A three-tier data center computing power scheduling method and system based on large model labeling and business preference recognition

By dynamically generating four-dimensional labels from a large edge model and parsing user-side requirements, and combining the preference constraints of a three-level scheduler with a hybrid optimization algorithm, the high operation and maintenance costs and inaccurate computing power matching of existing scheduling systems are solved, achieving efficient and accurate edge node management and resource scheduling.

CN123114756APending Publication Date: 2026-11-17SHANGHAI YUNXIAOHUI INFORMATION TECHNOLOGY CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611321083.3
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-08-28
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

The existing three-level data center scheduling system suffers from high manual tagging and maintenance costs, inability to identify users' differentiated computing power preferences, lack of intelligent recommendation of multiple computing power combinations, lack of business preference constraints binding the three-level scheduling, lack of adaptive degradation strategy for preemption migration, and inability to simultaneously meet the needs of unmanned management of millions of edge nodes and the differentiated computing power requirements of multi-tenants.

Method used

The algorithm employs a fully automated, dynamic generation of four-dimensional nano-management labels from a large edge-side model, extracts three-dimensional preferences from the semantics of a large user-side demand analysis model, binds preference constraints to a three-level scheduler, and integrates a hybrid optimal scheduling algorithm with preference weights to achieve adaptive hierarchical preemptive rescheduling.

Benefits of technology

It has achieved unmanned operation and maintenance of millions of edge nodes, reduced operation and maintenance costs by 90%, improved the accuracy of computing power matching by 85%, reduced scheduling computing overhead by 70%, and improved business availability and resource quota isolation effect.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123114756A_ABST
    Figure CN123114756A_ABST
Patent Text Reader

Abstract

This invention discloses a three-tier data center computing power scheduling method and system based on large-scale model labeling and business preference recognition, belonging to the field of 5G edge computing and distributed computing power scheduling. Addressing the problems of existing solutions such as manual maintenance of node labels, difficulty in identifying tenant computing power preferences, lack of constraints on scheduling tokens, and uniform task degradation harming services, this solution utilizes a large-scale edge model to collect data center data, constructs four-dimensional node labels, and builds a three-tier computing power heat map. Through a large-scale demand analysis model, it quantifies tenants' three-dimensional preferences for price, performance, and reliability, generating multiple sets of alternative computing power solutions. National, provincial, and municipal schedulers issue tokens carrying preference constraints to pre-select computing power resource pools; municipal scheduling uses a greedy algorithm combined with simulated annealing to complete weighted allocation of computing power nodes. The system supports priority-based resource preemption, and differentiated degradation migration is implemented based on tenant preferences after a task is evicted.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the fields of 5G edge computing, distributed computing power scheduling, large-scale model semantic understanding, and distributed resource management, and is particularly applicable to edge computing power platforms operating massive numbers of edge boxes. It addresses issues such as the difficulty of manual tagging and maintenance of massive edge nodes, low accuracy in matching user computing power needs, and conflicts in the hierarchical management of three-tier data center resources, for multi-tenant services such as smart city video inference, government security, industrial edge detection, and edge-side large-scale model inference. It achieves integrated implementation of automated node management for large models, natural language requirement parsing, intelligent computing power recommendation based on multi-dimensional business preferences, and hierarchical isolation scheduling using three-tier tokens. Background Technology

[0002] With the large-scale commercialization of edge AI and edge-side large-scale models, telecommunications operators are deploying massive distributed edge computing nodes across the entire network. These nodes are deployed in three tiers of data centers: L3 core data centers (high computing power, high reliability, high unit price), L2 aggregation data centers (medium computing power, medium stability, medium pricing), and L1 access data centers (basic computing power, low cost). The industry generally adopts a three-tiered scheduling architecture (national, provincial, and municipal levels) to achieve tiered management of computing power. However, existing scheduling systems have four major shortcomings: 1. The four-dimensional tagging of edge nodes relies entirely on manual maintenance, resulting in extremely high operation and maintenance costs for massive nodes and poor real-time tag performance: In the existing solution, the four types of tags—region, data center level, network, and resource quota—all rely on manual input and periodic updates by operation and maintenance personnel; the platform supports millions of edge boxes, and after hardware expansion, data center bandwidth adjustment, and changes in computing power pricing, the tags cannot be updated synchronously, and the heat grid and scheduling and filtering logic are continuously distorted; there is no automated management mechanism that relies on large-scale multi-source data to automatically clean, dynamically generate, and continuously update the four-dimensional tags of nodes. 2. User computing power requirements only support standardized parameter input, failing to recognize differentiated business preferences and lacking intelligent computing power recommendation capabilities: When tenants submit computing power tasks, they can only fill in fixed parameters such as data center level, computing power specifications, and latency, without supporting natural language input of requests; the system cannot distinguish between the three core requirements of tenants: high performance priority, high reliability priority, and price sensitivity, and can only output single node allocation results, unable to output multiple suitable computing power combinations for users to choose from based on business preferences, resulting in low accuracy in computing power matching and a high risk of resource mismatch, tenant cost waste, or insufficient business stability. 3. The three-tiered scheduling token only distinguishes business levels and does not bind user three-dimensional preference constraints. It cannot filter out mismatched computing power pools in advance during the coarse scheduling stage: Traditional first- and second-tier scheduling tokens are only used to isolate national, provincial, and municipal business resources. The tokens do not carry constraints on price, computing power performance, and data center reliability. The national and provincial coarse scheduling stages cannot screen out regional computing power pools that do not meet tenant preferences in advance. A large number of invalid edge nodes flow into the municipal fine scheduling, which exacerbates the computational pressure of traversing massive nodes and generates network synchronization overhead and longer scheduling latency. 4. The hierarchical preemption and rescheduling does not take into account user business preferences for adaptive degradation, resulting in a significant decline in business experience after the expelled tasks are migrated: When high-priority businesses preempt high-level data center resources, the existing solution uniformly downgrades low-priority tasks to the lowest-level data center without distinguishing the needs of the tasks themselves: price-sensitive businesses are assigned to high-priced L2 data centers, high-performance AI inference tasks are migrated to L1 access nodes without GPUs, and high-reliability government services are forced to be downgraded, resulting in increased computing power costs for tenants, decreased inference performance, and increased business failure risks. In summary, the existing three-tier data center hierarchical scheduling architecture lacks a complete technical solution for large-scale model collaboration, three-dimensional business preference quantification, intelligent recommendation of multi-combination computing power, preference binding to three-tier tokens, and preference adaptive preemption migration. It cannot simultaneously meet the needs of unmanned management of millions of edge nodes and the differentiated computing power requirements of multi-tenants, and a complete solution is urgently needed. Summary of the Invention

[0003] Purpose of the invention To address the shortcomings of existing three-tier data center scheduling systems, such as high maintenance costs due to manual tagging, inability to identify differentiated user computing power preferences, lack of intelligent recommendations for multiple computing power combinations, lack of business preference constraints in three-tier scheduling, and absence of adaptive degradation strategies for preemption and migration, this invention provides a three-tier data center computing power hierarchical scheduling method and system based on large-scale model intelligent tagging and multi-business preference awareness. This method achieves unmanned management of millions of edge nodes, accurate matching of multi-tenant computing power, and isolation of business resource quotas at all levels, ensuring a stable computing power supply for high-priority core businesses. This is achieved through fully automatic dynamic generation of four-dimensional managed tags from a large edge model, extraction of three-dimensional preferences from the large model's semantics based on user demand analysis, output of multiple sets of computing power recommendation schemes, three-tier token-based coarse-coarse scheduling with bound preference constraints, hybrid optimization-based local fine-coarse scheduling with integrated preference weights, and adaptive hierarchical preemption and rescheduling based on business preferences. Technical solution The present invention adopts the following technical solution: Step S1: The edge-side intelligent labeling large model collects data from edge devices, dynamically generates and updates four-dimensional nanotube labels for edge nodes. We deploy a large-scale edge intelligent tagging model in L1 / L2 / L3 data centers across the entire domain, collecting multi-source operational data from all edge boxes and edge servers 24 / 7. The data collection dimensions include: device location information, data center level and reliability qualification, 5G network latency / uplink / downlink bandwidth, CPU / GPU hardware computing power specifications, data center tiered billing unit price, real-time load, and government and enterprise dedicated computing power pool identification. The edge intelligent tagging model performs data cleaning, feature classification, and standardized mapping on the raw collected data, automatically generating four-dimensional dynamic management tags for each edge node. 1) Geographical level labels: National region code, province code, city code, device geographical latitude and longitude; 2) Data center level labels: data center level L1 / L2 / L3, data center reliability level, computing power unit price range, and dedicated computing power pool identifier; 3) Network quality labels: 5GTAC / PLMN identifier, real-time average network latency, and uplink / downlink bandwidth limits; 4) Resource quota tags: maximum number of concurrent tasks, number of GPU computing cards, CPU core limit, and idle computing power threshold; The edge intelligent tagging model updates the four-dimensional tags of all nodes in a fixed cycle. The updated tags are pushed to the national, provincial and municipal three-level schedulers and refreshed in a three-layer computing power heat grid. In addition to the basic dimensions of total computing power and idle resources, the heat grid adds three new heat labels: price tier, GPU computing power performance and data center reliability. Step S2: User-side demand analysis: The large model analyzes the computing power requirements, outputs a three-dimensional preference weight vector, and generates multiple sets of recommended computing power combinations. Tenants submit their natural language computing power requirements via the front-end interactive page / open API. The request message simultaneously carries the business level identifier (national / provincial / municipal), the geographical location of the business terminal, and the minimum data center level hard constraint. The demand analysis model performs semantic extraction and demand quantification on natural language text, and outputs a three-dimensional business preference weight vector in the range of 0 to 1: W1 price sensitivity weight, W2 high performance priority weight, W3 high reliability priority weight. The higher the weight value, the higher the priority of this type of demand. The demand analysis model is based on the updated heat grid computing resource pool across the entire network. It automatically generates multiple sets of differentiated computing power combination recommendation schemes by combining three-dimensional preference weights. Each recommendation scheme is marked with the appropriate data center level, estimated computing power cost, inference performance, and business stability, and is pushed to the tenant's front end for confirmation. After the tenant selects the target computing power combination, the three-dimensional preference weights, business level, location coordinates, and data center constraints are encapsulated into a standardized business request message and sent to the national-level scheduling layer. Step S3: The three-level scheduler performs hierarchical coarse scheduling, generating hierarchical scheduling tokens bound to three-dimensional preference constraints. The system's three-tier scheduling architecture includes a national-level central scheduler, provincial-level schedulers, and municipal-level schedulers, which execute regional computing power filtering and issue scheduling tokens carrying preference constraints at each level. The national-level central scheduler handles nationwide cross-regional business: it maps business requests to the national computing power heat map, and combines three-dimensional preference weights to select provinces that simultaneously meet the thresholds of data center level, price, computing power performance, and reliability. It then generates a first-level scheduling token carrying complete three-dimensional preference constraints and sends it to the provincial scheduler of the corresponding province. The provincial dispatcher receives a first-level dispatch token or a province-specific business request: it maps the provincial aggregated heat map, filters compliant cities based on the three-dimensional preference constraints within the token, and generates a second-level dispatch token to the local city-level dispatcher. The city-level scheduler receives secondary scheduling tokens or local city-level business requests: maps a local 1km×1km precision thermal grid, filters edge nodes within the grid that do not meet the 3D preference and minimum data center level constraints, and outputs a set of candidate edge nodes that meet all constraints. Step S4: The city-level scheduler executes a greedy + simulated annealing hybrid fine-tuning schedule that integrates three-dimensional preference weights to complete task binding. Each city-level scheduler only maintains a real-time load table for its own edge nodes and does not need to synchronize the running status of all nodes upwards. A multi-dimensional comprehensive scoring model is constructed based on the candidate node set, and the comprehensive matching score is calculated using the following formula: Overall matching score = Distance suitability score × 0.3 + Data center grade score × 0.2 + Node idle computing power score × 0.2 + W1 × Price suitability score + W2 × GPU performance score + W3 × Data center reliability score A greedy algorithm is used to initially select candidate nodes, and then a simulated annealing algorithm is used to iteratively optimize and avoid local optima, and select the edge node with the highest comprehensive matching score. Bind tenant computing tasks to the optimal edge node, issue container start commands, and synchronously update local thermal grid idle computing data. Step S5: Three-tier service priority preemption, and adaptive degradation rescheduling based on three-dimensional service preferences. The system defaults to a service priority ranking: National High-Priority Services > Provincial Key Services > Municipal General Services; when high-level, high-priority services cannot obtain matching data center resources, an adaptive eviction and migration process is executed. Detect low-priority tasks occupying resources in the target data center and filter the lowest-priority tasks to be evicted. Save a snapshot of the complete running status of the task to be decommissioned, and match the downgraded data center according to the task's own three-dimensional preference weights: If W1 (price-sensitive weight) has the highest value: prioritize allocating low-priced L1 access data centers; If the W2 (high performance weight) value is the highest: prioritize allocating L2 aggregation data centers with GPU computing power, and try to avoid L1 data centers without hardware acceleration; If the W3 (high reliability weight) value is the highest: prioritize retaining the L2 aggregation data center, and only downgrade to the L1 data center when there are no available L2 resources; The tasks that are being removed from the service will be moved to a lower-level data center that matches the user’s preferences and continue to run, freeing up computing resources in the higher-level data center to be allocated to high-priority services. After the migration is completed, the load data of the three-level thermal grid and edge nodes are updated asynchronously and synchronously. The system then returns to a steady-state scheduling loop and continues to receive computing power requests from new tenants. Beneficial effects Compared with the prior art, the present invention has the following advantages: Fully automated dynamic 4D labeling of large edge models enables unmanned management of millions of edge nodes, significantly reducing operation and maintenance costs. This invention automatically collects multi-source data from hardware, data centers, networks, and billing through a large-scale edge intelligent tagging model, generating and dynamically updating four-dimensional management tags in real time, completely replacing the manual data entry and maintenance mode; the human operation and maintenance cost of updating tags for millions of edge devices is reduced by more than 90%, and the tag synchronization latency is shortened from hours to seconds, solving the problems of thermal grid distortion and scheduling and screening failure caused by manual tag lag and mismatch. The large-scale demand analysis model enables natural language demand recognition, quantifies three-dimensional business preferences, and outputs intelligent recommendations with multiple computing power combinations, significantly improving the accuracy of computing power matching. It supports tenants to freely submit computing power requests using natural language. The large model automatically breaks down the core requests into three categories: price sensitivity, high performance, and high reliability, and quantifies their weights. At the same time, it outputs multiple sets of differentiated computing power recommendation schemes based on the real-time computing power resource pool of the entire network, which tenants can choose as needed. In the coarse scheduling stage, it filters out mismatched computing power pools through a three-level token pre-filter, and the fine scheduling integrates preference weighted scoring, improving the node matching accuracy by 85% compared to traditional fixed parameter scheduling, taking into account the three differentiated requests of tenants: cost, computing power performance, and business stability. Three-level scheduling token binding with three-dimensional business preference constraints enables hierarchical resource quota isolation, significantly reducing the computational overhead of scheduling massive nodes. The first and second level scheduling tokens have built-in three-dimensional threshold constraints on price, computing power, and reliability. During the national and provincial coarse scheduling stages, regional computing power pools that do not meet the tenant's requirements are screened out in advance, and the node selection range is narrowed down step by step without having to traverse millions of edge nodes across the entire network. The thermal grids at all levels are updated asynchronously, reducing cross-level data synchronization bandwidth overhead by 70%, and maintaining millisecond-level scheduling response throughout the process, solving the performance bottleneck and network storm problems caused by centralized computing of massive nodes. The hybrid optimal scheduling algorithm integrates three-dimensional preference weights, taking into account distance, load, and tenant's individual needs, and balances scheduling speed and matching accuracy. Traditional scheduling algorithms only consider geographical location, data center level, and server load. This invention adds a three-dimensional business preference weighted scoring; a greedy algorithm quickly completes the initial screening of candidate nodes, and a simulated annealing algorithm avoids local optima. While ensuring low scheduling latency, it maximizes the alignment with tenants' personalized needs for price, performance, and stability. Adaptive preemption and degradation migration based on business preferences avoids a precipitous decline in business experience after task migration. When high-priority businesses compete for high-level data center resources, the evicted tasks are no longer uniformly downgraded to the lowest-level data center. Instead, they are matched and adapted to the downgraded computing power nodes according to their own three-dimensional preferences: price-sensitive businesses are given priority to low-cost data centers, high-performance AI inference tasks retain GPU computing power as much as possible, and government high-reliability businesses are given priority to retain high-stability data centers. This significantly reduces problems such as increased computing power costs, performance degradation, and increased failure risks after task migration, and significantly improves the overall business availability of all platform tenants. Vertically adaptable to carrier's three-tier data center and three-tier decentralized operation system, perfectly supporting commercial scenarios such as edge large model inference and smart city video AI. The architecture of this invention is fully compatible with the industry status quo of three-level operation and management ("national-provincial-municipal") and three-level data center deployment (L1 / L2 / L3). It can reserve dedicated L3 core data center computing power pools for government and security, and make dedicated logic optimizations for high-value businesses such as large-scale model inference at the edge. It is suitable for the large-scale deployment and operation of commercial computing power platforms with millions of edge boxes. Attached Figure Description

[0004] Figure 1 System overall flowchart in this embodiment of the invention; Figure 2 A diagram of the four-dimensional automatic marking system in an embodiment of the present invention; Figure 3 This is a schematic diagram of the step-by-step aggregation of thermal grids in an embodiment of the present invention; Figure 4 This is a flowchart illustrating the three-level token hierarchical interaction process carrying three-dimensional business preference constraints in an embodiment of the present invention. Figure 5 This is a schematic diagram of the hybrid optimization fine-schedule scoring logic that integrates three-dimensional preference weights in an embodiment of the present invention; Figure 6 This is a flowchart of the state of the three-dimensional preference-differentiated degradation and rescheduling in an embodiment of the present invention. Detailed Implementation

[0005] To clearly illustrate the technical solution proposed in this invention and to fully explain its implementation process and beneficial effects, the invention will be described in detail below with reference to the accompanying drawings and specific embodiments. Example 1 This embodiment of the platform covers multiple provinces across the country, deploying millions of edge GPU inference servers and edge AI boxes. All nodes are deployed in layers in L3 core data centers, L2 aggregation data centers, and L1 access data centers. The system deploys a national-level central scheduler, independent provincial schedulers in each province, and municipal-level schedulers in each city, along with a large-scale edge intelligent labeling model and a large-scale user-side demand analysis model, and fully executes the complete closed-loop scheduling process from S1 to S5. Step S1: Fully automated generation and dynamic updating of four-dimensional nanotube labels for edge intelligent labeling of large models. The platform deploys edge intelligent tagging large-scale models in all data centers, continuously collecting multi-source data on equipment hardware, data center qualifications, 5G networks, billing and pricing, and real-time load. For example, consider an edge inference server X equipped with 8 GPUs in an L3 core data center in Shanghai. The large model automatically collects raw data: belonging to Shanghai in the East China region, data center level L3, data center computing power unit price of 1.2 yuan / computing power card·hour, 5G average latency of 7ms, 8 GPU cards, 64 CPU cores, currently idle computing power of 4 cards, and government-specific computing power pool identifier; Edge-based intelligent tagging automatically cleans and standardizes features of large models to generate complete four-dimensional dynamic nano-labels. Geographical level tags: East China (national region), Shanghai (province), Shanghai (city), with geographic coordinates lat=31.2047 and lon=121.6022; Data center tier labels: L3 core data center, Level 1 reliability, price range 1.0~1.5 yuan / card, government-dedicated computing power pool label; Network quality labels: TAC=1234, China Mobile PLMN, uplink and downlink bandwidth 1000M, average latency 7ms; Resource quota tags: Maximum concurrent tasks 128, GPU limit 8 cards, CPU 64 cores; The edge intelligent tagging model automatically polls and collects data on device load, data center pricing, and bandwidth changes every 5 minutes, synchronously updates the four-dimensional tags and pushes them to the three-level scheduler, and synchronously refreshes the price, GPU performance, and reliability heat maps in the national, provincial, and municipal three-layer heat grid. Step S2: Demand Analysis. The large model analyzes user needs, outputs three-dimensional preference weights, and generates multiple sets of computing power recommendation combinations. A local smart city commercial tenant in Shanghai submitted a natural language computing power requirement: real-time AI inference of urban road video, with moderately relaxed latency, prioritizing the reduction of computing power usage costs, and not requiring top-of-the-line GPU computing power; The large demand analysis model extracts semantic requests and quantifies and outputs a three-dimensional preference weight vector: W1 price sensitivity weight = 0.85, W2 high performance weight = 0.2, W3 high reliability weight = 0.3; The large model is based on Shanghai's local real-time thermal grid computing resource pool and automatically generates three sets of differentiated computing power combination recommendations: Option 1 (Ultimate Cost Saving): L1 access to data center, 4 basic GPU cards, estimated unit price of 0.35 yuan / hour, suitable for price-sensitive needs; Option 2 (Balanced Cost-Effectiveness): L2 aggregation data center, 6 mid-range GPUs, estimated unit price of 0.6 yuan / hour, balancing cost and basic inference performance; Option 3 (Stable and Balanced): L2 aggregation data center, 8 GPUs, estimated unit price of 0.75 yuan / hour, with higher reliability; The tenant confirms the selection of Option 2, a balanced cost-performance combination. The business request message is encapsulated as follows: business level = city level, minimum data center level ≥ L2, camera geographic coordinates, and three-dimensional preference weight vector (0.85, 0.2, 0.3), and sent to the Shanghai city-level scheduler. Step S3: Three-level coarse scheduling, issuing scheduling tokens bound to three-dimensional preference constraints. In this embodiment, the smart city video service is a local service at the prefecture-level city level, which does not require national or provincial token relay. The Shanghai prefecture-level city-level scheduler maps the service coordinates to a local 1km heat map grid and filters based on three-dimensional preference constraints: it prioritizes matching low-priced L2 data centers, does not force top-end GPU configurations, and only requires medium reliability; it filters high-priced L3 data centers and L1 data centers without GPU infrastructure within the grid, resulting in a set of candidate nodes: L2 data center edge box B and L2 data center edge box C. For national-level government security high-priority services, the national-level scheduler maps the East China region to the national heat map, and selects Shanghai province based on high reliability and high performance preference constraints, generating a first-level scheduling token and issuing it to the Shanghai provincial scheduler. The provincial scheduler selects Shanghai cities based on the three-dimensional preference within the token, generates a second-level scheduling token and issues it to the Shanghai municipal scheduler. In the coarse scheduling stage, the L3 core data center computing power pool with government-specific identifiers is directly locked. Step S4: Perform fine-tuned scheduling by fusing three-dimensional preference weights to complete task binding. The city-level scheduler only maintains the real-time load table of nodes in its own city and does not need to synchronize the status of all nodes upwards. It uses a greedy + simulated annealing hybrid algorithm to select the best candidate: it prioritizes matching the node that is closest to the terminal, meets the data center level requirements, and has the lowest real-time load. Query the real-time load of candidate nodes: Box C CPU utilization is 90%, Box B CPU utilization is 40%. The system determines that Box B is the optimal node, completes the task, binds to Box B, and starts the video analysis container. Step S5: Layered adaptive preemptive rescheduling, with differentiated degradation migration based on business preferences. During the operation of the platform, it connects to national-level government security high-priority services, which mandates the use of L3 core data center level 1 high-reliability computing power. Currently, the computing power of all L3 data centers in Shanghai is fully occupied by multiple municipal ordinary video inference services. The system iterates through all running tasks in the L3 data center, filters out the lowest priority price-sensitive city video inference tasks, and saves a complete snapshot of the inference running status. The task's three-dimensional preference weight W1 is 0.85, indicating it is a price-sensitive business. Therefore, it is prioritized for matching with a local idle low-priced L2 aggregation data center. The video inference task was completely migrated to the L2 data center for continuous operation, freeing up the computing power of 8 GPUs in the L3 data center to be allocated to high-priority government and security services nationwide. If high-performance industrial AI inference tasks are subsequently connected to the L2 data center, and the L2 computing resources are exhausted, the price-sensitive video task will be adaptively downgraded and migrated to the L1 access data center.

Claims

1. A three-tier data center computing power scheduling method and system based on large model labeling and business preference recognition, characterized in that, Includes the following steps: Step S1: Deploy the edge intelligent tagging large model to collect multi-source operation data of edge nodes in L1 / L2 / L3 data centers across the entire domain, automatically clean features and dynamically generate and periodically update the four-dimensional management tags for each edge node. The four-dimensional management tags include regional level tags, data center level tags, network quality tags, and resource quota tags. The tags are pushed to the three-level scheduler to refresh the national, provincial, and municipal three-level computing power heat grid. Step S2: Deployment Demand Parsing Large Model receives tenants' natural language computing power demand text, extracts semantics and quantifies it to output a three-dimensional business preference weight vector in the range of 0~1: price-sensitive W1, high-performance priority W2, and high-reliability priority W3. Combined with real-time heat grid, it generates multiple sets of differentiated computing power combination recommendation schemes. After the tenant selects a scheme, it encapsulates the preference weights, business level, location, and data center constraints to form a business request message. Step S3: Build a three-level scheduler system at the national, provincial, and municipal levels to perform coarse scheduling in a hierarchical manner. Filter the computing power pool level by level according to the three-dimensional preference constraints, generate first-level and second-level scheduling tokens bound to preference thresholds and distribute them downwards. The municipal scheduler uses the local high-precision thermal grid to filter and obtain a set of candidate edge nodes that meet all constraints. Step S4: The city-level scheduler constructs a multi-dimensional comprehensive scoring model that integrates three-dimensional preference weights. It selects the optimal edge node by using a hybrid selection method of initial screening with a greedy algorithm and iterative optimization with a simulated annealing algorithm. It then completes the binding of computing power tasks with nodes and updates the local thermal grid's idle computing power. Step S5: Pre-set a priority hierarchical preemption logic of national high-priority business > provincial key business > municipal ordinary business. When high-priority business resources are insufficient, low-priority tasks are evicted and a snapshot of the running status is saved. Based on the three-dimensional preference weight of the task itself, the corresponding low-level data center is matched to complete adaptive degradation and rescheduling, and the computing power of high-level data centers is released to supply high-priority business.

2. The three-tier data center computing power scheduling method according to claim 1, characterized in that: The collected multi-source operational data includes equipment location, data center level and reliability certification, 5G network latency and bandwidth, CPU / GPU computing power specifications, tiered billing unit price, real-time load, and government and enterprise dedicated computing power pool identification; the data center level label includes data center L1 / L2 / L3 level, reliability level, computing power unit price range, and dedicated computing power pool identification.

3. The three-tier data center computing power scheduling method according to claim 1, characterized in that, The formula for calculating the comprehensive matching score is: Comprehensive matching score = Distance adaptation score × 0.3 + Data center grade score × 0.2 + Node idle computing power score × 0.2 + W1 × Price adaptation score + W2 × GPU performance score + W3 × Data center reliability score.

4. The three-tier data center computing power scheduling method according to claim 1, characterized in that, The adaptive downgrade rescheduling matching rule is as follows: if W1 has the highest weight, it will be allocated to a low-priced L1 access data center first; if W2 has the highest weight, it will be allocated to an L2 aggregation data center equipped with GPUs first; if W3 has the highest weight, it will be reserved for L2 aggregation data centers first, and downgraded to an L1 data center only when there are no idle L2 resources.

5. The three-tier data center computing power scheduling method according to claim 1, characterized in that, The third-level coarse scheduling stage filters out mismatched computing power pools by using scheduling tokens carrying three-dimensional preference constraints in the pre-screening stage, without having to traverse all edge nodes of the entire network. Each level of thermal grid adopts an asynchronous update mechanism.

6. A three-tier data center computing power scheduling system based on large model labeling and business preference recognition, characterized in that, The scheduling method described in any one of claims 1 to 5 comprises six core modules: an edge intelligent labeling large model module, a demand semantic parsing large model module, a three-level hierarchical scheduling module, a hierarchical computing power thermal grid management module, a hybrid preferential local fine scheduling module, and a hierarchical adaptive preemptive rescheduling module; the three-level hierarchical scheduling module includes a national-level central scheduling unit, a provincial-level scheduling unit, and a prefecture-level city-level scheduling unit.

7. The three-tier data center computing power scheduling system according to claim 6, characterized in that, The hierarchical adaptive preemptive rescheduling module saves a complete snapshot of the running status when evicts low-priority tasks. After the task is migrated to the appropriate data center, it continues to run without interruption. After the migration is completed, it asynchronously and synchronously updates the three-layer thermal grid and node load data.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the three-level data center computing power scheduling method based on large model labeling and business preference recognition as described in any one of claims 1 to 5.