invention, scaling is horizontal scaling, which is achieved by increasing / decreasing the number of VM instances, for example, expanding from 2 to 5, or reducing from 5 to 1.

[0056] When training a resource allocation agent using a deep deterministic policy gradient algorithm, the network structure of the resource allocation agent includes an Actor network, a Critic network, and an experience replay and target network.

[0057] The Actor network corresponds to the policy function: input state vector, output deterministic action, and uses a fully connected network to achieve nonlinear mapping;

[0058] The Critic network's value function is: input state-action pairs, evaluate the value of actions, and use this value to guide policy updates;

[0059] Experience replay and target network: The experience replay pool is used to store samples, and the target network parameters are updated through a soft update mechanism to improve training stability; the specific parameters of the resource allocation agent are those of conventional schemes in this field, and the specific parameters are not limited or explained here;

[0060] Furthermore, a basic reward is constructed based on the hard constraint compliance rate, soft constraint compliance rate, and resource waste rate, involving the following expression: ;in, Basic reward coefficient; To achieve the hard constraint compliance rate, , The number of tasks with a delay of ≤1 second; This represents the total number of tasks. For soft constraint compliance rate, , M represents the number of nodes with a resource utilization rate of ≥80%, and M represents the total number of nodes. To reduce resource waste rate U represents the actual resource usage. To predict resource usage;

[0061] By leveraging priority weights in smart contracts, reward coefficients are dynamically adjusted to differentiate tenant priorities. The relevant expression is as follows: ;in, This is the adjusted reward coefficient; This is the SLA compliance correction factor. <0.95, =0.1, otherwise =-0.05;

[0062] The resource allocation agent is pre-trained using historical data, the policy network is optimized through multiple rounds of iteration, and then deployed to an edge-cloud collaborative environment. The trained policy is then solidified into an elastic scaling policy for cloud virtual machines and integrated with edge nodes and cloud resource managers through API interfaces.

[0063] The historical data specifically refers to historical scheduling logs, which include status, actions, rewards, and next status, with a cumulative sample size of ≥500,000 records. The training and optimization of the resource allocation agent can be achieved based on existing conventional solutions, and the specific implementation steps will not be elaborated here.

[0064] The resource allocation agent outputs the action space based on the current state, the edge nodes perform task offloading, and the cloud performs cloud virtual machine scaling; the scheduling cycle can be 5 seconds / time, synchronized with the SLA prediction cycle;

[0065] Monitor task execution results, calculate SLA compliance rate and resource waste rate, and generate reward signals;

[0066] During regular policy iteration, new samples are stored in the experience replay pool, triggering parameter updates for the resource allocation agent. New samples include state, action, reward, and next state. Parameter updates are paused when the average reward fluctuation over 100 consecutive scheduling cycles is ≤5%.

[0067] Furthermore, when conducting a regulatory evaluation of the dynamic balance effect of the resource allocation agent, the positive feedback value and negative feedback value of the resource allocation agent are calculated and obtained respectively. If the positive feedback value is greater than 0 and the negative feedback value is less than 1, the dynamic balance of the existing scheme is determined to be normal, and the implementation of the regular strategy iteration is maintained.

[0068] Conversely, if the dynamic balance of the existing solution is deemed abnormal, the conventional strategy iterative optimization mechanism will be triggered.

[0069] The calculation expressions for the positive feedback value and the negative feedback value are as follows:

[0070] ;in, It is a positive feedback value; This is the regulatory value for the edge offloading ratio, specifically the median of all edge offloading ratios; This represents the standard value for the edge offloading ratio; for both high-priority and low-priority tenant targets, the corresponding task offloading ratios should be as low as possible to avoid increased offloading costs due to edge resource contention; the larger the positive feedback value, the better the corresponding dynamic balancing effect.

[0071] ;in, The negative feedback value is represented by NL, which represents the resource waste rate. The total number of non-zero values; resource waste rate Total number of non-zero values ​​and resource waste rate The sum of all zeros; The allowable proportion of resource waste is defined as (0,1); the smaller the negative feedback value, the better the corresponding dynamic balance effect; the standard value of edge unloading ratio and the allowable proportion of resource waste can be determined based on the results of simulation tests using sample data in the early stage, or can be customized by professionals in the field according to the application and specification requirements of the actual application scenario, and the specific values ​​are not limited; the calculation of positive feedback value and negative feedback value can also be obtained based on the training and identification analysis of existing neural network models, and the calculation expression of the embodiment of the present invention is only a technical means.

[0072] In this embodiment of the invention, end-to-end learning of the resource allocation agent is achieved through a deep deterministic policy gradient algorithm, resolving the conflict between real-time performance and cost. While ensuring low latency for high-priority tenants, it can effectively improve resource utilization. The immutable weights recorded by the smart contract ensure fairness in rewards and punishments, avoiding scheduling bias caused by priority disputes among tenants, and effectively improving audit traceability efficiency. The penalty term for resource waste rate in the reward function can prompt the agent to dynamically adjust the cloud virtual machine scaling strategy, which can effectively reduce peak resource waste rate compared to traditional technical solutions. By monitoring and evaluating the dynamic balancing effect executed by the resource allocation agent, and using the evaluation results to dynamically trigger the conventional strategy iteration optimization mechanism, closed-loop monitoring of the dynamic balancing scheme is achieved, further improving the reliability and robustness of the dynamic balancing scheme implementation.

[0073] In the several embodiments provided by this invention, it should be understood that the disclosed system can be implemented in other ways. For example, the embodiments of the invention described above are merely illustrative; for example, the division of modules is only a logical functional division, and there may be other division methods in actual implementation.

[0074] The modules described as separate components may or may not be physically separate. The components shown as modules may or may not be physical modules; they may be located in one place or distributed across multiple network modules. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs.

[0075] Furthermore, the functional modules in the various embodiments of the present invention can be integrated into one processing module, or each module can exist physically separately, or two or more modules can be integrated into one module. The integrated module can be implemented in hardware or in the form of hardware plus software functional modules.

[0076] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the essential characteristics of the present invention.

[0077] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A cloud computing resource scheduling and monitoring system, characterized in that, include: The priority-based data analysis and processing module constructs a multi-tenant SLA dynamic prediction and priority-based model. Through the AI ​​time series prediction module, it makes short-term predictions of the tenant's task response time requirements, data processing accuracy threshold, and resource fluctuation characteristics for the next few minutes, generating a multi-dimensional SLA requirement matrix that includes real-time constraints, accuracy constraints, and availability constraints. Based on blockchain smart contracts, it records the weight coefficients of each tenant's SLA terms, forming an immutable basis for priority-based stratification. The dynamic balancing application closed-loop monitoring module designs an edge-cloud collaborative reinforcement learning scheduler based on the output SLA demand matrix and priority weights. It trains the resource allocation agent through a deep deterministic policy gradient algorithm, taking the edge node computing power, cloud remaining resources, and real-time SLA compliance rate as state inputs. It dynamically outputs the edge offloading ratio and cloud virtual machine elastic scaling strategy. The monitoring obtains the positive and negative feedback values ​​corresponding to the dynamic balancing effect executed by the resource allocation agent, and uses the positive and negative feedback values ​​to dynamically trigger the conventional policy iterative optimization mechanism. In designing the edge-cloud collaborative reinforcement learning scheduler, a state space S and an action space A are defined; the state space S serves as the input feature vector of the resource allocation agent; the action space is... ;in, This represents the edge unloading ratio. ; To scale the number of virtual machines in the cloud. ; A basic reward is constructed based on the hard constraint compliance rate, soft constraint compliance rate, and resource waste rate. The reward coefficient is dynamically adjusted using priority weights in smart contracts to differentiate tenant priorities. The relevant expression is as follows: ;in, This is the adjusted reward coefficient; Basic reward coefficient; Priority weights in smart contracts; This is the SLA compliance correction factor; k is the indicator index. , These are real-time performance metrics, accuracy metrics, and resource availability metrics.

2. The cloud computing resource scheduling and monitoring system according to claim 1, characterized in that, Collect historical task data from tenants and construct the corresponding input feature matrix. Where N is the number of tenants; T is the time window length; and F is the feature dimension, which includes task attributes, resource status, and environmental interference.

3. The cloud computing resource scheduling and monitoring system according to claim 2, characterized in that, The AI ​​time series prediction module builds a prediction model based on an attention-based LSTM network. ;in, To predict response time; To predict the accuracy of data processing; To predict resource availability.

4. The cloud computing resource scheduling and monitoring system according to claim 3, characterized in that, Based on the prediction results output by the prediction model Generate the SLA demand matrix for tenant j. : ; where the superscripts min and max are the lower and upper bounds of the 95% confidence interval for the corresponding predicted value, respectively.

5. A cloud computing resource scheduling and monitoring system according to claim 4, characterized in that, Define the base weighting coefficient based on historical SLA compliance rates. Adjusting the base weight coefficients involves the following expression: ;in, This represents the base weighting coefficient of tenant j under the SLA metric k; This represents the updated dynamic weight coefficients; This is a correction factor.

6. The cloud computing resource scheduling and monitoring system according to claim 1, characterized in that, The resource allocation agent is pre-trained using historical data, the policy network is optimized through multiple rounds of iteration, and then deployed to an edge-cloud collaborative environment. The trained policy is then solidified into an elastic scaling policy for cloud virtual machines and integrated with edge nodes and cloud resource managers through API interfaces.

7. A cloud computing resource scheduling and monitoring system according to claim 6, characterized in that, Calculate the positive and negative feedback values ​​of the resource allocation agent. If the positive feedback value is greater than 0 and the negative feedback value is less than 1, the dynamic balance of the existing scheme is determined to be normal, and the implementation of the regular strategy iteration is maintained. Conversely, if the dynamic balance of the existing scheme is abnormal, the conventional strategy iterative optimization mechanism will be triggered.

8. A cloud computing resource scheduling and monitoring system according to claim 7, characterized in that, The expressions for calculating positive and negative feedback values ​​are: ;in, It is a positive feedback value; This is the regulatory value for the edge unloading ratio; This is the standard value for the edge unloading ratio; ;in, This represents the negative feedback value; NL represents the total number of resources with a non-zero waste rate. It is the sum of the total number of resources with a non-zero waste rate and the total number of resources with a zero waste rate; The allowable proportion of resource waste, with a value range of (0,1).

Images