An AI computing power dynamic scheduling method and system based on a hybrid cloud architecture

Abstract

The application discloses an AI computing power dynamic scheduling method and system based on a hybrid cloud architecture and belongs to the technical field of cloud computing. By constructing a hybrid cloud resource real-time situation atlas, heterogeneous resource unified modeling is realized. In combination with a hybrid decision mechanism of deep reinforcement learning and heuristic rules, the scheduling intelligence and robustness are improved. Full-link monitoring and dynamic rescheduling are introduced to guarantee task SLA. A 'decision-execution-feedback-learning' closed loop is constructed to realize model continuous evolution. SDN and a cost model are fused to optimize cross-cloud transmission paths and costs. The system realizes global perception, intelligent decision, dynamic adaptation and autonomous optimization, and effectively improves AI computing power scheduling efficiency and economy.

Description

Technical Field

[0001] This invention relates to the field of cloud computing technology, and in particular to a method and system for dynamic scheduling of AI computing power based on a hybrid cloud architecture. Background Technology

[0002] Driven by the large-scale deployment of artificial intelligence applications, hybrid cloud architecture, which integrates the elasticity of public cloud, the security of private cloud, and the existing resources of local data centers, has become the mainstream infrastructure for supporting enterprise AI businesses. This architecture aims to achieve the best balance between resource utilization, cost control, and compliance and security, but its inherent heterogeneity, distribution, and dynamism pose new challenges to the efficient scheduling of AI computing power.

[0003] Existing technologies have significant limitations in addressing hybrid cloud scheduling scenarios. For example, the computing power scheduling system proposed in patent CN119576590A, while capable of optimizing resource allocation through intelligent matching, is primarily designed for homogeneous computing centers and struggles to characterize the heterogeneity and differentiated costs of resources across service providers in hybrid clouds, resulting in insufficient global optimization capabilities. Patent CN120743511A utilizes a dual-stream neural network and cat swarm algorithm to improve the adaptability of scheduling within a single cloud platform, but its strategy generation mechanism does not cover the collaborative orchestration and dynamic migration needs across public clouds, private clouds, and local data centers, making it unsuitable for the dynamic environment of hybrid clouds. Patents CN121070513A and CN121078045A focus on GPU cluster internal scheduling and computing network collaboration, respectively. The former lacks consideration for cross-cloud wide area network transmission factors, while the latter fails to deeply design an integrated decision-making framework for cost, performance, and security in hybrid clouds.

[0004] In summary, current technologies struggle to achieve refined, adaptive, and globally optimized scheduling of AI computing power in hybrid cloud environments, easily leading to uneven resource utilization, fluctuating task execution efficiency, and uncontrolled overall costs. Therefore, there is an urgent need for a new method and system capable of uniformly managing heterogeneous resources, dynamically adapting to complex tasks, intelligently optimizing cross-cloud network paths, and making collaborative decisions under multi-objective constraints. Summary of the Invention

[0005] To address the technical problems existing in the prior art, this invention provides a method for dynamic scheduling of AI computing power based on a hybrid cloud architecture, comprising the following steps: S101. Unified management and real-time situational awareness of hybrid cloud resources: By deploying monitoring probes in resource pools of public cloud, private cloud and local data center, real-time data collection of computing power resources, network status data, cost and billing data is carried out, and a globally unified real-time situational map of hybrid cloud resources is built and updated through data aggregation service. S102, AI Task Feature Extraction and Structured Profile Construction: Receive AI tasks to be scheduled, analyze task requirements through the feature extraction engine, extract computing power requirements, data features, performance and SLA requirements, as well as strategy and constraint features, and generate a structured task profile. S103. Intelligent cross-cloud scheduling decision based on multi-objective optimization: Based on the real-time situation map of the hybrid cloud resources and the structured task profile, candidate scheduling schemes are generated using a deep reinforcement learning policy network and a heuristic rule engine based on a preset multi-objective optimization function. The optimal scheduling decision is output through simulation evaluation and multi-objective decision arbitration. S104. Fine-grained strategy execution and cross-cloud routing configuration: Analyze the optimal scheduling decision, apply for and configure resources in each target cloud environment through the cloud resource operation adapter, configure cross-cloud data transmission path through the software-defined network controller, and deploy and start AI tasks through the task deployment coordinator. S105. Runtime end-to-end monitoring and dynamic rescheduling: During task execution, the end-to-end indicator collector monitors task performance and resource consumption, and the intelligent alarm and root cause analyzer detect anomalies. When the rescheduling conditions are triggered, the dynamic rescheduling process is initiated. S106. Closed-loop evaluation of scheduling effect and iterative optimization of strategy: After the task is completed, the actual execution data is collected and compared with the scheduling prediction value. Based on the analysis results, training samples are constructed to train the deep reinforcement learning policy network offline and update it online.

[0006] As a preferred embodiment of this application, step S101, which involves constructing the real-time situation map of the hybrid cloud resources, specifically includes: collecting CPU / GPU utilization, memory usage, network latency and bandwidth data, as well as the real-time unit price of public cloud and the converted cost of private resources, at fixed intervals using monitoring probes deployed in each resource pool; sending the collected data to a data aggregation service for cleaning, standardization, and timestamp alignment; synchronously writing the processed data into a time-series database and updating the global resource topology map maintained in memory or a graph database, wherein the global resource topology map represents resource entities with nodes, represents the connection relationships between entities with edges, and includes real-time updated resource attributes.

[0007] As a preferred embodiment of this application, step S102, generating the structured task profile specifically includes: parsing the task submission request and extracting basic information such as task type, container image, and data input / output path; extracting hardware accelerator type and quantity requirements, data geographical location and scale, latency tolerance and expected completion time, and security compliance and cost budget constraints through the computing power demand analyzer, data feature analyzer, performance and SLA analyzer, and policy and constraint analyzer in the feature extraction engine; and encoding the extracted multidimensional features into a structured task profile object and storing it persistently.

[0008] As a preferred embodiment of this application, step S103, specifically the intelligent cross-cloud scheduling decision, includes: formalizing the scheduling problem into a multi-objective constrained optimization problem with resource utilization, task completion time, and total cost of ownership as optimization objectives; calling a deep reinforcement learning policy network, using the concatenation of the graph neural network embedding vector of the resource situation map and the task profile feature vector as input to generate the first batch of candidate scheduling schemes; running a heuristic rule engine in parallel, executing predefined rules based on data localization and cost priority to generate the second batch of candidate scheduling schemes; using a lightweight simulator trained on historical data to predict the execution time, resource impact, and cost of each candidate scheme; and using a weighted scoring method or Pareto optimal frontier algorithm to comprehensively weigh the candidate schemes based on preset global policy preferences and task priorities, and outputting the optimal scheduling decision.

[0009] As a preferred embodiment of this application, in step S105, the dynamic rescheduling specifically includes: when the intelligent alarm and root cause analyzer detects a serious performance deviation, resource failure, or cost overrun risk, the dynamic rescheduling trigger is activated; the current task running status snapshot, root cause analysis results, and remaining SLA requirements are packaged into a rescheduling request and sent to the scheduling decision engine; the scheduling decision engine re-executes the scheduling decision process starting from the current moment to generate a revised scheduling scheme; the policy execution and routing module executes the revised scheduling scheme and achieves task migration through checkpoint recovery, instance migration, or rolling restart.

[0010] As a preferred embodiment of this application, step S106, specifically the closed-loop evaluation of scheduling effect and iterative optimization of strategy, includes: after the task is completed, summarizing the actual execution time, resource consumption, and total cost from monitoring data and billing; comparing the actual data with the predicted values ​​in step S103, calculating the prediction error, and evaluating the scheduling quality; based on the comparison results, constructing training samples containing a quadruple of state, action, reward, and new state according to the reinforcement learning sample format, and storing them in the experience replay buffer; periodically sampling data from the experience replay buffer, and using a deep reinforcement learning algorithm to train the policy network offline; and securely updating the new model after training and verification to the production environment through canary deployment.

[0011] This application also provides an AI computing power dynamic scheduling system based on a hybrid cloud architecture to implement the method described above. The system adopts a microservice architecture and includes the following core components that work together in a highly cohesive and loosely coupled manner: The resource situation awareness module is used to collect data through monitoring probes deployed in each resource pool, and to build and update a real-time situation map of hybrid cloud resources through the data aggregation service. The task profiling and modeling module is used to receive and parse AI tasks, extract multi-dimensional features through the feature extraction engine, and generate structured task profiles. A cross-cloud scheduling decision engine is used to generate optimal scheduling decisions based on resource status maps and task profiles, through a multi-objective optimization model, a deep reinforcement learning scheduler, a scheme simulation evaluator, and a decision arbitrator. The policy execution and routing module is used to parse scheduling decisions and execute resource provisioning, network configuration and task deployment through the cloud resource operation adapter, software-defined network controller and task deployment coordinator; The runtime monitoring and dynamic adjustment module is used to monitor, alarm, and reschedule running tasks through a full-link indicator collector, intelligent alarm and root cause analyzer, and dynamic rescheduling trigger. The cost analysis and optimization module is used to aggregate cost data, predict task expenses, and manage budget strategies. The policy feedback and model training module is used to collect scheduling effect data, construct training samples, and train and update the deep reinforcement learning model.

[0012] As a preferred embodiment of this application, the cross-cloud scheduling decision engine specifically includes: a multi-objective optimization model library, used to define and store mathematical models of optimization problems with resource utilization, task completion time, and total cost of ownership as objectives; a deep reinforcement learning scheduler, including a policy network and a value network, used to generate probability distributions of scheduling actions based on system state; a scheme simulation evaluator, used to load historical prediction models and quickly extrapolate and evaluate the execution effect of candidate scheduling schemes; and a decision arbitrator, used to perform comprehensive scoring or Pareto ranking of the evaluated candidate schemes based on global policy preferences, and output the final scheduling decision.

[0013] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. By constructing a real-time situational awareness map of hybrid cloud resources, this invention solves the problems of data silos and the lack of a unified view for heterogeneous resources, achieving logical centralization and dynamic perception of global resources. In step S101, this invention introduces a resource situational awareness module, deploying lightweight monitoring probes to collect real-time data on multi-source heterogeneous computing power (e.g., GPU utilization), network (e.g., bandwidth latency), and cost (e.g., real-time unit price). Through data cleaning and standardization, the data is unified into a structured data model, and a real-time situational awareness map of hybrid cloud resources is constructed based on graph database technology, expressing resource entities and their connections in the form of nodes and edges. This map supports second-level updates and topology reasoning, enabling dynamic modeling of cross-cloud resources. Thus, it achieves logical centralized management of physically distributed resources, providing a global, consistent, and high-fidelity data foundation, completely eliminating scheduling blind spots, significantly improving the visibility and accuracy of resource scheduling, and providing reliable support for upper-level intelligent decision-making.

[0014] 2. By employing a hybrid decision-making mechanism combining deep reinforcement learning and heuristic rules, this invention addresses the issues of poor robustness and decision imbalance in complex scenarios caused by single algorithms, achieving synergistic optimization of intelligent exploration and safety safeguards. In step S103, this invention innovatively designs a cross-cloud scheduling decision engine that integrates two decision paths: First, it uses a deep reinforcement learning strategy network, concatenating the embedding vectors of the resource situation map's graph neural network (GNN) with task profile features as input, combined with Monte Carlo Tree Search (MCTS) to generate forward-looking candidate solutions; second, it integrates a heuristic rule engine, incorporating expert rules such as cost priority, data localization, and SLA guarantees to generate a highly reliable baseline solution. A decision arbitrator performs sandbox simulations and weighted scoring on both types of solutions, comprehensively outputting the optimal scheduling strategy. Thus, it leverages the AI ​​model's ability to search for global optima in high-dimensional space while utilizing the rule engine to ensure stable system operation under data anomalies or emergency failures, significantly improving the accuracy, adaptability, and robustness of scheduling decisions, achieving a balance between intelligence and security.

[0015] 3. By employing a full-link monitoring and dynamic rescheduling mechanism, the problem of static scheduling's inability to respond to runtime anomalies is solved, achieving continuous support and dynamic SLA assurance throughout the task lifecycle. In step S105, this invention introduces a runtime monitoring and adaptive module. A full-link performance data acquisition device obtains real-time performance data from the infrastructure to the application layer, combined with intelligent alarms and root cause analyzers to accurately pinpoint the source of anomalies. When a severe deviation (such as GPU crashes or bandwidth congestion) or cost exceedance is detected, a dynamic rescheduling trigger is automatically activated, re-entering the current running status and remaining SLA into the decision-making process to generate migration or expansion plans. This achieves a leap from "static deployment" to "dynamic closed-loop," proactively identifying and automatically adjusting risks during task execution, significantly reducing task failure rates, effectively ensuring SLA achievement for high-priority AI tasks, and improving system fault tolerance and operational resilience.

[0016] 4. By constructing a closed-loop evolutionary system of "decision-execution-feedback-learning," the problem of scheduling model aging and reliance on manual parameter tuning is solved, enabling continuous intelligent iteration and self-optimization of the system. In step S106, this invention designs a policy feedback and model training module. After the task is completed, the predicted and actual execution data are automatically compared, a scheduling quality score is calculated, and reinforcement learning training samples (state-action-reward quadruplets) are constructed and stored in an experience replay buffer. The policy network is periodically trained offline, and new models are gradually deployed through canary deployment. Thus, the system can automatically extract experience and correct deviations from each scheduling practice, achieving continuous model evolution, avoiding frequent manual intervention, significantly reducing operation and maintenance costs, extending the system's lifespan, and possessing long-term adaptive capabilities.

[0017] 5. By employing a collaborative optimization mechanism combining SDN and cost models, the high cost and low path efficiency of cross-cloud data transmission are addressed, achieving coordinated scheduling of high performance and low cost. In step S104, this invention achieves deep collaboration between computing power and transportation capacity through a cross-cloud network orchestration module: during the decision-making phase, the cost analysis and optimization module uses cross-cloud traffic costs as a crucial constraint in a multi-objective optimization function; during the execution phase, the software-defined network controller dynamically configures the optimal route or SD-WAN link based on the data path planning in the work order, prioritizing transmission paths with ample bandwidth, low latency, and low unit price. Thus, while ensuring high throughput of AI training and inference data, it effectively avoids high traffic expenditures, significantly reducing the total cost of ownership (TCO) and achieving a win-win situation for both performance and cost.

[0018] In summary, this invention, through the systematic integration of the five innovative methods mentioned above, achieves a closed-loop scheduling process across the entire chain, from "resource perception—intelligent decision-making—dynamic execution—feedback learning—cost optimization." It overcomes multiple bottlenecks in the intelligence, dynamism, economy, and maintainability of traditional scheduling systems, demonstrating outstanding substantive features and significant progress. It not only improves the automation and intelligence level of AI computing power scheduling but also significantly enhances the system's stability, adaptability, and sustainability. It is suitable for large-scale, heterogeneous, and dynamically changing hybrid cloud AI computing scenarios, possessing promising industrial application prospects and widespread application value. Attached Figure Description

[0019] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with the invention and, together with the description, serve to explain the principles of the invention. 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, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0020] Figure 1 This is a flowchart of the AI ​​computing power dynamic scheduling method provided in the embodiments of the present invention.

[0021] Figure 2 This is a flowchart of the intelligent cross-cloud scheduling decision provided in an embodiment of the present invention.

[0022] Figure 3 This is a schematic diagram of policy execution and cross-cloud routing configuration provided in an embodiment of the present invention.

[0023] Figure 4 This is a flowchart of runtime monitoring and dynamic rescheduling provided in an embodiment of the present invention.

[0024] Figure 5 This is a flowchart of closed-loop evaluation and strategy iterative optimization provided in the embodiments of the present invention.

[0025] Figure 6 This is an architecture diagram of the AI ​​computing power dynamic scheduling system provided in the embodiments of the present invention. Detailed Implementation

[0026] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0027] It should be noted that all directional indications (such as up, down, left, right, front, back, etc.) in the embodiments of the present invention are only used to explain the relative positional relationship and movement of each component in a certain specific posture (as shown in the figure). If the specific posture changes, the directional indication will also change accordingly.

[0028] Furthermore, the use of terms such as "first" and "second" in this invention is for descriptive purposes only and should not be construed as indicating or implying their relative importance or implicitly specifying the number of technical features indicated. Therefore, features defined with "first" and "second" may explicitly or implicitly include at least one of those features. Additionally, the technical solutions of the various embodiments can be combined with each other, but only on the basis of being achievable by those skilled in the art. When the combination of technical solutions is contradictory or impossible to implement, such a combination of technical solutions should be considered non-existent and not within the scope of protection claimed by this invention.

[0029] Example 1 This invention provides a dynamic AI computing power scheduling method based on a hybrid cloud architecture. The core of this method lies in constructing an adaptive closed loop of "perception-decision-execution-optimization" to address the complexity of AI computing power scheduling in a hybrid cloud environment. Figure 1 As shown, the specific steps include: S101, Unified Management and Real-time Situational Awareness of Hybrid Cloud Resources This step aims to break down information silos among resource pools in a hybrid cloud environment and establish a global, real-time, and unified resource view, providing an accurate data foundation for intelligent scheduling.

[0030] First, lightweight monitoring probes are deployed in each resource pool of the target hybrid cloud environment. For public and virtual private clouds, the probes are deployed as virtual machine images or containers, using the official SDK provided by the cloud service provider to call its cloud monitoring API to obtain metrics such as instance status, vCPU / GPU utilization, and network inbound / outbound bandwidth. For private clouds or container platforms, the probes are deployed as DaemonSets on each worker node, using cAdvisor to collect container resource usage and node-exporter to collect node-level metrics. For traditional on-premises data center physical machines or virtual machines, the probes collect hardware information by executing system commands via SSH or by installing an agent. All monitoring probes are configured with a unified authentication and secure communication mechanism to ensure the legitimacy and security of data collection.

[0031] Next, each monitoring probe collects multi-dimensional data in parallel at a fixed collection cycle. The data collection covers four dimensions: 1) Computing resource data, collecting the total number and utilization of CPU cores, GPU card model and memory utilization, total memory capacity and utilization, and local storage I / O throughput and latency of each computing node (virtual machine, physical machine, container node); 2) Network status data, collecting network latency between nodes (measured via ICMP), bandwidth availability (by calling the cloud provider's VPC bandwidth monitoring interface), and link status and packet loss rate of cross-cloud dedicated lines or peering connections; 3) Cost and billing data, obtaining the real-time unit price, historical usage, and estimated cost of different instance specifications (such as on-demand instances, reserved instances, and preemptible instances) by calling the public cloud service provider's pricing API and billing API. For private clouds and IDCs, calculate the unit time cost of resources based on a preset fixed cost model (such as depreciation, electricity costs, and operation and maintenance amortization); 4) Policies and metadata: collect security policies (such as firewall rules and security groups), compliance tags (such as security level and data region) and resource tags (such as project and business department) of each resource pool.

[0032] Then, the monitoring probes send all the raw data they collect to a centralized data aggregation service via a secure channel. The data aggregation service cleans the data (for example, using sliding window and Z-score methods to filter out outliers caused by momentary fluctuations) and transforms the diverse data from different cloud vendors into an internally unified, standardized data model. After cleaning and standardization, the data aggregation service adds a precise timestamp and data source identifier to each data entry.

[0033] Finally, the data aggregation service writes the processed data to a time-series database for persistent storage and simultaneously updates the resource status map in memory. The time-series database efficiently stores and queries all timestamped indicator sequences, supporting high-performance aggregation and time-range queries. The resource status map is a dynamic data structure, typically existing in the form of a graph database. Its nodes represent various resource entities (such as virtual machine instances, physical servers, switches, and storage volumes), and edges represent network connections, membership relationships, or data flow relationships between entities. Each node and edge is attached with real-time updated performance, cost, and status attributes. Through this step, the system constructs a global, multi-dimensional, integrated real-time view of hybrid cloud resources, providing accurate data input for subsequent intelligent scheduling.

[0034] S102, AI Task Feature Extraction and Structured Profile Construction The goal of this step is to transform the AI ​​computing tasks submitted by users, which may be vaguely described, into standardized feature vectors that the system can accurately understand, quantify, and process, i.e., "task profiles".

[0035] When a user submits an AI task request through the system's RESTful API or graphical console, the task receiver and parser first processes the request. The task receiver and parser parses the HTTP request body, extracting basic fields such as task ID, task type (e.g., model training, batch inference, online service), program entry point, and dependent container image addresses. The system allows users to provide richer descriptions of their requirements by attaching YAML configuration files, JSON annotations, or specific tags.

[0036] Subsequently, the feature extraction engine performs deep analysis on the task to extract multi-dimensional features. The feature extraction engine adopts a pluggable architecture, coordinating multiple dedicated analyzers to work in parallel through a lightweight workflow. The computing power requirement analyzer scans the Dockerfile or dependency files of the task container image, analyzes the deep learning frameworks and CUDA versions it references, and intelligently infers the task's requirements for hardware accelerator type, number of GPUs, number of CPU cores, and memory size by combining a library of resource consumption patterns from similar historical tasks. The data feature analyzer parses the data input and output paths declared in the task configuration, identifies the geographical location and storage type of the data, estimates the data size by pre-reading metadata or based on historical jobs, and determines whether the access mode is sequential streaming or random reading. The performance and SLA analyzer extracts the task's latency tolerance (e.g., interactive tasks require <100ms, batch tasks can accept >1 hour), expected execution time, and required reliability metrics based on user-specified deadlines, service level agreement documents, or by performing k-nearest neighbor matching with a library of similar historical tasks. The strategy and constraint analyzer extracts hard rules that tasks must follow based on the rule engine, including data security and compliance constraints (such as "data processing must not cross borders" and "must meet the requirements of Level 3 Information Security Protection"), cost budget constraints (maximum budget for a single task and monthly budget for the project), and resource preferences or prohibitions (such as "prioritize the use of domestically produced chips" and "prohibit the use of preemptive instances").

[0037] After extracting the aforementioned multi-dimensional features, the system encodes these features into a structured, machine-readable task profile object. The task profile object uses a serializable format such as JSON or Protocol Buffers and contains explicit key-value pair fields. Each task profile object is assigned a globally unique ID and persistently stored in the task profile repository. The task profile object not only serves this immediate scheduling but also provides valuable data support for subsequent similar task recommendations, resource capacity planning, and historical analysis of scheduling strategies.

[0038] S103, Intelligent cross-cloud scheduling decision based on multi-objective optimization This step is the core intelligent component of the method. It comprehensively utilizes the global resource situation and specific task profiles, and through simulation and optimization, generates one or a set of optimal cross-cloud resource allocation schemes. For example... Figure 2 As shown, the specific implementation is as follows: The scheduling decision engine defines the current scheduling request as a multi-objective constrained optimization problem. The decision variable of the multi-objective constrained optimization problem is the amount of resources allocated to the current task on each candidate resource node. Its optimization objectives are usually set as four mutually constraining aspects: 1) Maximize the overall resource utilization of the hybrid cloud cluster, and make the load of each node as even as possible through load balancing algorithms to avoid resource fragmentation; 2) Minimize the end-to-end completion time of the task, which requires comprehensive estimation of the execution time on the target resources during the calculation phase, as well as the necessary data cross-cloud / cross-region transmission time; 3) Minimize the total cost of ownership of task execution, covering public cloud resource rental fees, cross-cloud / cross-region data transmission traffic fees, and the model-based cost of private resources; 4) Meet all hard constraints, such as data localization compliance, budget limits, and compatibility with specified hardware architectures.

[0039] Next, the scheduling decision engine employs a hybrid generation mechanism that combines a deep reinforcement learning policy network and a heuristic rule engine in parallel to balance exploration and utilization. On one hand, the pre-trained deep reinforcement learning policy network is invoked. This network takes as input the concatenation of the graph neural network embedding vector of the current resource situation map and the feature vector of the task profile, and outputs a probability distribution through forward propagation. This probability distribution indicates the recommendation degree for scheduling tasks to each resource node. Based on this distribution, the system performs Monte Carlo tree search or Top-K sampling to generate a batch of innovative and forward-looking candidate solutions. On the other hand, a domain-knowledge-based heuristic rule engine runs concurrently. It executes a series of predefined, interpretable rules (such as "prioritizing scheduling computational tasks to the availability zone where the data is located," "selecting the instance type with the lowest unit computing power cost within budget constraints," and "reserving high-performance resources for high-priority tasks") to generate another batch of reliable and intuitive candidate solutions. The results generated by both are merged to form a set of candidate scheduling solutions that combine diversity and reliability.

[0040] For each generated candidate scheduling scheme, the system launches a lightweight simulator based on historical data for rapid evaluation. This simulator loads a regression prediction model trained from historical task execution logs, and uses the candidate scheme as input to quickly simulate the task execution process under that scheme. The simulator continuously estimates: the execution time of the computation phase on the target resource specifications; the network transmission time (based on current network status and dataset size) if data needs to be transferred across clouds or availability zones; the cumulative impact on the future load of the target node during task execution; and the final detailed cost list (combining resource unit price and usage duration). The simulator outputs a structured scheme evaluation report containing the above prediction indicators and confidence intervals.

[0041] Subsequently, a multi-objective decision arbitrator intervenes to comprehensively weigh and rank the various candidate solutions. Based on the system administrator's preset global scheduling strategy (e.g., "prioritizing performance and latency during weekday office hours, and prioritizing cost reduction at night and on weekends") and the implicit priorities and constraints in the current task profile, the arbitrator assigns dynamic weights to multiple optimization objectives, or employs a Pareto optimal frontier search algorithm. If a weighted method is used, the comprehensive score for each solution is calculated as: Score = w1*F1 (utilization) + w2*F2 (time) + w3*F3 (cost), where F is the normalized utility function of each indicator, and w is the dynamic weight. If Pareto optimization is used, the set of solutions (non-dominated solutions) that are superior to other solutions on at least one objective and not inferior on other objectives is identified.

[0042] Ultimately, the multi-objective decision arbitrator outputs the optimal or near-optimal scheduling decision. If a weighted scoring method is used, the solution with the highest overall score is selected; if Pareto optimization is used, several optimal solutions may be provided for the user or upper-layer system to make the final decision. The output decision is a detailed, executable blueprint that explicitly specifies: whether the task needs to be split into subtasks and the logic for the split; which cloud platform, specific region, availability zone, and resource instance ID each subtask is assigned to; the specifications and quantity of resources to be requested; the optimized path and endpoints for data flow; and the suggested task startup sequence and dependencies. This decision is encapsulated as a scheduling work order and handed over to the execution layer.

[0043] S104, Fine-grained policy enforcement and cross-cloud routing configuration This step is responsible for the actual deployment and launch of the task in the hybrid cloud environment, ensuring that intelligent scheduling decisions are accurately, efficiently, and reliably translated into reality. For example... Figure 3 As shown, the specific implementation is as follows: The policy execution module first receives and parses the scheduling work order, identifying the different cloud environments and resource operation sequences involved. For resources that need to be created or configured on the public cloud, the policy execution module calls the corresponding driver plugin in the cloud resource operation adapter. This driver plugin transforms the abstract resource configuration description into a series of specific, idempotent cloud service provider API calls, and automatically handles authentication, parameter assembly, asynchronous operation polling, and error retry logic. For resources within a private cloud or Kubernetes cluster, it creates the corresponding namespace, deployment, stateful replica set, or job by calling the cluster's API Server or the REST API of the internal resource management platform, and configures the corresponding resource requests and limits.

[0044] Meanwhile, the software-defined network controller begins its work based on the data path planning in the work order. If the compute task and data source are located in different availability zones within the same region of the same cloud service provider, the controller ensures, via API, that the routing policy of the cloud provider's intranet high-speed channels (such as VPC peering connections or cloud enterprise networks) is optimal, avoiding detours through the public network. If data transmission involves cross-cloud service providers or between the cloud and the on-premises data center, the controller calls the cloud networking service or software-defined wide area network (SD-WAN) controller via API to dynamically select or establish the best path that meets bandwidth and latency requirements based on the current network quality and cost, and may temporarily request or adjust dedicated line bandwidth for this purpose. The controller injects the finalized data access endpoint (such as a domain name or intranet IP address that has undergone global load balancing and acceleration) and necessary authentication information into the runtime environment of the task about to start, in the form of configuration files or environment variables.

[0045] After confirming that all target resources are ready and network connectivity tests are successful, the task deployment coordinator executes the final application deployment instructions. The coordinator securely pulls the task container image from the specified repository to the target node and starts one or more container or virtual machine instances according to the work order description. The coordinator injects key contextual information for this scheduling (such as task ID, cost center tag, project, and scheduling decision version) into the runtime environment as Kubernetes tags, annotations, or environment variables, providing a globally consistent traceability for subsequent monitoring, metering and billing, log aggregation, and troubleshooting. At this point, the AI ​​task has been successfully deployed and is executing in the complex hybrid cloud environment, strictly following intelligent scheduling decisions.

[0046] S105, Runtime end-to-end monitoring and dynamic rescheduling This step involves continuous monitoring and proactive intervention throughout the entire task lifecycle to ensure stable, efficient, and economical task operation, and to effectively handle various unforeseen circumstances. For example... Figure 4 As shown, the specific implementation is as follows: The moment a task starts, the end-to-end metrics collector establishes a dedicated, fine-grained monitoring view for it. It not only relies on infrastructure monitoring data streams in S101, but also collects application-level metrics within the task in a non-intrusive manner through automatic injection via the service mesh Sidecar or Application Performance Management (APM) proxy. These metrics include: queries per second (SPS) for the model inference service, average / quantile request latency, and error code distribution; and the progress percentage, data throughput, and checkpoint retention status of batch processing tasks. All these metrics, from underlying GPU temperature and SM utilization to middleware queue depth and application-layer business success rate, are strongly correlated with a unique task ID and are aggregated, analyzed, and compared against baselines on a unified monitoring dashboard.

[0047] Intelligent Alerts and Root Cause Analyzer (RCA) continuously scan these multi-dimensional metric streams. Its built-in Complex Event Processing (CEP) rule engine triggers alerts based on task SLAs and dynamic baselines (such as historical data from the same period), configuring thresholds and pattern rules (e.g., "GPU utilization consistently above 90% for 5 minutes accompanied by temperature alerts," "Cross-cloud network latency 95th quantile (P95) suddenly increases by 50ms for more than 5 minutes"). More importantly, its root cause analysis model utilizes real-time resource landscape mapping and topology relationships to attempt to pinpoint the source of the problem. For example, when an increase in inference latency is detected for a task container, the analyzer automatically checks: the overall CPU / memory / IO load of the node hosting the container; resource contention among other adjacent containers on the same node (especially in GPU-shared scenarios); and the network connection quality from the node to its dependent data services or downstream services compared to historical data. Through this graph-based correlation analysis, it can quickly determine whether the problem is localized (task-specific), node-level (resource contention or hardware failure), or network / service dependency-level.

[0048] When issues that cannot be self-healed by elastic scaling, such as significant deviations from expected goals (e.g., performance failures), critical resource failures (e.g., node outages), or impending cost overruns, are detected, the dynamic rescheduling trigger is automatically activated. The dynamic rescheduling trigger packages a snapshot of the current task's runtime state (including completed workload, current resource usage details, and cost details), the root cause analysis results that triggered the alarm, and the remaining task SLA requirements (e.g., remaining workload, remaining budget, and final deadline), into a high-priority rescheduling request and sends it directly back to the scheduling decision engine's input queue. The decision engine uses this moment as a new starting point, aiming to "optimally complete the remaining work using the remaining available resources / budget," and re-executes a rapid scheduling decision process. The new decision may generate a corrective solution, such as: migrating the task from an overheating node to a cooler node in the same region; switching from on-demand instances to lower-priced preemptive instances to control costs; or migrating some subtasks back from the public cloud to recovered private cloud resources. The policy execution module then implements this corrective measure, which may be achieved through container checkpoint recovery technology, virtual machine hot migration, or graceful rolling restart, to smoothly guide the task onto a better and more stable operating track while minimizing business interruption and data loss.

[0049] S106: Closed-loop evaluation of scheduling performance and iterative optimization of strategy This step is the core of the system's self-learning and continuous evolution. By learning from the results of each scheduling practice, it continuously corrects its internal model, thereby improving the accuracy and intelligence of future decisions. Figure 5 As shown, the specific implementation is as follows: Once the task is finally completed (whether successfully, failed, or terminated by the user), an automated post-task performance evaluation process is immediately initiated. The system aggregates the actual runtime performance data of the task from multiple data sources: total execution time from the monitoring module; resource consumption and traffic costs accurate to the second from the cloud platform's detailed billing API and internal billing system; and the final SLA results achieved (such as whether it was completed before the deadline and the number of successfully processed records) from task logs and exit codes. This data is then consolidated into a unified task finality report.

[0050] The cost analysis and optimization module performs a detailed comparative analysis of the actual performance data with the predicted values ​​from the task scheduling decision in stage S103. The module calculates the prediction error for each item (e.g., "predicted cost 300 yuan, actual cost 320 yuan, absolute error +20 yuan, relative error +6.7%"; "predicted time 2 hours, actual time 1.8 hours, relative error -10%"). Simultaneously, the analysis module evaluates whether dynamic rescheduling is triggered throughout the task lifecycle, the type of triggering reason (performance / cost / failure), and the effectiveness of the rescheduling decision (e.g., percentage performance improvement after migration, cost savings). These analysis results are used to calculate the "quality score" for this scheduling.

[0051] Subsequently, the experience sample constructor utilizes task final state reports and scheduling quality analysis to construct high-quality training samples for reinforcement learning. A sample typically contains a standard reinforcement learning quadruple: (state st, action at, reward rt, new state st+1). Here, st is the system state just before task scheduling, representing the resource situation map and task profile; at is the initial scheduling decision (action) taken by the system; rt is a scalar reward value, calculated by a pre-defined reward function based on the ratio of actual to expected time consumption, cost ratio, SLA achievement, and rescheduling overhead (a higher reward indicates a better-than-expected result, and a lower or negative reward indicates a worse result); st+1 can be the final state when the task is completed in one go, and the system state at the moment rescheduling is triggered. These high-value samples, rich in successful and failed experiences, are stored in a fixed-capacity experience replay buffer and updated according to the first-in, first-out (FIFO) principle.

[0052] The system periodically initiates offline training tasks during off-peak periods or when the experience pool reaches a certain size. The model training platform randomly samples a batch of data from the experience replay buffer and uses advanced deep reinforcement learning algorithms (such as Proximal Policy Optimization (PPO) and Soft Actor-Critic (SAC)) to train and update the parameters of the policy network (Actor) and value network (Critic) in the cross-cloud scheduling decision engine. The training objective is to enable the policy network to learn to select actions (at) that are expected to yield higher long-term cumulative rewards when observing more complex states (st). After training, the new model is evaluated on a validation set containing historical scheduling scenarios to ensure that its overall performance (such as prediction accuracy and constraint satisfaction rate) is no lower than that of the current online model.

[0053] Ultimately, the model management and release service updates the production system in a secure and controlled manner. It employs a canary release or shadow mode: first, the new model is deployed to a small portion (e.g., 5%) of the scheduling request traffic, and the differences between its decisions and those of the old model or actual results are compared in a real production environment; once the new model is confirmed to be stable and effective, the traffic proportion is gradually increased until the old model is completely replaced. Through this complete technical loop of "practice-evaluation-learning-update," the system's scheduling strategy is continuously iterated and optimized, becoming increasingly intelligent and accurately adapting to the unique characteristics and patterns of the dynamically changing hybrid cloud environment it serves.

[0054] Example 2 This invention also provides an AI computing power dynamic scheduling system based on a hybrid cloud architecture. The system adopts a microservice architecture and modular design, with components working collaboratively through high cohesion and low coupling. For example... Figure 6 As shown, it specifically includes: Resource Situation Awareness Module The resource situation awareness module, acting as the system's "sensory nerves," is responsible for collecting and fusing data from heterogeneous environments. It consists of three sub-components: a monitoring probe cluster, a data aggregation service, and a time-series database and graph engine. The monitoring probe cluster acts as "sensors" throughout the hybrid cloud, actively capturing raw data such as computing power, network speed, and cost. The data aggregation service serves as a "cleaning and translation center," standardizing the messy, multi-source data. The time-series database and graph engine together act as the "memory and cognitive brain," with the former recording all indicator changes in a time series and the latter constructing and updating in real-time a topological graph representing resource entities and their interrelationships, providing a structured and relational global situational view for upper-level decision-making.

[0055] Task profiling and modeling module The task profiling and modeling module acts as the system's "requirement translator," converting user intent into machine-understandable language. Its core components are the task receiver and parser, the feature extraction engine, and the task profiling storage. The task receiver and parser provides a unified input interface. The feature extraction engine is the module's intelligent core; by integrating various analysis plugins such as code analysis, historical pattern matching, and natural language processing, it deeply dissects the task, extracting multi-dimensional features including computing power, data, performance, and strategy. The task profiling storage persistently saves the final, structured task profiling object, establishing a knowledge archive for continuous use in scheduling, analysis, and learning.

[0056] Cross-cloud scheduling decision engine The cross-cloud scheduling decision engine, acting as the system's "intelligent decision-making hub," is the core component for generating optimized scheduling schemes. It comprises a multi-objective optimization model library, a deep reinforcement learning scheduler, a scheme simulation evaluator, and a decision arbitrator. The multi-objective optimization model library defines the various objectives pursued by scheduling and their mathematical expressions. The deep reinforcement learning scheduler is a self-learning agent capable of generating innovative scheduling suggestions based on historical experience. The scheme simulation evaluator is a rapid "sandbox" used to predict the execution effects of candidate schemes. The decision arbitrator is the final "adjudicator," comprehensively evaluating and weighing all schemes based on the global strategy and optimization objectives, outputting the optimal scheduling decision instruction. These four components form a highly efficient decision-making pipeline.

[0057] Policy Enforcement and Routing Module The policy execution and routing module, acting as the system's "execution and connection arm," is responsible for translating decisions into action. It consists of a cloud resource operation adapter, a software-defined network controller, and a task deployment coordinator. The cloud resource operation adapter encapsulates all the details of interaction with various cloud platforms, enabling agile resource provisioning and recycling. The software-defined network controller specializes in network connectivity, intelligently configuring cross-cloud data transmission paths. The task deployment coordinator acts as a "conductor," coordinating the former two, as well as lower-level orchestration systems (such as Kubernetes), to complete the entire chain of operations from resource preparation and network configuration to application deployment according to the correct sequence and dependencies.

[0058] Runtime monitoring and dynamic adjustment module The runtime monitoring and dynamic adjustment module acts as the system's "autopilot and escort system," ensuring smooth mission operation. It includes a full-link metrics collector, an intelligent alarm and root cause analyzer, and a dynamic rescheduling trigger. The full-link metrics collector constructs a three-dimensional monitoring system from infrastructure to the application layer. The intelligent alarm and root cause analyzer not only identifies problems but also uses resource mapping to attempt to pinpoint the root cause. The dynamic rescheduling trigger intervenes decisively when necessary, initiating a correction process to ensure the mission does not deviate from its predetermined objectives.

[0059] Cost Analysis and Optimization Module The cost analysis and optimization module acts as the system's "financial manager," focusing on economic efficiency. This module includes a cost aggregation and calculator, a cost forecasting model, and a budget and strategy manager. The cost aggregation and calculator enables accurate cost calculation and allocation. The cost forecasting model provides future cost projections for decision-making. The budget and strategy manager sets cost limits and optimization strategies to ensure that all scheduling activities are economically sustainable.

[0060] Policy Feedback and Model Training Module The policy feedback and model training module, serving as the system's "learning and evolution center," is crucial for achieving continuous intelligence. It comprises an experience sample constructor, a model training platform, and a model management and deployment service. The experience sample constructor extracts learning materials from completed scheduling practices. The model training platform uses these materials to train and optimize deep reinforcement learning models in an offline environment. The model management and deployment service is responsible for safely and smoothly updating the trained new models to the scheduling decision engine in the production environment, thereby completing the iterative upgrade of the system's intelligence.

[0061] The modules mentioned above do not operate in isolation, but are tightly connected through an event-driven architecture and synchronous service calls. Data flows in from the perception module, is transformed into a control flow by the decision engine, is applied to the environment by the execution module, changes in the environment state are captured by the monitoring module and form a feedback flow, and finally flow into the training module to complete the learning loop.

[0062] Example 3 The present invention also provides an electronic device, including: a processor, a transmitting device, an input device, an output device, and a memory. The processor may be implemented using a general-purpose CPU (Central Processing Unit), a microprocessor, an application-specific integrated circuit, or one or more integrated circuits, and is used to execute relevant programs to implement the technical solutions provided in the embodiments of this application. The memory may be implemented using a read-only memory (ROM), a static storage device, a dynamic storage device, or a random access memory (RAM), and is used to store computer program code. The computer program code includes computer instructions. When the processor executes the computer instructions, the electronic device executes a method as described in any of the above possible implementation methods.

[0063] Example 4 The present invention also provides a computer-readable storage medium storing a computer program, the computer program including program instructions, which, when executed by a processor of an electronic device, cause the processor to perform a method as described in any of the above possible implementations.

[0064] In the description of this specification, the references to terms such as "an embodiment," "example," "specific example," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0065] The above description is merely a specific embodiment of the present invention, enabling those skilled in the art to understand or implement the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the present invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features claimed herein.

Claims

1. A method for dynamic scheduling of AI computing power based on a hybrid cloud architecture, characterized in that, Includes the following steps: S101. Unified management and real-time situational awareness of hybrid cloud resources: By deploying monitoring probes in resource pools of public cloud, private cloud and local data center, real-time data collection of computing power resources, network status data and cost and billing data is carried out, and a globally unified real-time situational map of hybrid cloud resources is built and updated through data aggregation service. S102, AI Task Feature Extraction and Structured Profile Construction: Receive AI tasks to be scheduled, analyze task requirements through the feature extraction engine, extract computing power requirements, data features, performance and SLA requirements, as well as strategy and constraint features, and generate a structured task profile. S103. Intelligent cross-cloud scheduling decision based on multi-objective optimization: Based on the real-time situation map of the hybrid cloud resources and the structured task profile, candidate scheduling schemes are generated using a deep reinforcement learning policy network and a heuristic rule engine based on a preset multi-objective optimization function. The optimal scheduling decision is output through simulation evaluation and multi-objective decision arbitration. S104. Fine-grained strategy execution and cross-cloud routing configuration: Analyze the optimal scheduling decision, apply for and configure resources in each target cloud environment through the cloud resource operation adapter, configure cross-cloud data transmission path through the software-defined network controller, and deploy and start AI tasks through the task deployment coordinator. S105. Runtime end-to-end monitoring and dynamic rescheduling: During task execution, the end-to-end indicator collector monitors task performance and resource consumption, and the intelligent alarm and root cause analyzer detect anomalies. When the rescheduling conditions are triggered, the dynamic rescheduling process is initiated. S106. Closed-loop evaluation of scheduling effect and iterative optimization of strategy: After the task is completed, the actual execution data is collected and compared with the scheduling prediction value. Based on the analysis results, training samples are constructed to train the deep reinforcement learning policy network offline and update it online.

2. The method according to claim 1, characterized in that, In step S101, constructing the real-time situation map of hybrid cloud resources specifically includes: collecting CPU / GPU utilization, memory usage, network latency and bandwidth data, as well as the real-time unit price of public cloud and the converted cost of private resources, at fixed intervals through monitoring probes deployed in each resource pool; sending the collected data to a data aggregation service for cleaning, standardization and timestamp alignment; synchronously writing the processed data into a time-series database and updating the global resource topology map maintained in memory or a graph database, wherein the global resource topology map represents resource entities with nodes, represents the connection relationship between entities with edges, and includes real-time updated resource attributes.

3. The method according to claim 1, characterized in that, In step S102, generating the structured task profile specifically includes: parsing the task submission request and extracting basic information such as task type, container image, and data input / output path; extracting hardware accelerator type and quantity requirements, data geographic location and scale, latency tolerance and expected completion time, and security compliance and cost budget constraints through the computing power demand analyzer, data feature analyzer, performance and SLA analyzer, and policy and constraint analyzer in the feature extraction engine; and encoding the extracted multidimensional features into a structured task profile object and storing it persistently.

4. The method according to claim 1, characterized in that, In step S103, the intelligent cross-cloud scheduling decision specifically includes: formalizing the scheduling problem into a multi-objective constrained optimization problem with resource utilization, task completion time, and total cost of ownership as optimization objectives; calling a deep reinforcement learning policy network, using the concatenation of the graph neural network embedding vector of the resource situation map and the task profile feature vector as input to generate the first batch of candidate scheduling schemes; running a heuristic rule engine in parallel, executing predefined rules based on data localization and cost priority to generate the second batch of candidate scheduling schemes; using a lightweight simulator trained on historical data to predict the execution time, resource impact, and cost of each candidate scheme; and using a weighted scoring method or Pareto optimal frontier algorithm to comprehensively weigh the candidate schemes based on preset global policy preferences and task priorities, and outputting the optimal scheduling decision.

5. The method according to claim 1, characterized in that, In step S105, the dynamic rescheduling specifically includes: when the intelligent alarm and root cause analyzer detects a serious performance deviation, resource failure, or cost overrun risk, the dynamic rescheduling trigger is activated; the current task running status snapshot, root cause analysis results, and remaining SLA requirements are packaged into a rescheduling request and sent to the scheduling decision engine; the scheduling decision engine re-executes the scheduling decision process starting from the current moment to generate a revised scheduling scheme; the policy execution and routing module executes the revised scheduling scheme and achieves task migration through checkpoint recovery, instance migration, or rolling restart.

6. The method according to claim 1, characterized in that, In step S106, the closed-loop evaluation of scheduling effect and iterative optimization of strategy specifically includes: after the task is completed, summarizing the actual execution time, resource consumption and total cost from monitoring data and billing; comparing the actual data with the predicted values ​​in step S103, calculating the prediction error and evaluating the scheduling quality; based on the comparison results, constructing training samples containing a quadruple of state, action, reward and new state according to the reinforcement learning sample format, and storing them in the experience replay buffer; periodically sampling data from the experience replay buffer, and using a deep reinforcement learning algorithm to train the policy network offline; and securely updating the new model after training and verification to the production environment through canary release.

7. A dynamic AI computing power scheduling system based on a hybrid cloud architecture, wherein the system applies the method described in any one of claims 1 to 6, characterized in that, The system adopts a microservice architecture and includes the following core components that work together in a highly cohesive and loosely coupled manner: The resource situation awareness module is used to collect data through monitoring probes deployed in each resource pool, and to build and update a real-time situation map of hybrid cloud resources through the data aggregation service. The task profiling and modeling module is used to receive and parse AI tasks, extract multi-dimensional features through the feature extraction engine, and generate structured task profiles. A cross-cloud scheduling decision engine is used to generate optimal scheduling decisions based on resource status maps and task profiles, through a multi-objective optimization model, a deep reinforcement learning scheduler, a scheme simulation evaluator, and a decision arbitrator. The policy execution and routing module is used to parse scheduling decisions and execute resource provisioning, network configuration and task deployment through the cloud resource operation adapter, software-defined network controller and task deployment coordinator; The runtime monitoring and dynamic adjustment module is used to monitor, alarm, and reschedule running tasks through a full-link indicator collector, intelligent alarm and root cause analyzer, and dynamic rescheduling trigger. The cost analysis and optimization module is used to aggregate cost data, predict task expenses, and manage budget strategies. The policy feedback and model training module is used to collect scheduling effect data, construct training samples, and train and update the deep reinforcement learning model.

8. The system according to claim 7, characterized in that, The cross-cloud scheduling decision engine specifically includes: a multi-objective optimization model library, used to define and store mathematical models of optimization problems with resource utilization, task completion time, and total cost of ownership as objectives; a deep reinforcement learning scheduler, including a policy network and a value network, used to generate probability distributions of scheduling actions based on system state; a scheme simulation evaluator, used to load historical prediction models to quickly extrapolate and evaluate the execution effect of candidate scheduling schemes; and a decision arbitrator, used to comprehensively score or Pareto rank the evaluated candidate schemes based on global policy preferences and output the final scheduling decision.

9. An electronic device comprising a processor, a memory, and a computer program stored in the memory, characterized in that, When the processor executes the computer program, it implements the method as described in any one of claims 1 to 6.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the method as described in any one of claims 1 to 6.

Images