Graph-based computing power network federated learning job scheduling method and device, electronic equipment and storage medium

Abstract

This invention provides a graph-based method, apparatus, electronic device, and storage medium for scheduling federated learning jobs in a computing power network, relating to the field of computing power network technology. The method includes inputting a computational view, a network view, a model staleness view, and jobs to be assigned into a scheduling model, and obtaining the scheduling strategy for the jobs to be assigned in the computing power network output by the scheduling model. This invention achieves a unified structured representation of multiple federated learning jobs and the relationships between devices through the computational view, network view, and model staleness view. By generating a fused feature vector, redundant information and noise in the computing power network can be effectively suppressed, which is beneficial to improving the accuracy of the scheduling strategy. The generation of the fused feature vector enables accurate perception of the computational state, network state, and model staleness state of the computing power network, thereby improving the accuracy of the scheduling strategy.

Description

Technical Field

[0001] This invention relates to the field of computing power network technology, and in particular to a graph-based computing power network federated learning job scheduling method, apparatus, electronic device and storage medium. Background Technology

[0002] In Computing Power Network (CPN) environments, asynchronous federated learning (AFL) is a promising distributed training solution to address the dual challenges of heterogeneous device resources and data privacy protection. AFL allows devices with varying performance to participate in training and upload updates at their own pace, effectively improving training efficiency in heterogeneous environments. However, when a CPN needs to concurrently process massive federated learning jobs, resource competition between jobs not only slows down their individual training progress and causes progress divergence, but also leads to severe model obsolescence. Gradients calculated by lagging nodes based on outdated models introduce significant biases during aggregation, interfering with the optimization direction of the global model, resulting in slow convergence and wasting valuable computing resources. This phenomenon is further exacerbated in scenarios where data is not independent and identically distributed, potentially triggering a vicious cycle at the system level: slow-converging jobs continue to consume resources due to longer training times, further intensifying resource competition and model obsolescence issues in other jobs, ultimately creating a cascading delay effect throughout the system.

[0003] Existing methods for scheduling federated learning jobs in computing networks include cloud-edge-device collaborative methods. These methods segment the global model and cluster it based on the KL divergence of client data distribution, thereby offloading some training tasks from edge devices to edge nodes. This coordinates the computing resources of the cloud, edge, and device to accelerate federated learning training. However, this method is a static scheduling strategy. Its decisions are based on offline, long-term historical states, making it unable to respond to instantaneous system changes and difficult to effectively manage real-time resource contention and cross-interference when multiple jobs are executed concurrently.

[0004] In summary, the existing federated learning job scheduling in computing power networks is inaccurate. Summary of the Invention

[0005] This invention provides a graph-based method, apparatus, electronic device, and storage medium for scheduling federated learning jobs in computing power networks, in order to address the inaccuracy of prior art in scheduling federated learning jobs in computing power networks and improve the accuracy of scheduling federated learning jobs in computing power networks.

[0006] This invention provides a graph-based federated learning job scheduling method for computing power networks, comprising: Input the computation view, network view, model obsolescence view, and jobs to be assigned into the scheduling model to obtain the scheduling strategy of the jobs to be assigned in the computing network output by the scheduling model. The scheduling model extracts features from the computation view, network view, model staleness view, and jobs to be assigned, obtaining multiple target feature vectors. Based on the context information of the computing network, the multiple target feature vectors are masked and weighted to generate a fused feature vector. Based on the fused feature vector, the allocation results of the jobs to be assigned in the computing network are obtained, and the scheduling strategy is derived. The jobs to be assigned include federated learning jobs.

[0007] According to the graph-based federated learning job scheduling method for computing power networks provided by the present invention, multiple target feature vectors include computational feature vectors, network feature vectors, and model staleness feature vectors, which are obtained in the following manner: The computation status of each device in the computation view is extracted in combination with the job to be assigned to obtain the computation status features of each device; based on all computation status features, a computation feature vector is obtained. In conjunction with the jobs to be assigned, feature extraction is performed on the network status of each communication link in the network view to obtain the network status features of each communication link; based on all network status features, a network feature vector is obtained. Feature extraction is performed on each local model of the job to be assigned in the model obsolescence view to obtain the model obsolescence status features of each job to be assigned; based on all model obsolescence status features, the model obsolescence feature vector is obtained.

[0008] According to the graph-based federated learning job scheduling method for computing power networks provided by this invention, the fused feature vector is generated in the following manner: The weights of each target feature vector and the retention rate of the control feature sparsity of each target feature vector are determined based on context information. Based on the importance scores of each dimension of the target feature vector, generate a binary mask for each dimension of the feature vector. Based on the retention rate of each target feature vector and the binarization mask of each target feature vector, the target feature vector is subjected to sparsification and masking to obtain each mask sparse feature vector. Normalize the sparse feature vectors of each mask to obtain the normalized feature vectors. The fused feature vector is obtained by weighted summation of each normalized feature vector based on the weights.

[0009] According to the graph-based federated learning job scheduling method for computing power networks provided by the present invention, the computing state characteristics include the remaining allocable computing resources of the device, the total job completion time of the device, and the computing time of a single job to be assigned. Network status characteristics include the remaining available bandwidth of communication links, the total link cleanup time of communication links, and the transmission delay of a single job to be assigned; The model obsolescence status features include the obsolescence of each local model in the job to be assigned, the maximum model obsolescence, and the model obsolescence mitigation speed.

[0010] According to the graph-based federated learning job scheduling method for computing power networks provided by this invention, the scheduling model is trained in the following manner: Training samples are obtained from the sample computation view, sample network view, sample model staleness view, and sample pending job assignment based on the computing power network. The training samples are input into a pre-defined model for training. The pre-defined model includes an execution network and an evaluation network. A predictive scheduling strategy based on the execution network to determine the current time slot of the training samples; The current reward score and the value prediction of the next time slot are determined based on the evaluation network to determine the predictive scheduling strategy for the current time slot. Based on the current reward score, the value prediction of the next time slot, and the value prediction of the current time slot, determine the gain factor of the prediction scheduling strategy for the current time slot; The parameters of the execution network are updated based on the gain factor until the preset model converges, thus obtaining the scheduling model.

[0011] According to the graph-based federated learning job scheduling method for computing power networks provided by this invention, the current reward score is determined as follows: By combining the sample calculation view with the sample to be assigned jobs, the sample calculation status features of each device are obtained. By combining the sample network view with the sample jobs to be assigned, the sample network state features of each communication link are obtained. By combining the sample to be assigned jobs, feature extraction is performed on the sample model antiquity view to obtain the sample model antiquity status features of each local model of the sample to be assigned jobs. The current reward score is calculated based on the state characteristics of each sample, the network state characteristics of each sample, and the model obsolescence state characteristics of each sample.

[0012] According to the graph-based federated learning job scheduling method for computing power networks provided by the present invention, the preset model further includes a graph encoder, and the value prediction for the next time slot and the value prediction for the current time slot are determined based on the following method: Based on the graph encoder, features are extracted from the sample computation view, sample network view, and sample model staleness view to obtain sample computation feature vectors, sample network feature vectors, and sample model staleness feature vectors. Input the sample computation feature vector, the sample network feature vector, and the sample model staleness feature vector into the evaluation network to obtain the value prediction of the current time slot; The sample computation view, sample network view, and sample model staleness view are updated based on the predictive scheduling strategy of the current time slot. Based on the graph encoder, features are extracted from the updated sample computation view, the updated sample network view, and the updated sample model staleness view to obtain the updated sample computation feature vector, the updated sample network feature vector, and the updated sample model staleness feature vector. The updated sample computation feature vector, the updated sample network feature vector, and the updated sample model staleness feature vector are input into the evaluation network to obtain the value prediction for the next time slot.

[0013] The present invention also provides a graph-based computing power network federated learning job scheduling device, comprising: The scheduling module is used to input the computation view, network view, model staleness view, and jobs to be assigned from the computing power network into the scheduling model, and obtain the scheduling strategy for the jobs to be assigned in the computing power network output by the scheduling model. Specifically, the scheduling model extracts features from the computation view, network view, model staleness view, and jobs to be assigned to obtain multiple target feature vectors. Based on the context information of the computing power network, the multiple target feature vectors are masked and weighted to generate a fused feature vector. Based on the fused feature vector, the allocation result of the jobs to be assigned in the computing power network is obtained, and the scheduling strategy is derived. The jobs to be assigned include federated learning jobs.

[0014] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements any of the graph-based computing power network federated learning job scheduling methods described above.

[0015] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements any of the graph-based computing power network federated learning job scheduling methods described above.

[0016] This invention provides a graph-based method, apparatus, electronic device, and storage medium for scheduling federated learning jobs in a computing power network. By inputting the computational view, network view, model staleness view, and jobs to be assigned into a scheduling model, the scheduling model outputs a scheduling strategy for the jobs to be assigned within the computing power network. Specifically, the scheduling model extracts features from the computational view, network view, model staleness view, and jobs to be assigned, obtaining multiple target feature vectors. Based on the context information of the computing power network, these target feature vectors are masked and weighted to generate a fused feature vector. The allocation result of the jobs to be assigned in the computing power network is obtained based on the fused feature vector, resulting in the scheduling strategy. The jobs to be assigned include federated learning jobs. This invention achieves a unified structured representation of multiple federated learning jobs and the relationships between devices through the computational view, network view, and model staleness view. By extracting multiple target feature vectors and masking and weighting them to generate a fused feature vector, redundant information and noise in the computing power network can be effectively suppressed, which is beneficial to improving the accuracy of the scheduling strategy. This invention achieves accurate perception of the computational state, network state, and model obsolescence of the computing power network by generating fused feature vectors, thereby improving the accuracy of scheduling strategies. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0018] Figure 1 This is one of the flowcharts illustrating the graph-based federated learning job scheduling method for computing power networks provided by this invention.

[0019] Figure 2 This is a schematic diagram of the training process of the scheduling model provided by the present invention.

[0020] Figure 3 This is the second flowchart of the graph-based computing power network federated learning job scheduling method provided by the present invention.

[0021] Figure 4 This is a schematic diagram of the structure of the graph-based computing power network federated learning job scheduling device provided by the present invention.

[0022] Figure 5 This is a schematic diagram of the structure of the electronic device provided by the present invention. Detailed Implementation

[0023] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0024] The embodiments of the present invention are applicable to typical edge-cloud collaborative federated learning environments, and can be specifically applied to the following scenarios.

[0025] Intelligent manufacturing monitoring platform. Each production line terminal device acts as a local node to train the equipment status prediction model, edge servers are responsible for regional aggregation, and the cloud center is responsible for global model maintenance. Through the graph-based computing power network federated learning job scheduling method of this invention, computing power and network resources can be dynamically allocated to ensure that the models of each terminal are updated synchronously, thereby improving prediction accuracy and response speed.

[0026] A smart city traffic management system. Each traffic camera terminal trains a target detection model locally, and edge nodes are aggregated in different regions. Due to uneven task load caused by differences in traffic flow in different areas, the graph-based federated learning job scheduling method of this invention can prioritize the scheduling of low-load nodes to participate in aggregation based on the joint analysis of network view and computation view, thereby improving the real-time performance and stability of the global model.

[0027] A collaborative medical image diagnostic system. Multiple hospitals' local servers jointly train a medical image recognition model while protecting privacy. By dynamically monitoring the update latency of each node through stale views, the system can automatically adjust its training frequency or computing power allocation when lagging nodes appear, ensuring the reliability and consistency of the aggregated medical image recognition model.

[0028] The following is combined Figures 1 to 5 This invention describes a graph-based computing power network federated learning job scheduling method, apparatus, and electronic device.

[0029] Figure 1 This is one of the flowcharts illustrating the graph-based federated learning job scheduling method for computing power networks provided by this invention, such as... Figure 1 As shown, the graph-based computing power network federated learning job scheduling method includes step S100.

[0030] S100: Input the computation view, network view, model staleness view, and jobs to be assigned into the scheduling model to obtain the scheduling strategy of the jobs to be assigned in the computing network output by the scheduling model.

[0031] The scheduling model extracts features from the computation view, network view, model staleness view, and jobs to be assigned, obtaining multiple target feature vectors. Based on the context information of the computing network, the multiple target feature vectors are masked and weighted to generate a fused feature vector. Based on the fused feature vector, the allocation results of the jobs to be assigned in the computing network are obtained, and the scheduling strategy is derived. The jobs to be assigned include federated learning jobs.

[0032] The computation view represents the computation status of devices in the current time-slot computing network. The computation view is constructed as a bipartite graph between federated learning jobs (job nodes) and devices (device nodes), aiming to quantify the competition for computing resources. The computation view contains two types of nodes: job nodes and device nodes. Device nodes are labeled with their remaining capacity and task queue attributes, while the edges between job nodes and device nodes carry the computational requirements and expected waiting times of the tasks. To maintain the sparsity of the graph and highlight the main competitive relationships, this invention performs edge pruning by retaining only the top L most feasible device connections for each job node. This design allows device-level resource competition relationships to be directly captured through the message passing mechanism of graph neural networks.

[0033] The network view represents the network transmission status of jobs in the current time slot of the computing network. The network view consists of a directed graph whose structure is consistent with the physical communication topology of the computing network. Nodes represent devices, edge gateways, and cloud servers, and edges represent physical links, with attributes such as bandwidth, backlog, and latency. The transmission path of each job is characterized by activating the corresponding edge sequence, with end-to-end latency as a path-level feature associated with the source node. To improve scalability, only active links and the top q most congested edges are retained in the network view. This allows the graph neural network to propagate congestion states across multi-hop paths, effectively capturing link-level interference.

[0034] The model staleness view focuses on the cascading effect of local model staleness in asynchronous federated learning jobs on a computing network simulating the current time slot. The nodes in the model staleness view include job nodes and device nodes. The model staleness view characterizes the staleness status of the job's local model on each device node.

[0035] The jobs to be assigned include federated learning jobs that need to be scheduled and assigned.

[0036] The scheduling model inputs the computation view, network view, model staleness view, and jobs to be assigned from the current time slot's computing power network into the scheduling model, and obtains the scheduling strategy for the jobs to be assigned in the computing power network, output by the scheduling model. The scheduling model includes a graph encoder, a hybrid view router, and an execution network. The graph encoder includes a computation view encoder, a network view encoder, and a model staleness view encoder.

[0037] The scheduling model extracts features from the computational view, network view, and model stagnation view to obtain multiple target feature vectors. The computational view is encoded using a computational view encoder to obtain computational feature vectors. The network view is encoded using a network view encoder to obtain network feature vectors. The model stagnation view is encoded using a model stagnation view encoder to obtain model stagnation feature vectors. These computational feature vectors, network feature vectors, and model stagnation feature vectors constitute multiple target feature vectors.

[0038] Multiple target feature vectors are input into the hybrid view router. The hybrid view router extracts the overall computing resource status, overall network resource status, and overall model staleness status of the computing power network to obtain context information. Based on the context information, the hybrid view router performs masking and weighted fusion of the multiple target feature vectors to generate a fused feature vector from a global perspective. The execution network generates a job-device affinity score matrix for the current time slot based on the fused feature vector and the jobs to be assigned. ,in, The number of jobs to be assigned. This represents the number of devices. The job-device affinity score matrix for the current time slot includes multiple job-device affinity scores for the current time slot. The job-device affinity score of each job to be assigned is normalized to obtain the scheduling probability of each job in the computing network.

[0039] ; in, For the first The pending jobs were scheduled to the first The scheduling probability of each device. To fuse feature vectors, For the current time slot The The number of pending tasks and the first The device's affinity score For the number of devices, For the first The scheduling action of a job to be assigned.

[0040] Based on the maximum scheduling probability of each job to be assigned, determine the matching scheduling device for each job. Based on the matching scheduling devices for all jobs to be assigned, derive the scheduling strategy.

[0041] The graph-based federated learning job scheduling method for computing power networks provided in this invention involves inputting the computation view, network view, model staleness view, and jobs to be assigned into a scheduling model to obtain the scheduling strategy for the jobs to be assigned in the computing power network, output by the scheduling model. Specifically, the scheduling model extracts features from the computation view, network view, model staleness view, and jobs to be assigned, obtaining multiple target feature vectors. Based on the context information of the computing power network, the multiple target feature vectors are masked and weighted to generate a fused feature vector. The allocation result of the jobs to be assigned in the computing power network is obtained based on the fused feature vector, resulting in the scheduling strategy. The jobs to be assigned include federated learning jobs. This invention achieves a unified structured representation of multiple federated learning jobs and the relationships between devices through the computation view, network view, and model staleness view. By extracting multiple target feature vectors and masking and weighting them to generate a fused feature vector, redundant information and noise in the computing power network can be effectively suppressed, which is beneficial to improving the accuracy of the scheduling strategy. This invention achieves accurate perception of the computational state, network state, and model obsolescence of the computing power network by generating fused feature vectors, thereby improving the accuracy of scheduling strategies.

[0042] like Figure 3 As shown, the execution entity of the graph-based federated learning job scheduling method for computing power networks of the present invention includes a scheduling system. The scheduling system includes a resource layer, a control layer, and an orchestration layer.

[0043] The resource layer consists of terminal devices, edge servers, and cloud data centers, providing heterogeneous computing, communication, and storage resources. Terminal devices are used to perform local model training; edge servers are responsible for local model aggregation and caching; and the cloud data center maintains the global model and coordinates the overall job progress. The resource layer is interconnected through the Computing Network Network (CPN) underlying architecture to achieve the collaborative use of multi-layered computing resources.

[0044] The control layer is deployed at the regional control nodes of the computing network. Based on the Software Defined Network (SDN) / Network Functions Virtualization (NFV) control plane, it realizes resource monitoring and command issuance. This layer periodically collects data such as device computing utilization (reflecting computing resource contention), bandwidth utilization (reflecting network resource contention), and task status information (reflecting job obsolescence interference), and reports the monitoring results (including computing status, network status, and obsolescence status of the local model) to the service layer.

[0045] The service layer, located in the cloud, is responsible for integrating real-time data from the control layer to build a unified resource view. Through data interfaces, the service layer generates three types of views: a computation view, a network view, and a model staleness view. These views reflect the device's computing power usage, network transmission load, and the differences in local model updates for each job within the computing network, respectively, providing input features for the scheduler.

[0046] The orchestration layer is the global decision-making center of the scheduling system. It includes the scheduling model, which comprises a graph encoder, a hybrid view router (MoV-R), and an execution network (Actor). The orchestration layer takes the computational view, network view, and model stagnation view provided by the service layer as input, outputs a dynamic job-device matching scheme (scheduling policy), and sends the scheduling policy to the control layer for execution, achieving efficient scheduling of multi-task federated learning jobs.

[0047] Through the above-mentioned hierarchical division, this invention realizes closed-loop management of the entire process from resource monitoring and status modeling to intelligent decision-making.

[0048] The graph-based computing power network federated learning job scheduling method of the present invention, based on the above scheduling system, mainly includes the following steps.

[0049] Step 1: State Collection and View Construction. The control layer collects resource usage information from each node in the computing network in real time, including CPU / GPU utilization, available bandwidth, link latency, task queue length, and local model update time. The service layer receives the real-time data output by the control layer and constructs the computing view, network view, and model stagnation view. Figure 3 The graph structures are used to describe the relationships of competition for computing resources, competition for network bandwidth, and propagation of model obsolescence.

[0050] Step S2: Multi-view graph encoding. The graph encoder of the scheduling model in the orchestration layer performs graph embedding calculations on the above three types of views and the jobs to be assigned. High-order features between nodes and edges in the computation view, network view, and model stale view are extracted by graph neural networks to obtain multiple target feature vectors, reflecting the current global operating status of the computing power network.

[0051] Step S3: Feature Fusion and Route Generation. MoV-R adaptively calculates the weights of each view (or each target feature vector) based on the context information of the computing network (including indicators such as average resource utilization, number of tasks, and maximum obsolescence), and performs sparsification on each target feature vector. After weighted fusion, a unified route representation (fused feature vector) is generated to describe the priority matching relationship between jobs and devices in a multi-dimensional state.

[0052] Step S4: Scheduling Decision Generation. The orchestration layer's actor network takes the fused feature vector as input and outputs a matching probability matrix between each job to be assigned and the candidate device. Based on the principle of maximizing probability, it generates a job allocation scheme (i.e., scheduling strategy) for the next time slot (or time slot).

[0053] Step S5: Task Execution and Feedback Update: The control layer allocates computing and communication resources according to the scheduling strategy and initiates the local training and model upload process. The cloud server distributes the global model to a predetermined subset of devices for local training and ultimately receives model updates uploaded by the devices. The execution results of each node are fed back to the service layer to update the view state and trigger the next round of scheduling decisions, achieving continuous adaptive optimization.

[0054] Through the above steps, the scheduling system can achieve comprehensive scheduling of computing power, bandwidth and latency in a multi-task concurrent execution environment, avoiding training delays and model obsolescence caused by resource conflicts.

[0055] Based on the above embodiments, the multiple target feature vectors include computed feature vectors, network feature vectors, and model staleness feature vectors, and the multiple target feature vectors are obtained in the following way: The computation status of each device in the computation view is extracted in combination with the job to be assigned to obtain the computation status features of each device; based on all computation status features, a computation feature vector is obtained. In conjunction with the jobs to be assigned, feature extraction is performed on the network status of each communication link in the network view to obtain the network status features of each communication link; based on all network status features, a network feature vector is obtained. Feature extraction is performed on each local model of the job to be assigned in the model obsolescence view to obtain the model obsolescence status features of each job to be assigned; based on all model obsolescence status features, the model obsolescence feature vector is obtained.

[0056] This invention includes a graph encoder. The graph encoder comprises a graph neural network that processes the view structures of the computational view, network view, and model stale view of this invention, and calculates node attributes and edge attributes through the message passing mechanism of the graph neural network.

[0057] The graph encoder encodes the computational view, network view, and model stagnation view separately, generating multiple corresponding target feature vectors. The message propagation mechanism varies depending on the graph structure: in the computational view, it's an exchange between jobs and devices; in the network view, it propagates along multi-hop links; and in the model stagnation view, it cascades along job-device-job paths. Nodes use an attention mechanism to weighted aggregate incoming messages to update their representations. The final output is a multi-dimensional computational feature vector. Network feature vectors and model staleness feature vectors ,in, Indicates the current time slot.

[0058] By combining the computational status of each device in the computation view with the jobs to be assigned, feature extraction is performed to obtain the computational status features of each device. The computational status features characterize the computational resource utilization of the device. Based on the computational status features of all devices, a computational feature vector of the computation view is obtained.

[0059] By combining the network status features of each communication link in the network view with the jobs to be assigned, the network status features of each communication link are obtained. These network status features characterize the network resource occupancy of the communication links. Based on the network status features of all communication links, the network feature vector of the network view is obtained.

[0060] Feature extraction is performed on each local model (located on the device or communication link) of the job to be assigned in the model staleness view to obtain the model staleness status features of each local model. Based on all model staleness status features, a model staleness feature vector is obtained.

[0061] This invention extracts the computing resource usage status of each device in a computing network by extracting the computing feature vector from the computing view. It also extracts the network resource usage status of each communication link in the network view by extracting the network feature vector from the network view. Finally, it extracts the staleness status of each local model in the model staleness view by extracting the model staleness feature vector from the model staleness view.

[0062] Based on the above embodiments, the computing status characteristics include the remaining allocable computing resources of the device, the total completion time of the device's jobs, and the computing time of a single job to be assigned. Network status characteristics include the remaining available bandwidth of communication links, the total link cleanup time of communication links, and the transmission delay of a single job to be assigned; The model obsolescence status features include the obsolescence of each local model in the job to be assigned, the maximum model obsolescence, and the model obsolescence mitigation speed.

[0063] First, the competition for computing resources among devices in the computing power network is quantified by considering the remaining allocable computing resources for each device, the total job completion time for each device, and the computation time of a single job awaiting assignment. The formulas for calculating the remaining allocable computing resources, the total job completion time for each device, and the computation time of a single job awaiting assignment are as follows.

[0064] ; in, For the current time slot The The remaining allocable computing resources of each device For the current time slot The Peak computing power of each device For the current time slot The The computing resources already used by each device. For the current time slot No. The number of pending jobs relative to the first The scheduling indication variables for each device (where, Taking 1 indicates the first The pending jobs were assigned to the first... One device, Taking 0 indicates the first The pending jobs were not assigned to the first... (one device) For the first The number of pending jobs is in the first The computing requirements of each device For the current time slot Assigned to the Total workload of each device For the current time slot The Total completion time for each device. For the current time slot The The device is assigned to the first The proportion of remaining computing resources for each pending job. , The smaller the value, the higher the value. The fewer remaining computing resources an unassigned job receives, the less computing resources it will have. For the current time slot The The number of pending jobs is in the first The computation time required for the first device to complete the task (i.e., the first...) (Calculation time for each pending job) This refers to the number of devices.

[0065] When multiple jobs share the same device, the remaining allocable computing resources are distributed among them, thus increasing the overall job completion time and the computation time of individual jobs due to competition for computing resources. Computational status characteristics include the remaining allocable computing resources of each device. Total completion time of each piece of equipment and the computation time of a single pending job. .

[0066] Secondly, the network resource competition of the computing power network is quantified by considering the remaining available bandwidth of each communication link, the total link cleanup time of each communication link, and the transmission delay of a single job to be assigned. The formulas for calculating the remaining available bandwidth of each communication link, the total link cleanup time of each communication link, and the transmission delay of a single job to be assigned are as follows.

[0067] ; in, For the current time slot Communication link The remaining available bandwidth, For the current time slot Communication link Peak bandwidth capacity, For the current time slot Communication link The bandwidth capacity already occupied by concurrent transmissions, For the current time slot The The pending tasks have been processed by the first... Each device communicates via a link Transmitted traffic, For the current time slot No. The number of pending jobs relative to the first Scheduling indicator variables for each device For the number of devices, The number of jobs to be assigned. For the current time slot Communication link The total traffic carried by the top For the current time slot Communication link Total link cleanup time, For the current time slot The Transmission latency of a single pending job (transmission latency of a single pending job). A collection of communication links. For the current time slot The Each job to be assigned passes through the communication link The amount of data transmitted.

[0068] When multiple jobs share the same communication link, the remaining bandwidth will be allocated among them. This leads to increased overall link clearing time and transmission delay of individual jobs due to network resource contention. These coupling effects collectively reflect the dynamic interference caused by network resource contention. Network state characteristics include the remaining available bandwidth of each communication link. Total link cleanup time for each communication link and the transmission delay of a single job to be assigned. .

[0069] Finally, the obsolescence status of each local model in the computing power network is quantified by considering the obsolescence of each local model, the maximum obsolescence, and the model obsolescence mitigation speed for each pending job. Local models include the models on the computing power network's devices used to process pending jobs. The formulas for calculating the obsolescence of each local model, the maximum obsolescence, and the model obsolescence mitigation speed are as follows.

[0070] ; in, For the current time slot The A global model of jobs to be assigned. For the current time slot The The number of pending jobs is in the first Local model for each device For the current time slot The The number of pending jobs is in the first The local model of the device is outdated, relatively small. Indicates the first Local models on individual devices can synchronize with the global model in a timely manner, typically when computing power is high or communication latency is low. Conversely, slow computing or network congestion will lead to delayed updates, exacerbating the obsolescence of the local model. For the number of devices, The number of jobs to be assigned. For the current time slot No. The number of pending jobs relative to the first Scheduling indicator variables for each device For the current time slot The The number of pending jobs is in the first Speed ​​of mitigating model obsolescence in local models of individual devices To maximize model obsolescence, For the first The sensitivity coefficient of each pending job to the aging of the local model. To update interval Before and after The first device The absolute reduction in the staleness of the local model for each pending job. This ensures that only the positive contribution of improving the obsolescence of the local model (i.e., reducing bias) is counted. If the obsolescence of the local model increases due to failure to update in a timely manner, this item is set to 0, indicating that the decision does not contribute to alleviating the obsolescence of the local model.

[0071] Model obsolescence status features include the obsolescence of each local model in the job to be assigned. Maximum model obsolescence And the speed of mitigating model obsolescence .

[0072] The model staleness status feature of this invention characterizes the staleness status of the local models of jobs to be assigned in the computing power network on each device. The network status feature of this invention characterizes the network resource status of the communication links in the computing power network. The computing status feature of this invention characterizes the computing resource status of the devices in the computing power network.

[0073] Based on the above embodiments, the fused feature vector is generated in the following manner: The weights of each target feature vector and the retention rate of the control feature sparsity of each target feature vector are determined based on context information. Based on the importance scores of each dimension of the target feature vector, generate a binary mask for each dimension of the feature vector. Based on the retention rate of each target feature vector and the binarization mask of each target feature vector, the target feature vector is subjected to sparsification and masking to obtain each mask sparse feature vector. Normalize the sparse feature vectors of each mask to obtain the normalized feature vectors. The fused feature vector is obtained by weighted summation of each normalized feature vector based on the weights.

[0074] In the multi-view framework, the computation view, network view, and model staleness view capture interference from different dimensions (such as computation contention, bandwidth contention, and staleness propagation). The importance of these views to scheduling decisions dynamically changes with the computing power and network state. If a static fusion strategy is adopted, the key views will not be highlighted due to fixed weights, and redundant information will be introduced. To address this, this invention designs a hybrid view router (MoV-R), which treats the encoders of each view as independent semantic experts and introduces contextual information to dynamically adjust the contribution weights and feature selections of each expert, thereby achieving adaptive feature fusion. The routing and sparsification process of the hybrid view router is as follows: like Figure 2 As shown, the routing weights are calculated by the gating network mapping contextual information to a set of target scores. These target scores are used to calculate two types of routing weights: the weights of each target feature vector that balance the view weights and the retention rate that controls feature sparsity. The specific calculation formulas are as follows.

[0075] ; in, For the current time slot The weights of each target feature vector, For the current time slot The weights of the eigenvectors are calculated. For the current time slot The weights of the network feature vectors, For the current time slot The weights of the model's stale eigenvectors. A set of target scores obtained based on contextual information. To calculate the target score corresponding to the feature vector, The target score corresponding to the network feature vector. The target score is the feature vector corresponding to the model's obsolescence. For the current time slot The retention rate of each target feature vector. To calculate the retention rate of the eigenvectors, The retention rate of network feature vectors. The retention rate of the model's stale feature vectors. Scaling factor For bias terms, To achieve the lowest retention rate, This is the activation function.

[0076] Based on the importance scores of each dimension of the target feature vector, a binary mask for each dimension of the feature vector is generated. The formula for calculating the binary mask for each dimension of the feature vector is as follows.

[0077] ; in, For the current time slot Target feature vector The Binarization mask for features in each dimension, For the current time slot Target feature vector The Importance scores of each dimension feature For the current time slot Target feature vector The scoring threshold of the dimensional features, The smoothing factor is used to scale the difference between the importance score and the scoring threshold during the feature selection process. This determines the slope of the activation function's transformation near the threshold. When When the value is small, the activation function tends to be a step function, causing the generated binarized mask to exhibit obvious binarization characteristics (tending to 0 or 1), thereby enhancing the removal of redundant features and generating a sparser view representation; when When the value is larger, the mapping process is smoother, which is beneficial for retaining more feature information in the early stages of training to allow for thorough exploration. For activation function, The total number of dimensional features.

[0078] like Figure 2 As shown, the retention rate of each target feature vector is used to analyze each target feature vector. Perform sparsification to obtain sparsified feature vectors. Based on the binarization mask of each target feature vector, each sparsed feature vector is... Perform masking to obtain the masked sparse feature vector. .in, For the mask sparse feature vector, For target feature vectors The binarization mask.

[0079] The sparse feature vectors of each mask are normalized to obtain normalized feature vectors. The sparse feature vectors of the mask are then processed... , and Projecting onto a unified space yields various normalized feature vectors. .in, For the current time slot The normalized eigenvectors, For the sparse feature vector of the mask Normalization processing.

[0080] The fused feature vector is obtained by weighted summation of the normalized feature vectors based on their respective weights. The fused feature vector represents the state of the computing network from a local perspective. The formula for calculating the fused feature vector is as follows.

[0081] ; in, For the current time slot The fused feature vector, For the current time slot The weights of each target feature vector, For the current time slot Calculate the eigenvector. For the current time slot The network feature vectors, For the current time slot The model's stale feature vectors For the current time slot The normalized eigenvectors, For the current time slot The normalized eigenvector corresponding to the calculated eigenvector. For the current time slot The normalized feature vector corresponding to the network feature vector, For the current time slot The normalized feature vector corresponding to the model's stale feature vector.

[0082] This invention determines the weights and retention rates of each target feature vector based on contextual information, allowing for a focus on key information and reducing errors caused by redundant information in each target feature vector. The normalized feature vectors are then weighted and summed according to their weights and retention rates to obtain a fused feature vector, achieving adaptive feature fusion.

[0083] Based on the above embodiments, the scheduling model is trained in the following manner: Training samples are obtained from the sample computation view, sample network view, sample model staleness view, and sample pending job assignment based on the computing power network. The training samples are input into a pre-defined model for training. The pre-defined model includes an execution network and an evaluation network. A predictive scheduling strategy based on the execution network to determine the current time slot of the training samples; The current reward score and the value prediction of the next time slot are determined based on the evaluation network to determine the predictive scheduling strategy for the current time slot. Based on the current reward score, the value prediction of the next time slot, and the value prediction of the current time slot, determine the gain factor of the prediction scheduling strategy for the current time slot; The parameters of the execution network are updated based on the gain factor until the preset model converges, thus obtaining the scheduling model.

[0084] A pre-defined model is constructed based on the graph encoder, hybrid view router, execution network, and evaluation network.

[0085] like Figure 2 As shown, training samples are input into a preset model. The graph encoder of the preset model, combined with the sample assignment task, extracts features from the sample computation view to obtain the sample computation state features of each device, and then obtains the sample computation feature vector. The graph encoder combines the sample network view with the sample assignment task to extract features, obtain the sample network state features of each communication link, and then obtain the sample network feature vector. The graph encoder combines the sample job to be assigned with the feature extraction of the sample model staleness view to obtain the sample model staleness status features of each local model of the sample job to be assigned, and then obtains the sample model staleness feature vector. .

[0086] The hybrid view router performs masking and weighted fusion of sample computation feature vectors, sample network feature vectors, and sample model staleness feature vectors based on the sample context information of the computing power network, generating a sample fused feature vector. The execution network determines the prediction and scheduling strategy for the current time slot based on the sample fusion feature vector.

[0087] The predictive scheduling strategy includes a binary sample job-device allocation matrix. ;in, For the current time slot Predictive scheduling strategy, For the allocation matrix, For the number of devices, The number of jobs to be assigned to the sample.

[0088] Predictive scheduling strategies will consume computing network equipment and network resources, affecting their subsequent availability, and the resulting execution delays will further affect the staleness of the models for each job.

[0089] Furthermore, the evaluation network calculates feature vectors based on the samples. Sample network feature vectors , Sample model staleness feature vector The evaluation network then determines the current reward score, the predicted value of the next time slot, and the predicted value of the current time slot based on the predicted scheduling strategy. The gain factor for the predicted scheduling strategy of the current time slot is further determined based on these factors. The formula for calculating the gain factor is as follows.

[0090] ; in, For the current time slot Gain factor, The current reward score, For the value prediction of the next time slot, For the value prediction of the current time slot, To balance the discount factor between immediate returns and long-term returns, This is the next time slot state of the computing network updated according to the predictive scheduling strategy of the current time slot. This represents the current state of the computing network.

[0091] The parameters of the execution network are updated based on the gain factor until the preset model converges, resulting in the scheduling model. The gain factor directly affects the parameter update process of the execution (Actor) network. The system constructs an Actor loss function based on the advantage function. .in, To execute the loss function of the (Actor) network, For the current time slot Predictive scheduling strategy The statistical scheduling probability value, This represents the current state of the computing network. For the current time slot The gain factor. During backpropagation, The policy gradient vector is scaled as a scalar weight. If This indicates that the current job assignment action (predictive scheduling strategy) is better than the average expectation. The system will amplify the gradient in this direction and increase the probability of this action being generated in the future. The system will correct the parameters in reverse to suppress inefficient scheduling schemes that lead to serious resource conflicts or outdated models.

[0092] This invention achieves collaborative perception, evaluation, and decision-making regarding predictive scheduling strategies by introducing an execution network and an evaluation network. The execution network rapidly allocates sample jobs, while the evaluation network performs long-term assessments of the predictive scheduling strategies output by the execution network based on the global state. The coordinated operation of the execution and evaluation networks simultaneously ensures both the agile response of the pre-defined model to dynamic workloads and the stable convergence of long-term performance.

[0093] Furthermore, to improve training stability and sample efficiency, this invention introduces a dual-timescale update mechanism based on the standard proximal policy optimization algorithm to adapt to the characteristics of the hybrid view router-driven Actor-Critic (Execution-Evaluation) network architecture. In traditional proximal policy optimization (PPO) algorithms, the Actor network and Critic network are usually updated at the same frequency. However, in a multi-job CPN environment, routing decisions frequently change the state distribution, which can easily lead to unstable policy gradient estimation. To address this, this invention divides the parameters of the pre-defined model into two categories: fast variables and slow variables, each corresponding to a different update rhythm.

[0094] Fast variables: The parameters θ of the Actor network are updated in each training iteration by optimizing and pruning the PPO objective function, enabling it to quickly respond to short-term dynamic changes such as device load, link congestion, and model obsolescence.

[0095] Slow variables: The routing parameters of the hybrid view router adopt an incremental update strategy, maintaining a copy of the target parameter through exponential moving average, and the update rules are as follows.

[0096] ; in, For the target routing parameters of the hybrid view router, For smoothing coefficients, These are the routing parameters currently being updated online.

[0097] This mechanism allows routing decisions to evolve slowly, avoiding the impact of drastic fluctuations in feature selection on policy gradient stability. The overall training process is organized in an alternating update manner: after collecting T steps of interaction data in each round, the Actor-Critic parameters are updated for one round based on mini-batch data; while the routing parameters are updated at a lower frequency and kept synchronized through exponential moving averages. This dual-timescale update mechanism separates rapid policy adaptation from incremental route optimization, effectively suppressing training fluctuations while maintaining scheduling agility, thereby achieving efficient and stable scheduling policy learning in a dynamic CPN environment.

[0098] Based on the above embodiments, the current reward score is determined in the following manner: By combining the sample calculation view with the sample to be assigned jobs, the sample calculation status features of each device are obtained. By combining the sample network view with the sample jobs to be assigned, the sample network state features of each communication link are obtained. By combining the sample to be assigned jobs, feature extraction is performed on the sample model antiquity view to obtain the sample model antiquity status features of each local model of the sample to be assigned jobs. The current reward score is calculated based on the state characteristics of each sample, the network state characteristics of each sample, and the model obsolescence state characteristics of each sample.

[0099] Within the Markov Decision Process (MDP) framework of this application, the current reward score is used to evaluate the current time slot. The predictive scheduling strategy assesses the quality of scheduling actions, guiding the algorithm to learn the optimal scheduling scheme by maximizing long-term cumulative rewards. Sample computational state characteristics include the remaining allocable computational resources of the devices in the computing network. Total completion time of equipment sample operation And the computation time for a single sample of pending jobs. The sample network state characteristics include the sample remaining available bandwidth of the communication links in the computing network. Total link cleanup time for communication links and the transmission delay of a single sample to be assigned job The sample model staleness status characteristics include the staleness of each local model for the sample job to be assigned. Maximum model obsolescence And the speed of mitigating model obsolescence .

[0100] The current reward score is calculated based on the state characteristics of each sample, the network state characteristics of each sample, and the model staleness state characteristics of each sample. The formula for calculating the current reward score is as follows.

[0101] ; in, For the current time slot The current reward score, For the current time slot The remaining available computing resources for the sample For the current time slot The total completion time of the sample job. For the current time slot The computation time for a single sample of jobs to be assigned. For the current time slot The remaining available bandwidth of the communication link sample. For the current time slot The total link cleanup time of the communication links in the sample. For the current time slot The transmission delay of a single sample of the job to be assigned. For the current time slot The staleness of the local model for the sample jobs to be assigned. For the current time slot Maximum model obsolescence For the current time slot The speed of mitigating model obsolescence This refers to the device number in the computing network. This is the number of the job to be assigned to the sample. This represents the total number of devices in the computing network. This represents the total number of jobs to be assigned to the sample.

[0102] The numerator of the current reward score calculation formula includes remaining allocable resources (including bandwidth and computing power, as well as obsolescence mitigation speed indicators, representing the effective benefits of the predictive scheduling strategy). The denominator consists of system latency costs (including total completion time and link cleanup time) and current obsolescence risks (device-level obsolescence and system-level maximum obsolescence), forming a penalty term. This ratio structure ensures that the current reward score is sensitive to system bottlenecks: if the predictive scheduling strategy triggers severe resource contention or model obsolescence, a sharp increase in the denominator will cause the current reward score to drop rapidly, forcing the Actor network to avoid such decisions during gradient updates. Ultimately, this current reward score is transformed into an advantage signal through the Critic network, enabling agile scheduling decisions that perceive disturbances.

[0103] Based on the above embodiments, the preset model further includes a graph encoder, and the value prediction for the next time slot and the value prediction for the current time slot are determined in the following manner: Based on the graph encoder, features are extracted from the sample computation view, sample network view, and sample model staleness view to obtain sample computation feature vectors, sample network feature vectors, and sample model staleness feature vectors. Input the sample computation feature vector, the sample network feature vector, and the sample model staleness feature vector into the evaluation network to obtain the value prediction of the current time slot; The sample computation view, sample network view, and sample model staleness view are updated based on the predictive scheduling strategy of the current time slot. Based on the graph encoder, features are extracted from the updated sample computation view, the updated sample network view, and the updated sample model staleness view to obtain the updated sample computation feature vector, the updated sample network feature vector, and the updated sample model staleness feature vector. The updated sample computation feature vector, the updated sample network feature vector, and the updated sample model staleness feature vector are input into the evaluation network to obtain the value prediction for the next time slot.

[0104] Calculate the feature vector of the sample Sample network feature vectors and the stale feature vector of the sample model Input the evaluation network to obtain the value prediction for the current time slot. Calculate the eigenvectors Sample network feature vectors and the stale feature vector of the sample model It presents the state of the computing network from a global perspective.

[0105] Calculate the feature vector of the updated sample. Updated sample network feature vectors and the stale feature vector of the updated sample model Input the evaluation network to obtain the value prediction for the next time slot. .

[0106] The graph-based computing power network federated learning job scheduling device provided by the present invention will be described below. The graph-based computing power network federated learning job scheduling device described below and the graph-based computing power network federated learning job scheduling method described above can be referred to in correspondence with each other.

[0107] like Figure 4 As shown, a graph-based computing power network federated learning job scheduling device includes: The scheduling module 401 is used to input the computation view, network view, model staleness view, and jobs to be assigned from the computing power network into the scheduling model, and obtain the scheduling strategy for the jobs to be assigned in the computing power network output by the scheduling model. Specifically, the scheduling model extracts features from the computation view, network view, model staleness view, and jobs to be assigned to obtain multiple target feature vectors. Based on the context information of the computing power network, the multiple target feature vectors are masked and weighted to generate a fused feature vector. Based on the fused feature vector, the allocation result of the jobs to be assigned in the computing power network is obtained, thus yielding the scheduling strategy. The jobs to be assigned include federated learning jobs.

[0108] The graph-based federated learning job scheduling device for computing power networks provided in this invention inputs the computational view, network view, model staleness view, and jobs to be assigned into a scheduling model to obtain the scheduling strategy for jobs to be assigned in the computing power network output by the scheduling model. Specifically, the scheduling model extracts features from the computational view, network view, model staleness view, and jobs to be assigned, obtaining multiple target feature vectors. Based on the context information of the computing power network, the multiple target feature vectors are masked and weighted to generate a fused feature vector. The allocation result of the jobs to be assigned in the computing power network is obtained based on the fused feature vector, thus obtaining the scheduling strategy. The jobs to be assigned include federated learning jobs. This invention achieves a unified structured representation of multiple federated learning jobs and the relationships between devices through the computational view, network view, and model staleness view. By extracting multiple target feature vectors and masking and weighting them to generate a fused feature vector, redundant information and noise in the computing power network can be effectively suppressed, which is beneficial to improving the accuracy of the scheduling strategy. This invention achieves accurate perception of the computational state, network state, and model obsolescence of the computing power network by generating fused feature vectors, thereby improving the accuracy of scheduling strategies.

[0109] All relevant content of each step involved in the above method embodiments can be referenced from the functional description of the corresponding functional module, and will not be repeated here.

[0110] Figure 5 An example is a schematic diagram of the physical structure of an electronic device, such as... Figure 5 As shown, the electronic device may include a processor 510, a communications interface 520, a memory 530, and a communication bus 540, wherein the processor 510, communications interface 520, and memory 530 communicate with each other via the communication bus 540. The processor 510 can call logical instructions in the memory 530 to execute a graph-based federated learning job scheduling method for computing power networks. This method includes: inputting a computational view, a network view, a model staleness view, and jobs to be assigned into a scheduling model; obtaining the scheduling strategy for jobs to be assigned in the computing power network output by the scheduling model; wherein the scheduling model extracts features from the computational view, network view, model staleness view, and jobs to be assigned to obtain multiple target feature vectors; based on the context information of the computing power network, masking and weighted fusion of the multiple target feature vectors to generate a fused feature vector; obtaining the allocation result of jobs to be assigned in the computing power network based on the fused feature vector, thus obtaining the scheduling strategy; the jobs to be assigned include federated learning jobs.

[0111] Furthermore, the logical instructions in the aforementioned memory 530 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0112] In another aspect, the present invention also provides a non-transitory computer-readable storage medium storing a computer program thereon. When executed by a processor, the computer program is implemented to perform the graph-based federated learning job scheduling method for computing power networks provided by the methods described above. The method includes: inputting a computational view, a network view, a model staleness view, and jobs to be assigned into a scheduling model to obtain a scheduling strategy for jobs to be assigned in the computing power network output by the scheduling model; wherein the scheduling model performs feature extraction on the computational view, network view, model staleness view, and jobs to be assigned to obtain multiple target feature vectors; based on the context information of the computing power network, the multiple target feature vectors are masked and weighted to generate a fused feature vector; based on the fused feature vector, the allocation result of jobs to be assigned in the computing power network is obtained to obtain the scheduling strategy; the jobs to be assigned include federated learning jobs.

[0113] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0114] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0115] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A graph-based federated learning job scheduling method for computing power networks, characterized in that, include: Input the computation view, network view, model staleness view, and jobs to be assigned into the scheduling model, and obtain the scheduling strategy of the jobs to be assigned in the computing network output by the scheduling model. The scheduling model extracts features from the computation view, the network view, the model staleness view, and the job to be assigned, obtaining multiple target feature vectors; based on the context information of the computing network, the multiple target feature vectors are masked and weighted to generate a fused feature vector; based on the fused feature vector, the allocation result of the job to be assigned in the computing network is obtained, thus obtaining the scheduling strategy; the job to be assigned includes federated learning jobs.

2. The graph-based federated learning job scheduling method for computing power networks according to claim 1, characterized in that, The multiple target feature vectors include computed feature vectors, network feature vectors, and model staleness feature vectors, and the multiple target feature vectors are obtained based on the following method: The computing status of each device in the computing view is extracted based on the job to be assigned, and the computing status features of each device are obtained; the computing feature vector is obtained based on all the computing status features. The network status of each communication link in the network view is extracted in conjunction with the job to be assigned, and the network status features of each communication link are obtained; based on all the network status features, the network feature vector is obtained. Feature extraction is performed on each local model of the job to be assigned in the model staleness view to obtain the model staleness state features of each job to be assigned; based on all the model staleness state features, the model staleness feature vector is obtained.

3. The graph-based federated learning job scheduling method for computing power networks according to claim 1, characterized in that, The fused feature vector is generated based on the following method: The weights of each target feature vector and the retention rate of the control feature sparsity of each target feature vector are determined based on the context information. Based on the importance scores of each dimension of the target feature vector, a binary mask for each dimension of the feature is generated. Based on the retention rate of each target feature vector and the binarization mask of each target feature vector, the target feature vectors are subjected to sparsification and masking to obtain each mask sparse feature vector. The sparse feature vectors of each mask are normalized to obtain the normalized feature vectors. The normalized feature vectors are weighted and summed based on the weights to obtain the fused feature vector.

4. The graph-based federated learning job scheduling method for computing power networks according to claim 2, characterized in that, The computing status characteristics include the remaining allocable computing resources of the device, the total completion time of the device's jobs, and the computing time of a single job to be assigned. The network status characteristics include the remaining available bandwidth of the communication links, the total link cleanup time of the communication links, and the transmission delay of a single job to be assigned. The model obsolescence status features include the obsolescence of each local model of the job to be assigned, the maximum model obsolescence, and the model obsolescence mitigation speed.

5. The graph-based federated learning job scheduling method for computing power networks according to claim 1, characterized in that, The scheduling model was trained in the following manner: Training samples are obtained from the sample computation view, sample network view, sample model staleness view, and sample pending job assignment based on the computing power network. The training samples are input into a preset model for training, the preset model including an execution network and an evaluation network; The predictive scheduling strategy for the current time slot of the training samples is determined based on the execution network. The current reward score and the value prediction for the next time slot are determined based on the evaluation network. Based on the current reward score, the value prediction of the next time slot, and the value prediction of the current time slot, determine the gain factor of the prediction scheduling strategy for the current time slot; The parameters of the execution network are updated based on the gain factor until the preset model converges, thus obtaining the scheduling model.

6. The graph-based federated learning job scheduling method for computing power networks according to claim 5, characterized in that, The current reward score is determined based on the following method: By combining the sample to be assigned jobs, feature extraction is performed on the sample calculation view to obtain the sample calculation status features of each device; By combining the sample to be assigned jobs, feature extraction is performed on the sample network view to obtain the sample network state features of each communication link; By combining the sample to be assigned jobs, feature extraction is performed on the sample model staleness view to obtain the sample model staleness status features of each local model of the sample to be assigned jobs. The current reward score is calculated based on the state features of each sample, the network state features of each sample, and the model obsolescence state features of each sample.

7. The graph-based federated learning job scheduling method for computing power networks according to claim 5, characterized in that, The preset model also includes a graph encoder, and the value prediction for the next time slot and the value prediction for the current time slot are determined based on the following method: Based on the graph encoder, features are extracted from the sample computation view, the sample network view, and the sample model staleness view to obtain sample computation feature vectors, sample network feature vectors, and sample model staleness feature vectors; The sample calculated feature vector, the sample network feature vector, and the sample model staleness feature vector are input into the evaluation network to obtain the value prediction of the current time slot. The sample computation view, the sample network view, and the sample model obsolescence view are updated based on the predictive scheduling strategy of the current time slot. Based on the graph encoder, features are extracted from the updated sample computation view, the updated sample network view, and the updated sample model staleness view to obtain the updated sample computation feature vector, the updated sample network feature vector, and the updated sample model staleness feature vector. The updated sample computation feature vector, the updated sample network feature vector, and the updated sample model staleness feature vector are input into the evaluation network to obtain the value prediction for the next time slot.

8. A graph-based computing power network federated learning job scheduling device, characterized in that, include: A scheduling module is used to input the computation view, network view, model staleness view, and jobs to be assigned from the computing power network into a scheduling model, and obtain the scheduling strategy for the jobs to be assigned in the computing power network output by the scheduling model. The scheduling model extracts features from the computation view, network view, model staleness view, and jobs to be assigned to obtain multiple target feature vectors. Based on the context information of the computing power network, the multiple target feature vectors are masked and weighted to generate a fused feature vector. Based on the fused feature vector, the allocation result of the jobs to be assigned in the computing power network is obtained, thus obtaining the scheduling strategy. The jobs to be assigned include federated learning jobs.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the graph-based computing power network federated learning job scheduling method as described in any one of claims 1 to 7.

10. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the graph-based computing power network federated learning job scheduling method as described in any one of claims 1 to 7.

Images