Distributed computing power grid-connected platform and self-adaptive load balancing method and device thereof

Abstract

The invention relates to the field of distributed computing, and particularly provides a distributed computing power grid-connected platform and a self-adaptive load balancing method thereof, and the method comprises the steps that a physical resource layer is a platform hardware foundation and carries out physical computing, storage and network resource use; the resource abstraction layer abstracts and standardizes the hardware of the physical resource layer, shields the bottom layer hardware difference, and converts the physical resources into scheduling and management logic resources; the core scheduling layer carries out intelligent scheduling and dynamic optimization of computing power resources based on the logic resources provided by the resource abstraction layer in combination with a self-adaptive load balancing mechanism; the application interface layer provides standardized and diversified access modes; the application scene layer is an actual application direction of platform computing power resources, and the platform performs scene application. Compared with the prior art, the application access threshold can be reduced, multi-field scene landing is supported, and cross-industry computing power collaboration and innovation are promoted.

Description

Technical Field

[0001] This invention relates to the field of distributed computing, specifically providing a distributed computing power grid platform and its adaptive load balancing method and apparatus. Background Technology

[0002] With the deepening of the digital economy, computing power has become a core productive force. However, traditional computing power supply models are gradually failing to meet the demand. The demand for computing power in fields such as artificial intelligence (large model training and inference), scientific computing (climate simulation, gene sequencing), blockchain (consensus computing, data storage), and high-definition rendering (film and animation, metaverse scenes) is growing exponentially. For example, training a large model with hundreds of billions of parameters requires millions of GPU hours of computing power, which is difficult for a single institution or centralized data center to handle independently. Distributed computing power grid platforms can integrate dispersed computing resources (such as idle enterprise servers, personal GPUs, edge nodes, etc.) to form a "computing power pool," breaking through the upper limit of single-point computing power and meeting the needs of large-scale, high-concurrency computing power.

[0003] The computing power demand fluctuates significantly across different scenarios. For example, during peak promotional periods on e-commerce platforms, the phased computing tasks of scientific research projects, and the model iteration cycles of AI companies, computing power demand may surge or plummet in a short period. If centralized data centers are built based on peak demand, resources will be idle during off-peak hours; if they are built based on average demand, they will be unable to cope with peak pressure. Distributed grid-connected platforms can dynamically schedule the computing power of the entire network, expanding or shrinking as needed to achieve "elastic computing power supply."

[0004] Currently, computing resources suffer from severe "distribution imbalance" and "idle waste," urgently requiring optimized allocation through a grid-connected platform. Computing resources are inherently dispersed: enterprise servers have significant idle computing power during off-peak hours (e.g., nighttime), personal devices (PCs, GPU workstations) have low utilization rates, and the computing power of edge nodes (such as 5G base stations and smart terminals) is not fully activated. Statistics show that the total amount of idle computing power globally exceeds three times the total computing power of centralized data centers. If this "dormant computing power" cannot be effectively integrated, it will result in enormous resource waste. One of the core values ​​of a distributed computing power grid-connected platform is breaking down the physical barriers of computing resources, connecting fragmented resources through standardized interfaces to achieve cross-regional and cross-entity computing power sharing.

[0005] Computing power demanders (such as SMEs and research teams) may lack the funds to build dedicated computing power clusters, while computing power owners (such as large enterprises and data centers) cannot convert their idle resources into value. Distributed grid-connected platforms can act as "computing power intermediaries," matching supply and demand through market mechanisms or collaborative models: for example, research institutions can rent idle computing power from enterprises at low prices, and enterprises can gain revenue by sharing computing power, forming a "computing power sharing ecosystem" and improving the overall utilization rate of computing power resources across society.

[0006] Distributed computing power grid platforms, lacking effective load balancing mechanisms, will fall into the dilemma of "resource integration but inefficient operation." Adaptive load balancing mechanisms are the core solution to this problem. Nodes in distributed computing power grid platforms exhibit significant heterogeneity: different nodes have vastly different hardware configurations (CPU / GPU models, memory, bandwidth), network environments (latency, stability), and operating states (load rate, energy consumption, failure rate). Blindly allocating tasks (such as assigning high-load tasks to low-performance nodes or pushing tasks to a single node) will lead to a situation where "busy nodes are overworked, and idle nodes are starved": some nodes crash due to overload, while others remain idle, resulting in a significant drop in overall computing power utilization and even task failures (such as timeouts and data loss).

[0007] The load of a distributed system changes dynamically in real time. Nodes may suddenly become less efficient due to hardware failures, network fluctuations, or preemption of local tasks. The computational load of tasks may also fluctuate due to changes in input data (such as increased computation time for complex samples in AI inference). Adaptive load balancing mechanisms, by sensing node status (load rate, response time, energy consumption) in real time and dynamically adjusting task allocation (such as migrating tasks from overloaded nodes to idle nodes and prioritizing the allocation of sensitive tasks to low-latency nodes), can prevent the propagation of single points of failure, ensure stable system operation amidst dynamic changes, and maximize overall computing power utilization.

[0008] Therefore, how to address the contradiction between the explosive growth in computing power demand and the uneven distribution and inefficient use of computing resources, while simultaneously solving the stability, efficiency, and cost issues brought about by dynamic loads in distributed systems, is a matter that urgently needs to be addressed by those skilled in the art. Summary of the Invention

[0009] This invention addresses the shortcomings of the prior art by providing a highly practical distributed computing power grid platform and its adaptive load balancing method.

[0010] A further technical objective of this invention is to provide a reasonably designed, safe, and applicable distributed computing power grid platform and its adaptive load balancing device.

[0011] The technical solution adopted by this invention to solve its technical problem is:

[0012] A distributed computing power grid-connected platform and its adaptive load balancing method include:

[0013] The physical resource layer forms the platform's hardware foundation, enabling the use of physical computing, storage, and network resources.

[0014] The resource abstraction layer abstracts and standardizes the hardware of the physical resource layer, shielding the differences in the underlying hardware and transforming physical resources into logical resources for scheduling and management.

[0015] The core scheduling layer uses the logical resources provided by the resource abstraction layer and combines them with an adaptive load balancing mechanism to intelligently schedule and dynamically optimize computing resources.

[0016] The application interface layer provides standardized and diverse access methods;

[0017] The application scenario layer represents the actual application direction of the platform's computing resources, where the platform applies these scenarios.

[0018] Furthermore, the physical resource layer performs parameter calculation and storage during AI large model training, and deploys it at the network edge to process local real-time computing tasks.

[0019] Deploy smart terminal devices to collect data and execute edge computing tasks, and also participate in data collection and preliminary processing, and schedule tasks according to their characteristics;

[0020] Deploy network infrastructure to provide data transmission channels for data interaction between physical resources and between resources and upper-layer modules.

[0021] Furthermore, the resource abstraction layer virtualizes the physical server hardware resources into multiple independent virtual resources, which are allocated to different tasks as needed;

[0022] Manage applications and dependencies in containers to enable application deployment, scaling, and scheduling;

[0023] The overall computing power is divided into different "slices", each of which can correspond to a specific task or user, thus achieving fine-grained allocation of computing power.

[0024] Furthermore, the resource abstraction layer abstracts physical storage devices into logical storage pools, allocates storage resources on demand, shields the differences in underlying network hardware, and builds a unified logical network.

[0025] Furthermore, the core scheduling layer collects status data of various resources in the physical resource layer and resource abstraction layer in real time, parses the input task, identifies the task type and requirements, and decomposes the task into schedulable sub-tasks or determines the resource specifications required by the task.

[0026] By combining resource monitoring data with task requirements, dynamic decision-making on task allocation strategies is made to reasonably schedule tasks to different resource nodes to achieve load balancing. By analyzing historical load data and task trends, future resource load conditions are predicted, and resource allocation is adjusted in advance to cope with sudden high loads.

[0027] Monitor the running status of resource nodes, and automatically migrate tasks to healthy nodes when a node fails or its performance degrades, ensuring the continuous operation of tasks;

[0028] Set task priorities based on task importance and user level, and prioritize scheduling high-priority tasks.

[0029] When the task load is low, schedule low-power nodes or dynamically adjust node operating parameters to reduce overall energy consumption;

[0030] Blockchain technology is used to record computing power usage and resource status, ensuring that the data is tamper-proof, enabling trusted computing power measurement and transactions, and also serving as a trust mechanism for resource scheduling.

[0031] The system automatically adjusts resource size based on load changes, expanding when the load is high and shrinking when the load is low.

[0032] Furthermore, the application interface layer standardizes application programming interfaces (APIs) so that external applications can access the platform by calling the APIs to submit tasks and query resources.

[0033] Software development kits for multiple programming languages; user interface for submitting computational tasks, setting task parameters, and viewing task status and results;

[0034] Provides resource usage monitoring data interfaces and billing data interfaces.

[0035] Furthermore, the application scenario layer utilizes the platform's large-scale, distributed computing power to support the training tasks of large AI models, accelerate the model training process, and process training data and complex model parameter calculations through multi-node parallel computing and heterogeneous computing power collaboration.

[0036] It provides computing power for equipment networking, real-time monitoring, and process optimization in industrial production;

[0037] By using distributed computing power for parallel computing, the computation cycle is shortened, the scientific research process is accelerated, real-time computing power is provided for autonomous driving systems, and various applications in smart cities are supported.

[0038] A distributed computing power grid-connected platform and its adaptive load balancing device include: at least one memory and at least one processor;

[0039] The at least one memory is used to store a machine-readable program;

[0040] The at least one processor is used to call the machine-readable program to execute a distributed computing power grid platform and its adaptive load balancing method.

[0041] Compared with existing technologies, the distributed computing power grid-connected platform and its adaptive load balancing method and apparatus of the present invention have the following outstanding advantages:

[0042] (a) Improve computing power scheduling efficiency and system stability;

[0043] Significantly improved computing power utilization: Through virtualization, containerization, and computing power slicing technologies in the resource abstraction layer, hardware resources can be "allocated on demand and reused and shared".

[0044] Enhanced system resilience: The fault tolerance and migration module can monitor node status in real time. When hardware failures or network fluctuations cause nodes to "degrade," it automatically migrates tasks to healthy nodes, reducing the task failure rate. The dynamic scaling engine automatically expands / shrinks based on load changes to cope with sudden scenarios such as e-commerce promotions and peak scientific research tasks, improving system stability.

[0045] Heterogeneous computing power collaboration efficiency optimization: Supports flexible scheduling of heterogeneous computing units such as CPU, GPU, and FPGA, and accurately allocates resources for scenarios such as AI inference (adapted to GPU), complex logic operations (adapted to CPU), and low-latency tasks (adapted to edge nodes).

[0046] Energy efficiency and cost balance: The energy efficiency optimization module can schedule low-power nodes or adjust device operating parameters (such as frequency reduction) during low-load periods to reduce system energy consumption; at the same time, it reduces the occupation of ineffective resources through "on-demand scaling", thereby reducing hardware investment costs and operation and maintenance costs.

[0047] (ii) Activate the value of idle computing power and lower the threshold for computing power use;

[0048] Accurate matching of computing power supply and demand: The platform acts as an intermediary to connect computing power owners (enterprises, data centers) with demanders (SMEs, research teams). Enterprises can obtain additional income by sharing idle servers at night, and research teams can reduce the cost of renting idle computing power, thereby lowering the threshold for innovation.

[0049] Avoiding redundant construction and waste: The problem of "resource idleness" in centralized data centers built according to peak demand is solved. Enterprises do not need to expand hardware separately for short-term peak demand. They can dynamically call up the computing power of the entire network through the platform, reduce redundant investment in hardware, and alleviate the pain points of "high cost of computing power construction and long payback period".

[0050] Trusted measurement ensures fair transactions: The blockchain trusted measurement module records data such as computing power usage and resource status, ensuring that the data is tamper-proof, avoiding the problem of "computing power measurement fraud", protecting the interests of both supply and demand sides, and promoting the market-oriented development of computing power transactions.

[0051] (III) Support innovation in multiple fields and promote the upgrading of computing infrastructure;

[0052] Empowering the digital transformation of key industries: Providing large-scale distributed computing power for AI large-scale model training, supporting the rapid iteration of GPT-type models and industrial large-scale models; providing low-latency computing power for edge nodes for autonomous driving, ensuring the real-time nature of environmental perception and decision-making planning; and providing full-domain computing power scheduling for smart cities, improving the efficiency of traffic optimization and security monitoring.

[0053] Promote cross-domain collaborative innovation in computing power: Standardized API gateways and multi-language SDKs support rapid access in fields such as education (virtual simulation), finance (risk calculation), and healthcare (data modeling), allowing different industries to share computing resources and technical capabilities (such as calling on idle GPUs in the AI ​​field for emergency medical data processing).

[0054] Improve the computing infrastructure system: Promote the construction of a distributed computing network consisting of "centralized data centers + edge nodes + smart terminals" to make up for the shortcomings of traditional centralized computing power, such as "limited coverage and high latency", and provide "elastic, efficient and low-cost" computing infrastructure support for the development of the digital economy. Attached Figure Description

[0055] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0056] Figure 1 This is a schematic diagram of the architecture of a distributed computing power grid-connected platform and its adaptive load balancing method.

[0057] Figure 2 A flowchart illustrating a distributed computing power grid-connected platform and its adaptive load balancing method. Figure 1 ;

[0058] Figure 3 A flowchart illustrating a distributed computing power grid-connected platform and its adaptive load balancing method. Figure 2 . Detailed Implementation

[0059] To enable those skilled in the art to better understand the present invention, the present invention will be further described in detail below with reference to specific embodiments. Obviously, the described embodiments are merely some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0060] The following is a preferred embodiment:

[0061] like Figure 1-3 As shown in this embodiment, a distributed computing power network platform and its adaptive load balancing method include: a physical resource layer, a resource abstraction layer, a core scheduling layer (adaptive load balancing), an application interface layer, and an application scenario layer. Each layer collaborates with the others, from providing hardware resources to resource abstraction and intelligent scheduling, and then to application access and scenario implementation, to jointly build a distributed computing power network platform, achieve adaptive load balancing, efficiently utilize computing power resources, and support diverse business needs.

[0062] The physical resource layer forms the hardware foundation for the entire platform's operation, providing diverse and distributed physical computing, storage, and network resources. It encompasses various types and locations of physical devices, providing computing power support for the upper layers. Specific module functions include:

[0063] Data center clusters: large-scale centralized computing power carriers, equipped with high-performance servers, storage and network equipment, capable of undertaking large-scale, highly complex computing tasks, such as the calculation and storage of massive parameters during the training of large AI models.

[0064] Edge computing nodes: Deployed at the network edge (such as near factories, traffic intersections, etc.) to process local real-time computing tasks and reduce data transmission latency, such as real-time status monitoring and control of equipment in industrial intelligent manufacturing.

[0065] Intelligent terminal devices, such as intelligent sensors and small intelligent computing devices, can collect data and perform simple edge computing tasks. They serve as a supplement to computing power and also participate in data collection and preliminary processing. For example, environmental monitoring terminals in smart cities collect data and perform simple analysis.

[0066] Heterogeneous computing units: These include computing hardware with different architectures such as CPU, GPU, and FPGA. They can be flexibly scheduled according to task characteristics (e.g., GPU acceleration is suitable for AI inference, while CPU is suitable for complex logic operations) to improve the computing efficiency of specific tasks.

[0067] Network infrastructure: Provides data transmission channels to ensure data interaction between physical resources and between resources and upper-layer modules, including 5G, fiber optic and other networks, to ensure smooth data transmission after computing power is connected to the network.

[0068] The resource abstraction layer abstracts and standardizes the hardware of the physical resource layer, shielding it from underlying hardware differences and transforming physical resources into logical resources that can be flexibly scheduled and managed. This facilitates unified computing power orchestration and allocation at the upper layer. Specific module functions include:

[0069] Virtualization engine: Virtualizes the hardware resources of a physical server, such as CPU and memory, into multiple independent virtual resources, which can be allocated to different tasks as needed, improving the reuse rate of hardware resources. For example, a physical server can be virtualized into multiple virtual machines to run different applications.

[0070] Container orchestration: Managing applications and dependencies in units of containers (such as Docker) enables rapid deployment, scaling, and scheduling of applications. Kubernetes, for example, can automatically schedule container instances based on load to ensure stable application operation.

[0071] Computing power slice management: The overall computing power is divided into different "slices". Each slice can correspond to a specific task or user, realizing fine-grained allocation of computing power. For example, a dedicated computing power slice can be allocated to scientific computing tasks to ensure the exclusivity of computing resources.

[0072] Storage virtualization: Abstracts physical storage devices (hard drives, storage arrays, etc.) into logical storage pools, allocates storage resources on demand, and can also realize functions such as data redundancy and snapshots, ensuring data storage security and flexible access.

[0073] Network abstraction: It shields the differences in underlying network hardware, builds a unified logical network, supports resource communication in different network environments, and can also realize network isolation, bandwidth allocation, etc., to ensure data transmission security and efficiency.

[0074] The core scheduling layer, based on the logical resources provided by the resource abstraction layer and combined with an adaptive load balancing mechanism, enables intelligent scheduling and dynamic optimization of computing resources, ensuring efficient and stable system operation. Specific module functions include:

[0075] Resource monitoring module: Collects real-time status data of various resources in the physical resource layer and resource abstraction layer, such as CPU utilization, memory usage, and network bandwidth usage, to provide data support for scheduling decisions.

[0076] Task parsing engine: Parses the input task, identifies the task type (such as AI training, scientific computing) and requirements (computing power, storage, latency requirements, etc.), and breaks down the task into schedulable subtasks or determines the resource specifications required for the task.

[0077] Intelligent scheduling algorithm: Based on intelligent algorithms such as reinforcement learning and genetic algorithms, combined with resource monitoring data and task requirements, the algorithm dynamically decides on task allocation strategies and rationally schedules tasks to different resource nodes to achieve load balancing. For example, based on node load and computing power, AI inference tasks are allocated to idle GPU nodes.

[0078] Load forecasting module: By analyzing historical load data and task trends, it predicts future resource load and adjusts resource allocation in advance to cope with sudden high loads (such as predicting computing power demand before e-commerce promotions and expanding resources in advance).

[0079] Fault tolerance and migration: Monitor the running status of resource nodes. When a node fails or its performance degrades, automatically migrate tasks to healthy nodes to ensure continuous task operation and improve system availability. For example, if a server fails, tasks are automatically transferred to other servers.

[0080] Priority queue management: Set task priorities based on task importance, user level, etc., and prioritize scheduling high-priority tasks (such as emergency medical data processing tasks over ordinary data statistics tasks) to ensure the operation of critical business.

[0081] Energy efficiency optimization module: Combines resource status and task requirements to optimize resource energy consumption, such as scheduling low-power nodes when the task load is low, or dynamically adjusting node operating parameters (such as frequency reduction) to reduce overall energy consumption.

[0082] Trusted blockchain metering: Utilizing blockchain technology to record data such as computing power usage and resource status, ensuring that the data is tamper-proof, enabling trusted computing power metering and transactions, and can also be used to establish a trust mechanism for resource scheduling, preventing data fraud from affecting scheduling decisions.

[0083] Dynamic scaling engine: Automatically adjusts resource scale according to load changes, expanding when the load is high (adding compute nodes, storage capacity, etc.) and shrinking when the load is low, optimizing resource costs, just like cloud servers automatically increasing or decreasing instances based on business volume.

[0084] The application interface layer provides standardized and diverse access methods, facilitating users' use of the platform's computing resources and enabling the platform to manage and monitor applications. Specific module functions include:

[0085] Standardized API Gateway: Provides standardized application programming interfaces (APIs), allowing external applications to access the platform by calling the APIs to perform operations such as task submission and resource query, shielding the platform's internal complexity and simplifying the application access process.

[0086] Multi-language SDK: Provides software development kits (SDKs) for multiple programming languages ​​(such as Python and Java), making it easy for developers with different technology stacks to quickly integrate platform functions into their own applications and lowering the development threshold.

[0087] Task submission platform: The user interface allows users to submit computing tasks, set task parameters (such as computing power requirements and runtime), and view task status and results. It is the entry point for users to access the platform's computing power.

[0088] Monitoring and Billing Interfaces: Provides resource usage monitoring data interfaces (for users or management systems to obtain resource usage information), and billing data interfaces (calculates fees based on resource usage, duration, etc., for operational billing).

[0089] The application scenario layer represents the actual application direction of the platform's computing power resources. The platform demonstrates its value by supporting these application scenarios. Different scenarios have different computing power requirements and characteristics. The specific role of each scenario is as follows:

[0090] AI Large Model Training: Utilize the platform's large-scale, distributed computing power to support the training tasks of large AI models (such as GPT-like models), accelerate the model training process, and process massive training data and complex model parameter calculations through multi-node parallel computing and heterogeneous computing power collaboration.

[0091] Industrial intelligent manufacturing: It provides computing power support for equipment networking, real-time monitoring, and process optimization in industrial production. For example, it uses edge computing nodes to realize real-time status monitoring and fault early warning of production equipment, and uses platform computing power to simulate and optimize production processes to improve production efficiency and quality.

[0092] Scientific computing: It assists in complex computational tasks in scientific research, such as astronomical simulations and molecular dynamics simulations. Through distributed computing power and parallel computing, it shortens the computation cycle and accelerates the scientific research process. For example, it uses the platform's computing power to perform molecular structure simulation calculations in new drug development.

[0093] Autonomous driving: Provides real-time computing power support for autonomous driving systems, processes massive amounts of data collected by vehicle sensors (cameras, radar, etc.), performs calculations such as environmental perception, decision-making and planning, and edge computing nodes can reduce data transmission latency, ensuring the real-time performance and safety of autonomous driving.

[0094] Smart Cities: Supporting various applications in smart cities, such as intelligent transportation (traffic flow analysis, signal optimization), intelligent security (video surveillance, behavior recognition), and smart energy (power grid monitoring, optimized scheduling), enabling real-time processing and intelligent decision-making of urban data through distributed computing power.

[0095] Other scenarios: Covering various computing power needs beyond those mentioned above, such as virtual simulation teaching in the education field and risk calculation in the financial field, the platform can flexibly adapt to different scenario needs and provide computing power services.

[0096] Based on the above method, a distributed computing power grid-connected platform and its adaptive load balancing device in this embodiment include: at least one memory and at least one processor;

[0097] The at least one memory is used to store a machine-readable program;

[0098] The at least one processor is used to call the machine-readable program to execute a distributed computing power grid platform and its adaptive load balancing method.

[0099] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A distributed computing power grid-connected platform and an adaptive load balancing method thereof, characterized in that, Comprise: The physical resource layer is the hardware foundation of the platform, which performs physical computing, storage and network resource usage; The resource abstraction layer abstracts and standardizes the hardware of the physical resource layer, shielding the differences between the underlying hardware, and converting physical resources into logical resources for scheduling and management; The core scheduling layer provides intelligent scheduling and dynamic optimization of computing resources based on the logical resources provided by the resource abstraction layer and combined with adaptive load balancing mechanisms; The application interface layer provides standardized and diversified access methods; The application scenario layer is the actual application direction of the platform computing resources, and the platform applies to the scene. 2.The distributed computing power grid-connected platform and the adaptive load balancing method thereof according to claim 1, wherein, The physical resource layer performs parameter calculation and storage during AI large model training, and is deployed at the network edge to handle local real-time computing tasks; Deploy intelligent terminal devices to collect data and perform edge computing tasks, and participate in data collection and preliminary processing according to task characteristics; Deploy network infrastructure to provide data transmission channels for data interaction between physical resources and between resources and upper modules. 3.The distributed computing power grid-connected platform and the adaptive load balancing method thereof according to claim 2, wherein, The resource abstraction layer virtualizes physical server hardware resources into multiple independent virtual resources and allocates them to different tasks as needed; Manage applications and dependencies in units of containers to achieve application deployment, expansion and scheduling; Cut the overall computing power into different "slices", each of which can correspond to a specific task or user to achieve fine-grained computing power allocation.

4. The distributed computing power grid-connected platform and its adaptive load balancing method according to claim 3, characterized in that, The resource abstraction layer abstracts physical storage devices into logical storage pools, allocates storage resources as needed, shields differences between underlying network hardware, and constructs a unified logical network.

5. The distributed computing power grid-connected platform and its adaptive load balancing method according to claim 4, characterized in that, The core scheduling layer collects state data of various resources in the physical resource layer and the resource abstraction layer in real time, analyzes input tasks, identifies task types and requirements, and decomposes tasks into schedulable subtasks or determines the required resource specifications of tasks; Combine resource monitoring data and task requirements to dynamically decide task allocation strategies, schedule tasks to different resource nodes reasonably to achieve load balancing, analyze historical load data and task trends, predict future resource load situations, adjust resource allocation in advance, and respond to sudden high loads; Monitor resource node operating status, automatically migrate tasks to healthy nodes when nodes fail or performance decreases, and ensure continuous task operation; Set task priorities based on task importance and user levels, and schedule high-priority tasks first; When task load is low, schedule low-power nodes or dynamically adjust node operating parameters to reduce overall energy consumption; Use blockchain technology to record computing power usage and resource status, ensure data cannot be tampered with, achieve trusted computing power measurement and transaction, and also be used for trust mechanism establishment for resource scheduling; Automatically adjust resource size according to load changes, expand when load is high, and shrink when load is low.

6. The distributed computing power grid-connected platform and its adaptive load balancing method according to claim 5, characterized in that, The application interface layer performs standardized application program interface, external applications access the platform by calling API, and implements task submission and resource query; Develop software development kits in multiple programming languages, user interaction interfaces, submit computing tasks, set task parameters, view task status and results; Provide resource usage monitoring data interfaces and billing data interfaces.

7. The distributed computing power grid-connected platform and its adaptive load balancing method according to claim 6, characterized in that, The application scenario layer utilizes the large-scale and distributed computing power of the platform to support the training task of the AI large model, accelerates the model training process, processes the training data and complex model parameter calculation through multi-node parallel computing and heterogeneous computing power collaboration; Provide computing power for the networking of equipment in industrial production, real-time monitoring and process optimization; Through distributed computing power parallel computing, shorten the calculation period, accelerate the scientific research process, provide real-time computing power for the automatic driving system, and support various applications in smart cities.

8. A distributed computing power grid-connected platform and an adaptive load balancing device thereof, characterized in that, Comprise: At least one memory and at least one processor; The at least one memory is used to store a machine readable program; The at least one processor is used to call the machine readable program and execute the method of any one of claims 1 to 7.

Images