Computing power scheduling platform and computing power leasing system comprising same

Abstract

Provided in the present application are a computing power scheduling platform and a computing power leasing system comprising same. The computing power scheduling platform comprises: an order management unit, which is configured to receive at least one computing task from each platform user; a storage space allocation unit, which is configured to allocate to the platform user a data storage space having a target capacity; a computing power allocation unit, which is configured to allocate, on the basis of computing power requirements of the computing task, a corresponding computing node to the computing task; and a cost calculation unit, which is configured to record the computing resource usage of the computing task and generate corresponding cost information. The computing task is packaged in an independent container, and the container is started on the computing node allocated by the computing power allocation unit, such that the computing task is executed. The embodiments of the present invention realize unified scheduling and management of computing power resources by means of containers and virtualization technology, effectively increase the utilization rate of computing power and reduce usage costs, and prevent unnecessary waste of computing power.

Description

Computing power scheduling platform and computing power leasing system thereof

[0001] CROSS-REFERENCE TO RELATED APPLICATIONS

[0002] This application claims priority to the Hong Kong Short-Term Patent Application No. 32024093689.0, filed on July 4, 2024, entitled “Computing power scheduling platform and computing power leasing system thereof”, the entire contents of which are incorporated herein by reference. TECHNICAL FIELD

[0003] The present application relates to the field of artificial intelligence technology, and in particular to a computing power scheduling platform and a computing power leasing system thereof. BACKGROUND

[0004] With the continuous development of artificial intelligence technology, the demand for computing power resources of many individuals and enterprise organizations is also growing. Since the cost of directly deploying and purchasing physical devices is very high, a leasing service for computing power resources has emerged to provide relatively low-cost available computing power resources for individuals and organizations. Typical computing power leasing businesses mainly include bare machine leasing and NVIDIA base system platform machine leasing.

[0005] Bare machine leasing refers to providing only physical devices without any system or software, and users install systems and software by themselves. In this application scenario, users need to have the ability to deploy from the bottom system. Moreover, physical devices can only be rented in the form of a whole machine (usually 8 cards per machine), and cannot provide rental of a single computing card.

[0006] Therefore, when the actual task computing power demand does not exceed 8 cards, a certain amount of computing power is wasted. The artificial intelligence development environment also needs to be built by the user, which is time-consuming and difficult. If only the operation and maintenance personnel are relied on to ensure the operation of the platform, the operation and maintenance difficulty is great, and the operation and maintenance work intensity is high.

[0007] NVIDIA base system platform machine leasing service refers to a server with a NVIDIA base system platform. Users can use computing power by adapting the interface of the system. In this application scenario, users do not need to have the ability to deploy from the bottom system, but only need to adapt the NVIDIA related interface.

[0008] Similarly, such leasing service can only provide the whole machine rental mode, and cannot realize the charging mode of using by quantity. The AI development environment also needs to be built by the user, which is time-consuming and difficult. If only the operation and maintenance personnel are relied on to ensure the operation of the platform, the operation and maintenance difficulty is great, and the operation and maintenance work intensity is high.

[0009] Therefore, there is an urgent need to provide a suitable method to solve the low utilization efficiency of existing computing power rental services and the inconvenience of managing and using them.

[0010] Application Content

[0011] The embodiments of the present application aim to provide a computing power scheduling platform and a computing power rental system, which can solve at least part of the defects of existing computing power rental services.

[0012] In a first aspect, the present application provides a computing power scheduling platform. The computing power scheduling platform comprises: an order management unit configured to receive at least one computing task from a platform user; a storage space allocation unit configured to allocate a data storage space with a target capacity to the platform user; a computing power allocation unit configured to allocate a corresponding computing node to the computing task according to the computing power requirement of the computing task; and a fee calculation unit configured to record the computing resource usage of the computing task and generate corresponding fee information; wherein the computing task is packaged in an independent container, the container is started on the computing node allocated by the computing power allocation unit, so that the computing task is executed; and the data information formed after the execution of the computing task is stored in the data storage space allocated by the storage space allocation unit.

[0013] Optionally, the computing power scheduling platform further comprises a monitoring unit configured to feed back the state information of the computing node during the execution of the computing task; the state information is selected from one or more of the following: CPU load, GPU load, and physical memory usage ratio.

[0014] Optionally, the computing power scheduling platform further comprises a platform interface presented on a user terminal; wherein the platform interface comprises an access page for accessing the container.

[0015] Optionally, the platform interface further comprises a user login page for platform user registration and login.

[0016] Optionally, the target capacity is determined by the platform user; and the data storage space is mapped to the container.

[0017] Optionally, the data storage space is provided with a plurality of storage space replicas.

[0018] Optionally, the data information formed after the execution of the computing task is stored based on a persistent storage mechanism, so that the data information is independent of the life cycle of the container.

[0019] Optionally, the computing power allocation unit is specifically configured to determine a computing power requirement of the computing task, and authorize a computing node satisfying the computing power requirement to a corresponding platform user, and bind the authorized computing node with the platform user.

[0020] In a second aspect, the embodiments of the present application further provide a computing power renting system. The system comprises: a physical device providing computing power resources and storage space; and a computing power scheduling platform as described above; wherein the computing power scheduling platform abstracts the computing power resources and storage space of the physical device by using a container and a virtualization technology to form corresponding virtual resources.

[0021] Optionally, the computing power scheduling platform forms the virtual resources corresponding to the physical device by using a Docker container and a KVM virtualization platform.

[0022] Compared with the prior art, the embodiments of the present application realize unified scheduling and management of computing power resources by using a container and a virtualization technology, so that users can conveniently use the computing power resources under the management of the platform and obtain corresponding computing results. Moreover, the renting fee can be calculated according to the amount of computing power used, effectively improving the utilization rate of computing power and reducing the use cost of customers, and avoiding unnecessary waste of computing power. BRIEF DESCRIPTION OF DRAWINGS

[0023] One or more embodiments are illustrated by way of example with reference to the accompanying drawings, which are schematic and not intended to be limiting of the embodiments, and in which like reference numerals designate similar elements in the figures and wherein the use of "for example", "e.g.", "of the embodiments", "an embodiment", "one embodiment", "some embodiments", "apparent to those skilled in the art", "clear to those skilled in the art" and "apparent from" or similar phrases in the description denotes differentiators and not an essential limitation.

[0024] Fig. 1 is an application scenario of a computing power scheduling platform according to an embodiment of the present application;

[0025] Fig. 2 is a data processing flowchart of a computing power scheduling platform according to an embodiment of the present application;

[0026] Fig. 3 is a functional block diagram of a computing power scheduling platform according to an embodiment of the present application;

[0027] Fig. 4 is a page diagram of a computing power scheduling platform according to an embodiment of the present application;

[0028] Fig. 5 is a schematic diagram of a computing power renting system according to an embodiment of the present application. DETAILED DESCRIPTION

[0029] For the convenience of understanding the present application, the present application will be described in more detail below in conjunction with the accompanying drawings and specific embodiments. It should be noted that when an element is described as "fixed to" another element, it can be directly on the other element or one or more intervening elements can be present therebetween. When an element is described as "connected to" another element, it can be directly connected to the other element or one or more intervening elements can be present therebetween. The terms "upper", "lower", "inner", "outer", "bottom", and the like used in the specification indicate the orientation or positional relationship shown in the drawings and are used only for the purpose of facilitating the description of the present application and simplifying the description, and therefore cannot be understood as indicating or implying that the device or element referred to must have a particular orientation, be constructed and operated in a particular orientation, and therefore cannot be understood as limiting the present application. In addition, the terms "first", "second", "third", and the like are only for descriptive purposes and cannot be understood as indicating or implying relative importance.

[0030] Unless otherwise defined, all technical and scientific terms used in the specification have the same meaning as commonly understood by one of ordinary skill in the art to which the present application belongs. The terms used in the specification of the present application are only for the purpose of describing the specific embodiments of the present application and are not used to limit the present application. The term "and / or" used in the specification includes any and all combinations of one or more related listed items.

[0031] In addition, the technical features involved in the different embodiments of the present application described below can be combined with each other as long as there is no conflict between them.

[0032] The technical solutions in the present application will be described below in conjunction with the accompanying drawings.

[0033] FIG. 1 is an application scenario of a computing power scheduling platform provided by an embodiment of the present application. As shown in FIG. 1, in the application scenario, a plurality of different user terminals 10, a computing power scheduling platform 20, physical devices 30, and network services 40 are included.

[0034] The user terminal 10 is a terminal device for user interaction with a computer system. It can allow a user to input instruction information, process the instruction information in one or more ways, and / or present feedback information formed after processing.

[0035] It can be any type of intelligent device with logical operation capability, including but not limited to smartphones, personal computers, laptop computers, and intelligent wearable devices. It should be noted that the user terminal is not specifically a single device, and it can be composed of a plurality of device components that establish communication connections with each other and are used to perform different functions, or it can be a plurality of functional modules integrated in the same device.

[0036] The user terminal 10 can also be configured with one or more different interactive devices to collect user instructions or show and feedback information to the user. These interactive devices include, but are not limited to, input keyboards, display screens, touch screens, and speakers, etc.

[0037] The computing power scheduling platform 20 is a system platform based on a network to provide various computing resources and services, and integrated with multiple functions. It can abstract the computing power and storage space resources provided by the physical device 30 into virtual resources through container and virtualization technology, and further provide on-demand allocation, elastic scaling, and high availability services.

[0038] Specifically, the computing power scheduling platform 20 can abstract the physical device 30 through Docker containers and KVM virtualization platforms to facilitate management thereof.

[0039] The physical device 30 is a real physical device that provides computing power resources and storage space resources. These real physical devices support the operation of computing systems and cloud platforms, and together constitute a computing infrastructure. It includes, but is not limited to, a processor (CPU) that provides computing resources, a graphics processing unit (GPU), a hard disk drive (HDD) and a solid state drive (SSD) that provide storage space for persistent storage of data, a power supply (PSU), and a cooling system, etc.

[0040] In this application, for the sake of simplicity, "first physical device" and "second physical device" are used to represent the device components for providing computing resources and storage space, respectively. However, those skilled in the art can understand that the "first physical device" and "second physical device" are only a classification description of hardware components according to the different functions to be performed, and are not used to specifically distinguish different devices. That is, the first physical device and the second physical device can be different components in the same device, or different devices.

[0041] The network service 40 is a wireless communication network for establishing a data transmission channel between two nodes, which is specifically based on any type of data transmission principle, such as a WiFi network located in a specific signal frequency band, a wireless cellular network, or a combination thereof, to establish a communication connection between the user terminal 10, the computing power scheduling platform 20, and the physical device 30.

[0042] In actual use, the user can use the user terminal 10 to log in to the computing power scheduling platform 20 by logging in to a portal website or starting a dedicated mobile application, and send a computing power rental request or similar order.

[0043] Subsequently, the computing power scheduling platform 20 allocates corresponding storage space and computing power resources to the user through an automated scheduling manner, so that the user can complete the computing power use request. The completed data information and related user data can be saved in an independent isolated storage space, so that the user can view and obtain the data through the user terminal 10.

[0044] The computing power scheduling platform 20 provided by the embodiments of the present application adopts an integrated docker container and KVM virtualization platform, does not need to install an agent program on the computing node, realizes one-network management, and reduces the use difficulty. The user or user can conveniently place an order, schedule and allocate computing resources, and realize unified management of orders, processes and billing through the user terminal 10. The computing power scheduling platform 20 is compatible with Nvidia DGX, Ubuntu, RHEL and other operating systems.

[0045] The computing power scheduling platform 20 defines partitioning and authorization, realizes fine-grained partitioning and license management, and can allocate corresponding computing resources to the user according to the computing task and authorize the user, realize authorized access from a bare machine to a single container, and guarantee the security of user data.

[0046] It should be noted that the application scenario shown in FIG. 1 is only used for illustrative purposes. One or more devices can be added or removed according to the actual needs of the situation, and is not limited to the devices shown in FIG. 1.

[0047] FIG. 2 is a schematic diagram of a data processing flow of the computing power scheduling platform provided by the embodiments of the present application. As shown in FIG. 2, the data processing method comprises the following steps:

[0048] S100, the platform user submits a computing task through a user terminal.

[0049] The "platform user" refers to the account information after the user information is registered and the account information is completed on the computing power scheduling platform. The user or user can register and generate an account on a portal website or a specific mobile application through a user terminal in advance, and the account can be managed by using any suitable type of platform user management method.

[0050] S200, the computing power scheduling platform verifies and audits the platform user and the computing task submitted by the platform user. After the verification and audit is passed, the computing task is bound and authorized to the corresponding platform user.

[0051] The verification and audit can be realized by any suitable method, including but not limited to account password, biological fingerprint information, etc. Through the permission management of the verification and audit, unauthorized access can be prevented, and the security and privacy of the data can be ensured.

[0052] S300, allocating a data storage space with a target capacity for the platform user in the storage resource of the distributed storage architecture.

[0053] The target capacity can be determined by the user or the user according to the actual needs, which is not limited here.

[0054] In addition, the data storage space is mapped to the container and used to store the data information of the platform user. Specifically, the mapping of the data storage space can be realized by the way of volume (Volumes) or bind mount (Bind Mounts).

[0055] Preferably, the data storage space can adopt a distributed redundant storage architecture, and multiple storage space copies are set. These storage space copies store user models, images, and calculation settlement data, which can be stored on different physical nodes. Thus, when the computing node being used fails, the image can be switched to the normal computing node in seconds, thereby solving the problem of task execution and data loss caused by computing node failure, and ensuring that the computing task can be continued to be executed in other normal running computing nodes when the computing node hardware fails.

[0056] S400, according to the computing power requirement of the computing task, the corresponding computing node is allocated and scheduled by the computing power scheduling platform, so that it can complete the computing task.

[0057] Each computing task is packaged in an independent container. Thus, with the start of the container on the computing node, the computing task can be executed on the corresponding computing node.

[0058] In this application, the term "container" is used to represent a lightweight, portable virtualization technology. It encapsulates the application and all its dependencies, so that the application can run consistently in any environment that supports container runtime. The term "computing node" is used to represent a physical device or virtual machine that provides actual computing resources. It can be considered as the basic unit of the computing power scheduling platform for executing various computing tasks.

[0059] The term "authorized computing node" means a computing node allocated for a specific platform user. The authorized computing node can be bound to the corresponding platform user, and the platform user who has logged in can access and control these computing nodes by accessing the access page of the container.

[0060] Specifically, based on the different computing power resources required by the computing task, the authorized computing nodes can also have corresponding different levels. For example, for a computing task that requires more computing power resources, a single machine can be authorized, and for a computing task that requires less computing power resources, a single computing card in a single machine can be authorized. Further, for a computing task that requires very little computing power resources, a certain container can also be authorized. As an example but not limitation, the computing power scheduling platform adopts a predictive resource scheduling algorithm, which can pre-allocate computing resources based on historical task execution data, current system load and time patterns, to reduce user waiting time. This algorithm can identify periodic task patterns and proactively adjust resource allocation strategies before peak periods to ensure balanced resource utilization.

[0061] S500, the computing power scheduling platform collects the specific use of computing power and calculates the corresponding computing power rental fee according to the collected computing power usage amount as the computing task starts running.

[0062] The computing power rental fee can be calculated according to a preset billing rule. The specific billing rule can be determined according to the actual needs, which is not limited here. Preferably, the billing rule can also be an adjustable rule, which can be adjusted according to different rental needs. For example, a specific fee discount is provided for long-term and stable computing power rental needs.

[0063] S600, the computing nodes allocated and scheduled by the computing power scheduling platform can be provided to the user terminal through the corresponding container access page, so that the user can conveniently operate and control the corresponding computing nodes.

[0064] S800, the data information generated after the computing node is calculated is stored in the data storage space, which can be read in response to the request of the user terminal.

[0065] The data information formed after the computing task is executed is stored based on the persistent storage mechanism, so that these data information is independent of the life cycle of the container, and can be accessed and obtained by the user terminal 10 even after the container is destroyed.

[0066] In some embodiments, in order to facilitate the user or user to know the running condition of the computing node in time, the computing power scheduling platform also provides a task monitoring function. Please continue to refer to FIG. 2, and the corresponding data data method also includes:

[0067] S700, in the process of executing the computing task, the state information of the computing node is fed back in real time.

[0068] The state information is selected from one or more of the following: CPU load, GPU load, and physical memory usage ratio. The state information can be displayed and presented in any suitable manner according to actual needs. For example, the number of computing power nodes, the number of available CPU cores and used CPU cores, the number of available GPU cores and used GPU cores, the available memory capacity and used memory capacity, and the like can be displayed in the manner shown in FIG. 4. Based on the data processing procedure provided in the above embodiments, the present embodiment further provides a specific structure of the computing power scheduling platform. FIG. 3 shows a functional block diagram of the computing power scheduling platform 20 according to the present embodiment.

[0069] As shown in FIG. 3, the computing power scheduling platform 20 includes an order management unit 21, a storage space allocation unit 22, a computing power allocation unit 23, and a fee calculation unit 24.

[0070] The order management unit 21 is configured to receive at least one computing task from a platform user. The storage space allocation unit 22 is configured to allocate a data storage space with a target capacity to the platform user. The computing power allocation unit 23 is configured to allocate a corresponding computing node to the computing task according to the computing power requirement of the computing task. The fee calculation unit 24 is configured to record the computing resource usage of the computing task and generate corresponding fee information.

[0071] In actual use, the platform user submits a rental order through a user terminal, and specifies the computing task and / or computing power requirement required by the user. The computing power scheduling platform automatically selects and allocates appropriate resources (e.g., computing nodes and storage spaces) for the user in an automatic scheduling manner. Subsequently, the computing power scheduling platform binds the allocated resources to the identity of the platform user, so as to ensure that only the platform user can access and use the resources.

[0072] When the platform user indicates the computing task to run, the computing power scheduling platform packs the computing task into a container, starts the container on the allocated computing node, and starts running the task. At the same time, the computing power scheduling platform starts recording the usage of the computing power and calculating the corresponding rental fee as the computing task starts to execute.

[0073] In some embodiments, the order management unit further has a task priority dynamic adjustment mechanism, which can intelligently adjust the task execution priority according to the user level, the task urgency, the system load condition, and the like, so as to ensure that critical tasks can be given priority processing in the case of limited resources, while maximizing the overall throughput of the system.

[0074] In some embodiments, please refer to FIG. 2, the computing power scheduling platform 20 can further comprise a monitoring unit 25. The monitoring unit 25 is configured to feed back the state information of the computing node during the execution of the computing task. The state information is selected from one or more of the following: CPU load, GPU load, and physical memory usage ratio.

[0075] Thus, the user or user can conveniently view the execution of the current computing task, the use of resources, and the rental fee information, etc. through the portal supported by the computing power scheduling platform.

[0076] Preferably, the computing power scheduling platform can also provide the function of automatic fault alarm. It can automatically warn before the fault (for example, when a certain parameter exceeds the threshold value) so that the operation and maintenance personnel can predict the fault point according to the parameter situation and handle it specifically, saving the inspection time.

[0077] Moreover, during the fault handling, the computing task can be conveniently scheduled to other physical devices, so that the platform user can continue to use without waiting for the fault to be handled. After the fault is solved, the physical device can be re-included in the abstracted virtual resource pool and wait for scheduling to execute other computing tasks.

[0078] In addition, the monitoring unit 25 is not only responsible for feeding back the state information of the computing node, but also undertakes the function of resource usage pattern mining. Through the analysis of historical monitoring data, the computing power scheduling platform 20 can identify the time pattern of resource utilization, the task type pattern, and the user behavior pattern, providing data support for predictive resource allocation. The data collected by the monitoring unit 25 will be used as the training set of the machine learning model, continuously optimizing the computing power scheduling strategy.

[0079] The computing power allocation unit 23 is further configured to predict the resource demand characteristics of the computing task based on a machine learning model trained by historical task execution data. The computing power scheduling platform 20 can intelligently match the most suitable computing node according to the multi-dimensional characteristics of the computing task type, input data size, expected execution time, etc., rather than simply judging the availability of resources. The computing power allocation unit 23 continuously learns the relationship between user task characteristics and resource utilization patterns, and continuously optimizes the allocation strategy.

[0080] Specifically, the computing power allocation unit 23 calculates the optimal resource allocation scheme based on the trained task-resource matching model according to the computing power demand feature vector of the computing task, and schedules and allocates the corresponding computing node. The task-resource matching model can comprehensively consider the multi-attribute of the computing task such as computing-intensive, memory-intensive, or IO-intensive, as well as the load balancing, energy efficiency, and other factors of the node, to achieve globally optimal resource allocation.

[0081] The computing power requirement feature vector is a multi-dimensional numerical representation of the resource requirements of a computing task, quantifying the computing characteristics of the task into structured data. This vector typically includes dimensions such as computing intensity, memory requirements, I / O intensity, parallelism, GPU acceleration suitability, expected execution duration, data size, and task type identification. For example, the feature vector of a deep learning training task may exhibit a combination of high computing intensity, high GPU suitability, and longer execution time.

[0082] The task-resource matching model is a trained machine learning model that establishes a mapping function from task features to optimal resource configurations by analyzing the relationship between historical task feature vectors and optimal resource configurations.

[0083] The core of the task-resource matching model lies in its data-driven training process, which differs from traditional rule-based configuration systems. This model is continuously trained and refined through large-scale historical execution data. The training dataset contains complete lifecycle records of thousands of historical computing tasks, each record including the task's feature vector, allocated resource configuration, actual execution time, resource utilization, energy efficiency, and other multi-dimensional indicators. The system automatically labels the optimal resource configuration, i.e., the configuration with the highest resource utilization under the premise of meeting performance requirements, as the target label for training.

[0084] The model training process adopts a multi-stage deep learning architecture, including feature extraction, pattern recognition, and decision generation. In the feature extraction stage, the system applies a convolutional neural network to process task code structure and data distribution features; in the pattern recognition stage, a recurrent neural network analyzes the resource usage timing patterns during task execution; in the decision generation stage, a reinforcement learning algorithm optimizes resource allocation strategies, balancing immediate performance and long-term system efficiency.

[0085] As the amount of accumulated training data grows, the task-resource matching model's understanding of resource requirement features for various tasks becomes increasingly accurate. For example, for deep learning training tasks, the model can predict GPU memory consumption curves based on information such as network architecture, batch size, and optimizer selection; for distributed data processing tasks, the model can accurately estimate network bandwidth requirements based on data sharding strategies and aggregation patterns.

[0086] In practical applications, as the number of tasks processed by the computing power scheduling platform grows, the prediction accuracy of the task-resource matching model shows a logarithmic growth trend, with accuracy greatly improving and stabilizing after a large amount of training. The intelligent matching model trained on massive data has completely changed the traditional computing resource allocation paradigm, evolving from static rules to dynamic learning, bringing significant technological breakthroughs and efficiency improvements to the computing power scheduling system.

[0087] There is a close input and inference relationship between the computing power demand feature vector and the task-resource matching model. The feature vector, as the core input of the model, is received and processed by the task-resource matching model, and finally outputs the optimal resource allocation scheme. This process essentially converts the abstract features of the task into specific resource allocation decisions. The feature vector provides a mathematical abstract representation of the task, while the task-resource matching model learns the complex mapping relationship from the feature space to the resource configuration space through a large number of historical samples, realizing the intelligent conversion from task description to resource allocation.

[0088] Taking deep learning training tasks as an example, the computing power scheduling platform will first analyze the submitted tasks, extract features such as network structure complexity, batch size, data set size, GPU affinity, and expected training duration, and form a feature vector input to the task-resource matching model. Based on the historical learning mode, the task-resource matching model may predict the optimal resource configuration of 4 high-performance GPUs, at least 32GB of video memory, 128GB of system memory, and high-speed NVMe storage. The computing power scheduling platform then allocates the corresponding computing nodes according to this decision and monitors the execution, feeding back the execution results for further optimization.

[0089] In some other embodiments, the computing power allocation unit 23 also has an adaptive resource elasticity scaling mechanism, which can dynamically adjust the allocated computing resources according to the actual resource usage during the execution of the computing task. When detecting changes in the resource demand of the computing task, the allocated computing resources can be increased or decreased in real time to ensure the maximum efficiency of resource usage while meeting the performance requirements of the task.

[0090] In some other embodiments of the present application, the computing power scheduling platform 20 also includes a task analysis unit configured to perform feature extraction and analysis on the computing tasks submitted by users, including but not limited to task type identification, resource consumption pattern prediction, execution duration estimation, etc. The task analysis unit uses a deep learning model to learn from the task code, data size, and historical execution records, providing decision-making basis for the computing power allocation unit 23.

[0091] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working processes of the devices and modules described above can refer to the corresponding processes in the foregoing method embodiments, which will not be described here. Those of ordinary skill in the art can realize that the units and algorithm steps of each example described in combination with the embodiments disclosed herein can be realized in electronic hardware, computer software or a combination of both. In order to clearly illustrate the interchangeability of hardware and software, the components and steps of each example have been described in the foregoing description in a general manner. Whether the functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. The skilled person can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of the present application.

[0092] The skilled person can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of the present application. The computer software described can be stored in a computer-readable storage medium, and the program can include the flow of the embodiments of the methods described above when executed. Among them, the storage medium can be a magnetic disk, an optical disk, a read-only memory or a random access memory, etc.

[0093] In summary, the computing power scheduling platform provided by the embodiments of the present application has the following characteristics:

[0094] 1) Intelligent resource matching: The computing power scheduling platform can accurately understand the resource demand characteristics of computing tasks through a data-driven task-resource matching model. This model is trained based on massive historical execution data and uses a multi-stage deep learning architecture, including feature extraction, pattern recognition and decision generation. Not only can it learn resource requirements from explicit features, but also can identify implicit computing patterns and predict resource consumption curves. As the number of tasks increases, the accuracy greatly improves and stabilizes after a large amount of training, realizing a technical breakthrough from static rules to dynamic learning.

[0095] 2) Computing power utilization rate improvement: Based on the intelligent scheduling model optimized by training, the computing power scheduling platform can deeply analyze the matching degree of the current load of the computing power server and the task characteristics, and realize optimal task allocation. The multi-objective optimization framework considers execution efficiency, resource utilization and energy consumption cost at the same time, so that the computing power utilization rate is improved from the highest 30% in the traditional way to more than 80%, with an improvement of more than 167%. Through the online learning mechanism, the predicted results are continuously compared with the actual execution data, and the resource allocation strategy is continuously optimized.

[0096] 3) Precise resource customization: Through feature vector extraction and model prediction, users do not need to manually specify detailed resource configurations. The computing power scheduling platform can automatically recommend the most suitable resource combination according to the task characteristics. Users can obtain resource allocation accurate to the level of a single computing card, realizing the transition from "whole machine rental" to "on-demand allocation". The model can intelligently recommend the most economical and efficient resource configuration according to the unique resource consumption patterns of different task types (such as deep learning training, inference, data analysis, etc.).

[0097] 4) User experience optimization: Based on the computing power scheduling platform system, users only need to submit computing tasks to obtain optimal resource configurations without professional knowledge. The computing power scheduling platform provides an AI development open source kit and an adaptive running environment. The system will analyze the task characteristics and automatically match the most suitable development environment. The transfer learning mechanism can identify new types of tasks and quickly adapt, ensuring that various user needs are met, saving environment configuration time by more than 1-2 days.

[0098] 5) Dynamic fine charging: The computing power scheduling platform's pay-per-use charging mode is based on real-time monitoring and prediction models of resource usage, which can accurately allocate resources when starting a computing task, avoiding cost waste caused by over-allocation. By continuously learning user task patterns, it can predict the resource consumption curve of computing tasks, enabling dynamic resource adjustment and fine charging based on the consumption curve, accurate to the minute level.

[0099] 6) Intelligent early warning and self-healing: The computing power scheduling platform integrates machine learning-based anomaly detection algorithms to analyze historical performance data and failure patterns and establish predictive maintenance models. The computing power scheduling platform implements intelligent monitoring of bare machines, computing power, services, and key processes around the clock, enabling early prediction of potential failures and early warning. Reinforcement learning algorithms optimize failure response strategies, automatically perform resource migration and recovery operations, and minimize the need for human intervention, significantly improving operational efficiency and simplifying and automating operations.

[0100] Unlike existing technologies, the computing power scheduling platform provided by the embodiments of the present application can provide intelligent resource allocation far superior to simple rule matching through data training model matching, achieving precise matching of resources and tasks. This data-driven decision-making approach can dynamically adapt to the resource demand characteristics of different types of tasks, maximize system resource utilization while ensuring task performance, greatly improve computing efficiency and reduce costs, and provide users with a better computing power service experience.

[0101] FIG. 5 shows a structural schematic diagram of a computing power rental system according to an embodiment of the present application. The specific implementation of the computing power rental system is not limited in the embodiments of the present application. It can be a hardware device for implementing the computing power scheduling platform or a physical device as shown in FIG. 1.

[0102] As shown in FIG. 5, the computing power renting system can include a processor, a communication interface, a memory, and a communication bus.

[0103] The processor, the communication interface, and the memory can communicate with each other through the communication bus. The communication interface is configured to communicate with network elements such as clients or other servers. The processor is configured to execute programs, and can execute one or more steps in the above method embodiments (for example, one or more steps in the above data processing procedure).

[0104] Specifically, the program can include program codes including computer operation instructions. In the embodiments of the present application, according to the type of hardware used, the processor can be a central processing unit (CPU), and can also be other general-purpose processors, digital signal processors (DSP), application specific integrated circuits (ASIC), field-programmable gate arrays (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc.

[0105] The memory is configured to store programs. The memory 1306 can include a high-speed RAM memory, and can also include a non-volatile memory such as at least one disk memory. The programs can be specifically configured to enable the processor 1302 to execute one or more steps in any of the above method embodiments.

[0106] Finally, it should be noted that: the above embodiments are only used to illustrate the technical solutions of the present application, and not to limit them; under the idea of the present application, the technical features in the above embodiments or different embodiments can also be combined, the steps can be implemented in any order, and there are many other changes of different aspects of the present application as described above. In order to be brief, they are not provided in details; although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that: it can still modify the technical solutions recorded in the foregoing embodiments, or make equivalent replacement for part of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the embodiments of the present application.

Claims

1. A computing power scheduling platform, characterized in that, Comprising: An order management unit configured to receive at least one computing task from a platform user; A storage space allocation unit configured to allocate a data storage space with a target capacity to the platform user; A computing power allocation unit configured to allocate a corresponding computing node to the computing task according to the computing power requirement of the computing task; A fee calculation unit configured to record the computing resource usage of the computing task and generate corresponding fee information; Wherein, the computing task is packaged in an independent container, and the container is started on the computing node allocated by the computing power allocation unit, so that the computing task is executed; The data information formed after the execution of the computing task is stored in the data storage space allocated by the storage space allocation unit. 2.The computing power scheduling platform of claim 1, wherein, Further comprising: A monitoring unit, the monitoring unit is configured to: During the execution of the computing task, feedback the state information of the computing node; the state information is selected from one or more of the following: CPU load; GPU load and physical memory usage ratio. 3.The computing power scheduling platform of claim 1, wherein, Further comprising: A platform interface presented on a user terminal; Wherein, the platform interface includes an access page for accessing the container. 4.The computing power scheduling platform of claim 3, wherein, The platform interface further includes a user login page for platform user registration and login. 5.The computing power scheduling platform of claim 1, wherein, The target capacity is determined by the platform user; the data storage space is mapped to the container.

6. The computing power scheduling platform of claim 5, wherein, The data storage space is provided with a plurality of storage space copies. 7.The computing power scheduling platform of claim 5, wherein, The data information formed after the execution of the computing task is stored through a persistent storage mechanism, so that the data information is independent of the life cycle of the container. 8.The computing power scheduling platform of claim 3, wherein, The computing power allocation unit is specifically configured to: Determine the computing power requirement of the computing task; Authorize the computing node meeting the computing power requirement to the corresponding platform user, and bind the authorized computing node with the platform user.

9. A hash power rental system, comprising: Comprising: Physical devices providing computing power resources and storage space; And The computing power scheduling platform of any one of claims 1-8; Wherein, the computing power scheduling platform abstracts the computing power resources and storage space of the physical devices through containers and virtualization technology, to form corresponding virtual resources. 10.The computing power leasing system of claim 9, wherein, The computing power scheduling platform forms virtual resources corresponding to the physical devices through Docker containers and KVM virtualization platforms.

Images