Computing power network scheduling method, device, system, equipment and medium

Abstract

The application provides a computing power network scheduling method, device, system, equipment and medium, and relates to the field of computing power network scheduling, which comprises the following steps: inputting a preset training task to a computing power center node to generate a model training task; receiving a model training result and selecting a target edge-end computing power node to construct an aggregated model; determining an affinity label value according to a model contribution degree, a successful rate of reporting a training task completion quantity and a successful rate of a time length spent in reporting a model training result; issuing all affinity label values to each edge-end computing power node to update the affinity label values; the model training result is determined according to the processing of the models related to each edge-end computing power node; and the model contribution degree is determined according to the model training result of each edge-end computing power node and the aggregated model. The application can improve the robustness of the model under the conditions of network congestion and unstable edge-end computing power nodes, and realize the overall optimization of application efficiency, model accuracy, network performance and energy consumption.

Description

Technical Field

[0001] This invention relates to the field of computing power network scheduling, and in particular to a computing power network scheduling method, apparatus, system, device and medium. Background Technology

[0002] Cloud-edge-device collaborative scheduling is a key application scenario for computing power networks. On one hand, edge computing nodes, through localized data processing, can meet the requirements of application response latency and network load, while effectively reducing the computing power and storage load on the central computing node. On the other hand, through localized data processing, the central computing node and edge computing nodes only transmit model parameters, which meets data security requirements. Against this backdrop, federated learning has become the main tool for achieving multi-level cloud-edge-device collaboration. Unlike the traditional cooperation model between data oligopolies, the computing power and network characteristics in cloud-edge-device collaborative scenarios place higher demands on federated learning frameworks. These demands are mainly reflected in the large number of edge computing nodes, diverse computing capacities, unstable states, and the complex network and low transmission capacity between edge computing nodes and the central computing node.

[0003] However, there are certain limitations in cloud-edge-device collaborative application scenarios, specifically: The star topology, deployed in a fixed form, requires pre-installation of the environment and algorithm adaptation on each edge computing node before deploying federated learning applications, resulting in a huge workload for maintenance and management; although unified scheduling and deployment can be achieved through container orchestrators (Kubernetes, K8s), container images lead to high system overhead, which is particularly prominent on edge computing nodes with limited computing power; on the other hand, the model training tasks on edge computing nodes run as daemons, resulting in long idle times when the runtime varies significantly among edge computing nodes, leading to a large waste of computing power and energy; while FaaS service scheduling strategies effectively reduce the idle time of edge computing nodes and optimize system energy consumption to some extent, the fixed and synchronous aggregation characteristics of the edge participating nodes result in a large amount of waiting time at the computing center node, leading to low overall task efficiency and wasted computing power and energy. Furthermore, due to the lack of an effective edge computing node selection mechanism, there is still significant room for improvement in model performance and system optimization. Summary of the Invention

[0004] This invention provides a computing power network scheduling method, apparatus, system, device, and medium to address the technical deficiencies in existing technologies that cannot complete edge node selection, edge service preheating, and algorithm optimization under asynchronous aggregation. This invention enables efficient cloud-edge-device collaborative scheduling and overall optimization.

[0005] In a first aspect, the present invention provides a computing power network scheduling method, comprising:

[0006] Based on the Function as a Service (FaaS) service scheduling platform, input the preset training tasks to the computing power center node to generate the model training tasks corresponding to each side computing power node.

[0007] Receive model training results from each edge computing node, and select the target edge computing node based on all model training results to build an aggregate model;

[0008] The affinity label value of all edge computing nodes is determined based on the model contribution of each edge computing node, the success rate of the number of reported training tasks completed, and the success rate of the time spent reporting model training results.

[0009] All affinity tag values ​​are distributed to each edge computing node to update the affinity tag value of each edge computing node;

[0010] The model training results of each edge computing node are determined based on the model of each edge computing node processing the related edge computing node.

[0011] The model contribution of each edge computing node is determined based on the model training results of each edge computing node and the aggregated model.

[0012] The model training results shall include at least the time spent reporting the model training results and the number of training tasks completed.

[0013] According to the computing power network scheduling method provided by the present invention, before inputting a preset training task to the computing power center node according to the Function as a Service (FaaS) service scheduling platform, the method includes:

[0014] Input the algorithm application data and interface environment data into the FaaS service scheduling platform to build the federated learning initial model in the computing power center node;

[0015] The algorithm application data includes at least the environment data for constructing federated learning and the algorithm data;

[0016] The interface environment data includes at least task scheduling interface data, model parameter assignment, and return model interface data.

[0017] According to a computing power network scheduling method provided by the present invention, the step of selecting a target edge computing power node based on all model training results includes:

[0018] The edge computing node whose time spent reporting model training results is greater than or equal to the preset time is identified as the first target edge node.

[0019] The percentage of training task completion reported by each edge computing node is determined based on the number of training tasks completed reported by each edge computing node and the total number of training tasks completed reported by all edge computing nodes.

[0020] The edge computing node that is determined to be the second target edge node is any edge computing node whose percentage of reported training task completions is greater than the preset percentage.

[0021] The target edge computing power node is determined based on the first target edge node and the second target edge node.

[0022] According to a computing power network scheduling method provided by the present invention, determining the affinity label value of all edge computing power nodes based on the model contribution of each edge computing power node, the success rate of reporting the number of completed training tasks, and the success rate of reporting the time spent on model training results includes:

[0023] The contribution value of each edge computing node is determined based on the model contribution of each edge computing node and the first weighting coefficient.

[0024] The reported training task completion value of each edge computing power node is determined based on the success rate of the number of training tasks completed reported by each edge computing power node and the second weighting coefficient.

[0025] The completion value of the time spent by each edge computing power node is determined based on the success rate of the time spent on the reported model training results of each edge computing power node and the third weighting coefficient.

[0026] The affinity label value of each edge computing node is determined based on the contribution value, the reported training task completion value, and the time spent completing the task.

[0027] The sum of the first weighting coefficient, the second weighting coefficient, and the third weighting coefficient is greater than or equal to a preset constant.

[0028] According to the computing power network scheduling method provided by the present invention, after updating the affinity label value of each edge computing power node, the method further includes:

[0029] The aggregation model corresponding to each iteration is generated based on the computing power center node, and the iteration stops when the accuracy of the constructed aggregation model is greater than the preset accuracy.

[0030] Alternatively, the aggregation model corresponding to each iteration can be generated based on the computing power center node, and the iteration can be stopped after the preset number of iterations is completed;

[0031] The iterative edge nodes are determined based on the relative relationship between random values ​​and preset variables, as well as the affinity label values ​​of each edge computing node.

[0032] According to the computing power network scheduling method provided by the present invention, before generating the aggregation model corresponding to each iteration based on the computing power center node, the method further includes:

[0033] When the random value is greater than the preset variable, the edge computing power node whose affinity label value is greater than the first preset label value is determined as the iterative edge node;

[0034] When the random value is less than or equal to the preset variable, the edge computing power node whose affinity label value is less than or equal to the first preset label value is determined as the iterative edge node;

[0035] The preset variables are determined based on the number of iterations and the preset linear equation.

[0036] According to a computing power network scheduling method provided by the present invention, after distributing all affinity label values ​​to each edge computing power node to update the affinity label value of each edge computing power node, the method further includes:

[0037] If the affinity tag value of any edge computing node is greater than the second preset tag value, the edge computing node is kept in a mirror preheating state.

[0038] If the affinity tag value of any edge computing node is less than or equal to the second preset tag value, the mirror preheating state of the edge computing node is cleared.

[0039] Secondly, a computing power network scheduling system is also provided, including: a computing power network, the computing power network including a computing power center node and multiple edge computing power nodes, the computing power center node including a computing power processor;

[0040] It also includes a memory and programs or instructions stored in the memory and executable on the computing power processor. When the program or instructions are executed by the computing power processor, the computing power network scheduling method is executed. The method includes: inputting preset training tasks to the computing power center node according to the Function as a Service (FaaS) service scheduling platform to generate model training tasks corresponding to each edge computing power node; receiving model training results from each edge computing power node and selecting a target edge computing power node based on all model training results to construct an aggregated model; and calculating the model contribution and the number of reported training tasks completed by each edge computing power node. The affinity label values ​​of all edge computing nodes are determined by the power consumption and the success rate of reporting model training results. All affinity label values ​​are distributed to each edge computing node to update the affinity label value of each edge computing node. The model training results of each edge computing node are determined based on the model related to each edge computing node. The model contribution of each edge computing node is determined based on the model training results of each edge computing node and the aggregated model. The model training results include at least the time spent reporting model training results and the number of reported training tasks completed.

[0041] Thirdly, a computing power network scheduling device is also provided, comprising:

[0042] Generation Unit: Used to generate model training tasks corresponding to each side computing power node by inputting preset training tasks to the computing power center node according to the Function as a Service (FaaS) service scheduling platform.

[0043] Selection unit: Used to receive model training results from each edge computing node, and select the target edge computing node based on all model training results to build an aggregated model;

[0044] Determining Unit: Used to determine the affinity label value of all edge computing nodes based on the model contribution of each edge computing node, the success rate of the number of reported training tasks completed, and the success rate of the time spent reporting model training results.

[0045] Update Unit: Used to distribute all affinity tag values ​​to each edge computing node to update the affinity tag value of each edge computing node;

[0046] The model training results of each edge computing node are determined based on the model of each edge computing node processing the related edge computing node.

[0047] The model contribution of each edge computing node is determined based on the model training results of each edge computing node and the aggregated model.

[0048] The model training results shall include at least the time spent reporting the model training results and the number of training tasks completed.

[0049] Fourthly, an electronic device is also provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the computing power network scheduling method.

[0050] Fifthly, a non-transitory computer-readable storage medium is also provided, on which a computer program is stored, which, when executed by a processor, implements the computing power network scheduling method.

[0051] This invention provides a computing power network scheduling method, apparatus, system, device, and medium. Combining containerized task runtime management capabilities and FaaS service dynamic scheduling and rapid deployment capabilities, it realizes the asynchronous aggregation capability of computing power center nodes based on the time spent reporting model training results and the number of reported training tasks completed under the historical status of edge computing power nodes. This effectively improves the robustness of the model under network congestion and unstable edge computing power nodes, thereby achieving overall optimization of application efficiency, model accuracy, network performance, and energy consumption. Attached Figure Description

[0052] 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.

[0053] Figure 1 This is one of the flowcharts of a computing power network scheduling method provided by the present invention;

[0054] Figure 2 This is a schematic diagram of a process for selecting target edge computing nodes based on the training results of all models, provided by the present invention.

[0055] Figure 3 This is a schematic diagram of a process for determining the affinity tag values ​​of all edge computing nodes provided by the present invention;

[0056] Figure 4 This is the second flowchart of a computing power network scheduling method provided by the present invention;

[0057] Figure 5 This is the third flowchart of a computing power network scheduling method provided by the present invention;

[0058] Figure 6This is a schematic diagram of the structure of a computing power network scheduling system provided by the present invention;

[0059] Figure 7 This is a schematic diagram of the structure of a computing power network scheduling device provided by the present invention;

[0060] Figure 8 This is a schematic diagram of the structure of the electronic device provided by the present invention. Detailed Implementation

[0061] 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.

[0062] Current mainstream solutions for computing power network scheduling include: deploying federated learning platforms directly on edge computing nodes in bare metal form, and deploying model aggregators on computing center nodes to achieve a star federated architecture of parameter servers and edge computing nodes; combining with mainstream cloud-native platforms, using a multi-level scheduling system of cloud-edge-endpoint, encapsulating federated learning capabilities in the form of containers and distributing them to edge computing nodes, with computing center nodes using container orchestration capabilities such as container orchestrators (Kubernetes, K8s) for unified scheduling management and model aggregation; and deploying basic platforms such as serverless frameworks on edge computing nodes to deploy federated learning as a FaaS service, with computing center nodes using Hypertext Transfer Protocol (HTTP) requests for unified scheduling and model aggregation.

[0063] Figure 1 This is one of the flowcharts illustrating a computing power network scheduling method provided by the present invention. The present invention provides a computing power network scheduling method, comprising:

[0064] Based on the Function as a Service (FaaS) service scheduling platform, input the preset training tasks to the computing power center node to generate the model training tasks corresponding to each side computing power node.

[0065] Receive model training results from each edge computing node, and select the target edge computing node based on all model training results to build an aggregate model;

[0066] The affinity label value of all edge computing nodes is determined based on the model contribution of each edge computing node, the success rate of the number of reported training tasks completed, and the success rate of the time spent reporting model training results.

[0067] All affinity tag values ​​are distributed to each edge computing node to update the affinity tag value of each edge computing node;

[0068] The model training results of each edge computing node are determined based on the model of each edge computing node processing the related edge computing node.

[0069] The model contribution of each edge computing node is determined based on the model training results of each edge computing node and the aggregated model.

[0070] The model training results shall include at least the time spent reporting the model training results and the number of training tasks completed.

[0071] In step 101, the Function as a Service (FaaS) service scheduling platform is a new way to build and deploy server-side software for cloud functions. In the FaaS service scheduling platform, the host instance layer and application software layer are removed and replaced by the cloud function platform, thereby greatly simplifying the development process. The preset training task is a federated learning task. At the computing power center node, the federated learning task process is scheduled to generate the model training task corresponding to each edge computing power node. The FaaS service scheduling platform distributes the federated learning task to the corresponding edge computing power node.

[0072] In step 102, the model training results of each edge computing node are determined based on the processing of the model related to each edge computing node. Specifically, each edge computing node performs localized model training. After the training task is completed, the locally updated model training results are returned to the computing center node through the interface, and the computing power is released.

[0073] The target edge computing node is determined based on the time spent reporting model training results and the number of training tasks completed in the model training results. After constructing the aggregated model based on the target edge computing node, this invention calculates the accuracy of the aggregated model for subsequent iterative operations. Although other edge computing nodes besides the target edge computing node do not participate in the construction of the aggregated model, they still record the time spent reporting model training results and the number of training tasks completed for subsequent affinity label value updates.

[0074] In step 103, the model contribution of each edge computing node is determined based on the model training results of each edge computing node and the aggregated model. In such an embodiment, at the computing center node, the model contribution of each edge computing node is calculated by the discrete values ​​or Shapley values ​​of the model and the aggregated model reported by the edge computing nodes.

[0075] The success rate of the number of reported training tasks is determined based on the number of reported training tasks completed by each edge computing node in the historical data participating in the construction of the aggregation model, and the total number of all reported training tasks in the historical data. The success rate of the time spent reporting model training results is determined based on the ratio of the average time spent on model training results participating in the construction of the aggregation model in the historical service reports of each edge computing node to the average time spent on all model training results. Then, the affinity label value of all edge computing nodes is determined based on the model contribution of each edge computing node, the success rate of the number of reported training tasks, and the success rate of the time spent reporting model training results.

[0076] In step 104, all affinity label values ​​are distributed to each edge computing node to update the affinity label value of each edge computing node. Those skilled in the art will understand that in the initial stage before each edge computing node has undergone model localization training, optionally, the affinity label of the edge computing node is constructed according to the initial rules. The initial value is 0 by default, and the default value is 1 after selection. The set of edge computing nodes participating in federated learning is selected to participate in the model training task distribution in the computing center stage. In step 104, the affinity label value of each edge computing node is further determined and updated according to the completion status of each edge computing node for the model training task and the construction status of the aggregated model.

[0077] Optionally, before inputting the preset training task to the computing center node according to the Function as a Service (FaaS) service scheduling platform, the following steps are included:

[0078] Input the algorithm application data and interface environment data into the FaaS service scheduling platform to build the federated learning initial model in the computing power center node;

[0079] The algorithm application data includes at least the environment data for constructing federated learning and the algorithm data;

[0080] The interface environment data includes at least task scheduling interface data, model parameter assignment, and return model interface data.

[0081] This invention constructs the basic environment and related algorithm application packages required for federated learning at the computing power center node through an open-source package management system, environment management system (Conda), and virtual environment (Venv), realizes the localized model training process, and defines interface services to implement task scheduling interface, model parameter assignment, and model return interface. At the computing power center node, the algorithm application and interface environment are compressed and registered to the FaaS service scheduling platform.

[0082] This invention utilizes a computing power network FaaS scheduling system based on asynchronous federation technology to achieve cloud-edge-device scheduling, optimizing application efficiency, model accuracy, network performance, and energy consumption. Based on containerization technology, it enables automatic deployment, distribution, and scheduling of FaaS services for cloud-edge-device federated learning applications, effectively reducing manual maintenance workload. Through asynchronous aggregation, the waiting latency of computing power center nodes is reduced by 30% compared to synchronous aggregation, significantly improving model training efficiency. Furthermore, based on node selection schemes, it increases the proportion of effective nodes, improving the node reporting success rate by 10% in the later stages of model iteration compared to traditional fixed-node strategies.

[0083] This invention provides a computing power network scheduling method, apparatus, system, device, and medium. Combining containerized task runtime management capabilities and FaaS service dynamic scheduling and rapid deployment capabilities, it realizes the asynchronous aggregation capability of computing power center nodes based on the time spent reporting model training results and the number of reported training tasks completed under the historical status of edge computing power nodes. This effectively improves the robustness of the model under network congestion and unstable edge computing power nodes, thereby achieving overall optimization of application efficiency, model accuracy, network performance, and energy consumption.

[0084] Figure 2 This is a flowchart illustrating a process for selecting a target edge computing node based on all model training results, provided by the present invention. The process of selecting a target edge computing node based on all model training results includes:

[0085] The edge computing node whose time spent reporting model training results is greater than or equal to the preset time is identified as the first target edge node.

[0086] The percentage of training task completion reported by each edge computing node is determined based on the number of training tasks completed reported by each edge computing node and the total number of training tasks completed reported by all edge computing nodes.

[0087] The edge computing node that is determined to be the second target edge node is any edge computing node whose percentage of reported training task completions is greater than the preset percentage.

[0088] The target edge computing power node is determined based on the first target edge node and the second target edge node.

[0089] In step 1021, all edge computing nodes are traversed to obtain the time spent by each edge computing node in reporting model training results. The time spent by each edge computing node in reporting model training results is compared with a preset time. If the time spent in reporting model training results is greater than or equal to the preset time, the edge computing node is determined as the first target edge node.

[0090] In step 1022, the success rate of the reported training task completion is determined based on the number of reported training tasks completed by each edge computing node in the historical data participating in the construction of the aggregation model and the total number of all reported training tasks in the historical data. Accordingly, in this implementation scheme, the proportion of the reported training task completion of each edge computing node can refer to the method for determining the success rate of the reported training task completion.

[0091] In step 1023, all edge computing nodes are traversed to obtain the percentage of reported training task completions of all edge computing nodes. The percentage of reported training task completions of each edge computing node is compared with a preset percentage. If the percentage of reported training task completions is greater than the preset percentage, the edge computing node is determined as the second target edge node.

[0092] In step 1024, the target edge computing node is determined based on the first target edge node and the second target edge node. Since there may be duplicate edge nodes in the first target edge node and the second target edge node, the target edge computing node can optionally be determined after deduplication.

[0093] Those skilled in the art will understand that, after determining the target edge computing power nodes based on the first target edge node and the second target edge node, the present invention constructs an aggregation model based on the target edge computing power nodes and calculates the model accuracy. Correspondingly, the computing power center node receives the target edge computing power nodes and weights them according to the proportion of the current reported service quantity. The aggregation model can be determined according to the following formula:

[0094] m1=(1-r)*m0+r*avg(M) (1)

[0095] In equation (1), m1 is the aggregated model, m0 is the current model, M is the set of models that receive edge computing power nodes, and r is the percentage of the number of training tasks reported by edge computing power nodes.

[0096] More specifically, the computing power center node in this invention determines the aggregation model based on the relevant models processed by the edge computing power nodes, including:

[0097] The current model parameters of the computing center node are determined to be the first model; the models reported by each edge computing node and received by the computing center node are determined to be the second model.

[0098] Based on the first model, the second model, and the percentage of tasks completed, the aggregation model for this iteration is determined using a linear weighting method.

[0099] Figure 3This is a flowchart illustrating a process for determining the affinity label values ​​of all edge computing power nodes provided by the present invention. The process of determining the affinity label values ​​of all edge computing power nodes based on the model contribution of each edge computing power node, the success rate of reporting the number of completed training tasks, and the success rate of reporting the time spent on model training results includes:

[0100] The contribution value of each edge computing node is determined based on the model contribution of each edge computing node and the first weighting coefficient.

[0101] The reported training task completion value of each edge computing power node is determined based on the success rate of the number of training tasks completed reported by each edge computing power node and the second weighting coefficient.

[0102] The completion value of the time spent by each edge computing power node is determined based on the success rate of the time spent on the reported model training results of each edge computing power node and the third weighting coefficient.

[0103] The affinity label value of each edge computing node is determined based on the contribution value, the reported training task completion value, and the time spent completing the task.

[0104] The sum of the first weighting coefficient, the second weighting coefficient, and the third weighting coefficient is greater than or equal to a preset constant.

[0105] In step 1031, the contribution value of each edge computing power node is determined based on the model contribution degree of each edge computing power node and the first weighting coefficient. The contribution value of each edge computing power node can be determined by multiplying the model contribution degree of each edge computing power node and the first weighting coefficient.

[0106] In step 1032, the reported training task completion value of each edge computing power node is determined based on the success rate of the reported training task completion number of each edge computing power node and the second weighting coefficient. The reported training task completion value of each edge computing power node can be determined by multiplying the success rate of the reported training task completion number of each edge computing power node and the second weighting coefficient.

[0107] In step 1033, the completion value of the time spent by each end computing power node is determined based on the success rate of the time spent reporting the model training results of each end computing power node and the third weighting coefficient. The completion value of the time spent by each end computing power node can be determined by multiplying the success rate of the time spent reporting the model training results of each end computing power node and the third weighting coefficient.

[0108] In step 1034, the affinity label value of each edge computing node is determined based on the contribution value, the reported training task completion value, and the time spent completion value. The affinity label value of each edge computing node can be determined based on the sum of the contribution value, the reported training task completion value, and the time spent completion value. Specifically, the following formula can be used as a reference:

[0109] e = ap + bs + cd (2)

[0110] In equation (2), e is the affinity label value of the edge computing node, p is the contribution value of the edge computing node, s is the success rate of the number of training tasks reported by the edge computing node, d is the success rate of the time spent reporting model training results by the edge computing node, a is the first weighting coefficient, b is the second weighting coefficient, and c is the third weighting coefficient. The preset constant is 1, that is, the first weighting coefficient a, the second weighting coefficient b, and the third weighting coefficient c should satisfy equation (3):

[0111] a+b+c≥1 (3)

[0112] In practical applications where the preset constant is 1, the present invention can be adjusted according to the actual application scenario. In scenarios where model performance is important, a can be set higher; in scenarios where system stability is important, b can be set higher; and in scenarios where system response latency is important, c can be set higher.

[0113] Figure 4 This is a second flowchart illustrating a computing power network scheduling method provided by the present invention. After updating the affinity tag value of each edge computing power node, it further includes:

[0114] The aggregation model corresponding to each iteration is generated based on the computing power center node, and the iteration stops when the accuracy of the constructed aggregation model is greater than the preset accuracy.

[0115] Alternatively, the aggregation model corresponding to each iteration can be generated based on the computing power center node, and the iteration can be stopped after the preset number of iterations is completed;

[0116] The iterative edge nodes are determined based on the relative relationship between random values ​​and preset variables, as well as the affinity label values ​​of each edge computing node.

[0117] In one optional embodiment, the preset accuracy rate can be 80%, while in other embodiments it can be 90%, 98%, 99%, etc. In each iteration, the accuracy of the constructed aggregate model will be improved. In this invention, the iteration edge node used in each iteration is associated with the affinity label value of each edge computing node after the previous iteration. Therefore, after each construction of the aggregate model, its aggregate model accuracy is calculated until the accuracy of the constructed aggregate model is greater than the preset accuracy rate, and then the iteration stops.

[0118] In another optional embodiment, the preset number of iterations can be 3, 5, 8 or even more. In this case, the preset training task is input to the computing power center node to generate the model training task corresponding to each edge node in each iteration, so as to generate the affinity label value of all edge computing power nodes in the current iteration. The above process is regarded as an iteration process. According to the preset number of iterations, the iteration stops after the preset number of iterations is completed.

[0119] The iterative edge nodes are determined based on the relative relationship between random values ​​and preset variables, as well as the affinity label values ​​of each edge computing power node. Specifically, before generating the aggregation model corresponding to each iteration based on the computing power center node, the following steps are also included:

[0120] When the random value is greater than the preset variable, the edge computing power node whose affinity label value is greater than the first preset label value is determined as the iterative edge node;

[0121] When the random value is less than or equal to the preset variable, the edge computing power node whose affinity label value is less than or equal to the first preset label value is determined as the iterative edge node;

[0122] The preset variables are determined based on the number of iterations and the preset linear equation.

[0123] In step 201, in the computing power center node of the present invention, in order to ensure the diversity of the changes in the computing power nodes participating in the calculation and the amount of data participating in the calculation, according to the Bandit algorithm for making real-time decisions in uncertain scenarios, the computing power nodes participating in the calculation are scattered. In an optional embodiment, a threshold e is defined, and random numbers are generated. When the random number is greater than e, the edge node is selected from the set with higher affinity as the iterative edge node.

[0124] The present invention also sets a first preset label value to define a set with high affinity and a set with low affinity. That is, in step 201, edge computing power nodes with affinity label values ​​greater than the first preset label value are determined as iterative edge nodes.

[0125] In step 202, correspondingly, when the random value is less than or equal to the preset variable, that is, when the random number is less than or equal to e, the edge node is selected from the set with lower affinity as the iterative edge node, that is, the edge computing power node with affinity label value less than or equal to the first preset label value is determined as the iterative edge node.

[0126] Optionally, the preset variable is determined based on the number of iterations and the preset linear equation. In the initial iteration, the value of e is relatively large, and subsequently, the value of e is linearly reduced according to the number of iterations and the preset linear equation to achieve a broad search of edge computing nodes. This invention optimizes edge computing nodes by considering model contribution and edge computing node response latency, and improves model performance by combining the Bandit algorithm's scattering method to achieve overall system optimization. By using the Bandit node selection strategy, the number of edge computing nodes participating in the computation in the cluster is increased by more than 3 times, and the area under the curve (AUC) of the computing center model is increased by 0.5%.

[0127] Figure 5 This is the third flowchart of a computing power network scheduling method provided by the present invention. After distributing all affinity tag values ​​to each edge computing power node to update the affinity tag value of each edge computing power node, it further includes:

[0128] If the affinity tag value of any edge computing node is greater than the second preset tag value, the edge computing node is kept in a mirror preheating state.

[0129] If the affinity tag value of any edge computing node is less than or equal to the second preset tag value, the mirror preheating state of the edge computing node is cleared.

[0130] The present invention will determine whether to maintain the computing mirror preheating state based on the received node affinity tag through parallel steps 301 and 302. This step will be implemented on each edge computing node. That is, if the affinity tag value of any edge computing node is greater than the second preset tag value, the edge computing node will be kept in the mirror preheating state; if the affinity tag value of any edge computing node is less than or equal to the second preset tag value, the mirror preheating state of the edge computing node will be cleared.

[0131] This invention optimizes the preheating strategy for edge computing node model training services based on node selection strategies, thereby reducing the cold start problem of edge computing node services. Through node prediction optimization strategies, the proportion of cold start time for edge computing nodes is reduced by 10% compared to the original strategy.

[0132] Figure 6 This is a schematic diagram of a computing power network scheduling system provided by the present invention. The present invention discloses a computing power network scheduling system, including: a computing power network, the computing power network including a computing power center node and multiple edge computing power nodes, the computing power center node including a computing power processor;

[0133] It also includes a memory and programs or instructions stored in the memory and executable on the computing power processor. When the program or instructions are executed by the computing power processor, the computing power network scheduling method is executed. The method includes: inputting preset training tasks to the computing power center node according to the Function as a Service (FaaS) service scheduling platform to generate model training tasks corresponding to each edge computing power node; receiving model training results from each edge computing power node and selecting a target edge computing power node based on all model training results to construct an aggregated model; and calculating the model contribution and the number of reported training tasks completed by each edge computing power node. The affinity label values ​​of all edge computing nodes are determined by the power consumption and the success rate of reporting model training results. All affinity label values ​​are distributed to each edge computing node to update the affinity label value of each edge computing node. The model training results of each edge computing node are determined based on the model related to each edge computing node. The model contribution of each edge computing node is determined based on the model training results of each edge computing node and the aggregated model. The model training results include at least the time spent reporting model training results and the number of reported training tasks completed.

[0134] This invention addresses the challenges of numerous and diverse edge computing nodes and unstable network load in cloud-edge-device collaborative computing network scenarios. It constructs an asynchronous federation technology, employing asynchronous aggregation, node selection, and FaaS service management, combined with algorithm optimization, to achieve collaborative model training and closed-loop system optimization between cloud and edge computing nodes. The FaaS service scheduling and management optimizes the preheating strategy for edge computing node model training services based on node selection strategies, avoiding cold starts, effectively improving the overall training efficiency and model accuracy of the system, and reducing network congestion and high energy consumption.

[0135] This invention provides a computing power network scheduling method, apparatus, system, device, and medium. Combining containerized task runtime management capabilities and FaaS service dynamic scheduling and rapid deployment capabilities, it realizes the asynchronous aggregation capability of computing power center nodes based on the time spent reporting model training results and the number of reported training tasks completed under the historical status of edge computing power nodes. This effectively improves the robustness of the model under network congestion and unstable edge computing power nodes, thereby achieving overall optimization of application efficiency, model accuracy, network performance, and energy consumption.

[0136] Figure 7This is a schematic diagram of a computing power network scheduling device provided by the present invention. The present invention discloses a computing power network scheduling device, including a generation unit 1: used to input preset training tasks to the computing power center node according to the Function as a Service (FaaS) service scheduling platform, and generate model training tasks corresponding to each side computing power node. The working principle of the generation unit 1 can be referred to the aforementioned step 101, and will not be repeated here.

[0137] The computing power network scheduling device further includes a selection unit 2: used to receive model training results from each edge computing power node, and select a target edge computing power node based on all model training results to construct an aggregated model. The working principle of the selection unit 2 can be referred to the aforementioned step 102, and will not be repeated here.

[0138] The computing power network scheduling device further includes a determination unit 3: used to determine the affinity label value of all edge computing power nodes based on the model contribution of each edge computing power node, the success rate of the number of reported training tasks completed, and the success rate of the time spent reporting model training results. The working principle of the determination unit 3 can be referred to the aforementioned step 103, and will not be repeated here.

[0139] The computing power network scheduling device also includes an update unit 4: used to send all affinity tag values ​​to each edge computing power node to update the affinity tag value of each edge computing power node. The working principle of the update unit 4 can be referred to the aforementioned step 104, and will not be repeated here.

[0140] The model training results for each edge computing node are determined based on the model that each edge computing node uses to process the model related to that edge computing node.

[0141] The model contribution of each edge computing node is determined based on the model training results of each edge computing node and the aggregated model.

[0142] The model training results shall include at least the time spent reporting the model training results and the number of training tasks completed.

[0143] This invention provides a computing power network scheduling method, apparatus, system, device, and medium. Combining containerized task runtime management capabilities and FaaS service dynamic scheduling and rapid deployment capabilities, it realizes the asynchronous aggregation capability of computing power center nodes based on the time spent reporting model training results and the number of reported training tasks completed under the historical status of edge computing power nodes. This effectively improves the robustness of the model under network congestion and unstable edge computing power nodes, thereby achieving overall optimization of application efficiency, model accuracy, network performance, and energy consumption.

[0144] Figure 8 This is a schematic diagram of the structure of the electronic device provided by the present invention. Figure 8An example is a schematic diagram of the physical structure of an electronic device, such as... Figure 8 As shown, the electronic device may include: a processor 810, a communication interface 820, a memory 830, and a communication bus 840, wherein the processor 810, the communication interface 820, and the memory 830 communicate with each other through the communication bus 840. The processor 810 can call logical instructions in the memory 830 to execute a computing power network scheduling method. This method includes: inputting preset training tasks to the computing power center node according to the Function as a Service (FaaS) service scheduling platform to generate model training tasks corresponding to each edge computing power node; receiving model training results from each edge computing power node and selecting a target edge computing power node based on all model training results to construct an aggregated model; and calculating the model contribution of each edge computing power node, the success rate of reporting the number of completed training tasks, and the time spent reporting model training results. A long success rate determines the affinity label value of all edge computing power nodes; all affinity label values ​​are distributed to each edge computing power node to update the affinity label value of each edge computing power node; the model training result of each edge computing power node is determined based on the model related to each edge computing power node; the model contribution of each edge computing power node is determined based on the model training result of each edge computing power node and the aggregated model; the model training result includes at least the time spent reporting the model training result and the number of reported training tasks completed.

[0145] Furthermore, the logical instructions in the aforementioned memory 830 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, essentially, 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.

[0146] On the other hand, the present invention also provides a computer program product, which includes a computer program that can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer can execute a computing power network scheduling method provided by the above methods. The method includes: inputting a preset training task to the computing power center node according to a Function as a Service (FaaS) service scheduling platform to generate a model training task corresponding to each edge computing power node; receiving model training results from each edge computing power node, and selecting a target edge computing power node according to all model training results to construct an aggregated model; and selecting a target edge computing power node according to each edge computing power node. The affinity label values ​​of all edge computing nodes are determined by the model contribution, the success rate of reporting the number of completed training tasks, and the success rate of reporting the time spent on model training results. All affinity label values ​​are distributed to each edge computing node to update the affinity label value of each edge computing node. The model training results of each edge computing node are determined based on the model related to each edge computing node. The model contribution of each edge computing node is determined based on the model training results of each edge computing node and the aggregated model. The model training results include at least the time spent on reporting model training results and the number of completed training tasks.

[0147] 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 implements the computing power network scheduling method provided by the above methods. The method includes: inputting a preset training task to the computing power center node according to a Function as a Service (FaaS) service scheduling platform to generate a model training task corresponding to each edge computing power node; receiving model training results from each edge computing power node and selecting a target edge computing power node based on all model training results to construct an aggregated model; and reporting the completion of the training task based on the model contribution of each edge computing power node. The affinity label values ​​of all edge computing nodes are determined by the success rate of the number of successes and the success rate of the time spent reporting model training results; all affinity label values ​​are distributed to each edge computing node to update the affinity label value of each edge computing node; the model training results of each edge computing node are determined based on the model related to each edge computing node; the model contribution of each edge computing node is determined based on the model training results of each edge computing node and the aggregated model; the model training results include at least the time spent reporting model training results and the number of reported training tasks completed.

[0148] 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.

[0149] 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.

[0150] 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 computing power network scheduling method, characterized in that, include: According to the Function as a Service (FaaS) service scheduling platform, preset training tasks are input to the computing power center node to generate model training tasks corresponding to each side computing power node. The computing power center node has a federated learning initial model built in it. Receive model training results from each edge computing node, and select the target edge computing node based on all model training results to build an aggregate model; The affinity label value of all edge computing nodes is determined based on the model contribution of each edge computing node, the success rate of the number of reported training tasks completed, and the success rate of the time spent reporting model training results. All affinity tag values ​​are distributed to each edge computing node to update the affinity tag value of each edge computing node; The model training results of each edge computing node are determined based on the model of each edge computing node processing the related edge computing node. The model contribution of each edge computing node is determined based on the model training results of each edge computing node and the aggregated model. The model training results shall include at least the time spent reporting the model training results and the number of training tasks completed. The step of selecting the target edge computing node based on the training results of all models includes: The edge computing node whose time spent reporting model training results is greater than or equal to the preset time is identified as the first target edge node. The percentage of training task completion reported by each edge computing node is determined based on the number of training tasks completed reported by each edge computing node and the total number of training tasks completed reported by all edge computing nodes. The edge computing node that is determined to be the second target edge node is any edge computing node whose percentage of reported training task completions is greater than the preset percentage. The target edge computing power node is determined based on the first target edge node and the second target edge node.

2. The computing power network scheduling method according to claim 1, characterized in that, Before the preset training task is input to the computing center node according to the Function as a Service (FaaS) service scheduling platform, the following steps are included: Input the algorithm application data and interface environment data into the FaaS service scheduling platform to build the federated learning initial model in the computing power center node; The algorithm application data includes at least the environment data for constructing federated learning and the algorithm data; The interface environment data includes at least task scheduling interface data, model parameter assignment, and return model interface data.

3. The computing power network scheduling method according to claim 1, characterized in that, The affinity label value for all edge computing nodes is determined based on the model contribution of each edge computing node, the success rate of reporting the number of completed training tasks, and the success rate of reporting the time spent on model training results. This includes: The contribution value of each edge computing node is determined based on the model contribution of each edge computing node and the first weighting coefficient. The reported training task completion value of each edge computing power node is determined based on the success rate of the number of training tasks completed reported by each edge computing power node and the second weighting coefficient. The completion value of the time spent by each edge computing power node is determined based on the success rate of the time spent on the reported model training results of each edge computing power node and the third weighting coefficient. The affinity label value of each edge computing node is determined based on the contribution value, the reported training task completion value, and the time spent completing the task. The sum of the first weighting coefficient, the second weighting coefficient, and the third weighting coefficient is greater than or equal to a preset constant.

4. The computing power network scheduling method according to claim 1, characterized in that, After updating the affinity tag value of each edge computing node, the following is also included: The aggregation model corresponding to each iteration is generated based on the computing power center node, and the iteration stops when the accuracy of the constructed aggregation model is greater than the preset accuracy. Alternatively, the aggregation model corresponding to each iteration can be generated based on the computing power center node, and the iteration can be stopped after the preset number of iterations is completed; The iterative edge nodes are determined based on the relative relationship between random values ​​and preset variables, as well as the affinity label values ​​of each edge computing node.

5. The computing power network scheduling method according to claim 4, characterized in that, Before generating the aggregation model corresponding to each iteration based on the computing power center node, the following is also included: When the random value is greater than the preset variable, the edge computing power node whose affinity label value is greater than the first preset label value is determined as the iterative edge node; When the random value is less than or equal to the preset variable, the edge computing power node whose affinity label value is less than or equal to the first preset label value is determined as the iterative edge node; The preset variables are determined based on the number of iterations and the preset linear equation.

6. The computing power network scheduling method according to claim 1, characterized in that, After distributing all affinity tag values ​​to each edge computing node to update the affinity tag value of each edge computing node, the process also includes: If the affinity tag value of any edge computing node is greater than the second preset tag value, the edge computing node is kept in a mirror preheating state. If the affinity tag value of any edge computing node is less than or equal to the second preset tag value, the mirror preheating state of the edge computing node is cleared.

7. A computing power network scheduling system, characterized in that, include: A computing power network, comprising a central computing power node and multiple edge computing power nodes, wherein the central computing power node includes a computing power processor; It also includes a memory and a program or instructions stored in the memory and executable on the computing power processor, wherein the program or instructions, when executed by the computing power processor, perform the computing power network scheduling method as described in any one of claims 1 to 6.

8. A computing power network scheduling device, characterized in that, include: Generation Unit: Used to input preset training tasks to the computing power center node according to the Function as a Service (FaaS) service scheduling platform, and generate model training tasks corresponding to each side computing power node. The computing power center node has a federated learning initial model built in it. Selection unit: Used to receive model training results from each edge computing node, and select the target edge computing node based on all model training results to build an aggregated model; Determining Unit: Used to determine the affinity label value of all edge computing nodes based on the model contribution of each edge computing node, the success rate of the number of reported training tasks completed, and the success rate of the time spent reporting model training results. Update Unit: Used to distribute all affinity tag values ​​to each edge computing node to update the affinity tag value of each edge computing node; The model training results of each edge computing node are determined based on the model of each edge computing node processing the related edge computing node. The model contribution of each edge computing node is determined based on the model training results of each edge computing node and the aggregated model. The model training results shall include at least the time spent reporting the model training results and the number of training tasks completed. The selection unit is specifically used for: The edge computing node whose time spent reporting model training results is greater than or equal to the preset time is identified as the first target edge node. The percentage of training task completion reported by each edge computing node is determined based on the number of training tasks completed reported by each edge computing node and the total number of training tasks completed reported by all edge computing nodes. The edge computing node that is determined to be the second target edge node is any edge computing node whose percentage of reported training task completions is greater than the preset percentage. The target edge computing power node is determined based on the first target edge node and the second target edge node.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the computing power network scheduling method as described in any one of claims 1 to 6.

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 computing power network scheduling method as described in any one of claims 1 to 6.

Images