A computing resource management system and method based on multi-platform integration

Abstract

This invention relates to the field of computing power management, specifically to a computing power resource management system and method based on multi-platform integration. The system includes a heterogeneous collaboration module, a computing power task module, a scheduling and processing module, a node selection module, and a quality sensing module. The heterogeneous collaboration module is used for computing power node adaptation and unifying computing power costs. The computing power task module is used for processing task requests. The scheduling and processing module is used for aggregating available resources and determining the task execution order. The node selection module is used for constructing data transmission paths. The quality sensing module is used for sensing resource status information and performing load balancing management. This invention can avoid overload and idleness of computing power equipment, improve the utilization rate of all computing power equipment, reduce operational complexity, enhance the flexibility of computing power networking, optimize computing power allocation performance, flexibly adapt to the task requirements of different users, improve the reliability and fault tolerance of computing power scheduling, and enhance the stability and reliability of computer systems.

Description

Technical Field

[0001] This invention relates to the field of computing power management, specifically to a computing power resource management system and method based on multi-platform integration. Background Technology

[0002] Computing resource management is a technology that monitors, allocates, schedules, and optimizes the computing power of computers. Its purpose is to distribute computing tasks to the most suitable computing resources, maximizing resource utilization and minimizing computational costs. To meet the real-time computing power needs of different users, a common approach is to sell network computing power to users, automatically settling the computing power cost after use, thus achieving intelligent scheduling of network computing power.

[0003] In the process of computer networking, computing devices are spatially dispersed, which can easily lead to underutilization of computing power. It is necessary to connect different types of devices through the network to form a unified virtual computing resource pool for collaborative work. However, for computer devices distributed in different regions, scheduling and allocating computing power is difficult, and cross-regional access to computing power incurs additional transmission costs, which can easily lead to data line overload.

[0004] Furthermore, due to the different data processing capabilities of the devices, the network computing resources provided by each device have different quality. Under the condition of a fixed scheduling cost, the mixing of data of different quality will increase the additional data processing overhead. It is difficult for general scheduling platforms to detect data quality, which greatly increases the computing power operation cost. Summary of the Invention

[0005] The purpose of this invention is to provide a computing resource management system and method based on multi-platform integration to solve the problems mentioned in the background art.

[0006] To address the aforementioned technical problems, the present invention provides the following technical solution: a computing resource management system based on multi-platform integration, comprising: a heterogeneous collaboration module, a computing task module, a scheduling and processing module, a node selection module, and a quality sensing module;

[0007] The heterogeneous collaboration module is used to adapt the cloud platform to computer devices in various locations through multi-level computing power resource nodes. By deploying lightweight adapters on each computing power node to connect to devices with different interfaces, it automatically discovers and registers newly connected heterogeneous computing power nodes, establishes a heterogeneous computing power collaboration platform, collects the computing power status of each node, displays the total network computing power, the number of physical hosts and the network contract status of each host in real time, dynamically calculates the available resources, and unifies the computing power of different architectures into computing power cost based on real-time resource utilization, operating costs, the number of queued tasks and network scheduling costs.

[0008] The computing power task module is used to process task requests when there is spare computing power resources available. It decomposes user tasks into sub-tasks with dependencies. The sub-tasks record resource requirements, data input volume and data output volume. It uses a spectral clustering algorithm to group computing power nodes, calculates the weighted cost of tasks and sends computing power bills.

[0009] The scheduling processing module is used to aggregate the available resources of all nodes in real time and centrally process all queued computing power call tasks within the same period. It uses each computing power node in the global resource view as a vertex and the computing power scheduling overhead as a connecting edge to model the computing power nodes into a directed acyclic graph structure. It uses a graph clustering algorithm to allocate batch subtasks to nodes and maps subtask clusters to different groups for execution, so as to minimize the total transmission overhead and computing power cost. After the user task is completed, the load balancing weight is dynamically adjusted with resource constraints, price costs and user historical contributions as constraints, and the task execution order is determined by topological sorting.

[0010] The node selection module is used to construct available transmission paths, connect computing power nodes to each other, mark the transmission speed, transmission conditions and maximum transmission volume of each path, connect the computing power devices running the subtasks to which the task belongs through the transmission paths to form the execution path of the task, take the longest execution path of each task as the critical path, plan the data transmission path of each task, minimize the completion time of service instances on the critical path, track the transmission status on each critical path, and perform stability tests on the path.

[0011] The quality sensing module is used to perceive network resource status information through INT metadata. When the INT metadata is transmitted in the network, each computing node writes real-time status information into the data packet. The real-time status information includes computation latency, jitter, path packet loss rate, and the computing resource status of the node. The computing quality of each node is determined by weighting according to the real-time status information. Through load balancing management, computing resources are scheduled among different nodes. Under the condition of a certain number of scheduled resources, the computing quality allocated to all users remains stable.

[0012] Furthermore, the heterogeneous collaboration module includes: a node adaptation unit, a computing power management unit, and a cost pricing unit;

[0013] The node adaptation unit is used to encapsulate cloud virtual machines, container clusters, edge servers and terminal devices located in different locations into computing power nodes;

[0014] The computing power management unit is used to network heterogeneous resources through an adapter, collect the resource status of each node in real time, and display the total computing power, number of hosts, network status and contract status in a visualization window.

[0015] The cost pricing unit is used to determine the computing power cost of each node in real time and to dynamically charge for the computing power output of the contracted equipment.

[0016] Furthermore, the computing power task module includes: a task processing unit and a collaborative allocation unit;

[0017] The task processing unit is used to receive task requests submitted by users and decompose the task into sub-tasks using a multi-granularity task decomposition method based on weighted graph clustering.

[0018] The collaborative allocation unit is used to perform cross-regional collaborative scheduling and matching based on constraints of resources, costs, and user contributions.

[0019] Furthermore, the scheduling processing module includes: a graph modeling unit and a resource constraint unit;

[0020] The graph modeling unit is used to model the decomposed subtasks as a directed acyclic graph, with computing power nodes as vertices and computing power scheduling overhead as connecting edges.

[0021] The resource constraint unit is used to allocate subtasks to computing power devices that can meet hardware requirements, while keeping the total cost of the tasks below the budget.

[0022] Furthermore, the node selection module includes: a node forwarding unit and a path testing unit;

[0023] The node forwarding unit is used to determine the task execution order using topology sorting, minimizing the critical path completion time;

[0024] The path testing unit is used to identify all subtasks located on the critical path and determine the computing nodes of each subtask and the data forwarding paths between nodes.

[0025] Furthermore, the quality sensing module includes: an INT sensing unit, a load balancing unit, and a quality management unit;

[0026] The INT sensing unit is used to generate INT data packets using in-band network telemetry technology. When the data packets are transmitted in the network, the network device writes real-time status information into the data packets.

[0027] The load balancing unit is used to provide a unified API gateway, track the execution status of each subtask, construct a network situation awareness map of the entire network, and perform load scheduling.

[0028] The quality management unit is used to sense the latency, jitter, and packet loss rate of each path, provide feedback on the computing power quality of each node, and adjust task allocation according to the quality.

[0029] A method for managing computing resources based on multi-platform integration includes the following steps:

[0030] Step S1. Adapt the cloud platform to computer equipment in various locations at multiple levels to establish a heterogeneous computing power collaboration platform. The platform collects the computing power status of nodes and dynamically calculates the available resources. It unifies the computing power of different architectures into computing power cost and dynamically bills the output computing power of contracted equipment.

[0031] Step S2. Process user task requests. Decompose the task into subtasks using a multi-granularity task decomposition algorithm based on weighted graph clustering. Use computing power nodes as vertices and computing power scheduling overhead as connecting edges to model the computing power network as a directed acyclic graph.

[0032] Step S3. Based on the directed acyclic graph, map each subtask cluster to different node groups for execution, minimizing the total transmission overhead and computing power cost. Using resource constraints, price costs, and user computing power contributions as constraints, dynamically adjust the load balancing weights and use topology sorting to determine the task execution order.

[0033] Step S4. Connect the computing power nodes to each other in the execution order, connect the computing power devices running the subtasks to which the task belongs through the transmission link to form the execution path of the task, plan the execution path of each task, and minimize the completion time of service instances on the critical path.

[0034] Step S5. Perceive network resource status information through INT metadata, determine the computing power quality of each node according to real-time status weighting, and schedule computing power resources among different nodes through load balancing management so that the variance of computing power quality allocated to all users is less than the threshold.

[0035] Furthermore, step S1 includes:

[0036] Step S11. Encapsulate cloud virtual machines, container clusters, edge servers and terminal devices located in different locations as computing power nodes. By deploying lightweight adapters on each computing power node to connect to devices with different interfaces, automatically discover and register newly connected heterogeneous computing power nodes. Network the heterogeneous resources through the adapters to establish a heterogeneous computing power collaboration platform.

[0037] Step S12. Collect the computing power status of each node, display the total network computing power, the number of physical hosts and the network contract status of each host in real time, dynamically count the available resources, and determine the real-time computing power cost of each node based on the real-time resource utilization, operating cost, number of queued tasks and network scheduling cost of the node.

[0038] Furthermore, step S2 includes:

[0039] Step S21. When there is spare computing power resources, receive the task application submitted by the user, decompose the task into sub-tasks using the multi-granularity task decomposition method based on weighted graph clustering, decompose the user task into sub-tasks with dependencies, and record the resource requirements, data input and data output in the sub-tasks. Use spectral clustering algorithm to group the computing power nodes.

[0040] Step S22. Aggregate the available resources of all nodes in real time, centrally process all queued computing power call tasks within the same period, and model the computing power nodes as a directed acyclic graph structure, taking each computing power node in the global resource view as a vertex and the computing power scheduling overhead as a connecting edge.

[0041] Furthermore, step S3 includes:

[0042] Step S31. Based on the directed acyclic graph, the subtasks are assigned to computing power device nodes that meet the hardware execution requirements through graph clustering algorithm. By adjusting the clustering parameters and clustering level, the size of the subtasks is controlled to match the computing power of the edge nodes of the subtasks, while ensuring that the total cost of task execution is lower than the budget.

[0043] Step S32. Based on the real-time collected node load information, which includes CPU, memory, network I / O and service queue length, the load balancing weight is dynamically adjusted through a reinforcement learning algorithm, and the Kahn algorithm is used for topology sorting to determine the task execution order.

[0044] Step S33. Perform node scheduling for new requests. When the overall load of some node regions is too high, the global coordination layer initiates cross-regional elastic scaling to migrate the task to the region with the lowest load.

[0045] Furthermore, step S4 includes:

[0046] Step S41. Construct available transmission paths, connect computing nodes to each other, and mark the transmission speed, transmission conditions and maximum transmission volume of each path. The longest execution path in each task is taken as the critical path.

[0047] Step S42. Identify all subtasks located on the critical path, and determine the computing nodes and data forwarding paths between nodes for each subtask, taking resource constraints, price costs and user computing power contributions as constraints. At the same time, track the transmission status on each critical path and perform stability tests on the paths.

[0048] Furthermore, step S5 includes:

[0049] Step S51. Generate INT data packets using in-band network telemetry technology. When the data packets are transmitted in the network, the network device writes real-time status information into the data packets. The real-time status information includes computation latency, jitter, path packet loss rate, and the computing resource status of the nodes.

[0050] Step S52. Sensing the latency, jitter, and packet loss rate of each path, and providing feedback on the computing power quality of each node. When a node is overloaded, the network is congested, or the computing power quality is below the threshold during operation, the load balancing mechanism is triggered to reschedule the task nodes. Through resource reservation and elastic quota mechanisms, the computing power quality allocated to all tasks is stabilized.

[0051] Compared with the prior art, the beneficial effects achieved by the present invention are:

[0052] 1. This invention can adapt cloud platforms and edge device platforms in various regions to multi-level computing resource nodes, establish a heterogeneous computing power collaboration platform for computing power scheduling and management and process user task requests. It can allocate tasks to the most suitable computing devices for execution according to the needs of task requests, avoid overload and idleness of computing power devices, improve the utilization rate of all computing power devices, reduce operation and maintenance complexity, enhance the flexibility of computing power networking, and optimize computing power allocation performance.

[0053] 2. This invention can decompose tasks and distribute them to different nodes for execution, minimizing transmission overhead and computing power costs. It models batch tasks as a directed acyclic graph structure and determines the forwarding path for each task through cross-regional resource collaborative scheduling and matching, achieving a low-latency and high-efficiency data processing process. It flexibly adapts to the task needs of different users and improves the reliability and fault tolerance of computing power scheduling.

[0054] 3. This invention can perceive the operating status information of computing resources through INT metadata. Under the condition of a certain scheduling resources, it can allocate computing power devices of different quality to keep the computing power quality allocated to all users stable, achieve computing power quality balance, avoid data processing disasters of single tasks, reduce data congestion of unit nodes, and improve the stability and reliability of computer systems. Attached Figure Description

[0055] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0056] Figure 1 This is a schematic diagram of the structure of a computing resource management system based on multi-platform integration according to the present invention;

[0057] Figure 2 This is a schematic diagram illustrating the steps of a computing resource management method based on multi-platform integration according to the present invention. Detailed Implementation

[0058] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only 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.

[0059] Please see Figure 1 The present invention provides a technical solution: a computing power resource management system based on multi-platform integration, comprising: a heterogeneous collaboration module, a computing power task module, a scheduling and processing module, a node selection module, and a quality sensing module;

[0060] The heterogeneous collaboration module is used to adapt the cloud platform to computer devices in various locations through multi-level computing power resource nodes. By deploying lightweight adapters on each computing power node to connect to devices with different interfaces, it automatically discovers and registers newly connected heterogeneous computing power nodes, establishes a heterogeneous computing power collaboration platform, collects the computing power status of each node, displays the total network computing power, the number of physical hosts and the network contract status of each host in real time, dynamically calculates the available resources, and unifies the computing power of different architectures into computing power cost based on real-time resource utilization, operating costs, the number of queued tasks and network scheduling costs.

[0061] The heterogeneous collaboration module includes: a node adaptation unit, a computing power management unit, and a cost pricing unit;

[0062] The node adaptation unit is used to encapsulate cloud virtual machines, container clusters, edge servers and terminal devices located in different locations into computing power nodes;

[0063] The computing power management unit is used to network heterogeneous resources through an adapter, collect the resource status of each node in real time, and display the total computing power, number of hosts, network status and contract status in a visualization window.

[0064] The cost pricing unit is used to determine the computing power cost of each node in real time and to dynamically charge for the computing power output of the contracted equipment.

[0065] The computing power task module is used to process task requests when there is spare computing power resources available. It decomposes user tasks into sub-tasks with dependencies. The sub-tasks record resource requirements, data input volume and data output volume. It uses a spectral clustering algorithm to group computing power nodes, calculates the weighted cost of tasks and sends computing power bills.

[0066] The computing power task module includes: a task processing unit and a collaborative allocation unit;

[0067] The task processing unit is used to receive task requests submitted by users and decompose the task into sub-tasks using a multi-granularity task decomposition method based on weighted graph clustering.

[0068] The collaborative allocation unit is used to perform cross-regional collaborative scheduling and matching based on constraints of resources, costs, and user contributions.

[0069] The scheduling processing module is used to aggregate the available resources of all nodes in real time and centrally process all queued computing power call tasks within the same period. It uses each computing power node in the global resource view as a vertex and the computing power scheduling overhead as a connecting edge to model the computing power nodes into a directed acyclic graph structure. It uses a graph clustering algorithm to allocate batch subtasks to nodes and maps subtask clusters to different groups for execution, so as to minimize the total transmission overhead and computing power cost. After the user task is completed, the load balancing weight is dynamically adjusted with resource constraints, price costs and user historical contributions as constraints, and the task execution order is determined by topological sorting.

[0070] The scheduling processing module includes: a graph modeling unit and a resource constraint unit;

[0071] The graph modeling unit is used to model the decomposed subtasks as a directed acyclic graph, with computing power nodes as vertices and computing power scheduling overhead as connecting edges.

[0072] The resource constraint unit is used to allocate subtasks to computing power devices that can meet hardware requirements, while keeping the total cost of the tasks below the budget.

[0073] The node selection module is used to construct available transmission paths, connect computing power nodes to each other, mark the transmission speed, transmission conditions and maximum transmission volume of each path, connect the computing power devices running the subtasks to which the task belongs through the transmission paths to form the execution path of the task, take the longest execution path of each task as the critical path, plan the data transmission path of each task, minimize the completion time of service instances on the critical path, track the transmission status on each critical path, and perform stability tests on the path.

[0074] The node selection module includes: a node forwarding unit and a path testing unit;

[0075] The node forwarding unit is used to determine the task execution order using topology sorting, minimizing the critical path completion time;

[0076] The path testing unit is used to identify all subtasks located on the critical path and determine the computing nodes of each subtask and the data forwarding paths between nodes.

[0077] The quality sensing module is used to perceive network resource status information through INT metadata. When the INT metadata is transmitted in the network, each computing node writes real-time status information into the data packet. The real-time status information includes computation latency, jitter, path packet loss rate, and the computing resource status of the node. The computing quality of each node is determined by weighting according to the real-time status information. Through load balancing management, computing resources are scheduled among different nodes. Under the condition of a certain number of scheduled resources, the computing quality allocated to all users remains stable.

[0078] The quality sensing module includes: an INT sensing unit, a load balancing unit, and a quality management unit;

[0079] The INT sensing unit is used to generate INT data packets using in-band network telemetry technology. When the data packets are transmitted in the network, the network device writes real-time status information into the data packets.

[0080] The load balancing unit is used to provide a unified API gateway, track the execution status of each subtask, construct a network situation awareness map of the entire network, and perform load scheduling.

[0081] The quality management unit is used to sense the latency, jitter, and packet loss rate of each path, provide feedback on the computing power quality of each node, and adjust task allocation according to the quality.

[0082] like Figure 2 As shown, a computing resource management method based on multi-platform integration includes the following steps:

[0083] Step S1. Adapt the cloud platform to computer equipment in various locations at multiple levels to establish a heterogeneous computing power collaboration platform. The platform collects the computing power status of nodes and dynamically calculates the available resources. It unifies the computing power of different architectures into computing power cost and dynamically bills the output computing power of contracted equipment.

[0084] Step S1 includes:

[0085] Step S11. Encapsulate cloud virtual machines, container clusters, edge servers and terminal devices located in different locations as computing power nodes. By deploying lightweight adapters on each computing power node to connect to devices with different interfaces, automatically discover and register newly connected heterogeneous computing power nodes. Network the heterogeneous resources through the adapters to establish a heterogeneous computing power collaboration platform.

[0086] Step S12. Collect the computing power status of each node, display the total network computing power, the number of physical hosts and the network contract status of each host in real time, dynamically count the available resources, and determine the real-time computing power cost of each node based on the real-time resource utilization, operating cost, number of queued tasks and network scheduling cost of the node.

[0087] Step S2. Process user task requests. Decompose the task into subtasks using a multi-granularity task decomposition algorithm based on weighted graph clustering. Use computing power nodes as vertices and computing power scheduling overhead as connecting edges to model the computing power network as a directed acyclic graph.

[0088] Step S2 includes:

[0089] Step S21. When there is spare computing power resources, receive the task application submitted by the user, decompose the task into sub-tasks using the multi-granularity task decomposition method based on weighted graph clustering, decompose the user task into sub-tasks with dependencies, and record the resource requirements, data input and data output in the sub-tasks. Use spectral clustering algorithm to group the computing power nodes.

[0090] Step S22. Aggregate the available resources of all nodes in real time, centrally process all queued computing power call tasks within the same period, and model the computing power nodes as a directed acyclic graph structure, taking each computing power node in the global resource view as a vertex and the computing power scheduling overhead as a connecting edge.

[0091] Step S3. Based on the directed acyclic graph, map each subtask cluster to different node groups for execution, minimizing the total transmission overhead and computing power cost. Using resource constraints, price costs, and user computing power contributions as constraints, dynamically adjust the load balancing weights and use topology sorting to determine the task execution order.

[0092] Step S3 includes:

[0093] Step S31. Based on the directed acyclic graph, the subtasks are assigned to computing power device nodes that meet the hardware execution requirements through graph clustering algorithm. By adjusting the clustering parameters and clustering level, the size of the subtasks is controlled to match the computing power of the edge nodes of the subtasks, while ensuring that the total cost of task execution is lower than the budget.

[0094] Step S32. Based on the real-time collected node load information, which includes CPU, memory, network I / O and service queue length, the load balancing weight is dynamically adjusted through a reinforcement learning algorithm, and the Kahn algorithm is used for topology sorting to determine the task execution order.

[0095] Step S33. Perform node scheduling for new requests. When the overall load of some node regions is too high, the global coordination layer initiates cross-regional elastic scaling to migrate the task to the region with the lowest load.

[0096] Step S4. Connect the computing power nodes to each other in the execution order, connect the computing power devices running the subtasks to which the task belongs through the transmission link to form the execution path of the task, plan the execution path of each task, and minimize the completion time of service instances on the critical path.

[0097] Step S4 includes:

[0098] Step S41. Construct available transmission paths, connect computing nodes to each other, and mark the transmission speed, transmission conditions and maximum transmission volume of each path. The longest execution path in each task is taken as the critical path.

[0099] Step S42. Identify all subtasks located on the critical path, and determine the computing nodes and data forwarding paths between nodes for each subtask, taking resource constraints, price costs and user computing power contributions as constraints. At the same time, track the transmission status on each critical path and perform stability tests on the paths.

[0100] Step S5. Perceive network resource status information through INT metadata, determine the computing power quality of each node according to real-time status weighting, and schedule computing power resources among different nodes through load balancing management so that the variance of computing power quality allocated to all users is less than the threshold.

[0101] Step S5 includes:

[0102] Step S51. Generate INT data packets using in-band network telemetry technology. When the data packets are transmitted in the network, the network device writes real-time status information into the data packets. The real-time status information includes computation latency, jitter, path packet loss rate, and the computing resource status of the nodes.

[0103] Step S52. Sensing the latency, jitter, and packet loss rate of each path, and providing feedback on the computing power quality of each node. When a node is overloaded, the network is congested, or the computing power quality is below the threshold during operation, the load balancing mechanism is triggered to reschedule the task nodes. Through resource reservation and elastic quota mechanisms, the computing power quality allocated to all tasks is stabilized.

[0104] Example: The computing network consists of 10 computers distributed at different addresses. Users submit batch computing task packages through API. The task packages are decomposed into sub-task sets. The global resource status database is queried. With the goal of minimizing the total weighted cost, the optimal target node is selected from the candidate set under the premise of meeting the constraints. The path of data transmission from the upstream task to the node is determined. The scheduling plan is issued to the corresponding edge nodes for execution. Dynamic rescheduling is triggered when there is an anomaly.

[0105] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0106] Finally, it should be noted that the above descriptions are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A computing power resource management method based on multi-platform fusion, characterized in that, The method includes the following steps: Step S1. Adapt the cloud platform to computer equipment in various locations at multiple levels to establish a heterogeneous computing power collaboration platform. The platform collects the computing power status of nodes and dynamically calculates the available resources. It unifies the computing power of different architectures into computing power cost and dynamically bills the output computing power of contracted equipment. Step S2. Process user task requests. Decompose the task into subtasks using a multi-granularity task decomposition algorithm based on weighted graph clustering. Use computing power nodes as vertices and computing power scheduling overhead as connecting edges to model the computing power network as a directed acyclic graph. Step S3. Based on the directed acyclic graph, map each subtask cluster to different node groups for execution, minimizing the total transmission overhead and computing power cost. Using resource constraints, price costs, and user computing power contributions as constraints, dynamically adjust the load balancing weights and use topology sorting to determine the task execution order. Step S4. Connect the computing power nodes to each other in the execution order, connect the computing power devices running the subtasks to which the task belongs through the transmission link to form the execution path of the task, plan the execution path of each task, and minimize the completion time of service instances on the critical path. Step S5. Perceive network resource status information through INT metadata, determine the computing power quality of each node according to real-time status weighting, and schedule computing power resources among different nodes through load balancing management so that the variance of computing power quality allocated to all users is less than the threshold. Step S3 includes: Step S31. Based on the directed acyclic graph, the subtasks are assigned to computing power device nodes that meet the hardware execution requirements through graph clustering algorithm. By adjusting the clustering parameters and clustering level, the size of the subtasks is controlled to match the computing power of the edge nodes of the subtasks, while ensuring that the total cost of task execution is lower than the budget. Step S32. Based on the real-time collected node load information, which includes CPU, memory, network I / O and service queue length, the load balancing weight is dynamically adjusted through a reinforcement learning algorithm, and the Kahn algorithm is used for topology sorting to determine the task execution order. Step S33. Perform node scheduling for new requests. When the overall load of some node regions is too high, the global coordination layer initiates cross-region elastic scaling to migrate the task to the region with the lowest load. Step S4 includes: Step S41. Construct available transmission paths, connect computing nodes to each other, and mark the transmission speed, transmission conditions and maximum transmission volume of each path. The longest execution path in each task is taken as the critical path. Step S42. Identify all subtasks located on the critical path, and determine the computing nodes and data forwarding paths between nodes for each subtask, taking resource constraints, price costs and user computing power contributions as constraints. At the same time, track the transmission status on each critical path and perform stability tests on the paths.

2. The computing power resource management method based on multi-platform fusion according to claim 1, characterized in that: Step S1 includes: Step S11. Encapsulate cloud virtual machines, container clusters, edge servers and terminal devices located in different locations as computing power nodes. By deploying lightweight adapters on each computing power node to connect to devices with different interfaces, automatically discover and register newly connected heterogeneous computing power nodes. Network the heterogeneous resources through the adapters to establish a heterogeneous computing power collaboration platform. Step S12. Collect the computing power status of each node, display the total network computing power, the number of physical hosts and the network contract status of each host in real time, dynamically count the available resources, and determine the real-time computing power cost of each node based on the real-time resource utilization, operating cost, number of queued tasks and network scheduling cost of the node.

3. The computing power resource management method based on multi-platform fusion according to claim 2, characterized in that: Step S2 includes: Step S21. When there is spare computing power resources, receive the task application submitted by the user, decompose the task into sub-tasks using the multi-granularity task decomposition method based on weighted graph clustering, decompose the user task into sub-tasks with dependencies, and record the resource requirements, data input and data output in the sub-tasks. Use spectral clustering algorithm to group the computing power nodes. Step S22. Aggregate the available resources of all nodes in real time, centrally process all queued computing power call tasks within the same period, and model the computing power nodes as a directed acyclic graph structure, taking each computing power node in the global resource view as a vertex and the computing power scheduling overhead as a connecting edge.

4. The computing power resource management method based on multi-platform fusion according to claim 3, characterized in that: Step S5 includes: Step S51. Generate INT data packets using in-band network telemetry technology. When the data packets are transmitted in the network, the network device writes real-time status information into the data packets. The real-time status information includes computation latency, jitter, path packet loss rate, and the computing resource status of the nodes. Step S52. Sensing the latency, jitter, and packet loss rate of each path, and providing feedback on the computing power quality of each node. When a node is overloaded, the network is congested, or the computing power quality is below the threshold during operation, the load balancing mechanism is triggered to reschedule the task nodes. Through resource reservation and elastic quota mechanisms, the computing power quality allocated to all tasks is stabilized.

5. A computing resource management system based on multi-platform integration, used to implement the computing resource management method based on multi-platform integration as described in claim 1, characterized in that, The system includes the following modules: heterogeneous collaboration module, computing power task module, scheduling and processing module, node selection module, and quality sensing module; The heterogeneous collaboration module is used to adapt the cloud platform to computer devices in various locations through multi-level computing power resource nodes. By deploying lightweight adapters on each computing power node to connect to devices with different interfaces, it automatically discovers and registers newly connected heterogeneous computing power nodes, establishes a heterogeneous computing power collaboration platform, collects the computing power status of each node, displays the total network computing power, the number of physical hosts and the network contract status of each host in real time, dynamically calculates the available resources, and unifies the computing power of different architectures into computing power cost based on real-time resource utilization, operating costs, the number of queued tasks and network scheduling costs. The computing power task module is used to process task requests when there is spare computing power resources available. It decomposes user tasks into sub-tasks with dependencies. The sub-tasks record resource requirements, data input volume and data output volume. It uses a spectral clustering algorithm to group computing power nodes, calculates the weighted cost of tasks and sends computing power bills. The scheduling processing module is used to aggregate the available resources of all nodes in real time and centrally process all queued computing power call tasks within the same period. It uses each computing power node in the global resource view as a vertex and the computing power scheduling overhead as a connecting edge to model the computing power nodes into a directed acyclic graph structure. It uses a graph clustering algorithm to allocate batch subtasks to nodes and maps subtask clusters to different groups for execution, so as to minimize the total transmission overhead and computing power cost. After the user task is completed, the load balancing weight is dynamically adjusted with resource constraints, price costs and user historical contributions as constraints, and the task execution order is determined by topological sorting. The node selection module is used to construct available transmission paths, connect computing power nodes to each other, mark the transmission speed, transmission conditions and maximum transmission volume of each path, connect the computing power devices running the subtasks to which the task belongs through the transmission paths to form the execution path of the task, take the longest execution path of each task as the critical path, plan the data transmission path of each task, minimize the completion time of service instances on the critical path, track the transmission status on each critical path, and perform stability tests on the path. The quality sensing module is used to perceive network resource status information through INT metadata. When the INT metadata is transmitted in the network, each computing node writes real-time status information into the data packet. The real-time status information includes computation latency, jitter, path packet loss rate, and the computing resource status of the node. The computing quality of each node is determined by weighting according to the real-time status information. Through load balancing management, computing resources are scheduled among different nodes. Under the condition of a certain number of scheduled resources, the computing quality allocated to all users remains stable.

6. A computing resource management system based on multi-platform integration according to claim 5, characterized in that: The heterogeneous collaboration module includes: a node adaptation unit, a computing power management unit, and a cost pricing unit; The node adaptation unit is used to encapsulate cloud virtual machines, container clusters, edge servers and terminal devices located in different locations into computing power nodes; The computing power management unit is used to network heterogeneous resources through an adapter, collect the resource status of each node in real time, and display the total computing power, number of hosts, network status and contract status in a visualization window. The cost pricing unit is used to determine the computing power cost of each node in real time and to dynamically charge for the computing power output of the contracted equipment.

7. A computing resource management system based on multi-platform integration according to claim 6, characterized in that: The computing power task module includes: a task processing unit and a collaborative allocation unit; The task processing unit is used to receive task requests submitted by users and decompose the task into sub-tasks using a multi-granularity task decomposition method based on weighted graph clustering. The collaborative allocation unit is used to perform cross-regional collaborative scheduling and matching based on constraints of resources, costs, and user contributions.

8. A computing resource management system based on multi-platform integration according to claim 7, characterized in that: The scheduling processing module includes: a graph modeling unit and a resource constraint unit; The graph modeling unit is used to model the decomposed subtasks as a directed acyclic graph, with computing power nodes as vertices and computing power scheduling overhead as connecting edges. The resource constraint unit is used to allocate subtasks to computing power devices that can meet hardware requirements, while keeping the total cost of the task below the budget. The node selection module includes: a node forwarding unit and a path testing unit; The node forwarding unit is used to determine the task execution order using topology sorting, minimizing the critical path completion time; The path testing unit is used to identify all subtasks located on the critical path and determine the computing nodes of each subtask and the data forwarding paths between nodes.

9. A computing resource management system based on multi-platform integration according to claim 8, characterized in that: The quality sensing module includes: an INT sensing unit, a load balancing unit, and a quality management unit; The INT sensing unit is used to generate INT data packets using in-band network telemetry technology. When the data packets are transmitted in the network, the network device writes real-time status information into the data packets. The load balancing unit is used to provide a unified API gateway, track the execution status of each subtask, construct a network situation awareness map of the entire network, and perform load scheduling. The quality management unit is used to sense the latency, jitter, and packet loss rate of each path, provide feedback on the computing power quality of each node, and adjust task allocation according to the quality.

Images