A multi-cloud heterogeneous computing power scheduling method, device, equipment and storage medium

Abstract

The present application relates to the technical field of computing power scheduling, and particularly relates to a multi-cloud heterogeneous computing power scheduling method and device, equipment and a storage medium; a computing power scheduling task is analyzed according to a preset computing power scheduling algorithm to obtain computing power characteristics and storage characteristics; a cloud space bandwidth of a cloud storage space, a heterogeneous computing power adaptation interface and a plurality of target computing power nodes are determined according to a preset bandwidth allocation strategy and the computing power characteristics; the computing power scheduling task is distributed to each target computing power node according to a global resource view, a preset graph correlation analysis method, the cloud space bandwidth and the heterogeneous computing power adaptation interface to obtain a plurality of task target nodes; the computing power characteristics and the storage characteristics of the task are extracted, multi-cloud heterogeneous resource data is collected, and a global resource view is constructed to complete resource adaptation and task distribution, improve scheduling efficiency, resource utilization and stability, and realize optimal computing power configuration.

Description

Technical Field

[0001] This invention relates to the field of computing power scheduling technology, and in particular to a multi-cloud heterogeneous computing power scheduling method, apparatus, device and storage medium. Background Technology

[0002] Currently, the accelerated digital transformation is driving multi-cloud heterogeneous architecture to become the mainstream mode of computing power scheduling. Cloud resources from multiple vendors, architectures, and regions are widely used in various computing scenarios, but they have many prominent shortcomings in practical applications. Some existing technologies lack comprehensive feature decomposition and analysis of computing power scheduling tasks, making it difficult to build a complete task requirement profile. There is a lack of effective global resource management methods, making it impossible to achieve unified perception and topology management of network resources, resulting in dispersed resources that are difficult to coordinate. The resource adaptation and task distribution mechanisms are imperfect, bandwidth allocation is unreasonable, and heterogeneous interface compatibility is poor, leading to low task-resource matching, easy node overload or resource idleness, low scheduling efficiency, and inability to meet the efficient and stable execution requirements of diverse computing power scheduling tasks, which seriously restricts the optimal allocation of computing power resources in multi-cloud heterogeneous environments. Summary of the Invention

[0003] In order to overcome the shortcomings of the prior art, the present invention aims to provide a multi-cloud heterogeneous computing power scheduling method, apparatus, device and storage medium.

[0004] A multi-cloud heterogeneous computing power scheduling method includes: collecting computing power scheduling tasks and performing feature analysis on the tasks according to a preset computing power scheduling algorithm to obtain computing power characteristics and storage characteristics; collecting data from a preset cloud storage space according to a preset time collection range and storage characteristics to obtain multi-cloud heterogeneous resource data; constructing a global resource view based on the multi-cloud heterogeneous resource data and a preset heterogeneous resource graph construction algorithm; determining the cloud space bandwidth, heterogeneous computing power adaptation interface, and multiple target computing power nodes of the cloud storage space according to a preset bandwidth allocation strategy and computing power characteristics; and distributing the computing power scheduling tasks to each target computing power node according to the global resource view, a preset graph association analysis method, cloud space bandwidth, and heterogeneous computing power adaptation interface to obtain multiple task target nodes.

[0005] Furthermore, the step of constructing a global resource view based on multi-cloud heterogeneous resource data and a preset heterogeneous resource graph construction algorithm includes: performing feature analysis on multi-cloud heterogeneous resource data based on preset cloud platform node attributes, preset computing power node attributes, preset network link node attributes, and preset resource pool node attributes to obtain correlation strength parameters, the number of computing power nodes, and resource specifications; determining the action space based on the number of computing power nodes and resource specifications; and constructing a global resource view based on the correlation strength parameters, action space, multi-cloud heterogeneous resource data, and the heterogeneous resource graph construction algorithm.

[0006] Furthermore, the step of distributing computing power scheduling tasks to various target computing power nodes based on a global resource view, a preset graph association analysis method, cloud space bandwidth, and heterogeneous computing power adaptation interface to obtain multiple task target nodes includes: dividing the computing power scheduling tasks according to a pre-trained scene recognition model, computing power characteristics, and storage characteristics to obtain multiple classification tasks; predicting the multiple classification tasks according to a preset long short-term memory model, a global resource view, and a graph association analysis method to obtain load prediction results; and distributing the multiple classification tasks to various target computing power nodes based on the load prediction results, cloud space bandwidth, and heterogeneous computing power adaptation interface to obtain multiple task target nodes.

[0007] Furthermore, the step of predicting multiple classified tasks based on a preset long short-term memory model, a global resource view, and a graph association analysis method to obtain load prediction results includes: matching the optimal network path from multiple preset network paths based on computing power characteristics; performing in-depth analysis of the global resource view based on the optimal network path and the graph association analysis method to obtain the exploration rate and scheduling count; and predicting multiple classified tasks based on the exploration rate, scheduling count, and the long short-term memory model to obtain load prediction results.

[0008] Furthermore, the step of predicting multiple classification tasks based on the exploration rate, scheduling count, and long short-term memory model to obtain load prediction results includes: configuring the long short-term memory model based on the exploration rate and scheduling count to obtain a load prediction model; and predicting multiple classification tasks based on the load prediction model to obtain load prediction results.

[0009] Furthermore, the step of distributing multiple categorized tasks to various target computing power nodes based on load prediction results, cloud space bandwidth, and heterogeneous computing power adaptation interfaces to obtain multiple task target nodes includes: collecting the operating status of multiple target computing power nodes; generating a scheduling scheme based on cloud space bandwidth, load prediction results, and operating status; and distributing multiple categorized tasks to multiple target computing power nodes based on the scheduling scheme, a preset cross-cloud API interface, and a heterogeneous computing power adaptation interface to obtain multiple task target nodes.

[0010] Furthermore, the step of determining the cloud space bandwidth, heterogeneous computing power adaptation interface, and multiple target computing power nodes of the cloud storage space according to the preset bandwidth allocation strategy and computing power characteristics includes: allocating bandwidth to multiple classified tasks according to the bandwidth allocation strategy and storage characteristics to obtain the cloud space bandwidth; and determining multiple target computing power nodes and heterogeneous computing power adaptation interfaces of the cloud storage space according to the computing power characteristics and storage characteristics.

[0011] Furthermore, a multi-cloud heterogeneous computing power scheduling device includes: a feature analysis module, used to collect computing power scheduling tasks and perform feature analysis on the computing power scheduling tasks according to a preset computing power scheduling algorithm to obtain computing power features and storage features; a data acquisition module, used to collect data from a preset cloud storage space according to a preset time acquisition range and storage features to obtain multi-cloud heterogeneous resource data; a resource view construction module, used to construct a global resource view according to the multi-cloud heterogeneous resource data and a preset heterogeneous resource graph construction algorithm; a data determination module, used to determine the cloud space bandwidth, heterogeneous computing power adaptation interface, and multiple target computing power nodes of the cloud storage space according to a preset bandwidth allocation strategy and computing power features; and a distribution module, used to distribute computing power scheduling tasks to each target computing power node according to the global resource view, a preset graph association analysis method, cloud space bandwidth, and heterogeneous computing power adaptation interface to obtain multiple task target nodes.

[0012] Furthermore, a multi-cloud heterogeneous computing power scheduling device is provided, comprising: a memory and at least one processor, wherein the memory stores instructions; the at least one processor invokes the instructions in the memory to cause the multi-cloud heterogeneous computing power scheduling device to execute the various steps of the multi-cloud heterogeneous computing power scheduling method described above.

[0013] Furthermore, a computer-readable storage medium stores instructions that, when executed by a processor, implement the steps of the multi-cloud heterogeneous computing power scheduling method described above.

[0014] The technical solution of this invention effectively addresses pain points such as resource dispersion, mismatch between demand and resources, and inefficient scheduling in multi-cloud heterogeneous environments through a full-link design encompassing computing power scheduling task collection and feature analysis, multi-cloud heterogeneous resource collection, global resource view construction, resource adaptation, and task distribution. It extracts computing power and storage features to construct task demand profiles, providing a reliable basis for end-to-end scheduling. It collects full-volume multi-cloud heterogeneous resource data by combining storage features and time ranges, ensuring the relevance and completeness of resource data. A global resource view is created through a heterogeneous resource graph construction algorithm, achieving global awareness and unified management of network resources. Based on computing power features, it adapts bandwidth strategies, determines adaptation interfaces and target nodes, and completes task distribution by combining the global resource view and graph association analysis. This shields architectural differences, ensures stable transmission, avoids node overload and resource idleness, and effectively improves computing power scheduling efficiency, resource utilization, and task execution stability in multi-cloud heterogeneous environments. This lays a solid foundation for subsequent scheduling and management, achieving optimized allocation of computing power resources and orderly and efficient advancement of the scheduling process. Attached Figure Description

[0015] The above and / or additional aspects and advantages of the present invention will become apparent and readily understood from the description of the embodiments taken in conjunction with the following drawings, in which: Figure 1 A first flowchart of a multi-cloud heterogeneous computing power scheduling method provided in an embodiment of the present invention; Figure 2 A second flowchart of a multi-cloud heterogeneous computing power scheduling method provided in an embodiment of the present invention; Figure 3 A third flowchart of a multi-cloud heterogeneous computing power scheduling method provided in an embodiment of the present invention; Figure 4 The fourth flowchart of a multi-cloud heterogeneous computing power scheduling method provided in this embodiment of the invention; Figure 5 The fifth flowchart of a multi-cloud heterogeneous computing power scheduling method provided in an embodiment of the present invention; Figure 6 The sixth flowchart of a multi-cloud heterogeneous computing power scheduling method provided in this embodiment of the invention; Figure 7 The seventh flowchart of a multi-cloud heterogeneous computing power scheduling method provided in an embodiment of the present invention; Figure 8 A schematic diagram of the structure of a multi-cloud heterogeneous computing power scheduling device provided in an embodiment of the present invention; Figure 9 This is a schematic diagram of the structure of a multi-cloud heterogeneous computing power scheduling device provided in an embodiment of the present invention. Detailed Implementation

[0016] The terms "first," "second," "third," "fourth," etc. (if present) in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" or "having" and any variations thereof are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0017] For ease of understanding, the specific process of the embodiments of the present invention is described below. Please refer to [link / reference]. Figure 1 One embodiment of a multi-cloud heterogeneous computing power scheduling method according to the present invention includes: 101. Collect computing power scheduling tasks and perform feature analysis on the computing power scheduling tasks according to the preset computing power scheduling algorithm to obtain computing power characteristics and storage characteristics; In this embodiment, after collecting computing power scheduling tasks, various types of accessed computing power scheduling tasks (including real-time computing tasks, offline processing tasks, high-concurrency tasks, etc.) are uniformly collected. A preset computing power scheduling algorithm (such as greedy scheduling algorithm, genetic scheduling algorithm, reinforcement learning scheduling algorithm, etc.) is used to perform comprehensive feature decomposition and analysis of the computing power scheduling tasks. The computing power features mainly involve the quantitative extraction of the task's computational requirements, specifically including indicators such as CPU computing power requirements, computational complexity, task concurrency, latency threshold, and task priority, reflecting the scale of computing resources and real-time requirements for task execution. Storage Features are extracted from task-related data, including indicators such as data volume, data read / write frequency, data storage type (structured / unstructured), data security requirements, and storage latency requirements, reflecting the degree of dependence on storage resources during task execution. Computing power features and storage features are interrelated and mutually influential, jointly forming a demand profile for computing power scheduling tasks. During feature analysis, it is necessary to ensure that both types of features are extracted comprehensively, providing reliable demand basis for all subsequent stages such as multi-cloud resource collection, resource matching, and task distribution, ensuring the accuracy and completeness of feature extraction, and guaranteeing the orderly progress of the entire scheduling process. 102. Collect data from the preset cloud storage space according to the preset time collection range and storage characteristics to obtain multi-cloud heterogeneous resource data; In this embodiment, after obtaining the storage characteristics of the computing power scheduling task, and combining a preset time collection range, comprehensive data collection is performed on preset multi-vendor, multi-architecture, and multi-region cloud storage spaces and supporting resources to ultimately obtain multi-cloud heterogeneous resource data. The preset time collection range needs to be flexibly set based on the storage characteristics. For example, for storage characteristics with high read / write frequency and high real-time requirements, the time collection range is set to the most recent 10 minutes to ensure that the collected resource data reflects the current storage resource status in real time. For storage characteristics with low read / write frequency and non-real-time requirements, the time collection range can be set to the most recent hour to ensure data validity while reducing resource consumption during the collection process. During the data collection process, the focus is on matching the storage characteristics with the corresponding collection granularity. For example, for storage characteristics of large data volumes, the focus is on collecting indicators such as the total capacity, remaining capacity, and storage IO rate of cloud storage space; for storage characteristics with high security requirements, the focus is on collecting indicators such as the encryption level, backup strategy, and access permissions of cloud storage; at the same time, in addition to cloud storage space data, it is also necessary to collect computing resources and network resources associated with storage resources, including the computing node configuration, CPU / GPU utilization, memory usage, network bandwidth, link latency, packet loss rate, etc. of each cloud platform, to form complete multi-cloud heterogeneous resource data, providing full basic data support for the subsequent construction of a global resource view; 103. A global resource view is constructed based on multi-cloud heterogeneous resource data and a preset heterogeneous resource graph construction algorithm; In this embodiment, a preset heterogeneous resource graph construction algorithm (such as a neural network construction algorithm or a topology graph generation algorithm) is used to structure and topologically integrate scattered and heterogeneous multi-cloud resource data, ultimately constructing a global resource view. During the construction process, the multi-cloud heterogeneous resource data is first classified into four categories: cloud platform node data, computing power node data, resource pool node data, and network link node data. The attribute parameters of each type of node are clearly defined (such as the CPU model and memory size of computing power nodes, the capacity and IO rate of storage nodes, and the latency and bandwidth of network link nodes). Subsequently, the heterogeneous resource graph construction algorithm... The system analyzes the relationships between various nodes, such as the connection links between computing power nodes and resource pool nodes, the affiliation between cloud platform nodes and their subordinate computing power nodes, and the network interconnection relationships between different cloud platforms. The resulting global resource view presents the distribution, real-time operating status, node relationships, and resource availability of all heterogeneous resources across the network in a topological and visual format, achieving global awareness and unified management of all resources in a multi-cloud heterogeneous environment. The global resource view enables the scheduling system to fully grasp the distribution and status of resources, providing a global perspective for subsequent scheduling decisions such as bandwidth allocation, target node selection, and task assignment. 104. Determine the cloud space bandwidth, heterogeneous computing power adaptation interface, and multiple target computing power nodes of the cloud storage space based on the preset bandwidth allocation strategy and computing power characteristics. In this embodiment, the preset bandwidth allocation strategy needs to be flexibly adapted to the computing power characteristics. For example, for high-concurrency, low-latency computing power characteristics (such as real-time data processing tasks), a dynamic bandwidth allocation strategy is adopted to prioritize the allocation of high-bandwidth resources and ensure the efficiency of task data transmission. For low-concurrency, non-real-time computing power characteristics (such as offline data computing tasks), a static bandwidth allocation strategy is adopted to allocate bandwidth resources reasonably and avoid bandwidth waste. The determination of cloud space bandwidth needs to match the data transmission requirements in the computing power characteristics. Combined with the concurrency of the task and the data volume, the minimum bandwidth threshold that meets the task execution is calculated, while reserving a certain amount of bandwidth redundancy to prevent link congestion. The determination of heterogeneous computing power adaptation interfaces is mainly based on the computing power architecture requirements in the computing power characteristics and the interface standards of various cloud platforms. For example, for GPU-intensive tasks, an interface adapted to GP is selected. For CPU-intensive tasks, the interface for U-level computing power is selected to be compatible with the CPU computing power, while ensuring that the interface is compatible with heterogeneous architectures of different cloud platforms (such as x86 architecture and ARM architecture) to avoid tasks failing to execute properly due to interface incompatibility. The selection of multiple target computing power nodes requires combining computing power characteristics and a global resource view to select computing power nodes with resource configuration matching task requirements, moderate resource utilization, and good network link quality. During the selection process, nodes with a high degree of matching with task computing power requirements and a load rate between 30% and 70% are given priority to avoid both node overload causing task execution lag and node idleness causing resource waste. The purpose of this step is to match task requirements with underlying resources, provide hardware and interface support for subsequent task deployment, and ensure that tasks can be executed stably and efficiently in an adapted resource environment. 105. Based on the global resource view, the preset graph association analysis method, cloud space bandwidth and heterogeneous computing power adaptation interface, the computing power scheduling task is distributed to each target computing power node to obtain multiple task target nodes; In this embodiment, relying on a global resource view, a pre-defined graph association analysis method, cloud space bandwidth, and heterogeneous computing power adaptation interfaces, the computing power scheduling tasks are efficiently distributed to multiple target computing power nodes. The global resource view enables global awareness of all network resources, and graph association analysis uncovers node association patterns, providing support for reasonable task allocation. Cloud space bandwidth ensures stable and efficient data transmission of tasks, avoiding lag caused by insufficient bandwidth. The heterogeneous computing power adaptation interface masks differences in node architecture, ensuring compatibility and interoperability between tasks and nodes. Ultimately, multiple target task nodes are formed, effectively improving computing power scheduling efficiency, resource utilization, and task execution stability in a multi-cloud heterogeneous environment, laying a solid foundation for subsequent scheduling and management. In this embodiment, a full-link design encompassing computing power scheduling task collection and feature analysis, multi-cloud heterogeneous resource collection, global resource view construction, resource adaptation, and task distribution effectively addresses pain points such as resource dispersion, mismatch between demand and resources, and inefficient scheduling in multi-cloud heterogeneous environments. By extracting computing power and storage features, a task demand profile is constructed, providing a reliable basis for the entire scheduling process. Collecting full-volume multi-cloud heterogeneous resource data in conjunction with storage features and time ranges ensures the relevance and completeness of resource data. A global resource view is created through a heterogeneous resource graph construction algorithm, achieving global awareness and unified management of network resources. Based on computing power features, bandwidth adaptation strategies are applied, and adaptation interfaces and target nodes are determined. Task distribution is completed by combining the global resource view and graph association analysis, masking architectural differences, ensuring stable transmission, and avoiding node overload and resource idleness. This effectively improves computing power scheduling efficiency, resource utilization, and task execution stability in multi-cloud heterogeneous environments, laying a solid foundation for subsequent scheduling and management, and achieving optimized allocation of computing power resources and orderly and efficient advancement of the scheduling process.

[0018] Please see Figure 2 A second embodiment of a multi-cloud heterogeneous computing power scheduling method according to the present invention includes: 201. Perform feature analysis on multi-cloud heterogeneous resource data based on preset cloud platform node attributes, preset computing power node attributes, preset network link node attributes, and preset resource pool node attributes to obtain correlation strength parameters, number of computing power nodes, and resource specifications. In this embodiment, the preset cloud platform node attributes specifically include cloud platform vendor identifier, architecture type (x86 / ARM / heterogeneous architecture), deployment geographical region, resource scheduling permissions, operation and maintenance level, fault tolerance capability, etc., used to define the resource boundaries and schedulable range of each cloud platform; the preset computing power node attributes focus on computing power output capability and operating status, specifically including the hardware configuration of the computing power node (CPU model, quantity, clock speed, GPU model, video memory capacity, memory size, storage interface), peak computing power output, current load rate, energy consumption parameters, operating status (normal, fault, idle), etc., determining the node's computing power and the amount of tasks it can handle; the preset network link node attributes are used to characterize the number of nodes. Based on transmission capacity and stability, specifically including link bandwidth, transmission latency, packet loss rate, nodes at both ends of the link connection (such as cloud platform and computing power nodes, computing power nodes and resource pools), link stability rating, link priority, etc., these factors affect the efficiency of data interaction between nodes. Preset resource pool node attributes reflect the resource pool's supply capacity and allocation rules, specifically including resource pool type (computing power pool, storage pool, hybrid resource pool), total capacity, remaining resources, resource allocation granularity, access permissions, data read / write rate, resource update frequency, etc. The feature analysis process revolves around four types of node attributes and is completed in three steps. The first step is data classification and filtering, matching multi-cloud heterogeneous resource data with the four types of node attributes one by one, and filtering... The first step is to extract valid resource data that meets the attribute standards to ensure the validity of the input data. The second step is parameter extraction and quantification. For the filtered valid data, the scope of nodes participating in the calculation has been clearly defined (only nodes that meet the attributes and are running normally). The focus is on extracting three types of parameters: First, the association strength parameter. By calculating the connection frequency, total data transmission volume, and resource dependence between different nodes, the degree of association between cloud platform nodes and computing power nodes, resource pool nodes, and computing power nodes and network link nodes is quantified. The larger the value of the association strength parameter, the more frequent the data interaction between nodes, providing a weighting basis for subsequent topology construction. Second, the number of computing power nodes. All nodes that meet the computing power node attributes and are in normal operation are counted one by one to determine the total scale of schedulable computing power nodes in the current multi-cloud environment, which is the basis for delineating the space for subsequent actions. The third step is resource specifications, which involves quantifying and summarizing the configuration parameters of computing power nodes, resource pool nodes, and network link nodes to form a standardized resource list, such as the upper limit of computing power per node, the upper limit of capacity of resource pool, and the upper limit of bandwidth of network links, thus clarifying the supply capacity boundaries of various resources. The third step is data verification and correction, which involves cross-verifying the extracted correlation strength parameters, the number of computing power nodes, and resource specifications to ensure that the three types of parameters can reflect the actual status of heterogeneous resources in the multi-cloud environment, providing reliable support for subsequent steps. 202. Determine the action space based on the number of computing nodes and resource specifications; In this embodiment, based on the number of computing nodes, the node selection range and scheduling mode boundaries of the action space are defined: If the number of computing nodes is large (e.g., more than 50), the node selection schemes included in the action space are richer, supporting multiple scheduling modes such as multi-node parallel scheduling, load balancing scheduling, and redundant backup scheduling. More scheduling partitions can also be divided to improve scheduling flexibility. If the number of computing nodes is small (e.g., less than 10), the action space focuses on the rational utilization of existing nodes, prioritizing nodes with high resource matching to avoid node overload. Action selection is relatively concentrated, with a focus on ensuring scheduling stability. Based on the resource specifications of the computing nodes (maximum computing power per node, load capacity), the task computing power level and number of concurrent tasks that a single computing node can handle are clearly defined to avoid scheduling actions exceeding the node's capacity. Capabilities can lead to task execution failures. Based on the resource pool's resource specifications (total capacity, remaining resources, allocation granularity), the maximum and minimum granularity of resource allocation are clearly defined to ensure reasonable resource allocation. Based on the network link's resource specifications (bandwidth, latency), the maximum data transmission rate and scheduling latency threshold are clearly defined to ensure scheduling actions meet data transmission requirements. Simultaneously, the determination of the action space has dynamic adaptability. When the number of computing nodes changes (e.g., new nodes, node failures) or resource specifications are adjusted (e.g., resource expansion, hardware upgrades), the action space is updated synchronously, invalid scheduling actions are removed, and new scheduling schemes that match the current resource status are added. This ensures that the action space always remains consistent with the actual state of multi-cloud heterogeneous resources, providing a clear and feasible operational boundary for subsequent global resource view construction and scheduling decisions. 203. A global resource view is constructed based on the correlation strength parameter, action space, multi-cloud heterogeneous resource data, and heterogeneous resource graph construction algorithm; In this embodiment, the global resource view serves as a topological, visual, and dynamic representation of all resources in a multi-cloud heterogeneous environment. Its function is to achieve global awareness, status monitoring, correlation display, and scheduling support for all network resources. It unifies and integrates multi-cloud heterogeneous resource data, correlation strength parameters, and action space, performs standardization processing, and eliminates duplicate and abnormal data. Simultaneously, based on the boundary requirements of the action space, it filters out valid resource data and correlation strength parameters that meet the scheduling scope, ensuring the rationality and relevance of the view construction. Based on a heterogeneous resource graph construction algorithm, it uses four types of nodes (cloud platform nodes, computing power nodes, network link nodes, and resource pool nodes) as topology nodes, and uses correlation strength parameters as the weights for connections between nodes (the higher the correlation strength, the thicker the connection lines and the more obvious the markings), constructing a complete resource topology structure. In the topology structure, each node is marked... This section includes information on various nodes, such as the architecture and region of cloud platform nodes, resource specifications and current load of computing power nodes, bandwidth and latency of network link nodes, and remaining resources and allocation strategies of resource pool nodes. It also marks the scheduling boundaries of the action space, clearly defining which nodes and resources can be included in the scheduling scope, allowing schedulers to intuitively grasp actionable tasks. A real-time update mechanism is configured for the global resource view, collecting real-time operational status data of multi-cloud heterogeneous resources (such as changes in computing power node load, resource pool remaining resources, and network link latency), and synchronously updating correlation strength parameters and node status labels to ensure the global resource view reflects dynamic resource changes in real time. Furthermore, when the action space is adjusted, the scheduling boundary labels in the view are updated synchronously to ensure the view remains consistent with actual scheduling needs and resource status, providing real-time support for subsequent scheduling decisions. In this embodiment, resource data is categorized and filtered based on four types of node attributes, extracting correlation strength parameters, the number of computing nodes, and resource specifications. Cross-validation ensures the accuracy and reliability of these parameters, providing high-quality data support for subsequent scheduling. The action space defined by the number of computing nodes and resource specifications can dynamically adapt to resource changes, avoiding ineffective scheduling and node overload, thus improving scheduling feasibility. The constructed global resource view realizes resource topology and visualization, clearly showing node associations and scheduling boundaries. Combined with a real-time update mechanism, it ensures that the global resource view is synchronized with the actual resource status, providing real-time support for subsequent scheduling decisions. This comprehensively improves the global management capability, scheduling efficiency, and resource utilization of multi-cloud heterogeneous resources, ensuring the stable and efficient operation of the scheduling system.

[0019] Please see Figure 3 A third embodiment of a multi-cloud heterogeneous computing power scheduling method in this invention includes: 301. Divide the computing power scheduling tasks according to the pre-trained scene recognition model, computing power features and storage features to obtain multiple classification tasks; In this embodiment, the scene recognition model is specifically a deep learning-based scene classification model. It relies on a pre-trained feature library of computing power scheduling tasks (covering feature parameters for various scenarios such as real-time data inference, offline model training, high-concurrency data processing, and ordinary data querying) to extract the demand features of computing power scheduling tasks and identify the application scenario to which the task belongs. The logic for dividing computing power scheduling tasks is as follows: based on the application scenario of the task identified by the scene recognition model, combined with the computing power features (such as CPU / GPU computing power requirements, concurrency, latency thresholds, etc.) and storage features (such as data volume, data read / write frequency, storage latency requirements, etc.) extracted from the computing power scheduling tasks in the previous stage, computing power scheduling tasks with consistent scenario requirements and similar computing power and storage features are grouped into one category, resulting in multiple homogeneous classification tasks. This ensures that the requirements of each classification task are unified, providing support for subsequent scheduling operations such as load prediction, bandwidth allocation, and task distribution. 302. Based on the preset long short-term memory model, global resource view and graph association analysis method, predict multiple classification tasks to obtain load prediction results; In this embodiment, the Long Short-Term Memory (LSTM) model has the advantage of time-series prediction and can capture the patterns of load changes; the global resource view provides a global awareness of all network resources, and the graph association analysis method is used to mine the patterns of node associations; based on the preset LSM model, global resource view, and graph association analysis method, multiple categories of tasks are predicted, which effectively improves the accuracy of load prediction, predicts the future load changes of each category of tasks, provides a reliable basis for subsequent task distribution and load balancing scheduling, avoids node overload and resource idleness problems, and improves the utilization rate of computing resources and scheduling stability in a multi-cloud heterogeneous environment; 303. Based on the load prediction results, cloud space bandwidth and heterogeneous computing power adaptation interface, multiple classification tasks are distributed to each target computing power node to obtain multiple task target nodes; In this embodiment, the load prediction result can predict the node load in advance, avoid task allocation overload or resource idleness, and achieve load balancing; cloud space bandwidth ensures stable and efficient data transmission of tasks, avoiding lag caused by insufficient bandwidth; the heterogeneous computing power adaptation interface shields the differences in node architecture, ensuring that tasks and nodes are compatible and connected, and finally successfully forming multiple task target nodes, providing support for subsequent scheduling and management, and effectively improving the computing power scheduling efficiency, resource utilization and task execution stability in a multi-cloud heterogeneous environment; In this embodiment, relying on a deep learning-based scene recognition model and combining computing power and storage characteristics, the computing power scheduling task is divided into homogeneous classification tasks, providing an accurate foundation for subsequent scheduling operations. Through long short-term memory models, global resource views, and graph association analysis methods, the accuracy of load prediction is improved, task load changes are anticipated, and node overload and resource idleness are avoided. Based on the load prediction results, cloud bandwidth, and heterogeneous computing power adaptation interfaces, tasks are distributed to ensure stable data transmission and compatible task-node integration. Multiple task target nodes are successfully formed, comprehensively improving the computing power scheduling efficiency, resource utilization, and task execution stability in a multi-cloud heterogeneous environment. This provides reliable support for subsequent scheduling and management, achieving reasonable optimization of computing power resources.

[0020] Please see Figure 4 The fourth embodiment of a multi-cloud heterogeneous computing power scheduling method in this invention includes: 401. Based on computing power characteristics, the optimal network path is obtained by matching from multiple preset network paths; In this embodiment, the preset multi-network paths refer to multiple alternative network transmission paths connecting cloud platform nodes, computing power nodes, and resource pool nodes in a multi-cloud heterogeneous environment. Each path corresponds to different transmission parameters such as bandwidth, latency, and packet loss rate, which can meet the transmission requirements of different computing power tasks. The computing power characteristics serve as the basis for path selection, determining the adaptability of the network path. The path selection process revolves around the computing power characteristics and is completed in three orderly steps: The first step is path screening, which combines the latency threshold and data transmission volume in the computing power characteristics to select candidate paths that meet the basic transmission requirements from the preset multi-network paths. For example, for high-concurrency, low-latency computing power characteristics (such as real-time inference tasks), candidate paths with transmission latency ≤10ms and packet loss rate ≤0.1% are selected; for large-volume, non-real-time computing power... The first step involves identifying candidate paths with bandwidth ≥1000Mbps and stability rating ≥A based on features (such as offline model training). The second step involves path quantification and comparison. For the selected candidate paths, multi-dimensional evaluation metrics (latency, bandwidth, packet loss rate, stability, and transmission cost) are constructed based on the computing power requirements. Corresponding weights are assigned to each metric (e.g., low-latency tasks have the highest weight), and a comprehensive score is calculated for each candidate path. The third step determines the optimal path. The candidate path with the highest comprehensive score is selected as the optimal network path. This path best matches the computing power requirements, ensuring efficient and stable data transmission for tasks, and providing a reliable transmission foundation for subsequent in-depth analysis of global resource views and load prediction. If multiple paths with similar scores exist, the path with higher correlation is prioritized. 402. Perform in-depth analysis of the global resource view based on the optimal network path and graph association analysis method to obtain the exploration rate and scheduling count; In this embodiment, the optimal network path clarifies the optimal channel for task data transmission, while the graph association analysis method is used to mine the relationships and operational patterns of various nodes in the global resource view. The combination of these two methods enables in-depth analysis of the global resource view, ultimately extracting two parameters: exploration rate and scheduling frequency. The specific analysis process is as follows: First, using the optimal network path as a clue, all nodes related to this path in the global resource view are located (including cloud platform nodes, computing power nodes, resource pool nodes at both ends of the path, and network link nodes along the path). The relationships and operational data of these nodes are then focused on. Second, the graph association analysis method is used to deeply mine the association strength, interaction frequency, and resource dependencies between these nodes. On the one hand, the historical relationships between these nodes in the current scheduling system are statistically analyzed. The scheduling frequency determines the scheduling popularity of a node (a higher scheduling frequency indicates frequent node use and potentially greater load pressure). On the other hand, the exploration rate is calculated. The exploration rate is the probability that the scheduling system will explore underutilized nodes (those with low association strength, few scheduling frequency, but sufficient resources). This rate measures the resource mining capability of the scheduling system; a higher exploration rate indicates that the system is more inclined to mine idle resources and improve resource utilization. Finally, the statistically obtained scheduling frequency and calculated exploration rate are verified to ensure that the data is consistent with the actual state of the global resource view and the transmission requirements of the optimal network path, providing parameter input for subsequent load prediction. The scheduling frequency reflects the historical load of a node, while the exploration rate reflects the resource mining potential; together, they constitute the reference basis for load prediction. 403. Based on the exploration rate, scheduling times, and long short-term memory model, predict the load for multiple classification tasks to obtain load prediction results; In this embodiment, multiple classification tasks are sub-tasks obtained by splitting them in the early stage based on computing power and storage characteristics. The Long Short-Term Memory (LSTM) model has the advantage of time-series prediction and can capture the temporal variation pattern of task load. The exploration rate and scheduling count are used as input parameters of the model to optimize the model's prediction accuracy and ensure that the load prediction results fit the actual scheduling scenario. The specific prediction process is divided into three stages: The first stage is model configuration, in which the exploration rate and scheduling count are used as input parameters and substituted into the LSTM model to adaptively configure the model. For example, the time window length of the model is adjusted according to the scheduling count (the more scheduling counts, the longer the time window, and the better it can capture the load variation pattern), and the prediction weight of the model is adjusted according to the exploration rate (the higher the exploration rate, the higher the model's load prediction weight for idle nodes), so that the model adapts to the dynamic characteristics of multi-cloud heterogeneous scheduling; The second stage is data input and... The training process involves inputting historical load data and computing power feature data from multiple classification tasks, along with exploration rate and scheduling counts, into a configured LSTM model for training and iteration. This optimizes model parameters and ensures the model can capture the correlation between classification task load and exploration rate / scheduling counts, as well as the temporal variation patterns of load. The third stage, load prediction and output, applies the trained model to multiple classification tasks, predicting future load trends for each task on different target computing power nodes. This yields quantified load prediction results (e.g., peak load, average load, and load fluctuation range for each task within the next hour). Simultaneously, the prediction results are validated. If the prediction deviation exceeds a preset threshold, the weights of exploration rate and scheduling counts are adjusted, and the prediction is repeated to ensure the accuracy and reliability of the load prediction results, providing a basis for subsequent task distribution and load balancing scheduling decisions. In this embodiment, through the end-to-end design of optimal network path selection, global resource in-depth analysis, and load prediction, the pain points of inefficient data transmission, insufficient resource mining, and inaccurate load prediction in multi-cloud heterogeneous environments are effectively addressed. This solution selects and determines the optimal network path based on computing power characteristics, achieving path matching with task requirements, reducing transmission latency and packet loss rate, and ensuring efficient and stable data transmission, laying the foundation for subsequent steps. Combining optimal path and graph association analysis methods, the exploration rate and scheduling frequency are extracted, clarifying node scheduling heat and mining the potential of idle resources, thereby improving resource utilization. Relying on the time-series prediction advantages of the LSTM model, the load changes of classified tasks are predicted, avoiding node overload and resource idleness problems in advance, providing a reliable basis for subsequent task distribution and load balancing scheduling, and comprehensively improving the efficiency, stability, and resource utilization of multi-cloud heterogeneous computing power scheduling.

[0021] Please see Figure 5 The fifth embodiment of a multi-cloud heterogeneous computing power scheduling method in this invention includes: 501. Configure the Long Short-Term Memory model based on the exploration rate and scheduling count to obtain the load prediction model; In this embodiment, the Long Short-Term Memory (LSTM) model, as a model with advantages in time series prediction, relies on reasonable parameter configuration for its predictive adaptability. The number of scheduling operations, reflecting the historical scheduling activity of a node, determines the model's time window length. A higher number of scheduling operations indicates richer historical load data for the node. Therefore, the model's time window length is adjusted to be longer (e.g., 60 minutes for scheduling operations ≥ 100 times; 30 minutes for scheduling operations < 50 times) to ensure the model can fully capture the long-term variation patterns of node load. The exploration rate, as a parameter measuring resource mining capability, adjusts the model's prediction weights accordingly. A higher exploration rate indicates that the system tends to mine idle resources, thus increasing the weighting of idle nodes (those with low association strength and few scheduling operations). The prediction weights are adjusted to reduce the weight ratio of high-frequency scheduling nodes, avoiding excessive reliance on historical high-frequency node data and neglecting the load potential of idle resources. Key parameters of the LSTM model, such as the number of hidden layer neurons, learning rate, and iteration count, are adaptively adjusted based on the dynamic characteristics of the exploration rate and scheduling count. For example, when the exploration rate is high (e.g., >0.6), the number of hidden layer neurons is appropriately increased to improve the model's ability to capture changes in the load of idle nodes. When the scheduling count fluctuates significantly, the model learning rate is reduced to slow down parameter updates, improve model stability, and prevent model prediction distortion due to scheduling count fluctuations. Simultaneously, some historical load data from previous classification tasks is introduced, and combined with the exploration rate and scheduling count, the initially configured model is preliminarily trained to verify the rationality of the configuration. 502. Based on the load prediction model, predict multiple classification tasks to obtain load prediction results; In this embodiment, multiple classification tasks are sub-tasks obtained by splitting them in the early stage based on computing power and storage characteristics. The load prediction result is the basis for subsequent task distribution and load balancing scheduling decisions. Relevant data from multiple classification tasks will be collected, including the computing power characteristics and historical load data of each classification task, as well as real-time data on the current exploration rate and scheduling count. This data will be standardized, abnormal data (such as sudden peaks in historical load or invalid scheduling records) will be removed, and missing data will be supplemented to ensure the completeness and standardization of the data input to the model. All preprocessed data will be input into the load prediction model. Based on the time-series prediction advantages of LSTM, and combined with the influence of exploration rate and scheduling count, the model will perform load prediction for each classification task. During the prediction process, the model will capture the prediction weights corresponding to the scheduling count and exploration rate of the corresponding nodes based on the computing power characteristics of each classification task. The system captures the temporal variation patterns of load, outputting the future load variation trends of each category of task on different target computing power nodes, and obtaining quantitative load prediction results. Specifically, this includes the peak load, average load, and load fluctuation range of each category of task within a preset future time period (e.g., 1 hour, 4 hours), as well as the load occupancy of different nodes after taking on the task. The output load prediction results are cross-validated by comparing the load prediction results with the historical load data of the category tasks and the current node running status to calculate the prediction deviation. If the deviation exceeds a preset threshold, the prediction weights of the model are adjusted in a timely manner (combining the real-time changes in the exploration rate and scheduling count), and the prediction is re-performed. If the deviation meets the requirements, the final load prediction result is output, ensuring that the results can accurately reflect the future load of each category of task, providing support for subsequent task deployment based on the load prediction results and achieving load balancing. In this embodiment, by configuring an LSTM model guided by exploration rate and scheduling counts for classification task prediction, the pain point of poor load prediction adaptability in multi-cloud heterogeneous environments is effectively addressed. The LSTM model's time window length is dynamically adjusted based on scheduling counts, and prediction weights are optimized in conjunction with the exploration rate. Key model parameters are simultaneously and adaptively adjusted to ensure the model fully adapts to differences in node scheduling activity and considers the load potential of idle resources, thereby improving model adaptability and stability. During prediction, data standardization and cross-validation ensure that input data is standardized and prediction results are accurate. The output quantified load data clearly reflects future load changes and node occupancy for classification tasks. This solution provides reliable support for subsequent task distribution and load balancing scheduling, effectively avoiding node overload and resource idleness issues, improving the utilization rate of multi-cloud heterogeneous computing resources, ensuring efficient and stable operation of the scheduling system, and meeting the scheduling needs of multi-classification tasks.

[0022] Please see Figure 6 The sixth embodiment of a multi-cloud heterogeneous computing power scheduling method in this invention includes: 601. Collect the operating status of multiple target computing nodes; In this embodiment, the process of collecting the operating status of multiple target computing power nodes revolves around node operating indicators, specifically as follows: The collected indicators include node hardware operating status (CPU utilization, GPU utilization, memory usage, disk read / write speed, hardware temperature), network operating status (node ​​access bandwidth, data transmission latency, packet loss rate), task execution status (current number of tasks, task execution progress, load occupancy), and abnormal status (hardware failure, link interruption, overload warning, etc.). The collection frequency is dynamically adjusted in conjunction with the load prediction results, clearly presenting the real-time carrying capacity and operating health of each target computing power node, providing a basis for the generation of scheduling schemes, and avoiding problems such as overload and lag after task distribution due to untimely node status perception. 602. Generate scheduling schemes based on cloud space bandwidth, load prediction results, and operational status; In this embodiment, the specific generation process of the scheduling scheme is as follows: First, element matching analysis: Matching the previously allocated cloud space bandwidth with the transmission requirements of each category of tasks to ensure that the bandwidth requirements of each task are met, avoiding data transmission lag due to insufficient bandwidth; combining load prediction results, clarifying the future load occupancy of each category of tasks on different target nodes, and predicting node load pressure; comparing the real-time operating status of nodes, filtering out nodes that are currently operating normally and whose load has not reached the upper limit, and eliminating faulty or overloaded nodes; Second, task allocation optimization: Based on the computing power characteristics and latency requirements of the category of tasks, combined with the node operating status and load prediction results, using a load balancing algorithm, reasonably allocating each category of tasks to target computing power nodes, for example… High-concurrency, low-latency tasks are assigned to nodes with low CPU / GPU utilization and low transmission latency, while large-volume tasks are assigned to nodes with sufficient bandwidth and compatible storage interfaces. Simultaneously, based on cloud bandwidth allocation, the task data transmission path is optimized to ensure transmission efficiency. The third step involves scheme verification and adjustment. The generated preliminary scheduling scheme is verified, and the total load on each node after accepting tasks is calculated. If the load on a node exceeds a preset threshold (e.g., CPU utilization ≥ 80%), task allocation is adjusted, migrating some tasks to nodes with lower loads. If bandwidth allocation does not match task transmission requirements, the bandwidth allocation ratio is adjusted accordingly. Ultimately, a scheduling scheme that balances efficiency, stability, and load balancing is formed, providing a clear execution basis for subsequent task deployment. 603. Based on the scheduling scheme, the preset cross-cloud API interface and heterogeneous computing power adaptation interface, multiple classified tasks are distributed to multiple target computing power nodes to obtain multiple task target nodes; In this embodiment, a pre-defined cross-cloud API interface is used to achieve interconnection and interoperability between different cloud platforms, ensuring that tasks can be transmitted and deployed across clouds. A heterogeneous computing power adaptation interface is used to solve the compatibility issues between computing power nodes with different architectures (such as x86 / ARM) and the categorized tasks, ensuring that tasks can be loaded and executed normally on the target nodes. The task deployment process strictly follows the scheduling scheme and is divided into three stages: The first stage is interface adaptation and debugging. Based on the target node type of each category of task in the scheduling scheme, the corresponding heterogeneous computing power adaptation interface is matched. Simultaneously, a communication connection is established between the scheduling system and each target node through the cross-cloud API interface, and interface connectivity is debugged to ensure that data transmission and task deployment are error-free. The second stage involves batch deployment of tasks, combined with the task allocation rules of the scheduling scheme. Multiple categorized tasks are distributed to the corresponding target computing power nodes in batches, prioritizing high-priority, low-latency tasks. During the distribution process, the task transmission status and node reception are monitored in real time. If issues such as interface adaptation anomalies or transmission interruptions occur, backup interfaces are switched or the distribution order is adjusted in a timely manner to ensure successful task distribution. In the third stage, task binding and node finalization occur. Once a categorized task is successfully distributed to the target computing power node, and the node confirms receipt and completes task initialization, the target computing power node is bound to the corresponding categorized task, forming a task target node. At the same time, the task acceptance status and interface running status of each task target node are recorded and updated synchronously to the global resource view, providing support for subsequent task monitoring and load adjustment, and ensuring that the task target node can stably accept and execute the corresponding categorized tasks. In this embodiment, by collecting multi-dimensional operational status data of target computing nodes, it is easy to grasp the real-time carrying capacity of each target computing node, avoiding problems such as task overload and lag caused by untimely status awareness. A scheduling scheme is generated based on cloud space bandwidth, load prediction results, and node operational status. Through element matching, task optimization, and verification adjustment, the matching of classified tasks with target nodes is achieved, load balancing is realized, and the utilization rate of computing resources is improved. Relying on cross-cloud API interfaces and heterogeneous computing power adaptation interfaces, it ensures that tasks are successfully deployed across clouds and architectures. At the same time, tasks are bound to nodes, and the global resource view is updated synchronously, providing support for subsequent task monitoring and load adjustment, and comprehensively improving the stability, efficiency, and compatibility of multi-cloud heterogeneous computing power scheduling.

[0023] Please see Figure 7 The seventh embodiment of a multi-cloud heterogeneous computing power scheduling method in this invention includes: 701. Allocate bandwidth to multiple classified tasks based on bandwidth allocation strategies and storage characteristics to obtain cloud space bandwidth; In this embodiment, the preset bandwidth allocation strategy includes two types: dynamic bandwidth allocation strategy and static bandwidth allocation strategy. The dynamic bandwidth allocation strategy is suitable for scenarios with high read / write frequency, large data transmission volume, and high latency requirements, while the static bandwidth allocation strategy is suitable for scenarios with stable read / write frequency and non-real-time requirements. Subsequently, by comparing the storage characteristics of each category task, key indicators are extracted, including data volume, data read / write frequency, storage latency requirements, and data transmission direction (upload / download), clarifying the differences in bandwidth requirements for each category task. For example, category tasks with storage characteristics of "large data volume, high read / write frequency, and low latency" (such as real-time data processing subtasks) are adapted to the dynamic bandwidth allocation strategy; category tasks with storage characteristics of "small data volume, low read / write frequency, and non-real-time" (such as offline data query subtasks) are adapted to the static bandwidth allocation strategy. The second step is to quantify the allocation and calculate the quota to obtain the cloud space bandwidth. Based on the matched strategy and storage characteristics, for each... Bandwidth allocation is calculated for each category of tasks: For tasks using a dynamic bandwidth allocation strategy, real-time bandwidth requirements are calculated based on data read / write frequency and data volume, reserving 10%-20% bandwidth redundancy to avoid congestion during peak transmission; for tasks using a static bandwidth allocation strategy, a fixed bandwidth quota is determined based on historical read / write data and data volume to ensure basic transmission needs are met while avoiding bandwidth waste; after calculation, the bandwidth requirements of all categories of tasks are summarized, and an overall balance adjustment is made in conjunction with the total bandwidth resources of the multi-cloud heterogeneous environment. If the total demand exceeds the total bandwidth, priority is given to ensuring the bandwidth requirements of high-priority, low-latency tasks, while the bandwidth quota for non-real-time tasks is appropriately reduced. Finally, the cloud space bandwidth corresponding to each category of task is determined, forming a standardized bandwidth allocation list to ensure that the bandwidth allocation is both in line with the storage characteristics of each task and in accordance with the bandwidth allocation strategy, balancing efficiency and resource conservation, and providing stable bandwidth support for subsequent task deployment and data transmission; 702. Determine multiple target computing power nodes and heterogeneous computing power adaptation interfaces for cloud storage space based on computing power characteristics and storage characteristics; In this embodiment, the task's dependence on storage resources and interface requirements are determined based on computing power and storage characteristics. Then, by comparing all computing power nodes in the global resource view, nodes meeting the following conditions are selected as candidate computing power nodes: First, hardware configuration matches computing power characteristics; for example, for GPU-intensive tasks (computing power characteristic is high GPU computing power requirement), nodes equipped with high-performance GPUs are selected, and for CPU-intensive tasks, nodes with a large number of CPUs and high clock speeds are selected. Second, storage adapts to storage characteristics; the node's storage interface type matches the task's storage requirements, and the storage read / write speed meets the task requirements. Third, the node is in normal operating condition, with a load rate within a reasonable range (30%-70%), and has the capacity to handle tasks. Finally, from the candidate computing power nodes, combined with task priority and node association strength, several optimal nodes are selected as target computing power nodes to ensure that each category of task has a corresponding adapted node, while achieving initial load balancing of nodes. The heterogeneous computing power adaptation interface serves to shield computing power nodes with different architectures (such as x86 architecture, ARM architecture, heterogeneous computing...). The underlying differences in architecture are considered to ensure that classification tasks can be loaded and executed normally on the target computing power nodes. During the determination process, the architecture type and interface standard of each target computing power node are first clarified. Then, combined with the computing power characteristics and storage characteristics of the classification tasks, corresponding heterogeneous computing power adaptation interfaces are matched. For example, for x86 architecture target computing power nodes, if the task is CPU-intensive (computing power characteristic) and structured data storage (storage characteristic), a general computing power adaptation interface corresponding to the x86 architecture is matched; for ARM architecture target computing power nodes, if the task is lightweight computing (computing power characteristic) and unstructured data storage (storage characteristic), a dedicated ARM architecture adaptation interface is matched; for heterogeneous architecture nodes (such as CPU+GPU hybrid architecture), an adaptation interface supporting multiple architectures is matched to ensure that the interface is compatible with the node architecture and task requirements, achieving smooth connection between tasks and nodes. Simultaneously, connectivity tests are performed on the determined heterogeneous computing power adaptation interfaces to verify whether the interfaces can transmit task data normally and support task execution, ensuring interface adaptability and reliability, and providing technical support for subsequent task distribution. In this embodiment, bandwidth allocation strategies and storage characteristics are combined to achieve differentiated and quantitative bandwidth allocation. Dynamic strategies adapt to high-demand tasks and reserve redundancy, while static strategies ensure basic needs are met and avoid waste, balancing total bandwidth resources, improving bandwidth utilization, and providing stable support for data transmission. Simultaneously, target computing nodes are selected based on computing power and storage characteristics to ensure hardware, storage, and task requirements are matched, while also considering node load balancing to guarantee stable execution of computing power scheduling tasks. By matching heterogeneous computing power adaptation interfaces, differences in node architecture are masked, and connectivity testing ensures interface reliability, enabling smooth task-node connection and laying a solid foundation for subsequent task deployment. This comprehensively improves the efficiency, compatibility, and resource utilization of multi-cloud heterogeneous computing power scheduling.

[0024] The above describes a multi-cloud heterogeneous computing power scheduling method according to an embodiment of the present invention. The following describes a multi-cloud heterogeneous computing power scheduling device according to an embodiment of the present invention. Please refer to [link / reference]. Figure 8 One embodiment of a multi-cloud heterogeneous computing power scheduling device according to the present invention includes: Feature analysis module 1 is used to collect computing power scheduling tasks and perform feature analysis on the computing power scheduling tasks according to the preset computing power scheduling algorithm to obtain computing power features and storage features. Data acquisition module 2 is used to collect data from a preset cloud storage space according to a preset time collection range and storage characteristics in order to obtain multi-cloud heterogeneous resource data. Resource view construction module 3 is used to construct a global resource view based on multi-cloud heterogeneous resource data and a preset heterogeneous resource graph construction algorithm; The data determination module 4 is used to determine the cloud space bandwidth, heterogeneous computing power adaptation interface and multiple target computing power nodes of the cloud storage space according to the preset bandwidth allocation strategy and computing power characteristics. The distribution module 5 is used to distribute computing power scheduling tasks to each target computing power node based on the global resource view, the preset graph association analysis method, cloud space bandwidth and heterogeneous computing power adaptation interface, so as to obtain multiple task target nodes; In this embodiment, a full-link design encompassing computing power scheduling task collection and feature analysis, multi-cloud heterogeneous resource collection, global resource view construction, resource adaptation, and task distribution effectively addresses pain points such as resource dispersion, mismatch between demand and resources, and inefficient scheduling in multi-cloud heterogeneous environments. By extracting computing power and storage features, a task demand profile is constructed, providing a reliable basis for the entire scheduling process. Collecting full-volume multi-cloud heterogeneous resource data in conjunction with storage features and time ranges ensures the relevance and completeness of resource data. A global resource view is created through a heterogeneous resource graph construction algorithm, achieving global awareness and unified management of network resources. Based on computing power features, bandwidth adaptation strategies are applied, and adaptation interfaces and target nodes are determined. Task distribution is completed by combining the global resource view and graph association analysis, masking architectural differences, ensuring stable transmission, and avoiding node overload and resource idleness. This effectively improves computing power scheduling efficiency, resource utilization, and task execution stability in multi-cloud heterogeneous environments, laying a solid foundation for subsequent scheduling and management, and achieving optimized allocation of computing power resources and orderly and efficient advancement of the scheduling process.

[0025] Figure 9This is a schematic diagram of a multi-cloud heterogeneous computing power scheduling device 900 provided in an embodiment of the present invention. This multi-cloud heterogeneous computing power scheduling device 900 can vary significantly due to different configurations or performance. It may include one or more central processing units (CPUs) 910 (e.g., one or more processors) and a memory 920, and one or more storage media 930 (e.g., one or more mass storage devices) for storing application programs 933 or data 932. The memory 920 and storage media 930 can be temporary or persistent storage. The program stored in the storage media 930 may include one or more modules (not shown in the diagram), each module including a series of instruction operations on the multi-cloud heterogeneous computing power scheduling device 900. Furthermore, the processor 910 may be configured to communicate with the storage media 930 and execute the series of instruction operations in the storage media 930 on the multi-cloud heterogeneous computing power scheduling device 900 to implement the steps of the multi-cloud heterogeneous computing power scheduling method provided in the above-described method embodiments.

[0026] A multi-cloud heterogeneous computing power scheduling device 900 may further include one or more power supplies 940, one or more wired or wireless network interfaces 950, one or more input / output interfaces 960, and / or one or more operating systems 931, such as Windows Server, MacOSX, Unix, Linux, FreeBSD, etc. Those skilled in the art will understand that... Figure 9 The illustrated multi-cloud heterogeneous computing power scheduling device structure does not constitute a limitation on a multi-cloud heterogeneous computing power scheduling device. It may include more or fewer components than illustrated, or combine certain components, or have different component arrangements.

[0027] The present invention also provides a computer-readable storage medium, which can be a non-volatile computer-readable storage medium or a volatile computer-readable storage medium, wherein the computer-readable storage medium stores instructions that, when executed on a computer, cause the computer to perform the steps of a multi-cloud heterogeneous computing power scheduling method.

[0028] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working process of the system, device, or unit described above can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0029] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0030] Finally, it should be noted that the above descriptions are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A multi-cloud heterogeneous computing power scheduling method, characterized in that, include: Collect computing power scheduling tasks and perform feature analysis on the computing power scheduling tasks according to the preset computing power scheduling algorithm to obtain computing power features and storage features; Multi-cloud heterogeneous resource data are collected from a preset cloud storage space based on a preset time collection range and storage characteristics; A global resource view is constructed based on multi-cloud heterogeneous resource data and a preset heterogeneous resource graph construction algorithm; The cloud storage space bandwidth, heterogeneous computing power adaptation interface, and multiple target computing power nodes are determined based on the preset bandwidth allocation strategy and computing power characteristics. Based on the global resource view, the preset graph association analysis method, cloud space bandwidth and heterogeneous computing power adaptation interface, the computing power scheduling task is distributed to each target computing power node to obtain multiple task target nodes.

2. The multi-cloud heterogeneous computing power scheduling method as described in claim 1, characterized in that, The process of constructing a global resource view based on multi-cloud heterogeneous resource data and a preset heterogeneous resource graph construction algorithm includes: Based on preset cloud platform node attributes, preset computing power node attributes, preset network link node attributes, and preset resource pool node attributes, feature analysis is performed on multi-cloud heterogeneous resource data to obtain correlation strength parameters, number of computing power nodes, and resource specifications. The action space is determined based on the number of computing nodes and resource specifications. A global resource view is constructed based on correlation strength parameters, action space, multi-cloud heterogeneous resource data, and heterogeneous resource graph construction algorithms.

3. The multi-cloud heterogeneous computing power scheduling method as described in claim 1, characterized in that, The computing power scheduling task is distributed to each target computing power node based on the global resource view, a preset graph association analysis method, cloud space bandwidth, and heterogeneous computing power adaptation interface, to obtain multiple task target nodes, including: The computing power scheduling tasks are divided based on the pre-trained scene recognition model, computing power features, and storage features to obtain multiple classification tasks; Based on a pre-defined long short-term memory model, a global resource view, and graph association analysis methods, predictions are made for multiple classification tasks to obtain load prediction results. Based on load prediction results, cloud space bandwidth, and heterogeneous computing power adaptation interfaces, multiple classification tasks are distributed to various target computing power nodes to obtain multiple task target nodes.

4. The multi-cloud heterogeneous computing power scheduling method as described in claim 3, characterized in that, The process of predicting multiple classification tasks based on a preset long short-term memory model, a global resource view, and graph association analysis to obtain load prediction results includes: The optimal network path is obtained by matching multiple preset network paths based on computing power characteristics; A deep analysis of the global resource view is performed based on the optimal network path and graph association analysis method to obtain the exploration rate and scheduling count; The load prediction results are obtained by predicting multiple classification tasks based on the exploration rate, scheduling times, and long short-term memory model.

5. The multi-cloud heterogeneous computing power scheduling method as described in claim 4, characterized in that, The process of predicting multiple classification tasks based on exploration rate, scheduling count, and long short-term memory model to obtain load prediction results includes: The long short-term memory model is configured based on the exploration rate and the number of scheduling attempts to obtain the load prediction model; The load prediction model is used to predict multiple classification tasks to obtain the load prediction results.

6. The multi-cloud heterogeneous computing power scheduling method as described in claim 3, characterized in that, The method, based on load prediction results, cloud bandwidth, and heterogeneous computing power adaptation interface, distributes multiple classification tasks to various target computing power nodes to obtain multiple task target nodes, including: Collect the operational status of multiple target computing nodes; A scheduling scheme is generated based on cloud space bandwidth, load prediction results, and operational status. Based on the scheduling scheme, the preset cross-cloud API interface and heterogeneous computing power adaptation interface, multiple classified tasks are distributed to various target computing power nodes to obtain multiple task target nodes.

7. The multi-cloud heterogeneous computing power scheduling method as described in claim 3, characterized in that, The process of determining the cloud storage space bandwidth, heterogeneous computing power adaptation interface, and multiple target computing power nodes based on a preset bandwidth allocation strategy and computing power characteristics includes: Bandwidth is allocated to multiple classified tasks based on bandwidth allocation strategies and storage characteristics to obtain cloud space bandwidth; Based on computing power characteristics and storage characteristics, determine multiple target computing power nodes and heterogeneous computing power adaptation interfaces for cloud storage space.

8. A multi-cloud heterogeneous computing power scheduling device, characterized in that, include: The feature analysis module is used to collect computing power scheduling tasks and perform feature analysis on the computing power scheduling tasks according to the preset computing power scheduling algorithm to obtain computing power features and storage features. The data acquisition module is used to collect data from a preset cloud storage space according to a preset time range and storage characteristics in order to obtain multi-cloud heterogeneous resource data. The resource view construction module is used to construct a global resource view based on multi-cloud heterogeneous resource data and a preset heterogeneous resource graph construction algorithm. The data determination module is used to determine the cloud space bandwidth, heterogeneous computing power adaptation interface and multiple target computing power nodes of the cloud storage space based on the preset bandwidth allocation strategy and computing power characteristics. The distribution module is used to distribute computing power scheduling tasks to various target computing power nodes based on the global resource view, preset graph association analysis methods, cloud space bandwidth, and heterogeneous computing power adaptation interface, so as to obtain multiple task target nodes.

9. A multi-cloud heterogeneous computing power scheduling device, characterized in that, The multi-cloud heterogeneous computing power scheduling device includes: a memory and at least one processor, wherein the memory stores instructions; At least one of the processors invokes the instructions in the memory to cause the multi-cloud heterogeneous computing power scheduling device to perform the steps of the multi-cloud heterogeneous computing power scheduling method as described in any one of claims 1-7.

10. A computer-readable storage medium storing instructions thereon, characterized in that, When the instructions are executed by the processor, they implement the various steps of the multi-cloud heterogeneous computing power scheduling method as described in any one of claims 1-7.

Images