A cloud platform resource elastic scheduling method and device

Abstract

This invention relates to the field of cloud computing and intelligent resource scheduling technology, specifically providing a method and apparatus for elastic resource scheduling on a cloud platform. A monitoring and acquisition module collects metrics and topology information of hosts and VMs; a feature preprocessing module performs short-term smoothing, anomaly denoising, and holiday and event marking to provide cleaned input for prediction; a hybrid prediction module outputs the resource demand prediction and confidence interval for each VM within a future window T; a decision optimization module generates candidate action sequences, which are then selected and fine-tuned by an RL strategy; a security policy engine performs compliance and risk verification on the actions; the Executor executes the actions step by step on the cloud platform, performing canary deployments, monitoring checks, and rollbacks during execution; the Execution Coordinator and the Feedback Learning module store execution data, supporting offline replay training and online fine-tuning. Compared with existing technologies, this invention maximizes resource utilization and reduces operational costs while ensuring business continuity.

Description

Technical Field

[0001] This invention relates to the field of cloud computing and intelligent resource scheduling technology, specifically providing a method and apparatus for elastic scheduling of cloud platform resources. Background Technology

[0002] With the widespread application of cloud computing and virtualization technologies, enterprise applications have highly dynamic and uncertain demands for computing, storage, and network resources. Existing resource scheduling methods typically rely on threshold triggering mechanisms or static rules. When business experiences sudden growth, the response speed is insufficient, easily leading to resource shortages and service performance degradation. Conversely, when business is idle for extended periods, untimely resource recovery results in resource waste and increased operational costs.

[0003] While some studies have attempted to use machine learning methods for resource usage prediction, single prediction methods struggle to cope with the complex and ever-changing business models in multi-tenant environments. Furthermore, while methods such as reinforcement learning can optimize strategies, they lack an understanding of future trends and are prone to fluctuations in short-term decision-making.

[0004] Therefore, a joint scheduling method that integrates prediction and reinforcement learning is needed to achieve intelligent elastic scaling of resources and seamless resource adjustment without the business's awareness. Summary of the Invention

[0005] This invention addresses the shortcomings of the prior art by providing a highly practical method for elastic scheduling of cloud platform resources.

[0006] A further technical objective of this invention is to provide a reasonably designed, safe, and applicable cloud platform resource elastic scheduling device.

[0007] The technical solution adopted by this invention to solve its technical problem is: A cloud platform resource elastic scheduling method, wherein a monitoring and acquisition module collects metrics and topology information of hosts and VMs; The feature preprocessing module performs short-term smoothing, anomaly denoising, and holiday and publication event labeling to provide cleaned input for prediction. The hybrid forecasting module outputs the resource demand forecast and confidence interval for each VM within a future window T. The decision optimization module generates a sequence of candidate actions, which are then selected and fine-tuned by the RL strategy to make the final action. The security policy engine performs compliance and risk verification on actions, and the Executor executes the actions step by step on the cloud platform, performing canary deployments, monitoring checks, and rollbacks during the execution process. The Executor and the Feedback Learning Module store execution data, supporting offline replay training and online fine-tuning.

[0008] Furthermore, the hybrid prediction module uses a weighted ensemble predictor that combines a statistical model and a deep model to simultaneously output the predicted mean and an uncertainty measure. The specific structure is as follows: The Stat submodule: Periodic decomposition + ARIMA / ExponentialSmoothing, is used to capture trends and seasonality; The Deep submodule primarily uses LSTM or Transformer to capture complex nonlinearities and burst patterns. Gaussian Processes (GP) submodule: used for uncertainty estimation and small sample generalization; The final prediction is based on a dynamic weighting combination: ; The weights Ws, Wd, and Wg are adaptively updated based on recent model errors, and confidence scores are output simultaneously. The prediction results include time series predictions and confidence intervals for three dimensions: CPU, Memory, and disk I / O capacity.

[0009] Furthermore, in the hybrid prediction module, MPC provides safe and interpretable candidate action sequences based on prediction and system constraints; RL performs long-term utility optimization and adaptive selection under uncertain conditions in the candidate space or directly on the current policy. Specifically, it includes: (1) The MPC submodule solves the following constrained optimization problem on the prediction window TTT; (2) The reinforcement learning submodule defines the agent; (3) In the decision-making process, MPC provides a candidate set Ucand; RL scores the candidate set and selects the action that maximizes long-term expectations, or, under high uncertainty, makes the RL strategy revert to the conservative action determined by MPC.

[0010] Furthermore, solve the following constrained optimization problem on the prediction window TTT: Target: ; Where ut+τ represents the scaling up / down action; Cres is the resource cost; CSLA is the predicted SLA violation cost; Cstab penalizes frequent actions to avoid jitter; the rolling solution only executes the first action and recalculates in the next decision cycle to ensure real-time performance and controllability.

[0011] Furthermore, a reinforcement learning agent is defined as: State St: includes the index sequence of the most recent K windows, the prediction window y^t+1..t+T and the confidence vector σ, VM and host topology and historical action effect features; Action At: Discrete or continuous set; Reward Rt: Multi-objective weighted, typically in the form of: ; Here, riskt represents the probability of interruption, rollback, or migration overhead caused by scaling up or down. The algorithm adopts Actor-Critic and combines a "safety barrier" mechanism: actions that violate hard constraints in the RL output need to be corrected by MPC or the rule engine.

[0012] Furthermore, the security policy engine includes: (1) Calculation of the multi-factor score IdleScore; Calculate the IdleScore for each VM: ; Where, f i It is a normalization function that maps the metric to [0,1], where 0 represents active and 1 represents completely idle; the weight wi is set according to the business type. If the IdleScore exceeds the threshold Ithr and continues within the cooling time window Tidle, it will enter the recycling candidate pool. (2) Graded safety operation; Warning phase: The system automatically notifies the user and marks it as "about to be recycled". The user has the option to appeal or extend the retention period. Non-destructive downsizing: Prioritize safe online downsizing methods: CPU is reduced by decreasing vCPUs or by merging / migrating; memory is reclaimed through ballooning; disk unused blocks are reclaimed through TRIM / Thin-provision. Merging and Migrating: If direct reduction is too risky, consolidation is used to merge multiple low-load VMs into fewer hosts, freeing up host resources and page tables; the migration scheme follows the path with the minimum MPC computation cost. Resource pooling and reallocation: Released resources are recorded as "hot pool", and the scheduler prioritizes allocating resources from the hot pool in subsequent expansion requests; Rollback and Recovery: If a sudden user activation or execution anomaly is detected at any step, the system will trigger a rollback or reverse expansion to restore the original resources; The recovery action is constrained by the minimum retention time Tmin, the cooling time Tcool, and the maximum operating frequency.

[0013] Furthermore, vCPU hot-plugging: CPU hotplug is called through the libvirt interface, and the guest-agent verifies the guest OS's support for vCPU hot-add / hot-remove; if the guest does not support hot-remove, a migration solution is selected. Dynamic memory adjustment: prioritizes using the balloon driver to dynamically reclaim and allocate memory, directly increases memory when hot-add is supported, and strictly checks the guest's available memory and swap status when reducing memory and executes it in stages; Disk expansion and shrinking: Expansion uses libvirt and storage pool to expand the volume and execute guest internal filesystem grow. Shrinking prioritizes using thin provision + TRIM reclamation. If online shrinking is not possible, a data migration and merging process is used. Live-migration: When migration is required to achieve seamless scaling or consolidation, pre-copy and live-migration techniques are used, and the process is completed in stages under network / IO constraints. Idempotency and State Machines: Each operation is managed by a task-id, and the operation is a state machine: pending → executing → verify → success / failure → rollback; Observation window and grayscale strategy: After each operation, a short observation window is entered. If the indicator is abnormal, it will automatically roll back or trigger manual intervention.

[0014] A cloud platform resource elastic scheduling device includes: at least one memory and at least one processor; The at least one memory is used to store a machine-readable program; The at least one processor is used to call the machine-readable program to execute a cloud platform resource elastic scheduling method.

[0015] Compared with the prior art, the cloud platform resource elastic scheduling method and apparatus of the present invention have the following outstanding advantages: 1. By introducing a fusion model of time series forecasting and deep learning, the system can detect business load trends in advance and complete resource expansion before the business peak arrives, effectively avoiding performance degradation and service interruption caused by delayed resource allocation.

[0016] 2. The reinforcement learning algorithm is used to make dynamic decisions on resource expansion and contraction. It can automatically select the optimal expansion, contraction or resource maintenance strategy based on real-time operating status, predicted load and business SLA requirements. This reduces the reliance on thresholds and rules in traditional operation and maintenance and greatly improves the system's intelligence and adaptability.

[0017] 3. This invention not only enables dynamic scaling of CPU and memory, but also includes hard disk resources in the scope of seamless expansion and contraction. Combined with the online resource adjustment capabilities of the Inspur Cloud Platform, it ensures uninterrupted business operation and truly achieves seamless full-stack adjustment of core resources such as computing and storage.

[0018] 4. When the system detects that a certain business virtual machine is running at low load for a long time, it can automatically reduce its CPU, memory and disk resource usage, and dynamically expand the released resources to other high-load businesses, thereby achieving resource pool rebalancing and maximizing overall utilization.

[0019] 5. By leveraging the virtualization capabilities provided by the cloud platform, including vCPU hot-swapping, dynamic memory adjustment, disk expansion / shrinking, and online migration, this invention can ensure the continuity and stability of services during resource expansion / shrinking and migration, and users are unaware of resource changes.

[0020] 6. This invention avoids long-term idleness or over-configuration of resources through intelligent prediction and automated scheduling, significantly improves the utilization rate of cluster resources, reduces server energy consumption and operation and maintenance costs, and enhances the economic benefits of cloud platforms in large-scale business scenarios.

[0021] 7. While ensuring high utilization, this invention designs a reward function with SLA indicators as the core, so that the system can optimize resource allocation while always ensuring business response time and service stability, thereby improving user experience.

[0022] 8. This invention significantly improves the interpretability of complex decisions by fine-tuning a large language model and using it for decision support and online learning. This enables operations and maintenance personnel to interact with the system and understand its behavior using natural language. Simultaneously, the system possesses the ability to quickly learn from human feedback and rare events, enhancing its generalization and robustness in dealing with unknown scenarios, thus achieving an evolution towards truly intelligent and human-centered operations and maintenance. Attached Figure Description

[0023] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0024] Figure 1 This is a flowchart illustrating a cloud platform resource elastic scheduling method; Figure 2 This is an architecture diagram of a cloud platform resource elastic scheduling method. Detailed Implementation

[0025] To enable those skilled in the art to better understand the present invention, the present invention will be further described in detail below with reference to specific embodiments. Obviously, the described embodiments are merely 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.

[0026] The following is a preferred embodiment: like Figure 1 , 2 As shown in this embodiment, a cloud platform resource elastic scheduling method is proposed, in which the monitoring and acquisition module collects the indicators and topology information of the host and VM; The feature preprocessing module performs short-term smoothing, anomaly denoising, and holiday and publication event labeling to provide cleaned input for prediction. The hybrid forecasting module outputs the resource demand forecast and confidence interval for each VM within a future window T. The decision optimization module generates a sequence of candidate actions, which are then selected and fine-tuned by the RL strategy to make the final action. The security policy engine performs compliance and risk verification on actions, and the Executor executes the actions step by step on the cloud platform, performing canary deployments, monitoring checks, and rollbacks during the execution process. The Executor and the Feedback Learning Module store execution data, supporting offline replay training and online fine-tuning.

[0027] In the hybrid prediction module, to achieve both forward-looking expansion and conservative shrinkage, this invention uses a weighted ensemble predictor, combining a statistical model and a deep model to simultaneously output the predicted mean and an uncertainty measure. The specific structure is as follows: The Stat submodule (Stat): Periodic decomposition + ARIMA / ExponentialSmoothing, used to capture trends and seasonality; Deep submodule: Primarily uses LSTM / Transformer to capture complex nonlinearities and burst patterns; Gaussian Processes (GP) submodule: used for uncertainty estimation and small sample generalization; The final prediction is based on a dynamic weighting combination: ; The weights Ws, Wd, and Wg are adaptively updated based on recent model errors, and confidence scores are output simultaneously. The prediction results include time-series predictions and confidence intervals for three dimensions: CPU, Memory, and disk I / O capacity, which are provided to downstream decision-making modules for risk perception and robust optimization.

[0028] In the decision optimization module, MPC provides safe and interpretable candidate action sequences based on prediction and system constraints; RL performs long-term utility optimization and adaptive selection under uncertain conditions in the candidate space or directly on the current policy.

[0029] The specific explanation is as follows: (1) MPC submodule (security assurance, rolling optimization); MPC solves the following constrained optimization problem over the prediction window TTT: Target: ; Where ut+τ represents scaling operations (vCPU hot-plugging, dynamic content adjustment, volume expansion, live_migrate directive, etc.); Cres is the resource cost (rental or energy consumption, etc.); CSLA is the predicted SLA violation cost; Cstab penalizes frequent operations to avoid jitter. Constraints include resource pool capacity, single scaling rate, tenant quotas, data consistency limits, etc. Rolling solver executes only the first step and recalculates in the next decision cycle to ensure real-time performance and controllability.

[0030] (2) Reinforcement Learning Submodule (Long-term Benefits and Strategy Optimization); A reinforcement learning agent is defined as: State St: includes the index sequence of the most recent K windows, the prediction window y^t+1..t+T and the confidence vector σ, VM and host topology, historical action effect features, etc. Action At: Discrete / continuous set, such as increasing / decreasing k vCPUs, increasing / decreasing mMB of memory, expanding / shrinking d GB of hard disk, live_migrate(to_host), no-op; Reward Rt: Multi-objective weighted, typically in the form of: ; Where risk represents the probability of interruption / rollback or migration overhead caused by scaling up or down. The algorithm adopts an Actor-Critic approach (such as PPO or CPO with constraints) and combines it with a "Safety Shield" mechanism: any action that may violate hard constraints in the RL output must be corrected by MPC or the rule engine.

[0031] (3) In the decision-making process, MPC provides a candidate set Ucand; RL scores the candidate set and selects the action that maximizes long-term expectations, or, under high uncertainty, causes the RL strategy to revert to the conservative action determined by MPC. This hybrid mechanism ensures both theoretical controllability and enhances long-term optimization capability.

[0032] To achieve the goal of "reducing configuration and expanding to other resources when business operations are not used for a long time", the security policy engine has designed a multi-factor IdleScore and a phased recycling process. (1) IdleScore calculation (multi-factor scoring); Calculate the IdleScore for each VM: ; Where f i It is a normalization function (mapping the metric to [0,1], where 0 represents active and 1 represents completely idle); the weight wi can be set according to the business type. If the IdleScore exceeds the threshold Ithr and continues within the cooldown time window Tidle, it will enter the recycling candidate pool.

[0033] (2) Recycling and reuse strategy (tiered safety operation); Recycling adopts a tiered strategy: Warning phase: The system will automatically notify the user (via portal message / email) and mark it "soon to be reclaimed". The user has a retention period and can appeal or extend the retention time.

[0034] Non-destructive downsizing: Prioritize safe online downsizing methods: CPU is reduced by decreasing vCPU (if hot-remove is supported) or by merging / migrating; memory is reclaimed by ballooning; disk is reclaimed by reclaiming unused blocks via TRIM / Thin-provision.

[0035] Merge / Migration Release: If direct reduction is too risky, use consolidation: merge multiple low-load VMs into fewer hosts (live-migration), releasing host resources and page tables; the migration scheme uses the path with the minimum MPC computation cost.

[0036] Resource pooling and reallocation: Released resources are recorded as "hot pool". The scheduler prioritizes allocating from the hot pool in subsequent expansion requests, reducing new allocation latency and lowering overall energy consumption.

[0037] Rollback and Recovery: If a sudden user activation (load recovery) or execution anomaly is detected at any step, the system can trigger a rollback or reverse expansion to restore the original resources.

[0038] The recovery action is constrained by the minimum retention time Tmin, the cooling time Tcool, and the maximum operating frequency to prevent jitter.

[0039] The actuator is responsible for action decomposition, canary execution, monitoring, and rollback. For the Inspur Cloud Platform (based on KVM virtualization) environment, the following key operations and safeguards are supported: vCPU hot-plugging: CPU hotplug is called through the libvirt interface, and the guest-agent is used to verify the guest OS's support for vCPU hot-add / hot-remove; if the guest does not support hot-remove, a migration solution is selected.

[0040] Dynamic memory adjustment: Prioritize the use of the balloon driver to dynamically reclaim / allocate memory; directly increase memory when hot-add is supported; strictly check guest available memory and swap status when reducing memory and execute in stages.

[0041] Disk expansion / shrinkage: Expansion uses libvirt / storage pool to expand the volume and executes guest internal filesystem grow; shrinkage prioritizes thin provision + TRIM reclamation; if online shrinkage is not possible, a data migration and merging process is used (creating new small volumes, incremental copying, and switching).

[0042] Live-migration: When migration is required to achieve seamless scaling or consolidation, pre-copy / live-migration technology is used, and the process is completed in stages under network / IO constraints.

[0043] Idempotency and State Machines: Each operation is managed by a task-id, and the operation is a state machine (pending→executing→verify→success / failure→rollback), supporting retries and idempotent execution.

[0044] Observation window and grayscale strategy: After each operation, a short observation window (e.g., 30s-5min) is entered. If the indicator is abnormal, it will automatically roll back or trigger manual intervention.

[0045] The system integrates a security policy engine, providing permission verification, quota control, audit logs, and compliance checks. High-risk actions (such as disk reduction and primary / secondary write switching) require mandatory manual approval or an access whitelist in the initial stages. All action records can be rolled back to snapshots / incremental backups with a single click to meet audit requirements.

[0046] Each decision and execution's inputs (monitoring, prediction, candidate actions), outputs (actual actions performed, costs, business impact), and environmental context (such as holidays, application release events, tenant types) are structured and recorded in the decision feedback repository. Two types of training are performed periodically: Offline replay training: Utilize massive historical execution data to improve the reinforcement learning model through offline policy and Monte Carlo simulation evaluation, avoiding dangerous exploration in real-world environments and stabilizing policy performance; Online security fine-tuning: Supervised fine-tuning (SFT) is performed using historical decision data, records of manual interventions by operations personnel (such as reasons for approval / rejection, manually executed scaling operations), and operations reports and event summaries written in natural language. The training model learns the inherent logic of scheduling strategies and expertise in the operations domain, enabling it to understand "why making a certain decision is good or bad in a certain state."

[0047] When there is uncertainty in the prediction, a variance penalty term is added to the objective function of MPC: ; When confidence is low, a conservative strategy (reserving redundancy) is adopted, and high-risk actions are penalized through the risk term of RL. The system uses multi-source confirmation (prediction + sudden indicators + application-level alarms) for peak detection to reduce false positives.

[0048] To verify the effectiveness, the key evaluation indicators suggested in this invention include: SLA achievement rate, scaling up / down response time (from decision to effect), scaling up / down success rate, rollback frequency, resource utilization improvement, resource reuse rate after scaling down, system trigger false reclamation rate, and system stability (number of jitters).

[0049] To mitigate risks, a phased rollout is recommended: first implement the expansion-only (safety) path—disk / CPU hot-add and memory hot-add—and verify it on non-critical instances; then roll out the shrinking and reuse path, starting with ballooning + thin-provisioning reclamation, gradually progressing to merging migration and automatic reduction. Manual approval thresholds and rollback mechanisms should be retained throughout all phases.

[0050] Based on the above method, a cloud platform resource elastic scheduling device in this embodiment includes: at least one memory and at least one processor; The at least one memory is used to store a machine-readable program; The at least one processor is used to call the machine-readable program to execute a cloud platform resource elastic scheduling method.

[0051] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for elastic scheduling of cloud platform resources, characterized in that, The monitoring and acquisition module collects metrics and topology information of the host and VMs; The feature preprocessing module performs short-term smoothing, anomaly denoising, and holiday and publication event labeling to provide cleaned input for prediction. The hybrid forecasting module outputs the resource demand forecast and confidence interval for each VM within a future window T. The decision optimization module generates a sequence of candidate actions, which are then selected and fine-tuned by the RL strategy to make the final action. The security policy engine performs compliance and risk verification on actions, and the Executor executes the actions step by step on the cloud platform, performing canary deployments, monitoring checks, and rollbacks during the execution process. The Executor and the Feedback Learning Module store execution data, supporting offline replay training and online fine-tuning.

2. The cloud platform resource elastic scheduling method according to claim 1, characterized in that, The hybrid prediction module uses a weighted ensemble predictor that combines a statistical model and a deep model to simultaneously output the predicted mean and an uncertainty measure. The specific structure is as follows: The Stat submodule: Periodic decomposition + ARIMA / ExponentialSmoothing, is used to capture trends and seasonality; The Deep submodule primarily uses LSTM or Transformer to capture complex nonlinearities and burst patterns. Gaussian Processes (GP) submodule: used for uncertainty estimation and small sample generalization; The final prediction is based on a dynamic weighting combination: ； The weights Ws, Wd, and Wg are adaptively updated based on recent model errors, and confidence scores are output simultaneously. The prediction results include time series predictions and confidence intervals for three dimensions: CPU, Memory, and disk I / O capacity.

3. The cloud platform resource elastic scheduling method according to claim 2, characterized in that, In the hybrid prediction module, MPC provides a safe and interpretable sequence of candidate actions based on prediction and system constraints; RL performs long-term utility optimization and adaptive selection under uncertain conditions in the candidate space or directly on the current policy. Specifically, it includes: (1) The MPC submodule solves the following constrained optimization problem on the prediction window TTT; (2) The reinforcement learning submodule defines the agent; (3) In the decision-making process, MPC provides a candidate set Ucand; RL scores the candidate set and selects the action that maximizes long-term expectations, or, under high uncertainty, makes the RL strategy revert to the conservative action determined by MPC.

4. The cloud platform resource elastic scheduling method according to claim 3, characterized in that, Solve the following constrained optimization problem on the prediction window TTT: Target: ； Where ut+τ represents the scaling up / down action; Cres is the resource cost; CSLA is the predicted SLA violation cost; Cstab penalizes frequent actions to avoid jitter; the rolling solution only executes the first action and recalculates in the next decision cycle to ensure real-time performance and controllability.

5. The cloud platform resource elastic scheduling method according to claim 4, characterized in that, A reinforcement learning agent is defined as: State St: includes the index sequence of the most recent K windows, the prediction window y^t+1..t+T and the confidence vector σ, VM and host topology and historical action effect features; Action At: Discrete or continuous set; Reward Rt: Multi-objective weighted, typically in the form of: ； Here, riskt represents the probability of interruption, rollback, or migration overhead caused by scaling up or down. The algorithm adopts Actor-Critic and combines a "safety barrier" mechanism: actions that violate hard constraints in the RL output need to be corrected by MPC or the rule engine.

6. The cloud platform resource elastic scheduling method according to claim 1, characterized in that, The security policy engine includes: (1) Calculation of the multi-factor score IdleScore; Calculate the IdleScore for each VM: ； Where, f i It is a normalization function that maps the metric to [0,1], where 0 represents active and 1 represents completely idle; the weight wi is set according to the business type. If the IdleScore exceeds the threshold Ithr and continues within the cooling time window Tidle, it will enter the recycling candidate pool. (2) Graded safety operation; Warning phase: The system automatically notifies the user and marks it as "about to be recycled". The user has the option to appeal or extend the retention period. Non-destructive downsizing: Prioritize safe online downsizing methods: CPU is reduced by decreasing vCPUs or by merging / migrating; memory is reclaimed through ballooning; disk unused blocks are reclaimed through TRIM / Thin-provision. Merging and Migrating: If direct reduction is too risky, consolidation is used to merge multiple low-load VMs into fewer hosts, freeing up host resources and page tables; the migration scheme follows the path with the minimum MPC computation cost. Resource pooling and reallocation: Released resources are recorded as "hot pool", and the scheduler prioritizes allocating resources from the hot pool in subsequent expansion requests; Rollback and Recovery: If a sudden user activation or execution anomaly is detected at any step, the system will trigger a rollback or reverse expansion to restore the original resources; The recovery action is constrained by the minimum retention time Tmin, the cooling time Tcool, and the maximum operating frequency.

7. The cloud platform resource elastic scheduling method according to claim 1, characterized in that, vCPU hot-plugging: CPU hotplug is called through the libvirt interface, and guest-agent is used to verify guestOS's support for vCPU hot-add / hot-remove; if guest does not support hot-remove, a migration solution is selected; Dynamic memory adjustment: prioritizes using the balloon driver to dynamically reclaim and allocate memory, directly increases memory when hot-add is supported, and strictly checks the guest's available memory and swap status when reducing memory and executes it in stages; Disk expansion and shrinking: Expansion uses libvirt and storage pool to expand the volume and execute guest internal filesystem grow. Shrinking prioritizes using thin provision + TRIM reclamation. If online shrinking is not possible, a data migration and merging process is used. Live-migration: When migration is required to achieve seamless scaling or consolidation, pre-copy and live-migration techniques are used, and the process is completed in stages under network / IO constraints. Idempotency and State Machines: Each operation is managed by a task-id, and the operation is a state machine: pending → executing → verify → success / failure → rollback; Observation window and grayscale strategy: After each operation, a short observation window is entered. If the indicator is abnormal, it will automatically roll back or trigger manual intervention.

8. A cloud platform resource elastic scheduling device, characterized in that, include: At least one memory and at least one processor; The at least one memory is used to store a machine-readable program; The at least one processor is configured to invoke the machine-readable program to perform the method according to any one of claims 1 to 7.

Images