Resource scheduling method, device, electronic device and storage medium

Abstract

The present application provides a resource scheduling method, device, electronic device and storage medium, the method comprising obtaining status data corresponding to each computing power node, screening all status data, and combining the screened status data into a status matrix; inputting the status matrix into a pre-trained fault identification model, and outputting a predicted fault computing power node matrix corresponding to the status matrix through the fault identification model; inputting the predicted fault computing power node matrix into a pre-trained migration decision model, obtaining resource-related information and pre-calculated current credibility of each computing power node through the migration decision model, and processing the predicted fault computing power node matrix separately using multiple first self-attention mechanisms; determining a predicted resource scheduling strategy, and outputting the predicted resource scheduling strategy through the migration decision model; and executing the predicted resource scheduling strategy, thereby solving the technical problem in the prior art that computing power nodes that are about to fail cannot achieve active fault tolerance.

Description

Technical Field

[0001] The present application relates to the field of data processing technology, and in particular to a resource scheduling method, device, electronic device and storage medium. Background Art

[0002] With the development of mobile communications and artificial intelligence technologies, the demand for computing power nodes in various technical businesses has exploded. The reliability of computing power nodes is crucial, as their failure can cause partial or complete functional failure of the systems corresponding to various technical businesses, seriously threatening the continuity and stability of these systems.

[0003] Detecting and recovering from computing node failures has always been a significant challenge across various technology businesses. Due to the complexity and diversity of the systems underlying these businesses, traditional methods have struggled to meet the demand for fast and accurate computing node failure detection. Furthermore, when a computing node is detected to be about to fail, existing technology typically issues a fault warning in the hope that technicians can resolve the issue. However, the time it takes for technicians to identify and resolve the issue wastes human resources and inefficiently schedules computing node resources. Summary of the Invention

[0004] In view of this, the purpose of this application is to propose a resource scheduling method, device, electronic device and storage medium to overcome all or part of the deficiencies in the prior art.

[0005] Based on the above-mentioned purpose, the present application provides a resource scheduling method, including: obtaining status data corresponding to each computing power node, and filtering all status data, and combining the filtered status data into a state matrix; inputting the state matrix into a pre-trained fault identification model, and outputting the predicted fault computing power node matrix corresponding to the state matrix through the fault identification model; inputting the predicted fault computing power node matrix into a pre-trained migration decision model, obtaining resource-related information and pre-calculated current credibility of each computing power node through the migration decision model, and based on the resource-related information of each computing power node, using multiple first self-attention mechanisms to process the predicted fault computing power node matrix separately to obtain multiple first attention vectors; based on the multiple first attention vectors and all current credibility, determining a predicted resource scheduling strategy, outputting the predicted resource scheduling strategy through the migration decision model; and executing the predicted resource scheduling strategy.

[0006] Optionally, the filtering of all status data and combining the filtered status data into a status matrix includes: classifying all status data according to predetermined categories, and for each category of status data, querying and determining the number of groups corresponding to the category in a pre-constructed storage table according to the category; grouping the status data corresponding to the category according to the number of groups to obtain multiple groups of status data corresponding to the category; calculating the information entropy of the status data corresponding to all categories based on the number of status data corresponding to the category and the number of all status data; calculating the information gain corresponding to each group of status data based on the information entropy, the number of status data corresponding to the category and the number of each group of status data; and filtering out a group of status data corresponding to the maximum information gain in each category from all status data, and combining the multiple groups of status data obtained by filtering into a status matrix.

[0007] Optionally, determining the predicted resource scheduling strategy based on the multiple first attention vectors and all current credibility includes: splicing the multiple first attention vectors to obtain a first splicing vector; and normalizing the first splicing vector based on all current credibility to obtain the predicted resource scheduling strategy.

[0008] Optionally, the fault identification model includes a multi-layer perceptron layer; inputting the state matrix into a pre-trained fault identification model, and outputting the predicted fault computing power node matrix corresponding to the state matrix through the fault identification model, includes: inputting the state matrix into the fault identification model, processing the state matrix respectively through multiple second self-attention mechanisms in the fault identification model to obtain multiple second attention vectors; splicing the multiple second attention vectors to obtain a second splicing vector; performing abnormality identification on the second splicing vector through the multi-layer perceptron layer to obtain the predicted fault computing power node matrix, and outputting the predicted fault computing power node matrix through the fault identification model.

[0009] Optionally, calculating the current credibility includes: obtaining the previous scheduling behavior and historical scheduling behavior scores of each computing power node and its current neighbors, the historical service quality, and the stability of the previous environment; for each computing power node, determining the previous scheduling behavior score based on the previous scheduling behavior and the historical scheduling behavior score, determining the previous service quality score based on the historical service quality, and determining the stability score of the previous environment based on the stability of the previous environment; determining the current credibility based on the previous scheduling behavior score, the previous service quality score, the stability score of the previous environment, and a preset weight coefficient.

[0010] Optionally, the resource-related information includes working resources and idle resources; and the resource-related information of each computing power node is based on, and multiple first self-attention mechanisms are used to process the predicted fault computing power node matrix respectively to obtain multiple first attention vectors, including: determining the first attention vector by the following formula: Among them, H d1i1 is the first attention vector with the recognition direction d1 after processing by the i1th feedforward layer, P i1 is the working resource after processing by the i1th feedforward layer, M i1 T is the idle resource after processing by the i1th feedforward layer, E i1 is the predicted fault computing power node matrix after processing by the i1th feed-forward layer, T is used to represent the transposed matrix, and softmax represents classification.

[0011] Optionally, the processing the state matrix respectively by multiple second self-attention mechanisms in the fault identification model to obtain multiple second attention vectors includes: determining the second attention vector by the following formula: in, is the second attention vector with recognition direction d2 after processing by the i2th feedforward layer, The state matrix is ​​converted into a predetermined query matrix after being processed by the i2-th feedforward layer, The state matrix is ​​converted into a predetermined query matrix after being processed by the i2-th feedforward layer, The state matrix is ​​converted into a predetermined content matrix after being processed by the i2th feedforward layer, T is used to represent the transposed matrix, and softmax represents the classification.

[0012] Based on the same inventive concept, the present application also provides a resource scheduling device, including: a combination module, configured to obtain status data corresponding to each computing power node, and filter all status data, and combine the filtered status data into a status matrix; an output module, configured to input the status matrix into a pre-trained fault identification model, and output the predicted fault computing power node matrix corresponding to the status matrix through the fault identification model; a processing module, configured to input the predicted fault computing power node matrix into a pre-trained migration decision model, obtain resource-related information and pre-calculated current credibility of each computing power node through the migration decision model, and based on the resource-related information of each computing power node, use multiple first self-attention mechanisms to process the predicted fault computing power node matrix respectively to obtain multiple first attention vectors; a determination module, configured to determine a predicted resource scheduling strategy based on the multiple first attention vectors and all current credibility, and output the predicted resource scheduling strategy through the migration decision model; an execution module, configured to execute the predicted resource scheduling strategy.

[0013] Based on the same inventive concept, the present application also provides an electronic device, comprising a memory, a processor, and a computer program stored in the memory and executable by the processor, wherein the processor implements the method described above when executing the computer program.

[0014] Based on the same inventive concept, the present application also provides a non-transitory computer-readable storage medium, which stores computer instructions, and the computer instructions are used to enable a computer to execute the method as described above.

[0015] From the above description, it can be seen that the resource scheduling method, device, electronic device and storage medium provided by the present application include obtaining the status data corresponding to each computing power node, screening all the status data, and combining the status data that have passed the screening into a status matrix, thereby reducing the data volume of all the status data. The status matrix is ​​input into a pre-trained fault identification model, and the fault identification model outputs the predicted fault computing power node matrix corresponding to the status matrix. On the one hand, the status matrix processed by the fault identification model is the matrix obtained after screening all the status data, which reduces the data volume processed by the fault identification model and improves the processing efficiency of the fault identification model for the status matrix. On the other hand, the predicted fault computing power node matrix output by the fault identification model includes all computing power nodes that are about to fail, thereby improving the efficiency of predicting fault computing power nodes. The predicted fault computing power node matrix is ​​input into a pre-trained migration decision model, and the resource-related information and pre-calculated current credibility of each computing power node are obtained through the migration decision model. Based on the resource-related information of each computing power node, the predicted fault computing power node matrix is ​​processed separately using multiple first self-attention mechanisms to obtain multiple first attention vectors, which help the migration decision model better understand the contextual information in the predicted fault computing power node matrix, thereby more accurately processing the predicted fault computing power node matrix. Based on the multiple first attention vectors and all current credibility, the predicted resource scheduling strategy is determined, and the predicted resource scheduling strategy is output through the migration decision model, so that the output predicted resource scheduling strategy is more accurate and stable. The predicted resource scheduling strategy is executed so that the computing power node has active fault tolerance and ensures the smooth operation of the task. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] In order to more clearly illustrate the technical solutions in this application or related technologies, the following briefly introduces the drawings required for use in the embodiments or related technical descriptions. Obviously, the drawings described below are merely embodiments of this application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without any creative work.

[0017] Figure 1 A schematic diagram of a resource scheduling method according to an embodiment of the present application;

[0018] Figure 2 This is a schematic diagram of the structure of a resource scheduling device according to an embodiment of the present application;

[0019] Figure 3 This is a schematic diagram of the hardware structure of the electronic device according to an embodiment of the present application. DETAILED DESCRIPTION

[0020] In order to make the objectives, technical solutions and advantages of this application more clear, this application is further described in detail below in combination with specific embodiments and with reference to the accompanying drawings.

[0021] It should be noted that, unless otherwise defined, the technical terms or scientific terms used in the embodiments of the present application should have the usual meanings understood by people with ordinary skills in the field to which this application belongs. The "first", "second" and similar words used in the embodiments of the present application do not indicate any order, quantity or importance, but are only used to distinguish different components. "Include" or "comprise" and similar words mean that the elements or objects appearing before the word cover the elements or objects listed after the word and their equivalents, without excluding other elements or objects. "Connect" or "connected" and similar words are not limited to physical or mechanical connections, but may include electrical connections, whether direct or indirect. "Up", "down", "left", "right" and the like are only used to indicate relative positional relationships. When the absolute position of the described object changes, the relative positional relationship may also change accordingly.

[0022] As mentioned in the background technology section, with the development of mobile communication technology and artificial intelligence technology, the demand for computing power nodes in various technical businesses has exploded. For example, in the case of the power business, it relies on widely distributed grid computing power nodes to realize its various functions, including real-time monitoring, data analysis, intelligent control, etc. Grid computing power nodes are the brains of the power business. They are connected to power equipment, sensors, and data centers through the Internet to support the operation and management of the power business. However, the reliability of these nodes is crucial, because their failure may cause partial or complete failure of the power grid, seriously threatening the continuity and stability of the power supply.

[0023] Detecting and recovering from computing node failures has always been a significant challenge across various technology businesses. Due to the complexity and diversity of the systems underlying these businesses, traditional methods have struggled to meet the demand for fast and accurate computing node failure detection. Furthermore, when a computing node is detected to be about to fail, existing technology typically issues a fault warning in the hope that technicians can resolve the issue. However, the time it takes for technicians to identify and resolve the problem wastes human resources, inefficiently scheduling computing node resources, and potentially leading to partial or complete system failure due to untimely resource scheduling.

[0024] In view of this, the embodiment of the present application proposes a resource scheduling method, referring to Figure 1 , including the following steps:

[0025] Step 101: obtain the status data corresponding to each computing power node, filter all the status data, and combine the filtered status data into a status matrix.

[0026] In this step, the technical business relies on computing nodes to implement various functions. Once a computing node fails, the technical business will not be able to proceed smoothly. Therefore, it is very necessary to predict whether the computing node will fail. The status data corresponding to each computing node is obtained. Through the status data, the current status of the computing node can be analyzed, and then it can be predicted whether the computing node has failed. The status data includes at least CPU usage data, memory data, and disk IO data. For example, the source of status data can be smart homes, smart cars, factories, and other users with different needs for edge computing. If the computing node failure cannot be predicted in time, the computing node that is about to fail will fail to execute the task. To avoid the above problems, the present application can simultaneously determine whether all computing nodes are about to fail. However, the total status data corresponding to all computing nodes is relatively large. Therefore, all status data needs to be filtered and the filtered data is combined into a state matrix. By combining the state matrix, the amount of data of all state data is reduced.

[0027] It should be noted that all computing nodes are divided into different domains based on their geographic location, organization, and other status, forming a heterogeneous edge network. Computing nodes in the same domain collect service requests and data from users in the domain and provide services to them. Computing nodes in different domains do not exchange information.

[0028] Step 102: input the state matrix into a pre-trained fault identification model, and output the predicted fault computing power node matrix corresponding to the state matrix via the fault identification model.

[0029] In this step, the state matrix is ​​input into the pre-trained fault identification model. The fault identification model processes the state matrix to obtain and output the predicted fault computing power node matrix. On the one hand, the state matrix processed by the fault identification model is a matrix obtained after all state data are screened, which reduces the amount of data processed by the fault identification model and improves the processing efficiency of the fault identification model on the state matrix. On the other hand, the predicted fault computing power node matrix output by the fault identification model includes all computing power nodes that are about to fail, which improves the efficiency of predicting fault computing power nodes.

[0030] It should be noted that the loss function used to train the fault recognition model is determined by the following formula:

[0031]

[0032] Among them, Loss is the loss function of the fault identification model, y i3,j is the label of the i3th fault of computing node j. For the i3th fault, y i3,j =1 means the i3th fault will occur, y i3,j =0 means that the i3th fault will not occur, x i3,j is the recognition score of the i3th fault, e is the total number of all fault types, and m is the number of all computing nodes.

[0033] Step 103: Input the predicted fault computing power node matrix into a pre-trained migration decision model, obtain resource-related information and pre-calculated current credibility of each computing power node through the migration decision model, and based on the resource-related information of each computing power node, use multiple first self-attention mechanisms to process the predicted fault computing power node matrix separately to obtain multiple first attention vectors.

[0034] In this step, when it is predicted that a computing power node is about to fail, it is also necessary to schedule resources for the tasks being performed by the computing power node. In order to enable timely resource scheduling, after the fault identification model outputs the predicted fault computing power node matrix, the predicted fault computing power node matrix is ​​automatically input into the pre-trained migration decision model. In addition, in order to determine which specific computing power node to schedule resources to, it is also necessary to obtain the resource-related information and current credibility of each computing power node from the pre-built blockchain, wherein the resource-related information can reflect the processing resource capabilities of the computing power node, and the current credibility can reflect the reliability of the computing power node. Based on the resource-related information of each computing power node, a plurality of first self-attention mechanisms are used to process the predicted fault computing power node matrix separately to obtain multiple first attention vectors, wherein the weight coefficients of each first self-attention are different, and different weight coefficients consider the relationship between each element in the predicted fault computing power node matrix and all other elements from different directions. The first self-attention mechanism allows the migration decision model to consider the relationship between each element in the predicted fault computing power node matrix and all other elements when processing the predicted fault computing power node matrix. This helps the migration decision model better understand the contextual information in the predicted fault computing power node matrix, thereby processing the predicted fault computing power node matrix more accurately.

[0035] It should be noted that the migration decision model includes a discriminator and a generator. During the training of the migration decision model, the discriminator and the generator form an adversarial network. The cross-entropy loss function will be used to optimize the weights of the generator and the discriminator, respectively, to encourage the generator to generate better migration decisions, and the discriminator to more accurately identify good decisions. Run the simulator to simulate the system operation after taking different resource scheduling decisions and obtain the predicted resource scheduling strategy score Sim(M) and the current resource scheduling strategy score Sim(S). When the discriminator outputs an improved decision, that is, D[0]=0, D[1]=1, it is necessary to ensure that Sim(M)≥Sim(S). Similarly, when the discriminator outputs the original decision, that is, D[0]=1, D[1]=0, it is necessary to ensure that Sim(M)<Sim(S). Based on this, the cross-entropy loss function L of the discriminator can be obtained. D , and based on this loss function, the discriminator weight is updated. The loss function used to train the discriminator of the fault recognition model is determined by the following formula:

[0036] L D =(Sim(M)≥Sim(S))(log(D[1])+log(1-D[0]))+(Sim(M)<Sim(S))(log(D[0])+log(1-D[1])) (2),

[0037] The above loss function indicates that when Sim(M) ≥ Sim(S), log(D[1]) needs to be as large as possible and log(D[0]) needs to be as small as possible. This is achieved by making log(1-D[0]) as large as possible. When Sim(M) < Sim(S), log(D[1]) needs to be as small as possible and log(D[0]) needs to be as large as possible. This is achieved by making log(1-D[1]) as large as possible. After the discriminator is trained, it needs to be fixed.

[0038] If the generator needs to generate better migration decisions, that is, to try to ensure that D[0] = 0 and D[1] = 1, the loss function used to train the generator of the fault recognition model is determined by the following formula:

[0039] L G =log(D[1])+log(1-D[0])(3), where the above loss function means: ensuring that log(D[1]) is as large as possible and log(D[0]) is as small as possible, and the smallness of log(D[0]) is reflected by making log(1-D[0]) as large as possible. The weights of the generator can be updated according to the above loss function.

[0040] It's also important to note that to ensure that tasks on computing nodes predicted to be about to fail are fully executed, cross-domain resource scheduling can effectively meet the fault tolerance requirements of nodes in resource-constrained domains. However, cross-domain resource scheduling requires consideration of inter-domain trust within heterogeneous edge networks. To this end, a blockchain must be pre-built to ensure trusted collaboration among computing nodes across domains. Computing nodes from different domains form a blockchain, enabling the secure and efficient migration of tasks between computing nodes. When a computing node joins the network, it registers its identity in the blockchain and regularly uploads its stored data summaries, operational status, and other information to the intelligent agent. Edge nodes within the same domain can share information and maintain mutual trust. Furthermore, the blockchain regularly evaluates the reputation of each computing node based on its performance, effectively identifying and eliminating malicious computing nodes. The blockchain fully records the identity, operational status, reputation, and resource scheduling decisions of computing nodes. As a distributed ledger, the information stored in the blockchain is jointly maintained by all computing nodes and cannot be tampered with, ensuring mutual trust among heterogeneous computing nodes.

[0041] Step 104 : Determine a predicted resource scheduling strategy based on the multiple first attention vectors and all current credibility scores, and output the predicted resource scheduling strategy via the migration decision model.

[0042] In this step, multiple first attention vectors more comprehensively reflect the characteristics of the predicted faulty computing node matrix in different directions, while the overall current reputation reflects the reliability of the computing node. Therefore, the migration decision model determines and outputs a predicted resource scheduling strategy based on multiple first attention vectors and the overall current reputation, making the output predicted resource scheduling strategy more accurate and stable.

[0043] Step 105: Execute the predicted resource scheduling strategy.

[0044] In this step, if a computing node is predicted to fail, a predicted resource scheduling strategy is determined and executed to prevent the computing node from being unable to execute tasks. This makes the computing node proactively fault-tolerant and ensures the smooth operation of tasks. It should be noted that when a computing node receives a resource scheduling decision, it locks its state on the blockchain and migrates its task data based on the resource scheduling decision.

[0045] Through the above scheme, the state data corresponding to each computing power node is obtained, and all state data are filtered, and the filtered state data are combined into a state matrix, thereby reducing the data volume of all state data. The state matrix is ​​input into a pre-trained fault identification model, and the fault identification model outputs the predicted fault computing power node matrix corresponding to the state matrix. On the one hand, the state matrix processed by the fault identification model is the matrix obtained after all state data are filtered, which reduces the amount of data processed by the fault identification model and improves the processing efficiency of the fault identification model for the state matrix. On the other hand, the predicted fault computing power node matrix output by the fault identification model includes all computing power nodes that are about to fail, thereby improving the efficiency of predicting fault computing power nodes. The predicted fault computing power node matrix is ​​input into a pre-trained migration decision model, and the migration decision model obtains the resource-related information and pre-calculated current reputation of each computing power node. Based on the resource-related information of each computing power node, the predicted fault computing power node matrix is ​​processed separately using multiple first self-attention mechanisms to obtain multiple first attention vectors, which helps the migration decision model better understand the contextual information in the predicted fault computing power node matrix, thereby more accurately processing the predicted fault computing power node matrix. Based on the multiple first attention vectors and all current reputations, a predicted resource scheduling strategy is determined, and the predicted resource scheduling strategy is output via the migration decision model, so that the output predicted resource scheduling strategy is more accurate and stable. Executing the predicted resource scheduling strategy enables the computing power node to have active fault tolerance and ensure the smooth operation of the task.

[0046] In some embodiments, the filtering of all state data and combining the filtered state data into a state matrix include: classifying all state data according to predetermined categories, and for each category of state data, querying and determining the number of groups corresponding to the category in a pre-constructed storage table according to the category; grouping the state data corresponding to the category according to the number of groups to obtain multiple groups of state data corresponding to the category; calculating the information entropy of the state data corresponding to all categories based on the number of state data corresponding to the category and the number of all state data; calculating the information gain corresponding to each group of state data based on the information entropy, the number of state data corresponding to the category and the number of each group of state data; and filtering out a group of state data corresponding to the maximum information gain in each category from all state data, and combining the multiple groups of state data obtained by filtering into a state matrix.

[0047] In this embodiment, due to the large volume and multi-dimensionality of all status data, representative data from all status data can be screened to improve fault detection efficiency while ensuring the accuracy of identifying the faulty computing node. First, all status data is categorized according to predetermined categories. For example, the predetermined categories may include CPU usage, memory, and disk I / O. The status data in each category can be further divided into multiple groups. Based on the category, a pre-built storage table is queried to determine the number of groups corresponding to that category. For example, if the category is CPU usage, the CPU usage category can be divided into two subcategories: system CPU usage subcategories and user CPU usage subcategories. The number of groups corresponding to the CPU usage category is two. If the category is memory, the memory category can be categorized by usage into primary memory subcategories and auxiliary memory subcategories. The number of groups corresponding to the memory category is two. If the category is disk I / O, the disk I / O category can be divided into high-level disk I / O subcategories and low-level disk I / O subcategories. The number of groups corresponding to the disk I / O category is two. The status data corresponding to each category is grouped according to the number of groups, resulting in multiple groups of status data corresponding to that category. The information entropy of the state data corresponding to all categories is calculated using the following formula:

[0048] Among them, the state data set of all computing power nodes is I, and D is all state data. This set can be divided into K categories, and the probability of the kth category is |f k | / |D|, where |f k | is the number of elements in the kth category.

[0049] The information gain corresponding to each set of state data is calculated using the following formula:

[0050] Among them, IG(C,D) is the information gain. There are n subcategories in the state data of each category, and the information entropy of subcategory C is E(C,D).

[0051] The greater the information gain, the greater the reduction in uncertainty in the data set, and the greater the amount of information contained in the data. The set of state data corresponding to the maximum information gain in each category is selected, and the resulting sets of state data are combined into a state matrix. Compared to the entire state data set, the state matrix contains relatively less data and more information, making it representative.

[0052] In some embodiments, determining the predicted resource scheduling strategy based on the multiple first attention vectors and all current credibility includes: splicing the multiple first attention vectors to obtain a first splicing vector; and normalizing the first splicing vector based on all current credibility to obtain the predicted resource scheduling strategy.

[0053] In this embodiment, because multiple attention vector coefficients reflect different perspectives on the relationship between each element and all other elements in the predicted fault computing node matrix, multiple first attention vectors are concatenated to produce a first concatenated vector. This first concatenated vector therefore reflects a comprehensive consideration of the relationship between each element and all other elements in the predicted fault computing node matrix, comprehensively exploring the relationships between computing nodes in the predicted fault computing node matrix. Credibility can be used to reflect the reliability of computing nodes. Based on all current credibility scores, the first concatenated vector is normalized to perform a linear mapping on it, resulting in an accurate and reliable prediction resource scheduling strategy.

[0054] In some embodiments, the fault identification model includes a multi-layer perceptron layer; inputting the state matrix into a pre-trained fault identification model and outputting the predicted fault computing power node matrix corresponding to the state matrix through the fault identification model includes: inputting the state matrix into the fault identification model, processing the state matrix separately through multiple second self-attention mechanisms in the fault identification model to obtain multiple second attention vectors; splicing the multiple second attention vectors to obtain a second splicing vector; performing abnormality identification on the second splicing vector through the multi-layer perceptron layer to obtain the predicted fault computing power node matrix, and outputting the predicted fault computing power node matrix through the fault identification model.

[0055] In this embodiment, the fault identification model processes the state matrix separately using multiple second self-attention mechanisms to obtain multiple second attention vectors. Each second self-attention mechanism has different weight coefficients, each considering the relationship between each element in the state matrix and all other elements from different perspectives. The second self-attention mechanism allows the fault identification model to consider the relationship between each element in the state matrix and all other elements when processing the state matrix, helping the fault identification model better understand the contextual information in the state matrix and thus process the state matrix more accurately. Because the multiple attention vector coefficients reflect the relationship between each element in the state matrix and all other elements from different perspectives, the multiple second attention vectors are concatenated to obtain a second concatenated vector. The second concatenated vector reflects the relationship between each element in the state matrix and all other elements from all perspectives, comprehensively mining the relationships between the state data in the state matrix. The second concatenated vector is used to identify anomalies using a multilayer perceptron layer. Due to continuous iterative training, the multilayer perceptron can automatically learn the complex relationships between input features to obtain a predicted fault computing power node matrix, thereby making the determined predicted fault computing power node matrix more accurate.

[0056] It should be noted that the Multilayer Perceptron (MLP) is a deep learning model based on a feedforward neural network. It consists of multiple neuron layers, each of which is fully connected to the previous layer. Multilayer Perceptrons can be used to solve various machine learning problems such as classification, regression, and clustering. Each neuron layer of the Multilayer Perceptron is composed of many neurons, where the input layer receives input features, the output layer gives the final prediction results, and the intermediate hidden layer is used to extract features and perform nonlinear transformations. Each neuron receives the output of the previous layer, performs a weighted sum and activation function operation, and obtains the output of the current layer. Through continuous iterative training, the Multilayer Perceptron can automatically learn the complex relationships between input features and make predictions.

[0057] In some embodiments, calculating the current credibility includes: obtaining the previous scheduling behavior and historical scheduling behavior scores of each computing power node and its current neighbors, the historical service quality, and the stability of the previous environment; for each computing power node, determining the previous scheduling behavior score based on the previous scheduling behavior and the historical scheduling behavior score, determining the previous service quality score based on the historical service quality, and determining the stability score of the previous environment based on the stability of the previous environment; determining the current credibility based on the previous scheduling behavior score, the previous service quality score, the stability score of the previous environment, and a preset weight coefficient.

[0058] In this embodiment, the blockchain will periodically evaluate the reputation of each computing power node. Therefore, the current reputation is determined by the previous data adjacent to the current one. The reputation determination includes the following three parts:

[0059] (1) Behavior evaluation: This part of the evaluation is mainly used to evaluate the performance of the computing power node during the resource migration process, and to punish malicious behaviors during the resource migration process, such as uploading the wrong status after resource migration, sending wrong data, etc. OK. Among them, Represents other computing power nodes at t k The trust level of computing power node n at any time, Represents other computing power nodes at t k The degree of distrust in computing power node n at any given moment, represents the uncertainty of other computing nodes about whether n is trustworthy, and t k The time is the time after the previous scheduling behavior adjacent to the current one is completed. The above parameters can be calculated by formula (7):

[0060]

[0061] Among them, the previous scheduling behaviors adjacent to the current one include normal behaviors and malicious behaviors. and They represent the computing power node n in time slot t k Normal and malicious behaviors during resource migration, Represents the time slot t k The number of computing nodes that interact with computing node n, Represents the time slot t k The total number of computing nodes at time t, the fewer resource migration transactions related to computing node n, the higher the uncertainty about n. k The behavioral evaluation score can be described as:

[0062]

[0063] Among them, α n ∈[0,1] represents the impact of uncertainty on the reputation of computing nodes.

[0064] Finally, the behavior evaluation of computing power node n can be obtained by t k The behavior evaluation at the moment and the historical scheduling behavior score are comprehensively determined:

[0065]

[0066] Among them, β n ∈[0,1] represents the attenuation factor of the behavior evaluation, that is, the newer the behavior evaluation, the greater the overall impact.

[0067] (2) Service Evaluation: This section is primarily used to evaluate the historical service quality provided by the computing power node to the user. After the service is completed, the user can leave a comment on their satisfaction with the service. This satisfaction is used to determine the previous service quality score. For the sake of fairness, users cannot determine whether the service was provided by the computing power node in their own domain or in another domain.

[0068] Let user i4 be in time slot t k The score of the previous historical service quality of the computing power node n and the current adjacent node is It can be calculated by the following formula:

[0069]

[0070]

[0071]

[0072] in, Represents user i4 in time slot t k Satisfaction with computing power node n and Represents user i4 in time slot t k The value of the service required, the more important the service, the greater the impact of satisfaction on the overall score. represents the deviation of the above value from the mean of all historical values ​​except the above value. θ is a constant used to control the influence of the value. Represents computing power node i5 in time slot t k The historical cumulative value of the value deviation can be calculated by formula (12), where γ is the attenuation parameter of the historical value deviation. By changing γ, the impact of the current value deviation on the overall score can be changed, thereby reducing the impact of a single extremely high value or extremely low value on the user score.

[0073] Assume that in time slot t k Total The services of users are completed by computing node n, where the 1st to pth services are requests from users in other domains, and the p+1th to Nth services are requests from users in other domains. user A service is a task requested by users in its own domain, where p∈[1,N user ], then computing power node n is in time slot t k The overall user rating is:

[0074]

[0075] Among them, λ∈[0,1] represents the weight of the rating of users in other domains. Generally, a value greater than 0.5 is assigned to λ, which means that users in other domains are given a greater weight to prevent computing power nodes from treating services from other domains negatively.

[0076] (III) Environmental evaluation: This part is used to evaluate the stability of the environment where the computing power node is located. If the service data is frequently hijacked, lost, or has other communication problems during the migration of resources to a computing power node, we will reduce the reputation of these computing power nodes and reduce the services migrated to them to reduce the losses caused by communication problems. k The stability score of the environment can be expressed as:

[0077]

[0078] in, For other computing power node i5 in time slot t k The number of times service data is sent to computing power node n, The number of times communication problems occurred.

[0079] Finally, computing node n in time slot t k The current reputation of is:

[0080]

[0081] in The weights of the previous dispatch behavior score, the previous service quality score, and the previous environment stability score are respectively: and

[0082] Reputation ratings will influence resource scheduling decisions, and scheduling resources from low-reputation computing nodes should be avoided. Current reputation is determined based on the previous scheduling behavior score, previous service quality score, previous environment stability score, and a preset weight coefficient. This considers a relatively comprehensive range of factors affecting credibility, making the determined reputation more accurate.

[0083] In some embodiments, the resource-related information includes working resources and idle resources; the resource-related information based on each computing power node is processed using multiple first self-attention mechanisms to process the predicted fault computing power node matrix respectively to obtain multiple first attention vectors, including: determining the first attention vector using the following formula: Among them, H d1i1 is the first attention vector with the recognition direction d1 after processing by the i1th feedforward layer, P i1 is the working resource after processing by the i1th feedforward layer, M i1 Tis the idle resource after processing by the i1th feedforward layer, E i1 is the predicted fault computing power node matrix after processing by the i1th feed-forward layer, T is used to represent the transposed matrix, and softmax represents classification.

[0084] In this embodiment, the feedforward layer can extract the complex features of the predicted fault computing power node matrix through linear transformation and activation function, which helps the migration decision model better understand the predicted fault computing power node matrix. In addition, when determining the first attention vector, the working resources and idle resources in the computing power node are taken into account, which helps the migration decision model determine a more accurate resource scheduling strategy. By increasing the number of hidden units in the feedforward layer, the complexity of the migration decision model can be increased, the expressive power of the migration decision model can be improved, and the first attention vector determined by the migration decision model can be made more accurate.

[0085] In some embodiments, the processing of the state matrix by multiple second self-attention mechanisms in the fault identification model to obtain multiple second attention vectors includes: determining the second attention vector by the following formula: in, is the second attention vector with recognition direction d2 after processing by the i2th feedforward layer, The state matrix is ​​converted into a predetermined query matrix after being processed by the i2-th feedforward layer, The state matrix is ​​converted into a predetermined query matrix after being processed by the i2-th feedforward layer, The state matrix is ​​converted into a predetermined content matrix after being processed by the i2th feedforward layer, T is used to represent the transposed matrix, and softmax represents the classification.

[0086] In this embodiment, the feedforward layer extracts complex features of the state matrix through linear transformations and activation functions, helping the fault identification model better understand the state matrix. Furthermore, the predetermined query matrix, predetermined query matrix, and predetermined content matrix are derived from the state matrix, expressing the state matrix from different perspectives. The predetermined query matrix is ​​multiplied by the transpose of the predetermined query matrix and then divided by the square root of dim to prevent the vanishing gradient problem. The resulting attention matrix is ​​then passed through a softmax layer and finally multiplied by the predetermined content matrix to obtain the final second attention vector, making the second attention vector determined by the fault identification model more accurate.

[0087] It should be noted that the method of the embodiment of the present application can be performed by a single device, such as a computer or server. The method of this embodiment can also be applied in a distributed scenario and performed by multiple devices working together. In such a distributed scenario, one of the multiple devices may only perform one or more steps of the method of the embodiment of the present application, and the multiple devices will interact with each other to complete the method.

[0088] It should be noted that the above description is limited to some embodiments of the present application. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recited in the claims may be performed in an order different from that described in the above embodiments and still achieve the desired results. Furthermore, the processes depicted in the accompanying drawings do not necessarily require the specific order or sequential order shown to achieve the desired results. In certain embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0089] Based on the same inventive concept, corresponding to any of the above-mentioned embodiment methods, the present application also provides a resource scheduling device.

[0090] refer to Figure 2 , the resource scheduling device includes:

[0091] The combining module 10 is configured to obtain the status data corresponding to each computing power node, filter all the status data, and combine the status data that pass the filtering into a status matrix.

[0092] The output module 20 is configured to input the state matrix into a pre-trained fault identification model, and output a predicted fault computing power node matrix corresponding to the state matrix via the fault identification model.

[0093] The processing module 30 is configured to input the predicted fault computing power node matrix into a pre-trained migration decision model, obtain resource-related information and pre-calculated current credibility of each computing power node through the migration decision model, and based on the resource-related information of each computing power node, use multiple first self-attention mechanisms to process the predicted fault computing power node matrix separately to obtain multiple first attention vectors.

[0094] The determination module 40 is configured to determine a predicted resource scheduling strategy based on the multiple first attention vectors and all current credibility, and output the predicted resource scheduling strategy via the migration decision model.

[0095] The execution module 50 is configured to execute the predicted resource scheduling strategy.

[0096] Through the above-mentioned device, the state data corresponding to each computing power node is obtained, and all state data are filtered. The filtered state data are combined into a state matrix, thereby reducing the data volume of all state data. The state matrix is ​​input into a pre-trained fault identification model, and the fault identification model outputs the predicted fault computing power node matrix corresponding to the state matrix. On the one hand, the state matrix processed by the fault identification model is the matrix obtained after all state data are filtered, which reduces the amount of data processed by the fault identification model and improves the processing efficiency of the fault identification model for the state matrix. On the other hand, the predicted fault computing power node matrix output by the fault identification model includes all computing power nodes that are about to fail, thereby improving the efficiency of predicting fault computing power nodes. The predicted fault computing power node matrix is ​​input into a pre-trained migration decision model. The migration decision model obtains the resource-related information and pre-calculated current reputation of each computing power node. Based on the resource-related information of each computing power node, the predicted fault computing power node matrix is ​​processed separately using multiple first self-attention mechanisms to obtain multiple first attention vectors, which help the migration decision model better understand the contextual information in the predicted fault computing power node matrix, thereby more accurately processing the predicted fault computing power node matrix. Based on the multiple first attention vectors and all current reputations, a predicted resource scheduling strategy is determined, and the predicted resource scheduling strategy is output via the migration decision model, so that the output predicted resource scheduling strategy is more accurate and stable. Executing the predicted resource scheduling strategy enables the computing power node to have active fault tolerance and ensure the smooth operation of the task.

[0097] In some embodiments, the combination module 10 is further configured to classify all state data according to predetermined categories, and for each category of state data, query and determine the number of groups corresponding to the category in a pre-constructed storage table according to the category; group the state data corresponding to the category according to the number of groups to obtain multiple groups of state data corresponding to the category; calculate the information entropy of the state data corresponding to all categories based on the number of state data corresponding to the category and the number of all state data; calculate the information gain corresponding to each group of state data based on the information entropy, the number of state data corresponding to the category and the number of each group of state data; and screen out a group of state data corresponding to the maximum information gain in each category from all the state data, and combine the multiple groups of state data obtained by screening into a state matrix.

[0098] In some embodiments, the determination module 40 is further configured to splice the multiple first attention vectors to obtain a first splicing vector; and normalize the first splicing vector based on all current credibility to obtain the predicted resource scheduling strategy.

[0099] In some embodiments, the output module 20 is further configured so that the fault identification model includes a multi-layer perceptron layer; the state matrix is ​​input into the fault identification model, and the state matrix is ​​processed separately by multiple second self-attention mechanisms in the fault identification model to obtain multiple second attention vectors; the multiple second attention vectors are spliced ​​to obtain a second splicing vector; the second splicing vector is subjected to abnormality identification through the multi-layer perceptron layer to obtain the predicted fault computing power node matrix, and the predicted fault computing power node matrix is ​​output via the fault identification model.

[0100] In some embodiments, a calculation module is also included, which is configured to obtain the previous scheduling behavior and historical scheduling behavior scores of each computing power node and its current neighbors, the historical service quality, and the stability of the previous environment; for each computing power node, the previous scheduling behavior score is determined based on the previous scheduling behavior and the historical scheduling behavior score, the previous service quality score is determined based on the historical service quality, and the stability score of the previous environment is determined based on the stability of the previous environment; the current credibility is determined based on the previous scheduling behavior score, the previous service quality score, the stability score of the previous environment, and a preset weight coefficient.

[0101] In some embodiments, the processing module 30 is further configured to: the resource-related information includes working resources and idle resources; determine the first attention vector by the following formula: Among them, H d1i1 is the first attention vector with the recognition direction d1 after processing by the i1th feedforward layer, P i1 is the working resource after processing by the i1th feedforward layer, M i1 T is the idle resource after processing by the i1th feedforward layer, E i1 is the predicted fault computing power node matrix after processing by the i1th feed-forward layer, T is used to represent the transposed matrix, and softmax represents classification.

[0102] In some embodiments, the output module 20 is further configured to determine the second attention vector using the following formula: in, is the second attention vector with recognition direction d2 after processing by the i2th feedforward layer, The state matrix is ​​converted into a predetermined query matrix after being processed by the i2-th feedforward layer, The state matrix is ​​converted into a predetermined query matrix after being processed by the i2-th feedforward layer, The state matrix is ​​converted into a predetermined content matrix after being processed by the i2th feedforward layer, T is used to represent the transposed matrix, and softmax represents the classification.

[0103] For the convenience of description, the above devices are described as being divided into various modules according to their functions. Of course, when implementing this application, the functions of each module can be implemented in the same or multiple software and / or hardware.

[0104] The apparatus of the above embodiment is used to implement the corresponding resource scheduling method in any of the above embodiments, and has the beneficial effects of the corresponding method embodiment, which will not be described in detail here.

[0105] Based on the same inventive concept, corresponding to any of the above-mentioned embodiments and methods, the present application also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and runnable on the processor, wherein when the processor executes the program, the resource scheduling method described in any of the above embodiments is implemented.

[0106] Figure 3 10 is a schematic diagram showing a more specific hardware structure of an electronic device provided in this embodiment. The device may include: a processor 1010, a memory 1020, an input / output interface 1030, a communication interface 1040, and a bus 1050. The processor 1010, the memory 1020, the input / output interface 1030, and the communication interface 1040 are communicatively connected to each other within the device via the bus 1050.

[0107] The processor 1010 can be implemented using a general-purpose CPU (Central Processing Unit), a microprocessor, an application-specific integrated circuit (ASIC), or one or more integrated circuits, and is used to execute relevant programs to implement the technical solutions provided in the embodiments of this specification.

[0108] The memory 1020 can be implemented in the form of ROM (Read Only Memory), RAM (Random Access Memory), static storage devices, dynamic storage devices, etc. The memory 1020 can store an operating system and other application programs. When the technical solutions provided in the embodiments of this specification are implemented through software or firmware, the relevant program code is stored in the memory 1020 and is called and executed by the processor 1010.

[0109] The input / output interface 1030 is used to connect an input / output module to implement information input and output. The input / output module can be configured as a component in the device (not shown in the figure) or can be externally connected to the device to provide corresponding functions. Input devices may include a keyboard, mouse, touch screen, microphone, various sensors, etc., and output devices may include a display, speaker, vibrator, indicator light, etc.

[0110] The communication interface 1040 is used to connect to a communication module (not shown) to enable communication between the device and other devices. The communication module can communicate via a wired method (such as USB, network cable, etc.) or a wireless method (such as mobile network, WiFi, Bluetooth, etc.).

[0111] The bus 1050 comprises a path for transmitting information between the various components of the device (eg, the processor 1010 , the memory 1020 , the input / output interface 1030 , and the communication interface 1040 ).

[0112] It should be noted that although the above device only shows the processor 1010, the memory 1020, the input / output interface 1030, the communication interface 1040, and the bus 1050, in a specific implementation, the device may also include other components necessary for normal operation. In addition, it will be understood by those skilled in the art that the above device may only include the components necessary to implement the embodiments of this specification, and does not necessarily include all the components shown in the figure.

[0113] The electronic device of the above embodiment is used to implement the corresponding resource scheduling method in any of the above embodiments, and has the beneficial effects of the corresponding method embodiment, which will not be described in detail here.

[0114] Based on the same inventive concept, corresponding to any of the above-mentioned embodiment methods, the present application also provides a non-transitory computer-readable storage medium, wherein the non-transitory computer-readable storage medium stores computer instructions, and the computer instructions are used to enable the computer to execute the resource scheduling method described in any of the above embodiments.

[0115] The computer-readable media of this embodiment include permanent and non-permanent, removable and non-removable media that can be used to store information by any method or technology. The information can be computer-readable instructions, data structures, program modules or other data. Examples of computer storage media include, but are not limited to, phase change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technology, read-only compact disc read-only memory (CD-ROM), digital versatile disc (DVD) or other optical storage, magnetic cassettes, tape disk storage or other magnetic storage devices or any other non-transmission media that can be used to store information that can be accessed by a computing device.

[0116] The computer instructions stored in the storage medium of the above embodiment are used to enable the computer to execute the resource scheduling method described in any of the above embodiments, and have the beneficial effects of the corresponding method embodiments, which will not be repeated here.

[0117] Those skilled in the art should understand that the discussion of any of the above embodiments is merely illustrative and is not intended to imply that the scope of the present application (including the claims) is limited to these examples. Within the scope of the present application, the technical features in the above embodiments or different embodiments may be combined, the steps may be implemented in any order, and there are many other variations of the different aspects of the embodiments of the present application as described above, which are not provided in detail for the sake of simplicity.

[0118] In addition, for simplicity of description and discussion, and in order not to make the embodiment of the application difficult to understand, the known power supply / ground connection with integrated circuit (IC) chip and other components may or may not be shown in the accompanying drawings provided. In addition, the device can be shown in the form of a block diagram to avoid making the embodiment of the application difficult to understand, and this also takes into account the following fact, that is, the details of the embodiment of these block diagram devices are highly dependent on the platform to be implemented in the embodiment of the application (that is, these details should be fully within the scope of understanding of those skilled in the art). When specific details (for example, circuit) are set forth to describe exemplary embodiments of the application, it will be apparent to those skilled in the art that the embodiment of the application can be implemented without these specific details or when these specific details are changed. Therefore, these descriptions should be considered to be illustrative rather than restrictive.

[0119] Although the present invention has been described in conjunction with specific embodiments thereof, many alternatives, modifications, and variations of these embodiments will be apparent to those skilled in the art based on the foregoing description. For example, other memory architectures (e.g., dynamic RAM (DRAM)) may utilize the embodiments discussed.

[0120] The embodiments of the present application are intended to cover all such substitutions, modifications, and variations that fall within the broad scope of the appended claims. Therefore, any omissions, modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the embodiments of the present application should be included in the scope of protection of this application.

Claims

1. A resource scheduling method, characterized in that: include: Obtain the status data corresponding to each computing power node, filter all the status data, and combine the filtered status data into a status matrix; Input the state matrix into a pre-trained fault identification model, and output the predicted fault computing power node matrix corresponding to the state matrix through the fault identification model; Inputting the predicted fault computing power node matrix into a pre-trained migration decision model, obtaining resource-related information and pre-calculated current reputation of each computing power node through the migration decision model, and processing the predicted fault computing power node matrix based on the resource-related information of each computing power node using multiple first self-attention mechanisms to obtain multiple first attention vectors; determining a predicted resource scheduling strategy based on the multiple first attention vectors and all current credibility scores, and outputting the predicted resource scheduling strategy via the migration decision model; executing the predicted resource scheduling strategy; The current reputation is calculated by the following steps: Obtain the previous and historical scheduling behavior scores of each computing power node and its neighbors, as well as the historical service quality and the stability of the previous environment; For each computing power node, based on the previous scheduling behavior and the historical scheduling behavior score, determine the previous scheduling behavior score; based on the historical service quality, determine the previous service quality score; based on the stability of the previous environment, determine the previous environment stability score; The current credibility is determined based on the previous scheduling behavior score, the previous service quality score, the previous environment stability score, and a preset weight coefficient.

2. The method according to claim 1, characterized in that The filtering of all state data and combining the filtered state data into a state matrix includes: Classify all status data according to predetermined categories, and for each category of status data, query and determine the number of groups corresponding to the category in a pre-built storage table according to the category; Grouping the status data corresponding to the category according to the number of groups to obtain multiple groups of status data corresponding to the category; Calculating the information entropy of the state data corresponding to all categories based on the number of state data corresponding to the category and the number of all state data; Based on the information entropy, the number of state data corresponding to the category and the number of each group of state data, the information gain corresponding to each group of state data is calculated; among all the state data, a group of state data corresponding to the maximum information gain in each category is screened out respectively, and the multiple groups of state data obtained by screening are combined into a state matrix.

3. The method according to claim 1, characterized in that The step of determining a prediction resource scheduling strategy based on the multiple first attention vectors and all current credibility scores includes: Splicing the multiple first attention vectors to obtain a first splicing vector; Based on all current credibility, the first splicing vector is normalized to obtain the predicted resource scheduling strategy.

4. The method according to claim 1, wherein The fault identification model includes a multi-layer perception machine layer; Inputting the state matrix into a pre-trained fault identification model and outputting a predicted fault computing power node matrix corresponding to the state matrix via the fault identification model includes: Inputting the state matrix into the fault recognition model, and processing the state matrix respectively through multiple second self-attention mechanisms in the fault recognition model to obtain multiple second attention vectors; Splicing the multiple second attention vectors to obtain a second splicing vector; The second splicing vector is subjected to abnormality identification through the multi-layer perceptron layer to obtain the predicted fault computing power node matrix, and the predicted fault computing power node matrix is ​​output through the fault identification model.

5. The method according to claim 1, wherein The resource-related information includes working resources and idle resources; Based on the resource-related information of each computing power node, the predicted fault computing power node matrix is ​​processed respectively using multiple first self-attention mechanisms to obtain multiple first attention vectors, including: The first attention vector is determined by the following formula: Among them, H d1i1 is the first attention vector with the recognition direction d1 after processing by the i1th feedforward layer, P i1 is the working resource after processing by the i1th feedforward layer, M i1 T is the idle resource after processing by the i1th feedforward layer, E i1 is the predicted fault computing power node matrix after processing by the i1th feed-forward layer, T is used to represent the transposed matrix, and softmax represents classification.

6. The method according to claim 1, characterized in that The state matrix is ​​processed respectively by multiple second self-attention mechanisms in the fault identification model to obtain multiple second attention vectors, including: The second attention vector is determined by the following formula: in, is the second attention vector with recognition direction d2 after processing by the i2th feedforward layer, The state matrix is ​​converted into a predetermined query matrix after being processed by the i2-th feedforward layer, The state matrix is ​​converted into a predetermined query matrix after being processed by the i2-th feedforward layer, The state matrix is ​​converted into a predetermined content matrix after being processed by the i2th feedforward layer, T is used to represent the transposed matrix, and softmax represents the classification.

7. A resource scheduling device, characterized in that: include: a combining module configured to obtain state data corresponding to each computing power node, filter all state data, and combine the filtered state data into a state matrix; an output module configured to input the state matrix into a pre-trained fault identification model, and output a predicted fault computing power node matrix corresponding to the state matrix via the fault identification model; a processing module configured to input the predicted fault computing power node matrix into a pre-trained migration decision model, obtain resource-related information and a pre-calculated current reputation of each computing power node through the migration decision model, and process the predicted fault computing power node matrix based on the resource-related information of each computing power node using multiple first self-attention mechanisms to obtain multiple first attention vectors; a determination module configured to determine a predicted resource scheduling strategy based on the multiple first attention vectors and all current credibility scores, and output the predicted resource scheduling strategy via the migration decision model; an execution module, configured to execute the predicted resource scheduling strategy; A calculation module is configured to obtain the previous scheduling behavior and historical scheduling behavior scores of each computing power node and its current neighbors, historical service quality, and stability of the previous environment; For each computing power node, based on the previous scheduling behavior and the historical scheduling behavior score, determine the previous scheduling behavior score; based on the historical service quality, determine the previous service quality score; based on the stability of the previous environment, determine the previous environment stability score; The current credibility is determined based on the previous scheduling behavior score, the previous service quality score, the previous environment stability score, and a preset weight coefficient.

8. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the program, the method according to any one of claims 1 to 6 is implemented.

9. A non-transitory computer-readable storage medium storing computer instructions, characterized in that: The computer instructions are used to enable a computer to execute the method according to any one of claims 1 to 6.

Images