Calculation power scheduling model training method and calculation power scheduling method

Abstract

The invention provides a federated learning-based computing power scheduling model training method and a computing power scheduling method, the method is applied to a processing node of a computing power scheduling system, the model training method comprises multiple training operations, and each training operation comprises the steps of obtaining a node feature sample of a computing node and a computing power route of the computing node; the node feature sample comprises a coordinate sample and a computing power resource sample of the computing node; training a local computing power scheduling model corresponding to the computing node by using the node feature sample and computing power routing, and obtaining a first model parameter of the trained local computing power scheduling model; and fusing all the first model parameters to obtain second model parameters of a global computing power scheduling model corresponding to the computing power scheduling system. The method can meet the requirements of high dynamic, isomerism and scalability of the cluster.

Description

Technical Field

[0001] This application belongs to the field of computing power scheduling technology in cloud computing, specifically involving a computing power scheduling model training method and computing power scheduling method based on federated learning. Background Technology

[0002] With the rapid development of technologies such as artificial intelligence and big data, computing clusters are being used more and more widely, and their scale and complexity are increasing.

[0003] Resource scheduling methods for computing clusters in related technologies are often insufficient to meet the requirements of real-time performance, high dynamism, heterogeneity, and scalability of computing clusters.

[0004] Therefore, there is an urgent need for a computing power scheduling method to optimize the computing power scheduling efficiency under computing clusters (such as hybrid cloud multi-clusters) so that the computing power scheduling method can meet the needs of computing clusters for high dynamics, heterogeneity and scalability. Summary of the Invention

[0005] This application aims to provide a method for training a computing power scheduling model based on federated learning and a computing power scheduling method, thereby solving at least some of the aforementioned problems in the prior art.

[0006] In a first aspect, embodiments of this application disclose a method for training a computing power scheduling model based on federated learning, applied to the processing nodes of a computing power scheduling system, the computing power scheduling system including multiple computing nodes; the method includes: performing multiple training operations to obtain a global computing power scheduling model; each training operation includes: obtaining node feature samples of the computing node and computing power routes of the computing node; the node feature samples include coordinate samples and computing power resource samples of the computing node; using the node feature samples and the computing power routes, training a local computing power scheduling model corresponding to the computing node, obtaining first model parameters of the trained local computing power scheduling model; fusing the first model parameters of all trained local computing power scheduling models to obtain second model parameters of the global computing power scheduling model corresponding to the computing power scheduling system.

[0007] Optionally, the processing node includes a local processing node corresponding to the computing node; after fusing the first model parameters corresponding to all computing nodes to obtain the second model parameters of the global computing power scheduling model corresponding to the computing power scheduling system, the method further includes: replacing the first model parameters of the local computing power scheduling model with the model parameter fusion value through the local processing node; the model parameter fusion value is obtained by fusing the first model parameters corresponding to all computing nodes; and determining the local computing power scheduling model after the model parameter replacement as the local computing power scheduling model corresponding to the computing node in the next iteration.

[0008] Optionally, the processing node includes local processing nodes corresponding to each of the computing nodes; obtaining the computing power route of the computing node includes: obtaining a topology information sample between the computing nodes through the local processing node; the topology information sample includes bandwidth and network latency; obtaining a weight sample between the computing nodes based on the topology information sample; and obtaining the computing power route of the computing node using the Dixtra algorithm based on the weight sample and the coordinate sample.

[0009] Optionally, obtaining the weights between the computing nodes based on the topology information sample includes: obtaining the weights between computing node i and computing node j through the local processing node according to the following method. :

[0010] in, The bandwidth between compute node i and compute node j, The network latency between compute node i and compute node j These are hyperparameters used to adjust the ratio of bandwidth to network latency.

[0011] Optionally, obtaining the topology information sample between the computing nodes includes: obtaining a physical topology map between the computing nodes through the local processing node using an open shortest path first routing algorithm; the nodes in the physical topology map are the computing nodes, and the edges represent the connection relationships between the computing nodes; and obtaining the topology information between the computing nodes by combining the connection relationships between the computing nodes and the computing power resource sample of the computing nodes.

[0012] Secondly, embodiments of this application disclose a computing power scheduling method based on federated learning, applied to the processing nodes of a computing power scheduling system, the computing power scheduling system including multiple computing nodes; the method includes: acquiring node characteristics of the computing node when a task to be processed is received; the node characteristics include the coordinates and computing power resources of the computing node; inputting the node characteristics into a global computing power scheduling model to obtain a target computing power route; the global computing power scheduling model is obtained by the method described in any one of the first aspects; and scheduling computing power for the task to be processed according to the target computing power route.

[0013] Optionally, after scheduling computing power for the task to be processed according to the target computing power route, the method further includes: updating the node feature samples of the computing nodes after scheduling computing power; the node feature samples before updating are the node feature samples described in any one of the first aspects; obtaining an updated global computing power scheduling model based on the updated node feature samples, so as to schedule computing power for the next task to be processed according to the updated global computing power scheduling model.

[0014] Thirdly, this application discloses a federated learning-based computing power scheduling model training device, applied to a computing power scheduling system. The computing power scheduling system includes multiple computing nodes. The device includes: a first acquisition module, used to perform multiple training operations to acquire a global computing power scheduling model; each training operation includes: acquiring node feature samples of the computing node and computing power routes of the computing node; the node feature samples include coordinate samples and computing power resource samples of the computing node; a second acquisition module, used to use the node feature samples and the computing power routes to train a local computing power scheduling model corresponding to the computing node, and acquire the first model parameters of the trained local computing power scheduling model; a third acquisition module, used to fuse the first model parameters of all trained local computing power scheduling models to acquire the second model parameters of the global computing power scheduling model corresponding to the computing power scheduling system.

[0015] Fourthly, embodiments of this application disclose a computing power scheduling device based on federated learning, applied to a computing power scheduling system, the computing power scheduling system including multiple computing nodes; the device includes: a fourth acquisition module, used to acquire node characteristics of the computing nodes when a task to be processed is received; the node characteristics include the coordinates and computing power resources of the computing nodes; a fifth acquisition module, used to input the node characteristics into a global computing power scheduling model to obtain a target computing power route; the global computing power scheduling model is obtained by the method described in any one of the first aspects; a sixth acquisition module, used to perform computing power scheduling on the task to be processed according to the target computing power route.

[0016] Fifthly, embodiments of this application also disclose an electronic device, including a processor and a memory, wherein the memory stores a program or instructions executable on the processor, and the program or instructions, when executed by the processor, implement the steps of the method described in the first or second aspect.

[0017] In a sixth aspect, embodiments of this application also disclose a readable storage medium storing a program or instructions that, when executed by a processor, implement the steps of the method described in the first or second aspect.

[0018] In this embodiment, node feature samples and computing power routes of computing nodes are obtained. The node feature samples include coordinate samples and computing power resource samples of the computing nodes. Using the node feature samples and the computing power routes, a local computing power scheduling model corresponding to the computing nodes is trained. The resulting local computing power scheduling model is sensitive to changes in computing power resources in the computing power scheduling system and can reflect the impact of the location of computing nodes on computing power scheduling. Even when computing nodes are relatively dispersed, it can promptly provide real-time feedback on computing power scheduling results based on resource changes, meeting the requirements of high dynamism, heterogeneity, and scalability. This embodiment does not collect the computing power resources of each computing node to train the global computing power scheduling model. Instead, it trains the local computing power scheduling model locally on the computing nodes, then merges the first model parameters of all trained local computing power scheduling models to obtain the second model parameters of the global computing power scheduling model corresponding to the computing power scheduling system. Through continuous iteration, the global computing power scheduling model is obtained. This embodiment implements federated learning-based computing power scheduling model training, which can ensure the data security of each computing node (e.g., server cluster) and prevent data leakage. Attached Figure Description

[0019] In the attached diagram: Figure 1 This is a flowchart illustrating the steps of a computing power scheduling model training method based on federated learning, as provided in an embodiment of this application. Figure 2 This is a flowchart illustrating the steps of another computing power scheduling model training method based on federated learning provided in this application embodiment; Figure 3 This is a flowchart illustrating the steps of a computing power scheduling method based on federated learning provided in an embodiment of this application. Figure 4 This is a flowchart illustrating the steps of another computing power scheduling model training method based on federated learning provided in the embodiments of this application; Figure 5 This is a flowchart of an adaptive computing power scheduling model training method provided in an embodiment of this application; Figure 6 This is a flowchart illustrating the steps of a federated learning-based computing power scheduling model training method and a computing power scheduling method provided in this application embodiment; Figure 7 This is a structural block diagram of a computing power scheduling model training device based on federated learning provided in an embodiment of this application; Figure 8 This is a structural block diagram of a computing power scheduling device based on federated learning provided in an embodiment of this application; Figure 9This is a block diagram of an electronic device provided in one embodiment of this application; Figure 10 This is a block diagram of an electronic device according to another embodiment of the present application. Detailed Implementation

[0020] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0021] The terms "first," "second," etc., used in the specification and claims of this application are used to distinguish similar objects and not to describe a specific order or sequence. It should be understood that such use of data can be interchanged where appropriate so that embodiments of this application can be implemented in orders other than those illustrated or described herein, and the objects distinguished by "first," "second," etc., are generally of the same class and the number of objects is not limited; for example, a first object can be one or more. Furthermore, in the specification and claims, "and / or" indicates at least one of the connected objects, and the character " / " generally indicates that the preceding and following objects are in an "or" relationship.

[0022] The following detailed description, with reference to the accompanying drawings, of the federated learning-based computing power scheduling model training method and the computing power scheduling method provided in this application, through specific embodiments, will be provided in detail.

[0023] Some of the nouns or terms appearing in this application shall be interpreted as follows: Federated learning, proposed by Google, is a machine learning algorithm that is easier to deploy and scale compared to deep learning. The main idea behind federated learning is to build a machine learning model from datasets distributed across multiple server clusters, while ensuring data security for each cluster and preventing data leaks. The federated learning model is a distributed algorithm model for multi-node collaborative learning. It can be trained locally on a large scale on the devices that generate the data, and the data is securely stored with the data owner. A central coordinator then updates the distributed model by obtaining the parameters from each node cluster.

[0024] Hybrid cloud is an IT infrastructure that can manage multiple cloud resources across regions and clouds, encompassing both public and private cloud products. Hybrid cloud can bridge the gap between private and public cloud products, providing a unified user service experience and offering users a multi-cloud management platform.

[0025] Compute First Networking (CFN) is a network that connects computing power, allowing demand-side devices to share the network's computing power, such as Graphics Processing Units (GPUs), Central Processing Units (CPUs), and storage resources.

[0026] Figure 1 This is a flowchart illustrating the steps of a federated learning-based computing power scheduling model training method provided in this embodiment. The method is applied to the processing nodes of a computing power scheduling system, which includes multiple computing nodes. Figure 1 As shown, the method may include performing multiple training operations to obtain a global computing power scheduling model; wherein each training operation includes the following steps: Step 101: Obtain node feature samples of computing nodes and computing power routing of computing nodes.

[0027] Among them, the node feature samples include the coordinate samples of computing nodes and the computing resource samples.

[0028] Node feature samples can be represented as .

[0029] in, This represents the characteristics of the CPU of computing node i (e.g., the number of cores). This describes the characteristics of the graphics processing unit (GPU) of computing node i. This represents the storage resources of compute node i; This represents the number of floating-point operations per second (FLOPS) of computing node i.

[0030] Step 102: Using node feature samples and computing power routing, train the local computing power scheduling model corresponding to the computing node, and obtain the first model parameters of the trained local computing power scheduling model.

[0031] Each computing node has its own local computing power scheduling model. For each computing node, based on the node feature samples of computing node i... and computing power routing Construct and compute the training data corresponding to node i :

[0032] training data Input the computing power scheduling model corresponding to computing node i, and obtain the trained local computing power scheduling model and its first model parameters. For example, the local computing power scheduling model is a federated learning model, and the first model parameter is the gradient.

[0033] For example, the computing power scheduling path between compute node i and compute node j is: In a computing power scheduling system with n computing nodes, the computing power routing for computing node i is... for A matrix, where the j-th element of the matrix is , where n is the total number of computing nodes in the computing power scheduling system.

[0034] Step 103: Merge the first model parameters of all trained local computing power scheduling models to obtain the second model parameters of the global computing power scheduling model corresponding to the computing power scheduling system.

[0035] For example, the model parameters (e.g., gradients) corresponding to each computing node are weighted, summed, and then averaged to obtain the fused model parameter value.

[0036] In this embodiment, the steps from obtaining node feature samples of computing nodes to obtaining the second model parameters of the global computing power scheduling model corresponding to the computing power scheduling system are iteratively executed until a preset iteration termination condition is met, resulting in the final second model parameters of the global computing power scheduling model. The global computing power scheduling model with the final second model parameters is used for computing power scheduling.

[0037] The computing power scheduling methods in related technologies have problems such as low resource utilization, high scheduling latency, and susceptibility to cluster location, making it difficult to meet the real-time, scalable, and comprehensive computing power scheduling requirements that need to take into account factors such as cluster location.

[0038] In this embodiment, node feature samples and computing power routes of computing nodes are acquired. The node feature samples include coordinate samples and computing power resource samples of the computing nodes. Using the node feature samples and computing power routes, a local computing power scheduling model corresponding to the computing nodes is trained. The resulting local computing power scheduling model is sensitive to changes in computing power resources in the computing power scheduling system and can reflect the impact of the location of computing nodes on computing power scheduling. Even when computing nodes are relatively dispersed, it can promptly provide real-time feedback on computing power scheduling results based on resource changes, meeting the requirements of high dynamism, heterogeneity, and scalability. This embodiment does not collect the computing power resources of each computing node to train the global computing power scheduling model. Instead, it trains the local computing power scheduling model locally on the computing nodes, then merges the first model parameters of all trained local computing power scheduling models to obtain the second model parameters of the global computing power scheduling model corresponding to the computing power scheduling system. Through continuous iteration, the global computing power scheduling model is obtained. This embodiment implements federated learning-based computing power scheduling model training, which can ensure the data security of each computing node (e.g., server cluster) and prevent data leakage.

[0039] Reference Figure 2 The method may include the following steps: Step 201: Obtain node feature samples of the computing nodes.

[0040] The node feature samples include coordinate samples and computing resource samples of the computing nodes. The local processing nodes corresponding to the computing nodes obtain the node feature samples of the computing nodes.

[0041] For example, the processing nodes include local processing nodes corresponding to each computing node, and global processing nodes corresponding to the computing power scheduling system. Each local processing node obtains node feature samples from the computing nodes.

[0042] Step 202: Obtain a sample of topological structure information between computing nodes through the local processing node.

[0043] The topology information sample includes bandwidth and network latency.

[0044] For example, step 202 may include the following sub-steps: Sub-step A1: Using the local processing node, obtain the physical topology map between computing nodes according to the open shortest path first routing algorithm.

[0045] In the physical topology graph, the nodes are computing nodes, and the edges represent the connection relationships between computing nodes.

[0046] For example, the physical topology map between multiple computing nodes is obtained by using the Open Shortest Path First (OSPE) routing algorithm combined with coordinate samples.

[0047] Sub-step A2 involves combining the connection relationships between computing nodes and the computing power resource samples of the computing nodes to obtain the topological structure information between the computing nodes.

[0048] In this embodiment, the OSPE routing algorithm can be used to generate a physical topology map between computing nodes (e.g., cluster nodes) and obtain topology information such as connection relationships, bandwidth, and network latency between each computing node.

[0049] After obtaining the physical topology graph, it can be mapped to a computing power graph. In this embodiment, based on the generated physical topology graph and node characteristics, the corresponding weight is calculated for each edge in the physical topology graph to ensure that the generated computing power network graph matches the actual environment. For example, it ensures that the generated computing power network graph matches the actual computing power resources of each computing node in the computing power scheduling system.

[0050] When adding or removing cluster nodes in the computing power scheduling system, a new network topology can be generated in a timely manner based on the updated cluster nodes, so that the entire computing power scheduling system has better scalability.

[0051] Step 203: Obtain weight samples between computing nodes based on the topology information samples; For example, step 203 may include the following sub-steps: Sub-step B1: Using the local processing node, obtain the weight samples between computing node i and computing node j according to the following method. :

[0052] in, The bandwidth between compute node i and compute node j, The network latency between compute node i and compute node j These are hyperparameters used to adjust the ratio of bandwidth to network latency.

[0053] In this embodiment, based on the weights between computing nodes for bandwidth and network latency calculation, the calculation method integrates the relationship between the distance between computing nodes and the network transmission rate, thus balancing the requirements of data transmission rate and response time.

[0054] Step 204: Based on the weight samples and coordinate samples, obtain the computing power route of the computing nodes using the Dixtra algorithm.

[0055] Computing power routing can be used to transmit the structural information of the computing power network and the perceived computing power resource data to the computing power scheduling engine of the computing power scheduling system, so that the computing power scheduling engine can receive the changes in these data in real time.

[0056] Based on the generated topology map (i.e., the computing power map), combined with the Dijkstra algorithm, the location information of different computing nodes (e.g., clusters) can be integrated.

[0057] For example, in cases where the cluster architecture is too distributed, the weight ratios at different locations can be adjusted, and corresponding computing power routes can be generated.

[0058] For example, the computing power route between computing node i and computing node j can be obtained using the following method. :

[0059] in, This represents the coordinates of node i. This represents the coordinates of node j. This represents the weight between computing node i and computing node j.

[0060] Step 205: Using node feature samples and computing power routing, train the local computing power scheduling model corresponding to the computing node, and obtain the first model parameters of the trained local computing power scheduling model.

[0061] The method for this step has been explained in step 102 above, and will not be repeated here.

[0062] Step 206: Merge the first model parameters of all trained local computing power scheduling models to obtain the second model parameters of the global computing power scheduling model corresponding to the computing power scheduling system.

[0063] The method for this step has been explained in step 103 above, and will not be repeated here.

[0064] The process iteratively executes steps from obtaining node feature samples of computing nodes to obtaining the second model parameters of the global computing power scheduling model corresponding to the computing power scheduling system, until the preset iteration termination condition is met, and the final model parameters of the global computing power scheduling model are obtained.

[0065] Following step 207, the method may further include the following steps: Step 208: Using the model parameter fusion value, replace the first model parameter of the local computing power scheduling model through the local processing node.

[0066] The model parameter fusion value is obtained by fusing the first model parameters corresponding to all computing nodes. For example, the first model parameters of all local computing power scheduling models can be fused by weighted summation and then averaging to obtain the model parameter fusion value.

[0067] In the current iteration, the fused model parameter value is determined as the second model parameter of the global computing power scheduling model. Then, the second model parameter of the global computing power scheduling model is distributed to each computing node, and the first model parameter of the local computing power scheduling model corresponding to each computing node is updated.

[0068] Step 209: The local computing power scheduling model after the model parameters are replaced is determined as the local computing power scheduling model corresponding to the computing node in the next iteration.

[0069] In this embodiment, an improved machine learning-based federated learning algorithm is applied to computing power scheduling. Based on this embodiment, data privacy is well protected across different clusters. Furthermore, compared to other training-type computing power scheduling algorithms in related technologies, the computing power scheduling method in this embodiment involves iterative optimization of local model parameters (i.e., the first model parameters) and global model parameters (i.e., the second model parameters). By fusing all local model parameters, the global model parameters are quantized when the aggregation server processes the global parameters. After obtaining the global model parameters, they are then distributed to the local computing nodes for training of the local computing power scheduling model. Through continuous iterative optimization, the prediction accuracy of the computing power scheduling model is improved, and it has the advantages of more lightweight model parameters, faster scheduling speed, and lower latency.

[0070] This embodiment also provides a federated learning-based computing power scheduling method, applied to a computing power scheduling system comprising multiple computing nodes. (Refer to...) Figure 3 The method may include the following steps: Step 301: Obtain the node characteristics of the computing node when the task to be processed is received.

[0071] Among them, node characteristics include the coordinates and computing resources of the computing node.

[0072] Step 302: Input the node features into the global computing power scheduling model to obtain the target computing power route.

[0073] The global computing power scheduling model is obtained through the method of any of the above embodiments.

[0074] Step 303: Based on the target computing power route, perform computing power scheduling for the task to be processed.

[0075] In this embodiment, the node characteristics of the computing node are obtained when a task to be processed is received. The node characteristics include the coordinates and computing resources of the computing node. The node characteristics are input into a global computing power scheduling model to obtain a target computing power route. The global computing power scheduling model is obtained through the method of any of the above embodiments. After obtaining the target computing power route, computing power scheduling can be performed on the task to be processed based on the target computing power route. This can meet the requirements of highly dynamic, heterogeneous, and scalable computing power scheduling.

[0076] For example, after step 303, the method may also include the following steps: Step 304: After computing power scheduling, update the node feature samples of the computing nodes.

[0077] The node feature samples before the update are obtained according to the method of any of the aforementioned time examples.

[0078] Step 305: Based on the updated node feature samples, obtain the updated global computing power scheduling model, and then schedule computing power for the next task to be processed according to the updated global computing power scheduling model.

[0079] Based on the updated node feature samples, the target model parameters of the global computing power scheduling model are fine-tuned to obtain the updated global computing power scheduling model.

[0080] In this embodiment, after the global computing power scheduling model is trained, dynamic scheduling of computing power resource allocation and continuous optimization of the computing power scheduling model are performed.

[0081] Specifically, based on changes in computing resource data and computing node coordinates in the computing power scheduling system, the allocation of computing resources is dynamically adjusted. The adjustment of computing resource allocation includes: migrating running tasks, adjusting resource allocation ratios, etc. After adjusting the computing resource allocation, the computing resources of each node are recalculated and the data is updated.

[0082] By continuously collecting the latest operational data and system status, the model parameters of the global computing power scheduling model are continuously updated iteratively to maintain and update computing power routing and generate positive feedback. This ensures that each step in the embodiment is constrained by resource changes, thereby improving the real-time performance of the computing power scheduling algorithm.

[0083] The hybrid cloud adaptive computing power scheduling method based on federated learning optimization provided in this embodiment can include a resource management phase and a computing power adaptive scheduling phase.

[0084] In the resource management phase, the computing power scheduling system can manage and integrate various types of heterogeneous resources and support the configuration and monitoring of these heterogeneous resources, including collecting and analyzing the status, deployment and connectivity of resources.

[0085] Furthermore, through automated methods, the computing power scheduling system can control and schedule heterogeneous resources, allocating resources and assigning tasks according to specific needs. By continuously monitoring, managing, and scheduling computing resources distributed in the hybrid cloud environment, the computing power scheduling system can perceive the distribution of computing resources, aggregate these resources, and map the physical topology to the virtual layer based on the physical deployment of each computing node in the computing power scheduling system. It then unifies the relevant data into virtualized resources, hiding the underlying implementation details during processing, thus forming a graph-based computing network view.

[0086] In the adaptive computing power scheduling phase, the computing power scheduling system employs intelligent scheduling algorithms and strategies to promote global optimization through local optimization, achieving automated allocation and scheduling of computing resources. Based on real-time demands and priorities, the system can dynamically adjust the quantity and scale of computing resources, supporting elastic scaling of business needs and providing more flexible resource services. Based on this embodiment, during the allocation of computing tasks, the utilization rate of computing resources can be optimized, improving the overall performance and efficiency of the system and reducing resource idleness and waste.

[0087] Based on the adaptive resource allocation and task scheduling strategy in this embodiment, it is easier to capture the dynamic changes in computing resources of each local computing node; and the computing power scheduling strategy can be dynamically adjusted according to the dynamic changes, making its processing speed faster and more intelligent. In addition, dynamically adjusting the computing power scheduling strategy according to the dynamic changes in computing resources can further improve the system's reliability, response speed and performance, providing users with a high-quality cloud computing service experience.

[0088] With the rapid development of technologies such as artificial intelligence and big data, the operation and processing models of many industries are changing, and computing application scenarios are constantly expanding and becoming more complex. Cloud computing models in these technologies often suffer from low resource utilization, high costs, and slow response times, making it difficult to meet the highly dynamic, heterogeneous, and scalable requirements of computing power scheduling systems, and they are also insensitive to changes in computing resources. Therefore, a new intelligent computing power scheduling algorithm is needed to optimize the computing power scheduling efficiency of hybrid cloud multi-cluster systems. However, the large scale and high complexity of hybrid cloud platforms make resource scheduling challenging.

[0089] For example, in related technologies, the accuracy of cluster computing power scheduling can be improved by determining the computing power scheduling result of the second cluster based on preset node scheduling information and computing power demand information. However, although this method improves the accuracy of computing power scheduling based on the second cluster, the resource utilization is not high, and the scheduling has a certain time delay, and it cannot provide real-time feedback on the scheduling results based on changes in cluster resources. Related technologies can also comprehensively consider the impact of multi-dimensional resources on tasks to achieve adaptive dynamic resource scheduling, thereby improving the task deployment completion and multi-dimensional resource utilization in AI computing power scheduling. However, although this method can dynamically allocate computing resources based on user requests, it does not consider the impact of cluster location on computing power scheduling, and when resources are too dispersed, the scheduling strategy of this method decreases, resulting in poor scalability of scheduling computing power resources.

[0090] This embodiment, based on the multi-environment management of hybrid cloud, includes all computing resources of multiple clusters under management. It generates a graph-based computing network through the physical topology of computing power, summarizes the computing resource data of each computing node in the computing power scheduling system in real time, and performs dynamic computing power scheduling through a computing power scheduling algorithm optimized based on federated learning.

[0091] In this embodiment, the core of the computing power scheduling system lies in its computing power scheduling engine. Besides the engine, the hybrid cloud architecture can be divided into several participants comprised of different cloud computing clusters. Each participant first optimizes the model parameters (e.g., gradient information) of its local computing power scheduling model within its local environment; then, the updated gradient information is uploaded to a centralized aggregation server, which optimizes the global computing power scheduling model. Based on the processing method of this embodiment, it can respond promptly to resource changes in each participant node and perform computing power scheduling efficiently.

[0092] This embodiment enables adaptive scheduling of computing resources, achieving dynamic allocation and management of these resources. This allows the system to flexibly adapt resources such as computing cores and memory based on real-time demands and task priorities. This embodiment improves the efficiency of computing resource utilization and system response speed, while also reducing operating costs. Furthermore, by treating each component in the hybrid cloud environment as an independent participant to handle different requests, this embodiment makes local training more closely resemble real-world application scenarios and can handle tasks of varying scales and complexities, thereby meeting the diverse needs of various industries.

[0093] The method of this embodiment will be further illustrated below.

[0094] In one embodiment, refer to Figure 4 Computing nodes can include computing clusters in private clouds and computing clusters in public clouds; computing nodes are participants.

[0095] The physical coordinates and physical resource information of the computing nodes are obtained, and a physical topology map is generated based on the physical coordinates and physical resource information. In this embodiment, the physical coordinates and physical resource information include: physical coordinates 1 and physical resource 1 of the private cloud, and physical coordinates 2 and physical resource 2 of the public cloud.

[0096] Based on physical coordinates and physical resource information, feature tuples are generated for each computing node. Then, based on the feature tuples and the physical topology graph, a computing power network graph is generated between the computing nodes. The feature tuples for the private cloud and the public cloud are feature tuple F1 and feature tuple F2, respectively.

[0097] Furthermore, the weights between nodes in the computing power network graph are calculated based on the computing power network graph; and computing power routes are generated based on the weights between nodes in the computing power network graph.

[0098] For each computing node (public or private cloud), a computing power scheduling model is trained locally based on its corresponding feature tuple; then, the parameters of the participating nodes are aggregated, and the global computing power scheduling model corresponding to the computing power scheduling system is globally iteratively optimized according to the computing power routing to obtain a global computing power scheduling model corresponding to the computing power scheduling system.

[0099] Computational power scheduling is performed based on a global computing power scheduling model. The system then fine-tunes the scheduling model and adjusts resources according to the physical coordinates and resource information of the compute nodes after scheduling. Specifically, during the fine-tuning process, participants download the computing power scheduling model locally and fine-tune it. The model parameters (e.g., gradients) of the fine-tuned model are then aggregated. Based on these aggregated parameters, the global computing power scheduling model is fine-tuned.

[0100] The method of this embodiment will be further illustrated below, with reference to... Figure 5 The computing power scheduling system includes multiple participants, which are equivalent to the computing nodes in the aforementioned embodiments.

[0101] For example, the participants include Figure 5 The participants are 1, 2, and k. The computing power scheduling system also includes the aggregation server of the hybrid cloud platform. The interaction process between each participant and the aggregation server of the hybrid cloud platform includes: Step 1: Participants perform local iterative optimization and updates to the machine learning model.

[0102] The machine learning model in this embodiment is the computing power scheduling model in the previous embodiment.

[0103] Step 2: Initialize the aggregation server of the hybrid cloud platform, or update the feedback coefficients.

[0104] Step 3: The aggregation server of the hybrid cloud platform performs iterative optimization and updates to the global computing power scheduling model.

[0105] Step 4: The hybrid cloud platform sends the combined feedback coefficients and aggregation parameters to each participant.

[0106] In this embodiment, based on the multi-environment management of hybrid cloud, it includes all computing resources of multiple clusters under management. A graph-based computing network is generated through the physical topology of computing power, and computing resource data is aggregated in real time. The computing power is dynamically scheduled by a computing power scheduling algorithm optimized based on federated learning.

[0107] The method of this embodiment will be further illustrated below, with reference to... Figure 6 The method may include the following steps: Step S1: Establish a computing resource pool.

[0108] All computing resources in a cluster can be managed through a computing resource pool, which consists of a computing resource layer and computing resource identifiers. For example, multi-cloud resources and heterogeneous resources can be managed based on the deployment of physical servers.

[0109] The computing resource layer is the foundation of the computing network, encompassing all available resources in the hybrid cloud platform and used to generate the topology for physical resource deployment. Computing resource identifiers are used to uniquely identify the managed computing resources for unified management in subsequent operations.

[0110] For example, a physical topology using a three-level identification strategy can be employed. First, a unique identifier (IdentityDocument, ID) is used to identify hardware computing resources. This includes information such as the Media Access Control (MAC) address of the CPU or network interface card. During deployment, a second level of identification can be applied to the physical cluster to record information such as its location. Finally, a third level of identification is applied to the virtual participants within the cluster.

[0111] Step S2: Generate the physical topology map.

[0112] The method for generating a physical topology graph may include the following sub-steps: Sub-step S21: Collect physical topology information, obtain the topology information of each physical node in the cloud computing environment, and construct the feature representation tuple of the node.

[0113] The physical topology information is equivalent to the topology information sample in the aforementioned embodiments. The topology information sample may include: the connection relationships between nodes, bandwidth, latency, etc. The constructed feature representation tuple for the i-th node... .

[0114] Sub-step S22: Based on the graph data structure and node characteristics, the OSPF routing algorithm is used to construct a complete network topology graph for each node, generating a graph-based computing network.

[0115] In the graph, nodes represent computing nodes, edges represent the connections between nodes, and weights represent the connection quality or resource allocation weights between nodes.

[0116] Sub-step S23: Calculate the weights of the edges in the graph based on the physical topology information and node characteristics.

[0117] For example, weighted samples can be calculated by combining bandwidth and latency metrics. For instance, the weights can be calculated using the following method: :

[0118] in, Characterizes the bandwidth between computing node i and computing node j. The network latency (i.e., time delay) between computing node i and computing node j is represented by α; α represents a hyperparameter that adjusts the ratio of bandwidth to latency. By combining these two metrics (bandwidth and latency) to calculate the weights, a balance can be struck between data transmission rate and response time.

[0119] Step S3: Construct the virtualization layer.

[0120] For example, Kubernetes automation tools and scripts can be used to automatically configure, deploy, and manage computing, storage, and network resources in a cloud environment. For instance, automated operations such as virtual machine creation can be performed.

[0121] Through orchestration tools and policies, tasks and workloads in the cloud environment are automatically scheduled and managed; for example, load balancing and resource optimization are performed to achieve efficient resource utilization and optimal performance. Simultaneously, the number of virtual machines can be automatically adjusted, and storage capacity can be automatically expanded to meet changing needs and provide elastic resource provisioning.

[0122] Step S4: Generate computing power routes.

[0123] For example, a pathfinding algorithm can be used to generate computing power routes. This can be used, such as tracing nodes in a computing power network graph. map i and map j The shortest path between them.

[0124] For example, an improved Dijkstra algorithm can be used to generate computational power routes. This improved Dijkstra algorithm determines the optimal computational power route by considering the edge weights W.

[0125] Among them, computing power routing is mainly used to transmit the topology information of the computing power network and the computing power data it senses to the computing power scheduling engine corresponding to the computing power scheduling system, so that the computing power scheduling engine can receive the changes in these data in real time.

[0126] In addition, this dynamically changing information will also be fed back to the computing power routing module, enabling the computing power routing module to continuously update the computing power routes in order to maintain a high degree of sensitivity to the current node resource status.

[0127] Example: Computational power routing between node i and node j It can be obtained through the following method:

[0128] By repeating the above steps, the optimal computing power scheduling path from node i to all other nodes in the computing power topology graph can be obtained. The computing power scheduling path data can be represented by a 1×n matrix Ri, where Ri is the computing power route for node i, and n is the number of computing nodes.

[0129] Step S5: Summarize and generate computing power data for each computing node.

[0130] Computing power data includes feature representation tuples and computing power routing As training data for the adaptive scheduling algorithm :

[0131] Step S6: Perform local iterative optimization of the corresponding computing power scheduling model parameters for the participating nodes.

[0132] For example, training data The system inputs the computing power scheduling model corresponding to the participating nodes and iteratively optimizes the model parameters for the participating nodes based on the local environment. The computing power scheduling model is a federated learning model.

[0133] Step S6 may include the following sub-steps: Sub-step S61: Participating nodes, based on the data allocated to their local nodes... The batch performs forward propagation to calculate the loss and backpropagation to calculate the gradient.

[0134] The training process can begin with initializing the model parameters W. k start.

[0135] In sub-step S62, the participating nodes calculate the gradient g corresponding to the current batch of data, and update the parameters W of the local model using gradient descent based on the gradient g. k .

[0136] After training on a batch of data is completed, a local iterative optimization is finished, and the gradient g after one iteration is obtained. k .

[0137] In sub-step S63, the participating nodes upload the gradients generated in this iteration to the aggregation server.

[0138] In sub-step S7, the aggregation server performs global iterative optimization based on the gradients uploaded by each participating node.

[0139] For example, sub-step S7 may include the following sub-steps: In sub-step S71, the aggregation server collects the gradients uploaded by all participating nodes, and performs weighted summation and average of the gradients uploaded by each participating node according to the pre-set weight coefficients.

[0140] Among them, the selection of weight coefficient P k The weight coefficient P can be determined based on the amount of data and data instructions of each participating node. k And through P k The final global gradient parameters are obtained by weighted summation and averaging of the gradients g.

[0141] In this embodiment, by performing a weighted summation and averaging of the gradients uploaded by each participating node, this aggregation algorithm can better reflect the contribution of each participating node in the global model training, thereby obtaining an accurate global gradient.

[0142] Sub-step S72 updates the global model parameters through the global gradient parameters to achieve global optimization.

[0143] In sub-step S73, the updated global model parameters from the aggregation server are distributed back to each participating node so that each participating node can perform the next round of local iterative optimization.

[0144] Step S8: Dynamically schedule the allocation of computing resources and continuously optimize the computing resource scheduling model.

[0145] This step is mainly used to continuously fine-tune the trained global model based on the changes in computing resource data after scheduling.

[0146] For example, step S8 may include the following sub-steps: Sub-step S81: Based on the prediction of the global optimization model, schedule the computing resources of the nodes output by the adaptive scheduling algorithm to execute the computing power scheduling task.

[0147] Sub-step S82: Dynamically adjust the allocation of computing resources; after dynamically adjusting the allocation of computing resources, recalculate the computing resources of each node and update the computing resource data.

[0148] For example, dynamic adjustments to computing resource allocation can include migrating running tasks and adjusting resource allocation ratios.

[0149] Sub-step S83 continuously collects the latest running data and system status, and iteratively updates the global model parameters.

[0150] To cope with the ever-changing computing power requirements and resource conditions in hybrid cloud environments, we continuously collect the latest operational data and system status, and iteratively update global model parameters, enabling more efficient and flexible adaptive resource scheduling.

[0151] In this embodiment, steps S1 to S5 belong to the resource management stage, and steps S6 to S8 belong to the computing power adaptive scheduling stage.

[0152] During the resource management phase, the computing power scheduling system can manage and integrate various types of heterogeneous resources, and support the configuration and monitoring of these resources, including collecting and analyzing resource status, deployment, and connectivity. Through automated processing, the computing power scheduling system can control and schedule heterogeneous resources, and allocate resources and assign tasks according to specific needs.

[0153] By continuously monitoring, managing, and scheduling computing resources distributed across a hybrid cloud environment, the computing power scheduling system achieves awareness of the distribution of computing resources. By sensing and aggregating these resources, the system maps the physical topology to the virtual layer based on their physical deployment, forming a graph-based computing network view.

[0154] During the adaptive computing power scheduling phase, local optimization promotes global optimization, enabling automated allocation and scheduling of computing resources. Based on real-time needs and priorities, the system can dynamically adjust the quantity and scale of computing resources, supporting elastic scaling of business requirements and providing more flexible resource services. In the process of allocating computing tasks, the utilization rate of computing resources is optimized, improving the overall performance and efficiency of the system and reducing resource idleness and waste.

[0155] This embodiment is based on an adaptive resource allocation and task scheduling strategy, which can capture the dynamic changes of each local computing node in a timely manner. Based on this, the computing power allocation strategy is dynamically adjusted to further improve the system's reliability, response speed and performance, and provide users with a high-quality cloud computing service experience.

[0156] In this embodiment, by collecting and analyzing historical data and real-time metrics, the optimal routing strategy for different workloads is learned through the reverse effect of the optimization training of the computing power scheduling engine. This strategy is then dynamically adjusted based on the current situation to improve routing efficiency and task execution speed. In selecting the optimal scheduling path, an improved Dijkstra algorithm is employed, taking into account the weights of the edges. Finding nodes in the computing power network graph and Find the shortest path between them to obtain the optimal computing power route.

[0157] The method of this embodiment will be further illustrated below. For example, the method may include the following steps: Step F1 involves bringing all computing resources under management to form a computing resource pool.

[0158] In this embodiment, computing resources are identified through an identification strategy. The identification strategy can be a three-level identification strategy. Specifically, the first level of identification is to uniquely represent hardware computing resources (such as CPU, network card MAC address, etc.) through hardware IDs; the second level of identification is to record information such as cluster location; and the third level of identification is to identify virtual participants within the cluster.

[0159] Based on the physical topology information of each computing resource, a topology structure for the physical resource deployment is generated to ensure the unique identification of the resources and facilitate unified management.

[0160] Step F2: Collect topological location information, obtain information such as the connection relationships, bandwidth, and latency between physical nodes, and construct node feature representation tuples. Using the OSPF routing algorithm, a complete network topology graph is constructed at each node to generate a graph-based computing network and build the network topology. Based on bandwidth and network latency, the weights of the edges in the graph are calculated to comprehensively evaluate connection quality and resource allocation.

[0161] Step F3 utilizes tools such as Kubernetes to automate the configuration and deployment of computing resources and manage resources in the cloud environment. This involves automatically scheduling tasks and workloads to achieve efficient resource utilization and performance optimization. Resources are dynamically adjusted to meet changing needs, such as automatically scaling the number of virtual machines and storage capacity.

[0162] Step F4: Using the improved Dijkstra algorithm, obtain the optimal path to determine the best computational power route between node i and node j.

[0163] For example, the computing network structure and resource data are transmitted to the computing scheduling engine, and data changes are updated in real time to maintain a high degree of sensitivity to node resource status.

[0164] Step F5: Obtain the feature representation tuple and computing power routing information for each node, and generate the training dataset for the adaptive scheduling algorithm. ,in:

[0165] Step F6, for each participating node During the local model parameter optimization process, the training data used is Initialize model parameters .

[0166] For example, global convergence can be set to require S iterations of optimization.

[0167] After initializing the model parameters, perform forward and backward propagation to calculate the gradients for each data batch, and update the local model parameters using gradient descent. For example, the calculation formula is as follows:

[0168] After training a batch of model parameters, the trained model parameters (e.g., gradients) are sent to the aggregation server.

[0169] Step F7: Collect the gradients uploaded by all participating nodes, perform weighted summation and then average to obtain the global gradient.

[0170] For example, the gradient parameters of the participants are aggregated by weighted summation and then averaging of the gradients of each participant. , This represents the first model parameter of the local computing power scheduling model in the s-th iteration. After receiving the gradient transmitted by the participants, the aggregation server calculates the gradient according to the pre-determined weight coefficients. gradient After weighted summation and averaging, the model parameter fusion result is obtained. The global model parameters are then updated based on the model parameter fusion result to achieve global optimization of the computing power scheduling model.

[0171] Furthermore, after global optimization, the updated global model parameters (i.e., the model parameter fusion result) are distributed to each participating node for each participating node to perform the next round of local iterative processing.

[0172] Step F8 involves predicting and scheduling computing resources based on the global model, and dynamically adjusting resource allocation.

[0173] Dynamic resource allocation adjustment includes migrating tasks and adjusting resource ratios; by continuously collecting the latest data and system status, iteratively updating global model parameters, efficient and flexible resource scheduling is achieved.

[0174] Furthermore, when a scheduled task arrives, it utilizes the computing resources of n clusters (i.e., compute nodes) under the hybrid cloud architecture, along with the model parameters trained through federated learning, to output a... The matrix is ​​used to select the cluster that will be last assigned to the scheduler by obtaining the maximum value of the matrix.

[0175] refer to Figure 7 This document illustrates a federated learning-based computing power scheduling model training device 40 provided in an embodiment of this application. The device is applied to a processing node in a computing power scheduling system, which includes multiple computing nodes. The device 40 includes: a first acquisition module 401, configured to perform multiple training operations to acquire a global computing power scheduling model; each training operation includes: acquiring node feature samples of the computing node and computing power routes of the computing node; the node feature samples include coordinate samples and computing power resource samples of the computing node; a second acquisition module 402, configured to use the node feature samples and the computing power routes to train a local computing power scheduling model corresponding to the computing node, acquiring first model parameters of the trained local computing power scheduling model; and a third acquisition module 403, configured to fuse the first model parameters of all trained local computing power scheduling models to acquire second model parameters of the global computing power scheduling model corresponding to the computing power scheduling system.

[0176] Optionally, the processing node includes a local processing node corresponding to the computing node; the computing power scheduling model training device 40 further includes: an eighth acquisition module, used to obtain the second model parameters of the global computing power scheduling model corresponding to the computing power scheduling system through the local processing node after fusing the first model parameters corresponding to all computing nodes, and then replace the first model parameters of the local computing power scheduling model with the model parameter fusion value; the model parameter fusion value is obtained by fusing the first model parameters corresponding to all computing nodes; and a determination module, used to determine the local computing power scheduling model after the model parameters are replaced as the local computing power scheduling model corresponding to the computing node in the next iteration.

[0177] Optionally, the first acquisition module 401 includes: a first acquisition submodule, used to acquire a topology information sample between the computing nodes through the local processing node; the topology information sample includes bandwidth and network latency; a second acquisition submodule, used to acquire a weight sample between the computing nodes based on the topology information sample; and a third acquisition submodule, used to acquire the computing power route of the computing nodes based on the weight sample and the coordinate sample using the Dixtra algorithm.

[0178] Optionally, the second acquisition submodule includes: a fourth acquisition submodule, used to acquire the weight between computing node i and computing node j through the local processing node according to the following method. :

[0179] in, The bandwidth between compute node i and compute node j, The network latency between compute node i and compute node j These are hyperparameters used to adjust the ratio of bandwidth to network latency.

[0180] Optionally, the first acquisition submodule includes: a fifth acquisition submodule, used to acquire a physical topology map between the computing nodes through the local processing node using an open shortest path first routing algorithm; the nodes in the physical topology map are computing nodes, and the edges represent the connection relationships between computing nodes; and a sixth acquisition submodule, used to acquire topology information between the computing nodes by combining the connection relationships between the computing nodes and the computing power resource samples of the computing nodes.

[0181] refer to Figure 8 This document illustrates a federated learning-based computing power scheduling device 50 provided in an embodiment of this application, applied to a processing node of a computing power scheduling system, the computing power scheduling system including multiple computing nodes; the device 50 includes: a fourth acquisition module 501, used to acquire node characteristics of the computing node when a task to be processed is received; the node characteristics include the coordinates and computing power resources of the computing node; a fifth acquisition module 502, used to input the node characteristics into a global computing power scheduling model to obtain a target computing power route; the global computing power scheduling model is obtained by the method described in any one of the above embodiments; and a sixth acquisition module 503, used to perform computing power scheduling on the task to be processed according to the target computing power route.

[0182] For example, the computing power scheduling device 50 further includes: an update module, used to update the node feature samples of computing nodes after scheduling computing power for the task to be processed according to the target computing power route; the node feature samples before the update are the node feature samples described in any of the foregoing embodiments; and a ninth acquisition module, used to acquire an updated global computing power scheduling model according to the updated node feature samples, so as to schedule computing power for the next task to be processed according to the updated global computing power scheduling model.

[0183] According to Figure 9The electronic device 600 may include one or more of the following components: processing component 602, memory 604, power supply component 606, multimedia component 608, audio component 610, input / output (I / O) interface 612, sensor component 614, and communication component 616.

[0184] Processing component 602 typically controls the overall operation of electronic device 600, such as operations associated with display, telephone calls, data communication, camera operation, and recording operations. Processing component 602 may include one or more processors 620 to execute instructions to perform all or part of the steps of the methods described above. Furthermore, processing component 602 may include one or more modules to facilitate interaction between processing component 602 and other components. For example, processing component 602 may include a multimedia module to facilitate interaction between multimedia component 608 and processing component 602.

[0185] Memory 604 is used to store various types of data to support the operation of electronic device 600. Examples of this data include instructions for any application or method operating on electronic device 600, contact data, phonebook data, messages, pictures, multimedia, etc. Memory 604 can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0186] Power supply component 606 provides power to various components of electronic device 600. Power supply component 606 may include a power management system, one or more power supplies, and other components associated with generating, managing, and distributing power to electronic device 600.

[0187] Multimedia component 608 includes an interface that provides an output interface between electronic device 600 and user. In some embodiments, the interface may include a liquid crystal display (LCD) and a touch panel (TP). If the interface includes a touch panel, the interface may be implemented as a touchscreen to receive input signals from the user. The touch panel includes one or more touch sensors to sense touches, swipes, and gestures on the touch panel. The touch sensors may not only sense the boundaries of touch or swipe actions but also detect the duration and pressure associated with the touch or swipe operation. In some embodiments, multimedia component 608 includes a front-facing camera and / or a rear-facing camera. When electronic device 600 is in an operating mode, such as shooting mode or multimedia mode, the front-facing camera and / or rear-facing camera may receive external multimedia data. Each front-facing camera and rear-facing camera may be a fixed optical lens system or have focal length and optical zoom capabilities.

[0188] Audio component 610 is used to output and / or input audio signals. For example, audio component 610 includes a microphone (MIC) used to receive external audio signals when electronic device 600 is in an operating mode, such as call mode, recording mode, and voice recognition mode. The received audio signals may be further stored in memory 604 or transmitted via communication component 616. In some embodiments, audio component 610 also includes a speaker for outputting audio signals.

[0189] Input / output (I / O) interface 612 provides an interface between processing component 602 and peripheral interface modules, such as keyboards, click wheels, buttons, etc. These buttons may include, but are not limited to, home buttons, volume buttons, power buttons, and lock buttons.

[0190] Sensor assembly 614 includes one or more sensors for providing state assessments of various aspects of electronic device 600. For example, sensor assembly 614 may detect the on / off state of electronic device 600, the relative positioning of components such as the display and keypad of electronic device 600, changes in position of electronic device 600 or a component of electronic device 600, the presence or absence of user contact with electronic device 600, orientation or acceleration / deceleration of electronic device 600, and temperature changes of electronic device 600. Sensor assembly 614 may include a proximity sensor configured to detect the presence of nearby objects without any physical contact. Sensor assembly 614 may also include a light sensor, such as a CMOS or CCD image sensor, for use in imaging applications. In some embodiments, sensor assembly 614 may also include an accelerometer, gyroscope, magnetometer, pressure sensor, or temperature sensor.

[0191] Communication component 616 facilitates wired or wireless communication between electronic device 600 and other devices. Electronic device 600 can access wireless networks based on communication standards, such as WiFi, carrier networks (such as 2G, 3G, 4G, or 5G), or combinations thereof. In one exemplary embodiment, communication component 616 receives broadcast signals or broadcast-related information from an external broadcast management system via a broadcast channel. In one exemplary embodiment, communication component 616 also includes a near-field communication (NFC) module to facilitate short-range communication. For example, the NFC module may be implemented based on radio frequency identification (RFID) technology, Infrared Data Association (IrDA) technology, ultra-wideband (UWB) technology, Bluetooth (BT) technology, and other technologies.

[0192] In an exemplary embodiment, the electronic device 600 may be implemented by one or more application-specific integrated circuits (ASICs), digital signal processors (DSPs), digital signal processing devices (DSPDs), programmable logic devices (PLDs), field-programmable gate arrays (FPGAs), controllers, microcontrollers, microprocessors, or other electronic components to implement a federated learning-based computing power scheduling model training method and / or a federated learning-based computing power scheduling method provided in the embodiments of this application.

[0193] In an exemplary embodiment, a non-transitory computer-readable storage medium including instructions is also provided, such as a memory 604 including instructions, which can be executed by a processor 620 of an electronic device 600 to perform the above-described method. For example, the non-transitory storage medium may be a ROM, random access memory (RAM), CD-ROM, magnetic tape, floppy disk, and optical data storage device, etc.

[0194] Figure 10 This is a block diagram of an electronic device 700 according to another embodiment of the present invention. For example, the electronic device 700 may be provided as a server. See also... Figure 10 The electronic device 700 includes a processing component 722, which further includes one or more processors, and memory resources represented by a memory 732 for storing instructions executable by the processing component 722, such as application programs. The application programs stored in the memory 732 may include one or more modules, each corresponding to a set of instructions. Furthermore, the processing component 722 is configured to execute instructions to perform a federated learning-based computing power scheduling model training method and / or a federated learning-based computing power scheduling method provided in embodiments of this application.

[0195] Electronic device 700 may also include a power supply component 726 configured to perform power management of electronic device 700, a wired or wireless network interface 750 configured to connect electronic device 700 to a network, and an input / output (I / O) interface 758. Electronic device 700 may operate on an operating system stored in memory 732, such as Windows Server™, Mac OS X™, Unix™, Linux™, FreeBSD™, or similar.

[0196] Other embodiments of this application will readily occur to those skilled in the art upon consideration of the specification and practice of the application disclosed herein. This application is intended to cover any variations, uses, or adaptations of this application that follow the general principles of this application and include common knowledge or customary techniques in the art not disclosed herein. The specification and examples are to be considered exemplary only, and the true scope and spirit of this application are indicated by the following claims.

[0197] It should be understood that this application is not limited to the precise structure described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The scope of this application is limited only by the appended claims.

Claims

1. A method for training a computing power scheduling model, characterized in that, A processing node applied to a computing power scheduling system, the computing power scheduling system comprising multiple computing nodes; the method includes: performing multiple training operations to obtain a global computing power scheduling model; each training operation includes: Obtain node feature samples of computing nodes and computing power routes of computing nodes; the node feature samples include coordinate samples and computing power resource samples of the computing nodes; Using the node feature samples and the computing power routing, train the local computing power scheduling model corresponding to the computing node, and obtain the first model parameters of the trained local computing power scheduling model; By integrating the first model parameters of all trained local computing power scheduling models, the second model parameters of the global computing power scheduling model corresponding to the computing power scheduling system are obtained.

2. The method according to claim 1, characterized in that, The processing node includes a local processing node corresponding to the computing node; after fusing the first model parameters of all trained local computing power scheduling models to obtain the second model parameters of the global computing power scheduling model corresponding to the computing power scheduling system, the method further includes: The local processing node uses the fused model parameter value to replace the first model parameter of the local computing power scheduling model; the fused model parameter value is obtained by fusing the first model parameter corresponding to all computing nodes. The local computing power scheduling model after replacing the model parameters is determined as the local computing power scheduling model corresponding to the computing node in the next iteration.

3. The method according to claim 1, characterized in that, The processing node includes a local processing node corresponding to each of the computing nodes; obtaining the computing power route for the computing node includes: The local processing node obtains a topology information sample between the computing nodes; the topology information sample includes bandwidth and network latency. Based on the topology information sample, obtain the weight sample between the computing nodes; Based on the weight samples and the coordinate samples, the computing power route of the computing nodes is obtained through the Dixtra algorithm.

4. The method according to claim 3, characterized in that, The step of obtaining weight samples between computing nodes based on the topology information sample includes: Through the local processing node, the weight samples between computing node i and computing node j are obtained according to the following method. : in, The bandwidth between compute node i and compute node j, The network latency between compute node i and compute node j These are hyperparameters used to adjust the ratio of bandwidth to network latency.

5. The method according to claim 3, characterized in that, The step of obtaining the topology information sample between the computing nodes includes: Using the local processing nodes, a physical topology graph between the computing nodes is obtained according to the open shortest path first routing algorithm; the nodes in the physical topology graph are the computing nodes, and the edges represent the connection relationships between the computing nodes. By combining the connection relationships between the computing nodes and the computing power resource samples of the computing nodes, the topological structure information between the computing nodes is obtained.

6. A computing power scheduling method, characterized in that, A processing node applied to a computing power scheduling system, wherein the computing power scheduling system includes multiple computing nodes; the method includes: When a task to be processed is received, the node characteristics of the computing node are obtained; the node characteristics include the coordinates and computing resources of the computing node. The node features are input into the global computing power scheduling model to obtain the target computing power route; the global computing power scheduling model is obtained by the method described in any one of claims 1 to 5. The computing power is scheduled for the task to be processed based on the target computing power route.

7. The method according to claim 6, characterized in that, After scheduling computing power for the task to be processed according to the target computing power route, the method further includes: Update the node feature samples of the computing nodes; the node feature samples before the update include the node feature samples of any one of claims 1 to 5; Based on the updated node feature samples, an updated global computing power scheduling model is obtained, and computing power is scheduled for the next task to be processed according to the updated global computing power scheduling model.

8. A computing power scheduling model training device, characterized in that, A processing node used in a computing power scheduling system, wherein the computing power scheduling system includes multiple computing nodes, the device comprising: The first acquisition module is used to perform multiple training operations to acquire a global computing power scheduling model; each training operation includes: acquiring node feature samples of computing nodes and computing power routes of computing nodes; the node feature samples include coordinate samples and computing power resource samples of the computing nodes; The second acquisition module is used to train a local computing power scheduling model corresponding to the computing node using the node feature samples and the computing power routing, and to acquire the first model parameters of the trained local computing power scheduling model. The third acquisition module is used to fuse the first model parameters of all trained local computing power scheduling models to obtain the second model parameters of the global computing power scheduling model corresponding to the computing power scheduling system.

9. A computing power scheduling device, characterized in that, A processing node used in a computing power scheduling system, wherein the computing power scheduling system includes multiple computing nodes; the device includes: The fourth acquisition module is used to acquire the node characteristics of the computing node when a task to be processed is received; the node characteristics include the coordinates and computing resources of the computing node; The fifth acquisition module is used to input the node features into the global computing power scheduling model to obtain the target computing power route; the global computing power scheduling model is obtained by the method described in any one of claims 1 to 5; The sixth acquisition module is used to perform computing power scheduling on the task to be processed according to the target computing power route.

10. An electronic device, characterized in that, It includes a processor and a memory, the memory storing a program or instructions that can run on the processor, the program or instructions being executed by the processor to implement the steps of the method as described in any one of claims 1 to 7.

Images