Intelligent collaborative scheduling system and method for scene integrating general computing and intelligent computing

Abstract

The invention discloses an intelligent collaborative scheduling system and method for a general computing and intelligent computing fusion scene, and relates to the technical field of computing power resource management and scheduling. In order to solve the problem that isomerous computing power resource islands are difficult to collaborate, the system comprises a computing power access layer used for executing specified operation on isomerous computing power resources through a standardized access template, abstracting the isomerous computing power resources into a unified logic computing power unit and registering the unified logic computing power unit into a computing power pool; the resource management layer is used for continuously monitoring and collecting static attributes and dynamic states of computing power resources to form a global resource real-time view; the business processing layer is used for receiving the business submitted by the user, analyzing and identifying the business demand, converting the business demand into a demand vector, and generating a dynamic arrangement scheme based on the real-time view and the multi-target strategy library; and the collaborative scheduling execution layer is used for converting the dynamic arrangement scheme into an instruction adaptive to various APIs and completing computing power resource allocation and service starting. According to the invention, unified management and intelligent cooperative scheduling of computing power resources can be realized.

Description

Technical Field

[0001] This invention relates to the field of computing resource management and scheduling technology, specifically an intelligent collaborative scheduling system and method for scenarios integrating general computing and intelligent computing. Background Technology

[0002] With the advancement of the "one-network computing power" strategy and the explosive growth of AI large-scale model technology, the demand for computing power has experienced explosive growth. However, the current computing power infrastructure faces three major challenges:

[0003] 1. Resource Heterogeneity and Silos: Data centers commonly contain general-purpose computing power (CPUs) with various architectures such as x86 and ARM, as well as intelligent computing power with various architectures such as NVIDIA GPUs, Ascend NPUs, and Hygon DCUs. Due to differences in architecture, drivers, and closed-source protocols, these computing powers are difficult to manage in a unified manner, forming serious "resource silos" and resulting in low overall resource utilization.

[0004] 2. Diverse Business Scenarios: Different business applications such as AI training, AI inference, scientific computing, and real-time control have vastly different requirements for computing resources (such as computing precision, GPU memory size, and network bandwidth). Traditional scheduling systems lack deep perception and intelligent adaptation capabilities for business needs, and cannot achieve optimal matching between business requirements and underlying heterogeneous computing resources.

[0005] 3. Challenges to the Domestic Chip Ecosystem: Under global chip regulations, domestically produced chips (such as Ascend 910B) urgently need large-scale application. However, their software ecosystem and development habits differ from mainstream GPUs, resulting in high application migration and deployment costs, which greatly limits their utilization and promotion speed.

[0006] In existing technologies, traditional cluster scheduling systems are mainly designed for homogeneous resources. Their scheduling strategies are relatively static and cannot effectively perceive the characteristics of accelerator cards such as NPUs / DCUs and the dynamic needs of AI services. Furthermore, they lack the ability to deeply integrate and collaboratively orchestrate general computing and intelligent computing resources. Therefore, there is an urgent need for a technical solution that can break down the barriers of heterogeneous computing power, achieve intelligent collaborative scheduling, and improve the utilization rate of domestic computing power. Summary of the Invention

[0007] This invention addresses the problem of isolated heterogeneous computing resources being difficult to coordinate, and provides an intelligent collaborative scheduling system and method for scenarios integrating general computing and intelligent computing. It enables unified management, intelligent orchestration and collaborative scheduling of heterogeneous computing resources, significantly improving the utilization rate of domestically produced chips and overall resource efficiency.

[0008] Firstly, the present invention provides an intelligent collaborative scheduling system for scenarios integrating general-purpose computing and intelligent computing, and the technical solution adopted to solve the above-mentioned technical problems is as follows:

[0009] An intelligent collaborative scheduling system for scenarios integrating general-purpose computing and intelligent computing, comprising:

[0010] The computing power access layer is used to perform automated driver installation, environment configuration and health check operations on three types of heterogeneous physical computing power resources: general computing power, intelligent computing power and social idle computing power, through predefined standardized access templates. Then, it abstracts the physical computing power resources that meet the access standards into logical computing power units with unified descriptors, and finally registers these logical computing power units into the global integrated computing power pool.

[0011] The resource management layer is used to continuously monitor and collect the static attributes and dynamic status of all physical computing resources in the global integrated computing power pool, form a real-time global resource view, and display it visually.

[0012] The business processing layer is used to receive business requests submitted by users. It first parses and identifies the business requirements and transforms them into a multi-dimensional requirement vector. Then, based on the global real-time resource view provided by the resource management layer and the pre-set multi-objective strategy library, it generates a dynamic orchestration scheme that adapts to the business requirements and the current resource status through the strategy engine.

[0013] The collaborative scheduling execution layer is used to convert dynamic orchestration schemes into instructions that are compatible with the APIs of various underlying cluster management systems, thereby enabling precise allocation of computing resources and the initiation of services.

[0014] Optionally, the predefined standardized access template includes adaptation rules for three types of heterogeneous physical computing power resources: general computing power, intelligent computing power, and social idle computing power, in order to cover the access needs of different computing power types.

[0015] Optionally, predefined standardized access templates support "plug-and-play" dual-mode access, including a plug-in template access mode and a customized computing power access mode, wherein:

[0016] The plugin template access mode is adapted to scenarios such as data centers and edge nodes where computing resources are concentrated and hardware configurations are relatively standardized.

[0017] Customized computing power access modes are adapted to scenarios with diverse hardware models and configurations of idle computing power in society.

[0018] Optionally, the business processing layer involved specifically includes:

[0019] The business parsing module is used to receive business submissions from users, parse them according to preset dimensions, automatically identify business requirements, including business scenario type, resource requirement specifications, software stack dependencies and priority strategies, and finally integrate the identification results into a structured requirement vector to provide data support for subsequent intelligent matching.

[0020] The dynamic orchestration module receives structured demand vectors and, based on the real-time global resource view provided by the resource management layer and the pre-set multi-objective strategy library, uses the strategy engine to perform multi-dimensional fitting and scoring of available computing resources through weighted evaluation or machine learning algorithms. Based on the optimal scoring result, a specific and feasible dynamic orchestration scheme is generated.

[0021] Further optional, the multi-objective strategy library involved includes efficiency-first, cost-first, domestic production-first, and resource affinity scheduling strategies.

[0022] Further optional, the dynamic orchestration scheme generated by the strategy engine specifies the target computing power node, decomposes the complex business into multiple sub-tasks, and dynamically decouples them according to the requirements of the sub-tasks and schedules them to be executed on logical computing power units of different architectures, and finally couples the results to complete the overall business.

[0023] Optionally, the collaborative scheduling execution layer involved supports fine-grained scheduling at the single-card level and cross-domain collaborative scheduling across centers and edge nodes, ensuring that services are accurately deployed to the designated target computing power nodes of the dynamic orchestration scheme.

[0024] Alternatively, the intelligent collaborative scheduling system may also include:

[0025] The strategy continuous optimization layer is used to collect system runtime performance data and result data during and after business execution. The collected data is compared and analyzed with the dynamic orchestration scheme to evaluate the effectiveness of the current scheduling strategy. Based on the analysis results, machine learning algorithms are used to iteratively optimize and adjust the decision logic and parameters of the strategy engine. The optimization results are fed back to the multi-objective strategy library of the business processing layer to continuously improve the quality of subsequent scheduling decisions, thereby realizing the system's continuous self-learning and optimization.

[0026] Secondly, this invention provides an intelligent collaborative scheduling method for scenarios integrating general-purpose computing and intelligent computing. The technical solution adopted to solve the above-mentioned technical problems is as follows:

[0027] An intelligent collaborative scheduling method for scenarios integrating general-purpose computing and intelligent computing includes the following steps:

[0028] Step 1: Using a predefined standardized access template, first perform automated driver installation, environment configuration, and health check operations on the three types of heterogeneous physical computing power resources: general computing power, intelligent computing power, and social idle computing power. Then, abstract the physical computing power resources that meet the access standards into logical computing power units with unified descriptors. Finally, register these logical computing power units into the global integrated computing power pool.

[0029] Step 2: Continuously monitor and collect the static attributes and dynamic status of all physical computing resources in the global integrated computing power pool to form a real-time global resource view and display it visually;

[0030] Step 3: Receive the business submitted by the user, first parse and identify the business requirements and transform them into a multi-dimensional requirement vector, and then, based on the global resource real-time view and the pre-set multi-objective strategy library provided in Step 2, generate a dynamic orchestration scheme that adapts to the business requirements and the current resource status through the strategy engine.

[0031] Step 4: Convert the dynamic orchestration scheme into instructions that are compatible with the APIs of various underlying cluster management systems to complete the precise allocation of computing resources and the startup of services;

[0032] Step 5: During and after business execution, collect system runtime performance data and result data, compare and analyze the collected data with the dynamic orchestration scheme, and evaluate the effectiveness of the scheduling strategy. Based on the analysis results, use machine learning algorithms to iteratively optimize and adjust the decision logic of the strategy engine, and feed the optimization results back to the multi-objective strategy library to continuously improve the quality of subsequent scheduling decisions, thereby realizing the system's continuous self-learning and optimization.

[0033] The intelligent collaborative scheduling system and method of the present invention, which is designed for scenarios integrating general-purpose computing and intelligent computing, has the following advantages compared with the prior art:

[0034] This invention solves the problem of the difficulty in coordinating isolated heterogeneous computing resources, and realizes the unified management and intelligent collaborative scheduling of general computing power and intelligent computing power resources. It significantly improves the utilization rate of domestic chips and the overall resource efficiency, and has strong compatibility, high resource utilization, excellent scheduling efficiency and good scalability. Attached Figure Description

[0035] Appendix Figure 1 This is a system architecture diagram of Embodiment 1 of the present invention;

[0036] Appendix Figure 2 This is a flowchart of the method according to Embodiment 2 of the present invention. Detailed Implementation

[0037] To make the technical solution, the technical problem solved, and the technical effect of the present invention clearer, the technical solution of the present invention will be clearly and completely described below in conjunction with specific embodiments.

[0038] Example 1:

[0039] Combined with appendix Figure 1 This embodiment proposes an intelligent collaborative scheduling system for scenarios integrating general computing and intelligent computing, which includes: a computing power access layer, a resource management layer, a business processing layer, and a collaborative scheduling execution layer.

[0040] 1. The computing power access layer is used to perform automated driver installation, environment configuration and health check operations on three types of heterogeneous physical computing power resources: general computing power (CPU), intelligent computing power (GPU / NPU / domestic chips such as Ascend 910B, Hygon DCU) and social idle computing power through predefined standardized access templates. Then, the physical computing power resources that meet the access standards are abstracted into logical computing power units with unified descriptors. Finally, these logical computing power units are registered to the global integrated computing power pool.

[0041] Among them, the predefined standardized access template includes adaptation rules for three types of heterogeneous physical computing power resources: computing power (CPU), intelligent computing power (GPU / NPU / domestic chips such as Ascend 910B, Hygon DCU) and social idle computing power, so as to cover the access needs of different computing power types.

[0042] The predefined standardized access templates support "plug-and-play" dual-mode access, including plug-in template access mode and customized computing power access mode. The plug-in template access mode is suitable for scenarios with concentrated computing power resources and relatively standardized hardware configurations, such as data centers and edge nodes. The customized computing power access mode is suitable for scenarios with diverse hardware models and large differences in configuration of social idle computing power (such as personal devices and scattered computing power of small organizations), and can achieve compatible access through flexible parameter configuration.

[0043] 2. Resource Management Layer: This layer continuously monitors and collects the static attributes (architecture, memory, video memory) and dynamic status (real-time utilization, health status, network topology) of all physical computing resources in the global converged computing pool, forming a real-time global resource view and displaying it visually.

[0044] 3. Business processing layer: This layer receives business requests submitted by users, parses and identifies the business requirements and transforms them into multi-dimensional requirement vectors. Then, based on the global real-time resource view provided by the resource management layer and the pre-built multi-objective strategy library, it generates a dynamic orchestration scheme that adapts to the business requirements and the current resource status through the strategy engine.

[0045] In this embodiment, the business processing layer specifically includes:

[0046] The business parsing module receives business submissions from users, parses them according to preset dimensions, and automatically identifies business requirements, including business scenario type (such as AI training, AI inference, scientific computing), resource requirement specifications (number of CPU cores, number of accelerator cards, memory size), software stack dependencies (such as CUDA version, PyTorch framework), and priority strategies (such as cost priority, latency sensitivity). Finally, the identification results are fused and transformed into a structured requirement vector to provide data support for subsequent intelligent matching.

[0047] The dynamic orchestration module receives structured demand vectors and, based on the real-time global resource view provided by the resource management layer and the pre-set multi-objective strategy library, uses the strategy engine to perform multi-dimensional fitting and scoring of available computing resources through weighted evaluation or machine learning algorithms. Based on the optimal scoring result, a specific and feasible dynamic orchestration scheme is generated.

[0048] The multi-objective strategy library includes efficiency-first, cost-first, domestic production-first, and resource affinity scheduling strategies.

[0049] The dynamic orchestration scheme generated by the strategy engine specifies the target computing power nodes, decomposes complex business (such as AI training) into multiple sub-tasks, and dynamically decouples them according to the requirements and characteristics of the sub-tasks and schedules them to logical computing power units of different architectures for execution. Finally, the results are coupled to complete the overall business.

[0050] 4. The collaborative scheduling execution layer is used to convert dynamic orchestration schemes into instructions that are compatible with the APIs of various underlying cluster management systems (such as Kubernetes and Slurm), so as to complete the precise allocation of computing resources and the start-up of services.

[0051] In this embodiment, the collaborative scheduling execution layer supports fine-grained scheduling at the single-card level and cross-domain collaborative scheduling across centers and edge nodes, ensuring that services are accurately deployed to the designated target computing power nodes of the dynamic orchestration scheme.

[0052] Based on the above, the intelligent collaborative scheduling system of this embodiment also includes:

[0053] 5. The strategy continuous optimization layer is used to collect system runtime performance data (such as business execution time, computing unit utilization, and energy consumption) and result data during and after business execution. The collected data is compared and analyzed with the dynamic orchestration scheme to evaluate the effectiveness of the current scheduling strategy. Based on the analysis results, machine learning algorithms are used to iteratively optimize and adjust the decision logic of the strategy engine (i.e., resource matching rules, priority algorithms, etc. in the multi-objective strategy library), and the optimization results are fed back to the multi-objective strategy library of the business processing layer to continuously improve the quality of subsequent scheduling decisions, thereby realizing the system's continuous self-learning and optimization.

[0054] Example 2:

[0055] Combined with appendix Figure 2 This embodiment proposes an intelligent collaborative scheduling method for scenarios integrating general-purpose computing and intelligent computing, which includes the following steps:

[0056] Step 1: Using a predefined standardized access template, first perform automated driver installation, environment configuration, and health check operations on three types of heterogeneous physical computing power resources: general computing power (CPU), intelligent computing power (GPU / NPU / domestic chips such as Ascend 910B, Hygon DCU), and idle social computing power. Then, abstract the physical computing power resources that meet the access standards into logical computing power units with unified descriptors. Finally, register these logical computing power units into the global integrated computing power pool.

[0057] In this step, the predefined standardized access template includes adaptation rules for three types of heterogeneous physical computing power resources: computing power (CPU), intelligent computing power (GPU / NPU / domestic chips such as Ascend 910B and Hygon DCU), and idle computing power in society, in order to cover the access needs of different computing power types.

[0058] The predefined standardized access templates support "plug-and-play" dual-mode access, including plug-in template access mode and customized computing power access mode. The plug-in template access mode is suitable for scenarios with concentrated computing power resources and relatively standardized hardware configurations, such as data centers and edge nodes. The customized computing power access mode is suitable for scenarios with diverse hardware models and large differences in configuration of social idle computing power (such as personal devices and scattered computing power of small organizations), and can achieve compatible access through flexible parameter configuration.

[0059] Step 2: Continuously monitor and collect the static attributes (architecture, memory, video memory) and dynamic status (real-time utilization, health status, network topology) of all physical computing resources in the global integrated computing power pool to form a real-time global resource view and display it visually.

[0060] Step 3: Receive the business requirements submitted by the user, first parse and identify the business requirements and transform them into a multi-dimensional requirement vector, and then, based on the real-time global resource view and the pre-set multi-objective strategy library provided in Step 2, generate a dynamic orchestration scheme that adapts to the business requirements and the current resource status through the strategy engine.

[0061] This step includes the following detailed process:

[0062] Step 3.1: Receive the business submitted by the user, parse it according to the preset dimensions, and automatically identify the business requirements, including the business scenario type (such as AI training, AI inference, scientific computing), resource requirement specifications (number of CPU cores, number of accelerator cards, memory size), software stack dependencies (such as CUDA version, PyTorch framework), and priority strategy (such as cost priority, latency sensitivity). Finally, the identification results are integrated and transformed into a structured requirement vector to provide data support for subsequent intelligent matching.

[0063] Step 3.2: Receive the structured demand vector. Based on the real-time global resource view provided by the resource management layer and the pre-built multi-objective strategy library, the strategy engine uses weighted evaluation or machine learning algorithms to perform multi-dimensional fitting and scoring of available computing resources. Based on the optimal scoring result, a specific and feasible dynamic orchestration scheme is generated. This dynamic orchestration scheme specifies the target computing nodes, decomposes complex business processes (such as AI training) into multiple sub-tasks, and dynamically decouples them according to the requirements of each sub-task, scheduling them to logical computing units with different architectures for execution. Finally, the results are coupled to complete the overall business process.

[0064] The multi-objective strategy library includes efficiency-first, cost-first, domestic production-first, and resource affinity scheduling strategies.

[0065] Step 4: Convert the dynamic orchestration scheme into instructions that are compatible with the APIs of various underlying cluster management systems (such as Kubernetes and Slurm) to complete the precise allocation of computing resources and the startup of services.

[0066] This step supports fine-grained scheduling at the single-card level and cross-domain collaborative scheduling across centers and edge nodes, ensuring that services are accurately deployed to the designated target computing nodes of the dynamic orchestration scheme.

[0067] Step 5: During and after business execution, collect system runtime performance data (such as business execution time, computing unit utilization, and energy consumption) and result data. Compare and analyze the collected data with the dynamic orchestration scheme to evaluate the effectiveness of the current scheduling strategy. Based on the analysis results, use machine learning algorithms to iteratively optimize and adjust the decision logic of the strategy engine (i.e., resource matching rules, priority algorithms, etc. in the multi-objective strategy library), and feed the optimization results back to the multi-objective strategy library to continuously improve the quality of subsequent scheduling decisions, thereby achieving continuous self-learning and optimization of the system.

[0068] In summary, the intelligent collaborative scheduling system and method of the present invention, which is designed for the integration of general computing and intelligent computing, has the following advantages:

[0069] 1. It can seamlessly integrate dispersed and heterogeneous general-purpose computing power, intelligent computing power, and idle social computing power resources to form a unified global converged computing power pool, providing consistent computing power services to the upper layer. It also supports unified management of mainstream heterogeneous devices such as CPUs, GPUs, NPUs (Ascend), and DCUs (Hygon), with a coverage rate of 68%. At the same time, through standardized access templates, access efficiency is improved by 55%.

[0070] 2. Standardized access templates support "plug-and-play" dual-mode access, including plug-in template access mode and customized computing power access mode. The plug-in template access mode is suitable for scenarios with concentrated computing power resources and relatively standardized hardware configurations, such as data centers and edge nodes, while the customized computing power access mode is suitable for scenarios with diverse hardware models and large configuration differences among social idle computing power. This increases the average resource utilization rate of the entire computing power pool from the industry average of 30% to over 65%, the intelligent computing power resource utilization rate from below 40% to over 75%, and the utilization rate of social idle computing power by over 60%.

[0071] 3. The accuracy of matching business needs with computing resources has improved by over 40%, avoiding performance losses caused by resource mismatch. For example, in practical applications, the completion time of similar AI training tasks has been shortened by an average of 15%-25% due to the matching with more suitable computing resources; the latency of some computationally intensive tasks has been reduced by 30%.

[0072] 4. Based on business awareness and intelligent matching, the time for environment configuration and resource allocation for complex AI training tasks is reduced from 4-8 hours of traditional manual operation to 5-10 minutes, with an automation deployment rate of 99%. The instance start-up and shutdown time for AI inference applications is less than 60 seconds, greatly improving customers' R&D efficiency and business response speed.

[0073] The above specific examples illustrate the principles and implementation methods of the present invention in detail. These embodiments are merely for the purpose of helping to understand the core technical content of the present invention. Based on the above specific embodiments of the present invention, any improvements and modifications made to the present invention by those skilled in the art without departing from the principles of the present invention should fall within the patent protection scope of the present invention.

Claims

1. An intelligent collaborative scheduling system for scenarios integrating general-purpose computing and intelligent computing, characterized in that, It includes: The computing power access layer is used to perform automated driver installation, environment configuration and health check operations on three types of heterogeneous physical computing power resources: general computing power, intelligent computing power and social idle computing power, through predefined standardized access templates. Then, it abstracts the physical computing power resources that meet the access standards into logical computing power units with unified descriptors, and finally registers these logical computing power units into the global integrated computing power pool. The resource management layer is used to continuously monitor and collect the static attributes and dynamic status of all physical computing resources in the global integrated computing power pool, form a real-time global resource view, and display it visually. The business processing layer is used to receive business requests submitted by users. It first parses and identifies the business requirements and transforms them into a multi-dimensional requirement vector. Then, based on the global real-time resource view provided by the resource management layer and the pre-set multi-objective strategy library, it generates a dynamic orchestration scheme that adapts to the business requirements and the current resource status through the strategy engine. The collaborative scheduling execution layer is used to convert dynamic orchestration schemes into instructions that are compatible with the APIs of various underlying cluster management systems, thereby enabling precise allocation of computing resources and the initiation of services.

2. The intelligent collaborative scheduling system for the fusion of general-purpose computing and intelligent computing scenarios according to claim 1, characterized in that, The predefined standardized access template includes adaptation rules for three types of heterogeneous physical computing power resources: general computing power, intelligent computing power, and social idle computing power, in order to cover the access needs of different computing power types.

3. The intelligent collaborative scheduling system for the fusion of general-purpose computing and intelligent computing scenarios according to claim 2, characterized in that, The predefined standardized access templates support "plug-and-play" dual-mode access, including a plugin template access mode and a customized computing power access mode, wherein: The plugin template access mode is adapted to scenarios such as data centers and edge nodes where computing resources are concentrated and hardware configurations are relatively standardized. Customized computing power access modes are adapted to scenarios with diverse hardware models and configurations of idle computing power in society.

4. The intelligent collaborative scheduling system for the fusion of general-purpose computing and intelligent computing scenarios according to claim 1, characterized in that, The business processing layer specifically includes: The business parsing module is used to receive business submissions from users, parse them according to preset dimensions, automatically identify business requirements, including business scenario type, resource requirement specifications, software stack dependencies and priority strategies, and finally integrate the identification results into a structured requirement vector to provide data support for subsequent intelligent matching. The dynamic orchestration module receives structured demand vectors and, based on the real-time global resource view provided by the resource management layer and the pre-set multi-objective strategy library, uses the strategy engine to perform multi-dimensional fitting and scoring of available computing resources through weighted evaluation or machine learning algorithms. Based on the optimal scoring result, a specific and feasible dynamic orchestration scheme is generated.

5. The intelligent collaborative scheduling system for the fusion of general-purpose computing and intelligent computing scenarios according to claim 4, characterized in that, The multi-objective strategy library includes efficiency-first, cost-first, domestic production-first, and resource affinity scheduling strategies.

6. The intelligent collaborative scheduling system for the fusion of general-purpose computing and intelligent computing scenarios according to claim 4, characterized in that, The dynamic orchestration scheme generated by the strategy engine specifies the target computing power nodes, decomposes complex business into multiple sub-tasks, and dynamically decouples them according to the requirements and characteristics of the sub-tasks, scheduling them to logical computing power units of different architectures for execution, and finally coupling the results to complete the overall business.

7. The intelligent collaborative scheduling system for the fusion of general-purpose computing and intelligent computing scenarios according to claim 6, characterized in that, The collaborative scheduling execution layer supports fine-grained scheduling at the single-card level and cross-domain collaborative scheduling across centers and edge nodes, ensuring that services are accurately deployed to the designated target computing power nodes of the dynamic orchestration scheme.

8. The intelligent collaborative scheduling system for the fusion of general-purpose computing and intelligent computing scenarios according to claim 4, characterized in that, The system also includes: The strategy continuous optimization layer is used to collect system runtime performance data and result data during and after business execution. The collected data is compared and analyzed with the dynamic orchestration scheme to evaluate the effectiveness of the current scheduling strategy. Based on the analysis results, machine learning algorithms are used to iteratively optimize and adjust the decision logic and parameters of the strategy engine. The optimization results are fed back to the multi-objective strategy library of the business processing layer to continuously improve the quality of subsequent scheduling decisions, thereby realizing the system's continuous self-learning and optimization.

9. An intelligent collaborative scheduling method for scenarios integrating general-purpose computing and intelligent computing, characterized in that, Includes the following steps: Step 1: Using a predefined standardized access template, first perform automated driver installation, environment configuration, and health check operations on the three types of heterogeneous physical computing power resources: general computing power, intelligent computing power, and social idle computing power. Then, abstract the physical computing power resources that meet the access standards into logical computing power units with unified descriptors. Finally, register these logical computing power units into the global integrated computing power pool. Step 2: Continuously monitor and collect the static attributes and dynamic status of all physical computing resources in the global integrated computing power pool to form a real-time global resource view and display it visually; Step 3: Receive the business submitted by the user, first parse and identify the business requirements and transform them into a multi-dimensional requirement vector, and then, based on the global resource real-time view and the pre-set multi-objective strategy library provided in Step 2, generate a dynamic orchestration scheme that adapts to the business requirements and the current resource status through the strategy engine. Step 4: Convert the dynamic orchestration scheme into instructions that are compatible with the APIs of various underlying cluster management systems to complete the precise allocation of computing resources and the startup of services; Step 5: During and after business execution, collect system runtime performance data and result data, compare and analyze the collected data with the dynamic orchestration scheme, and evaluate the effectiveness of the scheduling strategy. Based on the analysis results, use machine learning algorithms to iteratively optimize and adjust the decision logic of the strategy engine, and feed the optimization results back to the multi-objective strategy library to continuously improve the quality of subsequent scheduling decisions, thereby realizing the system's continuous self-learning and optimization.

Images