A cloud resource dynamic scheduling method and system based on virtualization technology

Abstract

The application provides a cloud resource dynamic scheduling method and system based on virtualization technology, comprising: collecting the running state of each resource node in the cloud platform and extracting multi-dimensional time sequence features; constructing a resource topology intention perception model based on a graph neural network, identifying the load dependency relationship and potential bottleneck intention between resource nodes; taking the load dependency relationship and bottleneck intention as input, combining a reinforcement learning strategy network, and pre-generating a resource migration and elastic scaling decision sequence for a future time window before scheduling triggering; according to the decision sequence, performing hot migration or in-place elastic scaling operation on the target resource node, and updating the corresponding traffic forwarding rules in the virtual network topology; collecting the actual resource utilization rate and service quality indicators after scheduling execution, and feeding back to the reinforcement learning strategy network to optimize the feature weight of the intention perception model and the generation strategy of the scheduling decision. The application can improve the scheduling timeliness and resource utilization efficiency.

Description

Technical Field

[0001] This invention relates to the field of cloud computing technology, and specifically to a method and system for dynamic scheduling of cloud resources based on virtualization technology. Background Technology

[0002] With the rapid development of information technology, cloud computing has become one of the core technologies supporting modern information infrastructure. Cloud computing abstracts physical computing resources into dynamically allocable and elastically scalable virtual resource pools through virtualization technology, providing users with on-demand computing, storage, and network services. Virtualization technology, as the fundamental support for cloud computing, mainly includes layers such as server virtualization, storage virtualization, and network virtualization. In the fields of industrial internet and intelligent manufacturing, new service models such as cloud manufacturing platforms and industrial PaaS platforms place higher demands on cloud computing resource scheduling, posing severe challenges to the real-time performance and reliability of resource scheduling. Summary of the Invention

[0003] Based on the above problems, this invention proposes a dynamic scheduling method and system for cloud resources based on virtualization technology. Through this invention, the timeliness of scheduling and the efficiency of resource utilization can be improved, and the risk of service quality degradation caused by scheduling delays can be reduced.

[0004] In view of this, one aspect of the present invention proposes a dynamic scheduling method for cloud resources based on virtualization technology, comprising: The operating status of each resource node in the cloud platform is collected, and multi-dimensional timing features of CPU, memory, bandwidth and I / O are extracted. A resource topology intent-aware model is constructed based on graph neural networks to identify load dependencies and potential bottleneck intents among resource nodes. Using load dependency and bottleneck intent as input, and combining reinforcement learning policy network, a resource migration and elastic scaling decision sequence for future time windows is pre-generated before scheduling is triggered. The decision sequence includes the type of target virtualized resource, migration path and scaling level. Based on the decision sequence, hot migration or in-situ elastic scaling operations are performed on the target resource nodes specified in the decision sequence, and the corresponding traffic forwarding rules in the virtual network topology are updated synchronously. The actual resource utilization rate and service quality indicators after the scheduling execution are collected and fed back to the reinforcement learning policy network as reward and punishment signals to optimize the feature weights of the intent-aware model and the generation strategy of scheduling decisions.

[0005] Optionally, the step of collecting the operating status of each resource node in the cloud platform and extracting multi-dimensional timing features of CPU, memory, bandwidth, and I / O includes: Monitoring agents are deployed in each of the resource nodes, including virtual machines, containers, and virtualized instances that carry virtual network functions. The monitoring agents periodically read the CPU utilization, memory usage, bandwidth throughput, and I / O read / write rate of the nodes they are in using a unified collection protocol. The raw indicator data reported by the monitoring agents in each resource node are aligned with a unified timestamp to eliminate data misalignment caused by node clock deviation, and the indicators of each dimension are normalized to form a standardized multi-dimensional time series sample for each resource node. For missing values ​​and abrupt noise in standardized multidimensional time series samples, interpolation methods between adjacent time points are used to filter and complete them, resulting in continuous time series samples. A fixed-length sliding window is used to truncate continuous time-series samples, and the mean, variance, and trend slope of each dimension index within each window are extracted to form a multi-dimensional feature vector that characterizes the current load dynamics of the resource node. The multidimensional feature vectors of all resource nodes in the cloud platform are aggregated according to the resource topology to form a structured feature set that can be input into the intent-aware model.

[0006] Optionally, the step of constructing a resource topology intent-aware model based on a graph neural network to identify load dependencies and potential bottleneck intents among resource nodes includes: Using each resource node in the cloud platform as a graph node, and the network connections, service calls, and resource sharing relationships between resource nodes as graph edges, a structured feature set is used as the initial feature vector of each graph node to construct a directed attribute graph that reflects the current resource topology of the cloud platform. Multi-round graph neural network message passing is performed on the directed attribute graph. In each round, each graph node aggregates its own feature vector with the feature vectors of its neighboring nodes according to the edge weights, updates its own hidden layer representation, and gradually integrates the load state information in its topological neighborhood into the representation of each graph node. Based on the hidden layer representation of each graph node after multi-round message aggregation, the attention weight between graph node pairs is calculated through the graph attention mechanism. Graph node pairs with attention weights exceeding a preset threshold are marked as associated node pairs with significant load dependencies. Using the hidden layer representation of each graph node as input, the fully connected classification layer outputs the probability of resource saturation of each graph node within a future time window. Resource nodes with probabilities exceeding a preset threshold are marked as potential bottleneck nodes. The bottleneck propagation path is determined by combining the relationship between the associated nodes of the bottleneck nodes, thus forming the output of the resource topology intent perception model.

[0007] Optionally, the step of pre-generating a resource migration and elastic scaling decision sequence oriented towards future time windows before scheduling is triggered, using load dependencies and bottleneck intentions as input and combining a reinforcement learning policy network, includes: The current environment state is defined by the set of potential bottleneck nodes, the bottleneck propagation path, and the load dependency relationship of the associated node pairs. The current environment state is encoded as a state vector of the reinforcement learning policy network. The state vector also contains information on the remaining CPU, memory, and bandwidth capacity of each host machine in the cloud platform. The executable scheduling operations are defined as the action space, and the feasible constraint boundaries of each type of action are set according to the remaining carrying capacity of each host machine. The state vector is input into the reinforcement learning policy network. Under the condition of satisfying the feasible constraint boundary, the reinforcement learning policy network infers and outputs an action sequence for future time windows in sequence. Each action includes the type of target virtualization resource, the migration path of the target host machine, and the elastic scaling level. The action sequence is simulated in multiple steps using a virtualized environment simulator to predict the load relief of each bottleneck node and the overall resource utilization balance after the action sequence is executed. The simulation evaluation results are used as the reward signal of the policy network to select the optimal decision sequence with the best expected return.

[0008] Optionally, the step of performing hot migration or in-situ elastic scaling operations on the target resource nodes specified in the decision sequence based on the decision sequence, and synchronously updating the corresponding traffic forwarding rules in the virtual network topology, includes: The optimal decision sequence is arranged according to the target resource type and the execution causal constraint relationship between each scheduling action. The execution causal constraint relationship is extracted from the load dependency relationship. The scheduling action of the upstream virtualized resource instance of the bottleneck propagation path is executed first to ensure that the migration or scaling operation of the downstream node is triggered sequentially after the upstream resources are ready. Based on the arrangement order, a hot migration operation is initiated for virtual machines marked as migration type in the decision sequence, transferring their memory state and runtime context to the target host machine without interruption; For containers marked as scaling type, synchronously perform replica expansion or resource quota adjustment according to the magnitude specified in the decision; After each virtual machine hot migration or container scaling operation is completed, a traffic forwarding rule update instruction is issued to the centralized network control component responsible for managing virtual network forwarding policies. The centralized network control component updates the flow table of the corresponding virtual switch or the routing policy of the virtual network function instance according to the instruction, and redirects the traffic path originally pointing to the node address before migration to the new node address. After all scheduling actions are completed, the actual connection status of the current virtualized resource topology is compared with the target topology status expected by the decision sequence. Compensation and correction operations are triggered for node connection relationships with deviations until the actual topology is completely consistent with the target topology.

[0009] Optionally, the step of feeding back the actual resource utilization rate and service quality index after the collection and scheduling execution as reward and punishment signals to the reinforcement learning policy network to optimize the feature weights of the intent-aware model and the generation strategy of scheduling decisions includes: After all scheduling actions are completed and topology consistency verification is finished, the actual resource utilization, service response latency and request success rate during the stable period after scheduling are collected by the monitoring agent of each resource node, which serve as the raw observation data for evaluating the scheduling effect. The original observation data is compared with the bottleneck relief expectation predicted by the pre-scheduling intention perception model. A scalar reward and punishment signal is constructed by the actual improvement in resource balance, the degree of latency improvement, and the change in request success rate. If the expectation is consistent with the actual situation, a positive reward is given; if the deviation exceeds the tolerance range, a penalty is given. The reward and punishment signals, along with the state vector and action sequence corresponding to this scheduling, are stored in the experience replay buffer. Historical experience batches are periodically sampled from the buffer, and gradient updates are performed on the parameters of the reinforcement learning policy network so that the policy network is biased towards generating action sequences with higher historical rewards in subsequent inference. Extract node samples that were predicted as bottlenecks but did not actually saturate, and those that were not predicted but actually saturated, from the current scheduling. Use these node samples as correction training samples for the graph neural network to fine-tune the feature weights of the graph attention mechanism in the intent-aware model, thereby improving the model's prediction accuracy of load dependency and bottleneck intent. Based on the updated intent-aware model and policy network, the current cloud platform resource topology is re-performed in terms of perception, decision-making, and execution processes, forming a continuously operating closed-loop dynamic scheduling mechanism.

[0010] Optionally, when weighting and fusing the multi-dimensional temporal characteristics of each virtualized resource node, the real-time stress score of each node is calculated using the following formula:

[0011] in, This represents the overall load stress score for a virtualized resource node; D represents the total number of resource dimensions collected, covering CPU, memory, bandwidth, and I / O. Let d be the current normalized utilization rate of the d-th resource dimension; The adaptive weight for the d-th resource dimension is dynamically learned from the contribution frequency of each dimension to the bottleneck event in the historical scheduling samples. Let be the saturation sensitivity coefficient of the d-th resource dimension, which characterizes the nonlinear growth rate of the pressure score when the utilization rate of this dimension approaches saturation; Let d be the standard deviation of the utilization rate of the d-th resource dimension within the current sliding window, representing the degree of load fluctuation of that dimension; is the fluctuation amplification factor for the d-th resource dimension, used to distinguish the different impacts of stable high load and drastic high load on scheduling urgency; The over-threshold duration penalty amplification factor is a preset fixed constant used to control the node's overall load pressure score as the number of cumulative over-threshold cycles increases. The rate of increase and amplification; The larger the value, the more significant the amplification effect of the pressure score obtained by the node that maintains a high pressure state for a long time; The duration for which the node continuously exceeds the preset alarm threshold in terms of pressure score, i.e., the cumulative number of periods exceeding the threshold; As a reference time base, the system scheduling cycle length is used for... Perform normalization; This is a historical bottleneck recurrence rate correction factor for the node, which is the ratio of the number of bottleneck events that occurred on the node in the recent historical window to the average number of bottleneck events on all nodes in the cloud platform. Nodes that have frequently become bottlenecks in the past are given additional pressure scores.

[0012] Optionally, when the graph neural network performs neighborhood message aggregation, the edge weights between nodes are calculated using the following formula:

[0013] in, The attention edge weight of resource node m to neighbor node n reflects the strength of the influence of node n's load state on node m. , These are the current hidden layer feature vectors of node m and node n, respectively; Cosine similarity is the similarity between two vectors, which measures the degree of similarity in the load patterns of two nodes. This is a temperature scaling factor that controls the degree of concentration of attention distribution; This is a normalized value of the historical call frequency between node m and node n, reflecting the actual interaction density between the two nodes in the service topology. Let m be the set of neighboring nodes of node m in the resource topology graph; The link quality coefficient between node m and node n is calculated jointly by the real-time network packet loss rate and latency jitter between the two nodes. The worse the link quality, the higher the link quality coefficient. The smaller the value, the lower the attention weight of that edge, thus suppressing the interference of dependencies propagating through unstable links on the aggregation result; This refers to the frequency of actual resource sharing events between node m and node n within the current time window, including underlying sharing behaviors such as memory page sharing, storage volume mounting, and CPU group scheduling. A sensitivity coefficient for resource sharing frequency, controlling the growth rate of the contribution of sharing behavior to edge weights; product term This constitutes the resource sharing intensity modulation factor.

[0014] Optionally, the reward function of the reinforcement learning policy network is calculated as follows:

[0015] in, This represents the immediate reward value obtained after the execution of the t-th scheduling action; The improvement in the overall resource utilization balance of the cloud platform before and after scheduling is measured by the decrease in the variance of the utilization rate of each node. The migration overhead consumed by the t-th scheduling action includes the weighted sum of hot migration data transfer volume and container reconstruction latency; The residual fluctuation of the load of each node after the execution of step t is measured by the mean of the standard deviation of utilization during the stabilization period after scheduling. The overall service quality score after scheduling is calculated by weighting response latency and request success rate. The preset service quality assurance threshold; This is an indicator function that takes a value of 1 and applies a penalty when the service quality is below a threshold. These are the weighting coefficients for equilibrium revenue, migration costs, volatility penalties, and service quality breach penalties, respectively. The improvement in the overall energy efficiency of the cloud platform after the execution of the t-th scheduling action is measured by the difference in host power consumption consumed by the unit computing task completed before and after scheduling. A positive value indicates that scheduling reduces energy consumption. The weighting coefficient for energy efficiency benefits is used to incorporate green energy-saving targets into the scheduling optimization direction; This is the number of cascaded reschedulings triggered by the t-th scheduling action. This value is incremented when a node's migration or scaling operation causes its neighboring nodes to become overloaded and triggers a new round of scheduling. This is the penalty coefficient for cascading scheduling, used to suppress aggressive scheduling actions that trigger cascading reactions when generated by the policy network.

[0016] Another aspect of the present invention provides a cloud resource dynamic scheduling system based on virtualization technology, for executing a cloud resource dynamic scheduling method based on virtualization technology, comprising: a control processing module and resource nodes; The control processing module is configured as follows: The operating status of each resource node in the cloud platform is collected, and multi-dimensional timing features of CPU, memory, bandwidth and I / O are extracted. A resource topology intent-aware model is constructed based on graph neural networks to identify load dependencies and potential bottleneck intents among resource nodes. Using load dependency and bottleneck intent as input, and combining reinforcement learning policy network, a resource migration and elastic scaling decision sequence for future time windows is pre-generated before scheduling is triggered. The decision sequence includes the type of target virtualized resource, migration path and scaling level. Based on the decision sequence, hot migration or in-situ elastic scaling operations are performed on the target resource nodes specified in the decision sequence, and the corresponding traffic forwarding rules in the virtual network topology are updated synchronously. The actual resource utilization rate and service quality indicators after the scheduling execution are collected and fed back to the reinforcement learning policy network as reward and punishment signals to optimize the feature weights of the intent-aware model and the generation strategy of scheduling decisions.

[0017] The technical solution of this invention provides a cloud resource dynamic scheduling method based on virtualization technology, comprising: collecting the operating status of each resource node in the cloud platform and extracting multi-dimensional temporal features of CPU, memory, bandwidth, and I / O; constructing a resource topology intent-aware model based on a graph neural network to identify load dependencies and potential bottleneck intents among resource nodes; using load dependencies and bottleneck intents as inputs, and combining a reinforcement learning policy network, pre-generating a resource migration and elastic scaling decision sequence oriented towards future time windows before scheduling is triggered, wherein the decision sequence includes the type, migration path, and scaling level of the target virtualized resource; performing hot migration or in-situ elastic scaling operations on the target resource nodes specified in the decision sequence according to the decision sequence, and synchronously updating the corresponding traffic forwarding rules in the virtual network topology; collecting the actual resource utilization and service quality indicators after scheduling execution, and feeding them back as reward and punishment signals to the reinforcement learning policy network to optimize the feature weights of the intent-aware model and the generation strategy of scheduling decisions. By using graph neural networks to deeply model the dependencies between nodes in the virtualized resource topology, and combining this with the predictive decision-making mechanism of reinforcement learning, resource scheduling is transformed from a passive response to an active prediction-driven process. This significantly improves scheduling timeliness and resource utilization efficiency. Furthermore, through closed-loop feedback, the strategy is continuously and adaptively optimized, reducing the risk of service quality degradation caused by scheduling delays. Attached Figure Description

[0018] Figure 1 This is a flowchart of a cloud resource dynamic scheduling method based on virtualization technology provided in an embodiment of the present invention; Figure 2 This is a schematic block diagram of a cloud resource dynamic scheduling system based on virtualization technology provided in an embodiment of the present invention. Detailed Implementation

[0019] To better understand the above-mentioned objectives, features, and advantages of the present invention, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be noted that, unless otherwise specified, the embodiments and features described in these embodiments can be combined with each other.

[0020] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and therefore the scope of protection of the invention is not limited to the specific embodiments disclosed below.

[0021] The terms "first," "second," etc., in the specification, claims, and accompanying drawings of this application are used to distinguish different objects, not to describe a specific order. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or apparatus that includes a series of steps or units is not limited to the listed steps or units, but may optionally include steps or units not listed, or may optionally include other steps or units inherent to these processes, methods, products, or apparatuses.

[0022] In this document, the term "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of this application. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0023] The following reference Figures 1 to 2 This invention describes a cloud resource dynamic scheduling method and system based on virtualization technology, provided by some embodiments of the present invention.

[0024] like Figure 1 As shown, one embodiment of the present invention provides a method for dynamic scheduling of cloud resources based on virtualization technology, including: The operating status of each resource node in the cloud platform is collected, and multi-dimensional timing features of CPU, memory, bandwidth and I / O are extracted. It is understood that the resource nodes include virtual machine instances, container instances, and virtual network function instances deployed in the form of virtual machines or containers, collectively referred to as virtualized resource nodes.

[0025] A resource topology intent-aware model is constructed based on graph neural networks to identify load dependencies and potential bottleneck intents among resource nodes. Using load dependency and bottleneck intent as input, and combining reinforcement learning policy network, a resource migration and elastic scaling decision sequence for future time windows is pre-generated before scheduling is triggered. The decision sequence includes the type of target virtualized resource, migration path and scaling level. Based on the decision sequence, hot migration or in-situ elastic scaling operations are performed on the target resource nodes specified in the decision sequence, and the corresponding traffic forwarding rules in the virtual network topology are updated synchronously to maintain the overall consistency of the virtualized resource topology. The actual resource utilization rate and service quality indicators after the scheduling execution are collected and fed back to the reinforcement learning policy network as reward and punishment signals to optimize the feature weights of the intent-aware model and the generation strategy of scheduling decisions.

[0026] The technical solution of this embodiment uses graph neural networks to deeply model the dependencies between nodes in the virtualized resource topology, and combines the predictive decision-making mechanism of reinforcement learning to transform resource scheduling from passive response to active prediction-driven, significantly improving scheduling timeliness and resource utilization efficiency. Furthermore, it achieves continuous adaptive optimization of the strategy through closed-loop feedback, reducing the risk of service quality degradation caused by scheduling lag.

[0027] In some possible embodiments of the present invention, the step of collecting the operating status of each resource node in the cloud platform and extracting multi-dimensional timing features of CPU, memory, bandwidth, and I / O includes: Monitoring agents are deployed in each of the resource nodes, including virtual machines, containers, and virtualized instances that carry virtual network functions. The monitoring agents periodically read the CPU utilization, memory usage, bandwidth throughput, and I / O read / write rate of the nodes they are in using a unified collection protocol. The raw indicator data reported by the monitoring agents in each resource node are aligned with a unified timestamp to eliminate data misalignment caused by node clock deviation, and the indicators of each dimension are normalized to form a standardized multi-dimensional time series sample for each resource node. For missing values ​​and abrupt noise in standardized multidimensional time series samples (caused by network jitter or proxy anomalies), interpolation methods between adjacent time points are used for filtering and completion (to ensure the continuity and reliability of time series data) to obtain continuous time series samples; A fixed-length sliding window is used to truncate continuous time-series samples, and the mean, variance, and trend slope of each dimension index within each window are extracted to form a multi-dimensional feature vector that characterizes the current load dynamics of the resource node. The multidimensional feature vectors of all resource nodes in the cloud platform are aggregated according to the resource topology to form a structured feature set that can be input into the intent-aware model.

[0028] In this embodiment, the consistency and integrity of multi-source heterogeneous node data are ensured through agent-based unified collection, time-series alignment, and anomaly completion. Through sliding window extraction and cross-node aggregation, high-quality and structured load feature inputs are provided for subsequent graph neural network modeling, fundamentally improving the accuracy of scheduling decisions.

[0029] In some possible embodiments of the present invention, the step of constructing a resource topology intent-aware model based on a graph neural network to identify load dependencies and potential bottleneck intents among resource nodes includes: Using each resource node in the cloud platform as a graph node, and the network connections, service calls, and resource sharing relationships between resource nodes as graph edges, a structured feature set is used as the initial feature vector of each graph node to construct a directed attribute graph that reflects the current resource topology of the cloud platform. Multi-round graph neural network message passing is performed on the directed attribute graph. In each round, each graph node aggregates its own feature vector with the feature vectors of its neighboring nodes according to the edge weights, updates its own hidden layer representation, and gradually integrates the load state information in its topological neighborhood into the representation of each graph node. Based on the hidden layer representation of each graph node after multi-round message aggregation, the attention weight between graph node pairs is calculated through the graph attention mechanism. Graph node pairs with attention weights exceeding a preset threshold are marked as associated node pairs with significant load dependencies. Using the hidden layer representation of each graph node as input, the fully connected classification layer outputs the probability of resource saturation of each graph node within a future time window. Resource nodes with probabilities exceeding a preset threshold are marked as potential bottleneck nodes. The bottleneck propagation path is determined by combining the relationship between the associated nodes of the bottleneck nodes, thus forming the output of the resource topology intent perception model.

[0030] It is understood that there is a one-to-one mapping relationship between the graph nodes and their corresponding virtualized resource instances (i.e., resource nodes), and the bottleneck prediction conclusion of the graph nodes directly corresponds to the scheduling priority determination of the corresponding virtualized resource nodes.

[0031] In this embodiment, by structuring the cloud resource topology into a directed attribute graph and introducing a graph attention mechanism, the model can explicitly capture the load dependencies between nodes, achieve early perception of bottleneck nodes and their propagation paths, and provide accurate intent input with topological semantics for subsequent predictive scheduling decisions.

[0032] In some possible embodiments of the present invention, the step of pre-generating a resource migration and elastic scaling decision sequence oriented towards future time windows before scheduling is triggered, using load dependency and bottleneck intent as input and combining a reinforcement learning policy network, includes: The current environment state is defined by the set of potential bottleneck nodes, the bottleneck propagation path, and the load dependency relationship of the associated node pairs. The current environment state is encoded as a state vector of the reinforcement learning policy network. The state vector also contains information on the remaining CPU, memory, and bandwidth capacity of each host machine in the cloud platform. The action space includes the following three types of atomic actions: hot migration actions for virtualized resource instances, specifying the source virtual machine instance and the target host machine; elastic scaling actions for container instances, specifying the target container group and the scaling level; and traffic rerouting actions for virtual networks, specifying the virtual network function instance whose forwarding path needs to be adjusted and the new routing policy.

[0033] The executable scheduling operations are defined as the action space, and the feasible constraint boundaries of each type of action are set according to the remaining carrying capacity of each host machine. The state vector is input into the reinforcement learning policy network. Under the condition of satisfying the feasible constraint boundary, the reinforcement learning policy network infers and outputs an action sequence for future time windows in sequence. Each action includes the type of target virtualization resource, the migration path of the target host machine, and the elastic scaling level. The action sequence is simulated in multiple steps using a virtualized environment simulator to predict the load relief of each bottleneck node and the overall resource utilization balance after the action sequence is executed. The simulation evaluation results are used as the reward signal of the policy network to select the optimal decision sequence with the best expected return.

[0034] In this embodiment, by directly encoding the topology intent perception result into a reinforcement learning state space and combining it with multi-step look-ahead evaluation by a simulator, the scheduling decision completes feasibility verification and benefit optimization before execution, avoiding blind trial-and-error scheduling and significantly improving the foresight of the decision and the success rate of execution.

[0035] In some possible embodiments of the present invention, the step of performing hot migration or in-situ elastic scaling operations on the target resource nodes specified in the decision sequence according to the decision sequence, and synchronously updating the corresponding traffic forwarding rules in the virtual network topology to maintain the overall consistency of the virtualized resource topology, includes: The optimal decision sequence is arranged according to the target resource type and the execution causal constraint relationship between each scheduling action. The execution causal constraint relationship is extracted from the load dependency relationship. The scheduling action of the upstream virtualized resource instance of the bottleneck propagation path is executed first to ensure that the migration or scaling operation of the downstream node is triggered sequentially after the upstream resources are ready, avoiding resource competition caused by concurrent execution. Based on the arrangement order, a hot migration operation is initiated for virtual machines marked as migration type in the decision sequence, transferring their memory state and runtime context to the target host machine without interruption; For containers marked as scaling type, synchronously perform replica expansion or resource quota adjustment according to the magnitude specified in the decision; After each virtual machine hot migration or container scaling operation is completed, a traffic forwarding rule update instruction is issued to the centralized network control component responsible for managing virtual network forwarding policies. The centralized network control component updates the flow table of the corresponding virtual switch or the routing policy of the virtual network function instance according to the instruction, redirecting the traffic path originally pointing to the node address before migration to the new node address, ensuring the continuity of service traffic during the topology change process. After all scheduling actions are completed, the actual connection status of the current virtualized resource topology is compared with the target topology status expected by the decision sequence. Compensation and correction operations are triggered for node connection relationships with deviations until the actual topology is completely consistent with the target topology.

[0036] In this embodiment, by orchestrating the execution order according to dependencies, synchronizing migration and network rule updates, and implementing a topology consistency verification and compensation mechanism after execution, it is ensured that service traffic is not interrupted during scheduling and that virtualized resource topology does not generate islands or routing black holes, thus achieving smooth resource reorganization with zero awareness.

[0037] In some possible embodiments of the present invention, the step of feeding back the actual resource utilization rate and service quality index after the collection and scheduling execution as reward and punishment signals to the reinforcement learning policy network to optimize the feature weights of the intent-aware model and the generation strategy of scheduling decisions includes: After all scheduling actions are completed and topology consistency verification is finished, the actual resource utilization, service response latency and request success rate during the stable period after scheduling are collected by the monitoring agent of each resource node, which serve as the raw observation data for evaluating the scheduling effect. The original observation data is compared with the bottleneck relief expectation predicted by the pre-scheduling intention perception model. A scalar reward and punishment signal is constructed by the actual improvement in resource balance, the degree of latency improvement, and the change in request success rate. If the expectation is consistent with the actual situation, a positive reward is given; if the deviation exceeds the tolerance range, a penalty is given. The reward and punishment signals, along with the state vector and action sequence corresponding to this scheduling, are stored in the experience replay buffer. Historical experience batches are periodically sampled from the buffer, and gradient updates are performed on the parameters of the reinforcement learning policy network so that the policy network is biased towards generating action sequences with higher historical rewards in subsequent inference. Extract node samples that were predicted as bottlenecks but did not actually saturate, and those that were not predicted but actually saturated, from the current scheduling. Use these node samples as correction training samples for the graph neural network to fine-tune the feature weights of the graph attention mechanism in the intent-aware model, thereby improving the model's prediction accuracy of load dependency and bottleneck intent. Based on the updated intent-aware model and policy network, the current cloud platform resource topology is re-performed in terms of perception, decision-making, and execution processes, forming a continuously operating closed-loop dynamic scheduling mechanism.

[0038] In this embodiment, a precise reward and punishment signal is constructed by comparing the actual effect after scheduling with the expected result. This signal drives the targeted optimization of the policy network and the intent-aware model, respectively. As the system continues to run, it continuously corrects prediction and decision-making biases, thereby enabling the scheduling capability to adaptively enhance as the cloud platform's load pattern evolves.

[0039] In some possible embodiments of the present invention, when weighted and fused with the multidimensional temporal characteristics of each virtualized resource node, the real-time stress score of each node is calculated using the following formula:

[0040] in, This represents the overall load stress score for a virtualized resource node; D represents the total number of resource dimensions collected, covering CPU, memory, bandwidth, and I / O. Let d be the current normalized utilization rate of the d-th resource dimension; The adaptive weight for the d-th resource dimension is dynamically learned from the contribution frequency of each dimension to the bottleneck event in the historical scheduling samples. Let be the saturation sensitivity coefficient of the d-th resource dimension, which characterizes the nonlinear growth rate of the pressure score when the utilization rate of this dimension approaches saturation; Let d be the standard deviation of the utilization rate of the d-th resource dimension within the current sliding window, representing the degree of load fluctuation of that dimension; is the fluctuation amplification factor for the d-th resource dimension, used to distinguish the different impacts of stable high load and drastic high load on scheduling urgency; The over-threshold duration penalty amplification factor is a preset fixed constant used to control the node's overall load pressure score as the number of cumulative over-threshold cycles increases. The rate of increase and amplification; The larger the value, the more significant the amplification effect of the pressure score obtained by the node that maintains a high pressure state for a long time; The duration for which the node continuously exceeds the preset alarm threshold in terms of pressure score, i.e., the cumulative number of periods exceeding the threshold; As a reference time base, the system scheduling cycle length is used for... Perform normalization; This is a historical bottleneck recurrence rate correction factor for the node, which is the ratio of the number of bottleneck events that occurred on the node in the recent historical window to the average number of bottleneck events on all nodes in the cloud platform. Nodes that have frequently become bottlenecks in the past are given additional pressure scores.

[0041] In some possible embodiments of the present invention, when the graph neural network performs neighborhood message aggregation, the edge weights between nodes are calculated using the following formula:

[0042] in, The attention edge weight of resource node m to neighbor node n reflects the strength of the influence of node n's load state on node m. , These are the current hidden layer feature vectors of node m and node n, respectively; Cosine similarity is the similarity between two vectors, which measures the degree of similarity in the load patterns of two nodes. This is a temperature scaling factor that controls the degree of concentration of attention distribution; This is a normalized value of the historical call frequency between node m and node n, reflecting the actual interaction density between the two nodes in the service topology. Let m be the set of neighboring nodes of node m in the resource topology graph; The link quality coefficient between node m and node n is calculated jointly by the real-time network packet loss rate and latency jitter between the two nodes. The worse the link quality, the higher the link quality coefficient. The smaller the value, the lower the attention weight of that edge, thus suppressing the interference of dependencies propagating through unstable links on the aggregation result; This refers to the frequency of actual resource sharing events between node m and node n within the current time window, including underlying sharing behaviors such as memory page sharing, storage volume mounting, and CPU group scheduling. A sensitivity coefficient for resource sharing frequency, controlling the growth rate of the contribution of sharing behavior to edge weights; product term This constitutes a resource sharing intensity modulation factor, enabling node pairs that frequently share underlying resources to obtain higher dependency edge weights.

[0043] In some possible embodiments of the present invention, the reward function of the reinforcement learning policy network is calculated as follows:

[0044] in, This represents the immediate reward value obtained after the execution of the t-th scheduling action; The improvement in the overall resource utilization balance of the cloud platform before and after scheduling is measured by the decrease in the variance of the utilization rate of each node. The migration overhead consumed by the t-th scheduling action includes the weighted sum of hot migration data transfer volume and container reconstruction latency; The residual fluctuation of the load of each node after the execution of step t is measured by the mean of the standard deviation of utilization during the stabilization period after scheduling. The overall service quality score after scheduling is calculated by weighting response latency and request success rate. The preset service quality assurance threshold; This is an indicator function that takes a value of 1 and applies a penalty when the service quality is below a threshold. These are the weighting coefficients for equilibrium revenue, migration costs, volatility penalties, and service quality breach penalties, respectively. The improvement in the overall energy efficiency of the cloud platform after the execution of the t-th scheduling action is measured by the difference in host power consumption consumed by the unit computing task completed before and after scheduling. A positive value indicates that scheduling reduces energy consumption. The weighting coefficient for energy efficiency benefits is used to incorporate green energy-saving targets into the scheduling optimization direction; This is the number of cascaded reschedulings triggered by the t-th scheduling action. This value is incremented when a node's migration or scaling operation causes its neighboring nodes to become overloaded and triggers a new round of scheduling. This is the penalty coefficient for cascading scheduling, used to suppress aggressive scheduling actions that trigger cascading reactions when generated by the policy network.

[0045] In some possible embodiments of the present invention, in the step of performing hot migration or in-situ elastic scaling operations on the target resource nodes specified in the decision sequence according to the decision sequence, and synchronously updating the corresponding traffic forwarding rules in the virtual network topology to maintain the overall consistency of the virtualized resource topology, the following host machine bearing adaptability function is used for the selection of the hot migration target host machine:

[0046] in, The candidate host machine is given a performance score, and the higher the score, the more suitable the host machine is for accepting the migrated virtualization resources. This represents the current normalized utilization rate of the candidate host in the d-th resource dimension. This represents the total number of virtualized resource instances currently hosted on the candidate host, reflecting its management complexity. This is the instance density penalty coefficient, used to suppress excessive concentration of resources on a small number of hosts; To reduce network transmission latency between the source node and the candidate host; This represents the maximum network latency between all host pairs within the cloud platform, used for normalization. The resource reservation elasticity coefficient for candidate hosts is defined as the proportion of the remaining capacity of the host under the current load to the limit of triggering local elastic scaling. The value range is (0,1]. The larger the value, the more local elastic space the host still has after accepting the migrated resources to cope with subsequent load surges. The recent failure probability estimate of the candidate host is obtained by inputting the host's hardware health monitoring indicators (including CPU temperature, memory error correction event frequency and disk bad sector rate) into the reliability prediction model. The value range is [0,1). The higher the failure probability, the lower the fit score, thereby avoiding the risk of migrating resources to a potentially faulty host.

[0047] In some possible embodiments of the present invention, when correcting the graph attention feature weights of the intent-aware model, the correction magnitude of each node is determined using the following formula:

[0048] in, This is the incremental correction for the graph attention feature weights of a specific resource node on the d-th resource dimension; The learning step size for feature weight correction controls the magnitude of each correction. This is the predicted value of the bottleneck probability for this node by the intent-aware model; This is a binary label representing the actual bottleneck occurrence at the node as observed after scheduling; the larger the prediction deviation, the larger the correction increment. This represents the current normalized utilization rate of the node in the d-th resource dimension. is the mean utilization rate of all nodes in the cloud platform in the d-th resource dimension; the fractional term represents the relative significance of the deviation of this dimension from the mean, and is used to preferentially allocate the correction increment to the resource dimension that contributes the most to the prediction error; The historical correction stability coefficient of the d-th resource dimension at this node is defined as the inverse normalized value of the magnitude of change of the weight of this dimension in recent rounds of corrections. The cumulative prediction error across cycles for this node is defined as the exponentially weighted moving average of the absolute values ​​of the prediction errors for this node over several recent scheduling cycles, reflecting the difficulty of long-term prediction for this node. To accumulate error amplification factor, a larger correction gain is applied to nodes that are difficult to predict accurately in the long term, so that the model can concentrate more learning resources on feature optimization of high-difficulty nodes.

[0049] in, The historical correction stability coefficient of the d-th resource dimension at this node is calculated as follows: take the most recent... Change in weight of this dimension during round correction Given an absolute value sequence, calculate the number of sign flips between adjacent rounds of change. (i.e., the number of times two adjacent rounds of corrections are in opposite directions) and the mean of the absolute values ​​of the changes. ,but:

[0050] in, The number of historical window rounds for stability assessment; In order to be in The number of times the weight correction direction for this dimension reverses within the cycle reflects the oscillation frequency of the correction direction; for The mean of the absolute values ​​of the changes in the weights of this dimension within a round reflects the average magnitude of the correction in each round; the product of the two comprehensively characterizes the frequency and amplitude of the oscillations, and the larger the product, the more unstable the weights of this dimension. The smaller the value, the stronger the suppression of the current correction increment in that dimension.

[0051] Please refer to Figure 2 Another embodiment of the present invention provides a cloud resource dynamic scheduling system based on virtualization technology, used to execute a cloud resource dynamic scheduling method based on virtualization technology, comprising: a control processing module and resource nodes; The control processing module is configured as follows: The operating status of each resource node in the cloud platform is collected, and multi-dimensional timing features of CPU, memory, bandwidth and I / O are extracted. A resource topology intent-aware model is constructed based on graph neural networks to identify load dependencies and potential bottleneck intents among resource nodes. Using load dependency and bottleneck intent as input, and combining reinforcement learning policy network, a resource migration and elastic scaling decision sequence for future time windows is pre-generated before scheduling is triggered. The decision sequence includes the type of target virtualized resource, migration path and scaling level. Based on the decision sequence, hot migration or in-situ elastic scaling operations are performed on the target resource nodes specified in the decision sequence, and the corresponding traffic forwarding rules in the virtual network topology are updated synchronously. The actual resource utilization rate and service quality indicators after the scheduling execution are collected and fed back to the reinforcement learning policy network as reward and punishment signals to optimize the feature weights of the intent-aware model and the generation strategy of scheduling decisions.

[0052] It should be known that, Figure 2 The block diagram of the cloud resource dynamic scheduling system based on virtualization technology shown is for illustrative purposes only, and the number of modules shown does not limit the scope of protection of this invention. The cloud resource dynamic scheduling system based on virtualization technology provided in this embodiment can be used to execute various embodiments of the corresponding cloud resource dynamic scheduling method based on virtualization technology. For specific implementation details, please refer to the descriptions of the various method embodiments, which will not be repeated here.

[0053] It should be noted that, for the sake of simplicity, the foregoing method embodiments are all described as a series of actions. However, those skilled in the art should understand that this application is not limited to the described order of actions, as some steps may be performed in other orders or simultaneously according to this application. Furthermore, those skilled in the art should also understand that the embodiments described in the specification are preferred embodiments, and the actions and modules involved are not necessarily essential to this application.

[0054] In the above embodiments, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.

[0055] In the several embodiments provided in this application, it should be understood that the disclosed apparatus can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of the units described above is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between devices or units may be electrical or other forms.

[0056] The units described above as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0057] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0058] If the integrated units described above are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage device (CMD). Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a memory and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned memory includes various media capable of storing program code, such as USB flash drives, read-only memory (ROM), random access memory (RAM), portable hard drives, magnetic disks, or optical disks.

[0059] Those skilled in the art will understand that all or part of the steps in the various methods of the above embodiments can be implemented by a program instructing related hardware. The program can be stored in a computer-readable storage medium, which may include: flash drive, read-only memory (ROM), random access memory (RAM), disk or optical disk, etc.

[0060] The embodiments of this application have been described in detail above. Specific examples have been used to illustrate the principles and implementation methods of this application. The description of the above embodiments is only for the purpose of helping to understand the method and core ideas of this application. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of this application. Therefore, the content of this specification should not be construed as a limitation of this application.

[0061] While the present invention has been disclosed above, it is not limited thereto. Any person skilled in the art can easily conceive of variations or substitutions without departing from the spirit and scope of the present invention, and various modifications and alterations can be made, including combinations of the different functions and implementation steps described above, as well as software and hardware implementation methods, all of which are within the protection scope of the present invention.

Claims

1. A method for dynamic scheduling of cloud resources based on virtualization technology, characterized in that, include: The operating status of each resource node in the cloud platform is collected, and multi-dimensional timing features of CPU, memory, bandwidth and I / O are extracted. A resource topology intent-aware model is built based on graph neural networks to identify load dependencies and potential bottleneck intents between resource nodes. Using load dependency and bottleneck intent as input, and combining reinforcement learning policy network, a resource migration and elastic scaling decision sequence for future time windows is pre-generated before scheduling is triggered. The decision sequence includes the type of target virtualized resource, migration path and scaling level. Based on the decision sequence, hot migration or in-situ elastic scaling operations are performed on the target resource nodes specified in the decision sequence, and the corresponding traffic forwarding rules in the virtual network topology are updated synchronously. The actual resource utilization rate and service quality indicators after the scheduling execution are collected and fed back to the reinforcement learning policy network as reward and punishment signals to optimize the feature weights of the intent-aware model and the generation strategy of scheduling decisions.

2. The cloud resource dynamic scheduling method based on virtualization technology according to claim 1, characterized in that, The steps of collecting the operating status of each resource node in the cloud platform and extracting multi-dimensional timing features of CPU, memory, bandwidth, and I / O include: Monitoring agents are deployed in each of the resource nodes, including virtual machines, containers, and virtualized instances that carry virtual network functions. The monitoring agents periodically read the CPU utilization, memory usage, bandwidth throughput, and I / O read / write rate of the nodes they are in using a unified collection protocol. The raw indicator data reported by the monitoring agents in each resource node are aligned with a unified timestamp to eliminate data misalignment caused by node clock deviation, and the indicators of each dimension are normalized to form a standardized multi-dimensional time series sample for each resource node. For missing values ​​and abrupt noise in standardized multidimensional time series samples, interpolation methods between adjacent time points are used to filter and complete them, resulting in continuous time series samples. A fixed-length sliding window is used to truncate continuous time-series samples, and the mean, variance, and trend slope of each dimension index within each window are extracted to form a multi-dimensional feature vector that characterizes the current load dynamics of the resource node. The multidimensional feature vectors of all resource nodes in the cloud platform are aggregated according to the resource topology to form a structured feature set that can be input into the intent-aware model.

3. The cloud resource dynamic scheduling method based on virtualization technology according to claim 2, characterized in that, The steps of constructing a resource topology intent-aware model based on graph neural networks to identify load dependencies and potential bottleneck intents among resource nodes include: Using each resource node in the cloud platform as a graph node, and the network connections, service calls, and resource sharing relationships between resource nodes as graph edges, a structured feature set is used as the initial feature vector of each graph node to construct a directed attribute graph that reflects the current resource topology of the cloud platform. Multi-round graph neural network message passing is performed on the directed attribute graph. In each round, each graph node aggregates its own feature vector with the feature vectors of its neighboring nodes according to the edge weights, updates its own hidden layer representation, and gradually integrates the load state information in its topological neighborhood into the representation of each graph node. Based on the hidden layer representation of each graph node after multi-round message aggregation, the attention weight between graph node pairs is calculated through the graph attention mechanism. Graph node pairs with attention weights exceeding a preset threshold are marked as associated node pairs with significant load dependencies. Using the hidden layer representation of each graph node as input, the fully connected classification layer outputs the probability of resource saturation of each graph node within a future time window. Resource nodes with probabilities exceeding a preset threshold are marked as potential bottleneck nodes. The bottleneck propagation path is determined by combining the relationship between the associated nodes of the bottleneck nodes, thus forming the output of the resource topology intent perception model.

4. The cloud resource dynamic scheduling method based on virtualization technology according to claim 3, characterized in that, The step of pre-generating a resource migration and elastic scaling decision sequence oriented towards future time windows before scheduling is triggered, using load dependency and bottleneck intent as input and combining a reinforcement learning policy network, includes: The current environment state is defined by the set of potential bottleneck nodes, the bottleneck propagation path, and the load dependency relationship of the associated node pairs. The current environment state is encoded as a state vector of the reinforcement learning policy network. The state vector also contains information on the remaining CPU, memory, and bandwidth capacity of each host machine in the cloud platform. The executable scheduling operations are defined as the action space, and the feasible constraint boundaries of each type of action are set according to the remaining carrying capacity of each host machine. The state vector is input into the reinforcement learning policy network. Under the condition of satisfying the feasible constraint boundary, the reinforcement learning policy network infers and outputs an action sequence for future time windows in sequence. Each action includes the type of target virtualization resource, the migration path of the target host machine, and the elastic scaling level. The action sequence is simulated in multiple steps using a virtualized environment simulator to predict the load relief of each bottleneck node and the overall resource utilization balance after the action sequence is executed. The simulation evaluation results are used as the reward signal of the policy network to select the optimal decision sequence with the best expected return.

5. The cloud resource dynamic scheduling method based on virtualization technology according to claim 4, characterized in that, The step of performing hot migration or in-situ elastic scaling operations on the target resource nodes specified in the decision sequence based on the decision sequence, and synchronously updating the corresponding traffic forwarding rules in the virtual network topology, includes: The optimal decision sequence is arranged according to the target resource type and the execution causal constraint relationship between each scheduling action. The execution causal constraint relationship is extracted from the load dependency relationship. The scheduling action of the upstream virtualized resource instance of the bottleneck propagation path is executed first to ensure that the migration or scaling operation of the downstream node is triggered sequentially after the upstream resources are ready. Based on the arrangement order, a hot migration operation is initiated for virtual machines marked as migration type in the decision sequence, transferring their memory state and runtime context to the target host machine without interruption; For containers marked as scaling type, synchronously perform replica expansion or resource quota adjustment according to the magnitude specified in the decision; After each virtual machine hot migration or container scaling operation is completed, a traffic forwarding rule update instruction is issued to the centralized network control component responsible for managing virtual network forwarding policies. The centralized network control component updates the flow table of the corresponding virtual switch or the routing policy of the virtual network function instance according to the instruction, and redirects the traffic path originally pointing to the node address before migration to the new node address. After all scheduling actions are completed, the actual connection status of the current virtualized resource topology is compared with the target topology status expected by the decision sequence. Compensation and correction operations are triggered for node connection relationships with deviations until the actual topology is completely consistent with the target topology.

6. The cloud resource dynamic scheduling method based on virtualization technology according to claim 5, characterized in that, The steps of collecting and scheduling data, and then feeding back the actual resource utilization and service quality indicators as reward / penalty signals to the reinforcement learning policy network to optimize the feature weights of the intent-aware model and the generation strategy for scheduling decisions include: After all scheduling actions are completed and topology consistency verification is finished, the actual resource utilization, service response latency and request success rate during the stable period after scheduling are collected by the monitoring agent of each resource node, which serve as the raw observation data for evaluating the scheduling effect. The original observation data is compared with the bottleneck relief expectation predicted by the pre-scheduling intention perception model. A scalar reward and punishment signal is constructed by the actual improvement in resource balance, the degree of latency improvement, and the change in request success rate. If the expectation is consistent with the actual situation, a positive reward is given; if the deviation exceeds the tolerance range, a penalty is given. The reward and punishment signals, along with the state vector and action sequence corresponding to this scheduling, are stored in the experience replay buffer. Historical experience batches are periodically sampled from the buffer, and gradient updates are performed on the parameters of the reinforcement learning policy network so that the policy network is biased towards generating action sequences with higher historical rewards in subsequent inference. Extract node samples that were predicted as bottlenecks but did not actually saturate, and those that were not predicted but actually saturated, from the current scheduling. Use these node samples as correction training samples for the graph neural network to fine-tune the feature weights of the graph attention mechanism in the intent-aware model, thereby improving the model's prediction accuracy of load dependency and bottleneck intent. Based on the updated intent-aware model and policy network, the current cloud platform resource topology is re-performed in terms of perception, decision-making, and execution processes, forming a continuously operating closed-loop dynamic scheduling mechanism.

7. The cloud resource dynamic scheduling method based on virtualization technology according to claim 6, characterized in that, When weighted and fused the multi-dimensional time-series characteristics of each virtualized resource node, the real-time stress score of each node is calculated using the following formula: ； in, This represents the overall load stress score for a virtualized resource node; D represents the total number of resource dimensions collected, covering CPU, memory, bandwidth, and I / O. Let d be the current normalized utilization rate of the d-th resource dimension; The adaptive weight for the d-th resource dimension is dynamically learned from the contribution frequency of each dimension to the bottleneck event in the historical scheduling samples. Let be the saturation sensitivity coefficient of the d-th resource dimension, which characterizes the nonlinear growth rate of the pressure score when the utilization rate of this dimension approaches saturation; Let d be the standard deviation of the utilization rate of the d-th resource dimension within the current sliding window, representing the degree of load fluctuation of that dimension; is the fluctuation amplification factor for the d-th resource dimension, used to distinguish the different impacts of stable high load and drastic high load on scheduling urgency; The over-threshold duration penalty amplification factor is a preset fixed constant used to control the node's overall load pressure score as the number of cumulative over-threshold cycles increases. The rate of increase and amplification; The larger the value, the more significant the amplification effect of the pressure score obtained by the node that maintains a high pressure state for a long time; The duration for which the node continuously exceeds the preset alarm threshold in terms of pressure score, i.e., the cumulative number of periods exceeding the threshold; As a reference time base, the system scheduling cycle length is used for... Perform normalization; This is a historical bottleneck recurrence rate correction factor for the node, which is the ratio of the number of bottleneck events that occurred on the node in the recent historical window to the average number of bottleneck events on all nodes in the cloud platform. Nodes that have frequently become bottlenecks in the past are given additional pressure scores.

8. The cloud resource dynamic scheduling method based on virtualization technology according to claim 7, characterized in that, When a graph neural network performs neighborhood message aggregation, the edge weights between nodes are calculated using the following formula: ； in, The attention edge weight of resource node m to neighbor node n reflects the strength of the influence of node n's load state on node m. , These are the current hidden layer feature vectors of node m and node n, respectively; Cosine similarity is the similarity between two vectors, which measures the degree of similarity in the load patterns of two nodes. This is a temperature scaling factor that controls the degree of concentration of attention distribution; This is a normalized value of the historical call frequency between node m and node n, reflecting the actual interaction density between the two nodes in the service topology. Let m be the set of neighboring nodes of node m in the resource topology graph; The link quality coefficient between node m and node n is calculated jointly by the real-time network packet loss rate and latency jitter between the two nodes. The worse the link quality, the higher the link quality coefficient. The smaller the value, the lower the attention weight of that edge, thus suppressing the interference of dependencies propagating through unstable links on the aggregation result; This refers to the frequency of actual resource sharing events between node m and node n within the current time window, including underlying sharing behaviors such as memory page sharing, storage volume mounting, and CPU group scheduling. A sensitivity coefficient for resource sharing frequency, controlling the growth rate of the contribution of sharing behavior to edge weights; product term This constitutes the resource sharing intensity modulation factor.

9. The cloud resource dynamic scheduling method based on virtualization technology according to claim 8, characterized in that, The reward function of a reinforcement learning policy network is calculated as follows: ； in, This represents the immediate reward value obtained after the execution of the t-th scheduling action; The improvement in the overall resource utilization balance of the cloud platform before and after scheduling is measured by the decrease in the variance of the utilization rate of each node. The migration overhead consumed by the t-th scheduling action includes the weighted sum of hot migration data transfer volume and container reconstruction latency; The residual fluctuation of the load of each node after the execution of step t is measured by the mean of the standard deviation of utilization during the stabilization period after scheduling. The overall service quality score after scheduling is calculated by weighting response latency and request success rate. The preset service quality assurance threshold; This is an indicator function that takes a value of 1 and applies a penalty when the service quality is below a threshold. These are the weighting coefficients for equilibrium revenue, migration costs, volatility penalties, and service quality breach penalties, respectively. The improvement in the overall energy efficiency of the cloud platform after the execution of the t-th scheduling action is measured by the difference in host power consumption consumed by the unit computing task completed before and after scheduling. A positive value indicates that scheduling reduces energy consumption. The weighting coefficient for energy efficiency benefits is used to incorporate green energy-saving targets into the scheduling optimization direction; This is the number of cascaded reschedulings triggered by the t-th scheduling action. This value is incremented when a node's migration or scaling operation causes its neighboring nodes to become overloaded and triggers a new round of scheduling. This is the penalty coefficient for cascading scheduling, used to suppress aggressive scheduling actions that trigger cascading reactions when generated by the policy network.

10. A cloud resource dynamic scheduling system based on virtualization technology, used to execute the cloud resource dynamic scheduling method based on virtualization technology as described in any one of claims 1 to 9, characterized in that, include: Control processing module and resource nodes; The control processing module is configured as follows: The operating status of each resource node in the cloud platform is collected, and multi-dimensional timing features of CPU, memory, bandwidth and I / O are extracted. A resource topology intent-aware model is built based on graph neural networks to identify load dependencies and potential bottleneck intents between resource nodes. Using load dependency and bottleneck intent as input, and combining reinforcement learning policy network, a resource migration and elastic scaling decision sequence for future time windows is pre-generated before scheduling is triggered. The decision sequence includes the type of target virtualized resource, migration path and scaling level. Based on the decision sequence, hot migration or in-situ elastic scaling operations are performed on the target resource nodes specified in the decision sequence, and the corresponding traffic forwarding rules in the virtual network topology are updated synchronously. The actual resource utilization rate and service quality indicators after the scheduling execution are collected and fed back to the reinforcement learning policy network as reward and punishment signals to optimize the feature weights of the intent-aware model and the generation strategy of scheduling decisions.

Images