Dynamic computing power scheduling method and device based on deep learning and medium

Abstract

The embodiment of the invention discloses a dynamic computing power scheduling method and device based on deep learning and a medium, belongs to the technical field of cloud computing, and solves the problem that cloud computing resources are seriously wasted due to the fact that a cloud computing scheduling method in the prior art is prone to unbalanced resource allocation. Comprising the following steps: preprocessing acquired cloud computing data to generate a structured feature vector; wherein the cloud computing data at least comprises cloud node state data, computing task attributes and environment data; through a multi-head attention mechanism, task resource space-time correlation features corresponding to the structured feature vectors are extracted; the task resource space-time correlation features are input into a weight dynamic prediction module and a resource matching degree scoring module at the same time so as to carry out parallel processing of weight prediction and matching degree scoring calculation; and solving a multi-objective optimization function corresponding to computing power scheduling based on a parallel processing result, and generating a computing power scheduling instruction based on a solving result.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of cloud computing, and particularly relates to a dynamic computing power scheduling method based on deep learning, a device and a medium. BACKGROUND

[0002] Cloud computing refers to a shared resource pool formed by integrating a large number of distributed computing resources through a network, and users can obtain computing, storage, network and other services on demand. In the cloud computing environment, the scheduling efficiency of computing resources directly affects the service quality and operating cost.

[0003] Traditional static scheduling based on predefined rules allocates resources according to fixed rules such as node affinity and resource request amount. For example, according to the preference of a task for a node or the current resource idle amount of the node, the task is allocated to a node that meets the rules. However, when facing a sudden traffic, due to the lack of prediction ability, only passive expansion can be triggered, and the response delay time is long, resulting in task accumulation.

[0004] Secondly, the existing technology usually uses the expansion and contraction scheduling based on threshold triggering, which sets the resource utilization threshold by artificial means, and when the node or cluster resource reaches the threshold, the resource instance can be increased or decreased to cope with the load change. However, due to the lack of unified optimization framework for mixed resource pool, resource allocation imbalance easily occurs, resulting in serious waste of cloud computing resources. SUMMARY

[0005] The embodiments of the present application provide a dynamic computing power scheduling method based on deep learning, a device and a medium, which are used to solve the technical problem that the cloud computing scheduling method in the prior art is prone to resource allocation imbalance, resulting in serious waste of cloud computing resources.

[0006] The embodiments of the present application adopt the following technical solutions:

[0007] The embodiments of the present application provide a cloud computing computing power intelligent scheduling method based on deep learning. The method comprises: preprocessing the obtained cloud computing data to generate a structured feature vector; wherein the cloud computing data at least includes cloud node state data, computing task attributes and environment data; extracting task resource space-time correlation features corresponding to the structured feature vector through a multi-head attention mechanism; inputting the task resource space-time correlation features into a weight dynamic prediction module and a resource matching degree scoring module at the same time to perform parallel processing of weight prediction and matching degree scoring calculation; based on the parallel processing result, solving a multi-objective optimization function corresponding to the computing power scheduling, and generating a computing power scheduling instruction based on the solving result.

[0008] The embodiment of the application generates a structured feature vector through preprocessing, unifies the formats of multiple heterogeneous data sources, eliminates noise and outliers, and avoids analysis deviation caused by data problems. Secondly, the multi-head attention mechanism is used to extract spatio-temporal correlation features, which can capture the deep correlation of tasks and resources in the time and space dimensions in parallel, and take into account the time dynamic change and spatial distribution adaptability. In addition, the embodiment of the application processes weight prediction and matching degree scoring in parallel, reduces time consumption, improves decision efficiency, ensures that the two modules are calculated based on the same features, reduces result deviation, and meets the real-time and consistency requirements. The embodiment of the application solves a multi-objective optimization function based on parallel results, balances resource utilization, delay, energy consumption and other objectives, narrows the search range to improve efficiency, generates scheduling instructions that accurately adapt to dynamic demand, relieves multi-objective conflicts, and improves cloud computing utilization.

[0009] In an implementation manner of the application, the obtained cloud computing data is preprocessed to generate a structured feature vector, specifically including: obtaining cloud node state data, the cloud node state data at least including CPU utilization, GPU utilization, memory occupancy and network bandwidth; obtaining computing task attributes, the computing task attributes at least including computing type, data size and SLA constraints; obtaining environment data, the environment data at least including regional electricity price, network delay and cooling cost; mapping the cloud computing data into a preset data interval for normalization processing, and performing mean value filling processing on missing values of the cloud computing data; and converting the processed cloud computing data into a structured feature vector.

[0010] In an implementation manner of the application, the multi-head attention mechanism is used to extract the spatio-temporal correlation features of the structured feature vector, specifically including: based on the feature dimension attribute, the structured feature vector is divided into multiple feature subsets, and each feature subset is linearly transformed and mapped into different subspaces to generate corresponding query vectors, key vectors and value vectors; the multiple parallel attention sublayers corresponding to the multi-head attention mechanism are used to perform correlation processing on the query vectors, key vectors and value vectors corresponding to each feature subset respectively, to obtain correlation feature components of each feature subset; based on different time points in the time dimension, the correlation feature components are spliced and fused to generate task resource spatio-temporal correlation features containing time dimension and space dimension correlation information, and a cloud computing task fusion feature matrix is generated according to the task resource spatio-temporal correlation features.

[0011] In an implementation manner of the application, after the task resource spatio-temporal correlation features are input into the weight dynamic prediction module and the resource matching degree scoring module, the method further includes: in the weight dynamic prediction module, inputting historical cloud computing data into a preset parameter prediction model to output initial weight parameters; and inputting the initial weight parameters and the cloud computing task fusion feature matrix into a preset dynamic adjustment model to output optimized weights.

[0012] In an implementation form of the present application, after the task resource space-time correlation features are input into the weight dynamic prediction module and the resource matching degree scoring module, the method further comprises: in the resource matching degree scoring module, constructing a cloud node topology graph based on the connection relationship between each cloud node; inputting the cloud computing task fusion feature matrix and the cloud node topology graph into a preset graph neural network; performing aggregation processing on the features corresponding to each node in the cloud node topology graph through the graph convolution layer in the preset graph neural network to output a comprehensive feature vector of each cloud node for the current task; wherein the comprehensive feature vector at least includes the current cloud node resource state, the communication cost between other nodes and the task demand correlation information; and obtaining the task resource matching degree score based on the comprehensive feature vector and the preset scoring function.

[0013] In an implementation form of the present application, before the task resource matching degree score is obtained based on the comprehensive feature vector and the preset scoring function, the method further comprises: determining the index data for calculating the task resource matching degree score; wherein the index data at least includes the resource compliance rate, the resource redundancy rate, the communication delay proportion and the data transmission energy consumption; performing quantitative processing on each index data to obtain a quantitative value corresponding to each index data; and constructing the preset scoring function based on the quantitative value, the index weight corresponding to the index data and the correction coefficient.

[0014] In an implementation form of the present application, the multi-objective optimization function corresponding to the computing power scheduling is solved based on the parallel processing result, specifically comprising: taking the maximization of the resource task matching degree, the minimization of the resource allocation cost and the minimization of the task execution delay as the optimization target, constructing a multi-objective optimization function; taking the task execution deadline, the maximum load capacity of the cloud node and the network bandwidth as the constraint condition; solving the multi-objective optimization function under the constraint condition through a heuristic optimization algorithm to generate an optimal solution set; determining the computing power scheduling strategy in the optimal solution set based on a preset decision rule; wherein the preset decision rule is related to the task priority and the actual business demand.

[0015] In an implementation form of the present application, after the solving result is used to generate the computing power scheduling instruction, the method further comprises: sending the computing power scheduling instruction to the cloud resource pool to execute the computing power scheduling instruction; obtaining the task execution result; and performing parameter optimization on the weight dynamic prediction module based on the execution result.

[0016] The embodiment of the application provides a cloud computing computing power intelligent scheduling device based on deep learning, comprising: at least one processor; and a memory in communication connection with the at least one processor; wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to: pre-process acquired cloud computing data to generate a structured feature vector; wherein the cloud computing data at least comprises cloud node state data, computing task attributes and environment data; extract task resource space-time correlation features corresponding to the structured feature vector through a multi-head attention mechanism; input the task resource space-time correlation features into a weight dynamic prediction module and a resource matching degree scoring module at the same time to perform parallel processing of weight prediction and matching degree scoring calculation; based on the parallel processing result, solve a multi-objective optimization function corresponding to computing power scheduling, and generate a computing power scheduling instruction based on a solving result.

[0017] The embodiment of the application provides a nonvolatile computer storage medium, which stores computer executable instructions, and the computer executable instructions are configured to: pre-process acquired cloud computing data to generate a structured feature vector; wherein the cloud computing data at least comprises cloud node state data, computing task attributes and environment data; extract task resource space-time correlation features corresponding to the structured feature vector through a multi-head attention mechanism; input the task resource space-time correlation features into a weight dynamic prediction module and a resource matching degree scoring module at the same time to perform parallel processing of weight prediction and matching degree scoring calculation; based on the parallel processing result, solve a multi-objective optimization function corresponding to computing power scheduling, and generate a computing power scheduling instruction based on a solving result.

[0018] The above at least one technical scheme adopted by the embodiment of the application can achieve the following beneficial effects: the embodiment of the application generates a structured feature vector through preprocessing, unifies the formats of multiple source heterogeneous data, eliminates noise and abnormal values, and avoids analysis deviation caused by data problems. Secondly, the space-time correlation features are extracted through the multi-head attention mechanism, which can capture the deep correlation of tasks and resources in the time-space dimension in parallel, and take into account the time dynamic change and the space distribution adaptability. In addition, the embodiment of the application performs parallel processing of weight prediction and matching degree scoring, reduces time consumption, improves decision efficiency, ensures that the two modules are calculated based on the same features, reduces result deviation, and meets the real-time and consistency requirements. The embodiment of the application solves a multi-objective optimization function based on the parallel result, balances resource utilization rate, delay, energy consumption and other targets, narrows the search range to improve efficiency, and generates a scheduling instruction that accurately adapts to dynamic demand, relieves multi-objective conflicts, and improves cloud computing utilization rate. BRIEF DESCRIPTION OF DRAWINGS

[0019] In order to make the technical solutions in the present application or prior art clearer, the drawings needed to be used in the embodiments or prior art description will be briefly introduced below. Obviously, the drawings described below are only some embodiments described in the present application, and other drawings can be obtained by those skilled in the art without creative labor under the premise of not paying creative labor. In the drawings:

[0020] Figure 1 A structure diagram of a cloud computing computing power intelligent scheduling system based on deep learning is provided for the embodiments of the present application.

[0021] Figure 2 A flowchart of a cloud computing computing power intelligent scheduling method based on deep learning is provided for the embodiments of the present application.

[0022] Figure 3 A structure diagram of a cloud computing computing power intelligent scheduling device based on deep learning is provided for the embodiments of the present application.

[0023] Reference signs:

[0024] 200: cloud computing computing power intelligent scheduling device based on deep learning, 201: processor, 202: memory. DETAILED DESCRIPTION

[0025] The embodiments of the present application provide a cloud computing computing power intelligent scheduling method, device and medium based on deep learning.

[0026] In order to make the technical solutions in the present application or prior art clearer, the drawings needed to be used in the embodiments or prior art description will be briefly introduced below. Obviously, the drawings described below are only some embodiments described in the present application, and other drawings can be obtained by those skilled in the art without creative labor under the premise of not paying creative labor.

[0027] The technical solutions of the embodiments of the present application will be described in detail below with the help of the drawings.

[0028] Figure 1 A structure diagram of a cloud computing computing power intelligent scheduling system based on deep learning is provided for the embodiments of the present application. Figure 1 As shown in the figure, the cloud computing computing power intelligent scheduling system based on deep learning includes a resource perception layer, a deep learning model layer and a scheduling decision layer.

[0029] Resource perception layer: The data acquisition module obtains the cloud node state in real time; the task analysis module extracts the task attributes; and the environment monitoring module collects external variables. The resource perception layer preprocesses the multi-source heterogeneous data and generates a structured feature vector.

[0030] Deep learning model layer: The spatio-temporal feature extraction module inputs the feature vector output by the resource perception layer, captures the long-term spatio-temporal dependence between the task and the resource through a multi-head attention mechanism algorithm, and outputs a fusion feature matrix. The dynamic weight prediction module combines the fusion feature matrix and the historical weight feedback through an LSTM-A3C hybrid model to generate multi-objective weight parameters. The LSTM (Long Short-Term Memory) is a recurrent neural network. The A3C (Asynchronous Advantage Actor-Critic) is a reinforcement learning model combining the policy gradient method and the value function estimation. The heterogeneous resource adaptation module uses the fusion feature matrix to combine the cloud node topology graph, models the communication cost between nodes through a graph neural network algorithm, and outputs a task-resource matching score.

[0031] Scheduling decision layer: The scheduling decision layer, the strategy generation module solves a multi-objective optimization function according to the weight and the matching score; the execution engine issues a task allocation instruction to the cloud resource pool; and the feedback loop collects the task execution result for online training of the model.

[0032] Figure 2 A flowchart of a cloud computing algorithm intelligence scheduling method based on deep learning provided by an embodiment of the present application. As shown in Figure 2 The cloud computing algorithm intelligence scheduling method based on deep learning includes the following steps:

[0033] S101, preprocessing the obtained cloud computing data to generate a structured feature vector.

[0034] In an implementation manner of the present application, cloud node state data is obtained, and the cloud node state data at least includes CPU utilization, GPU utilization, memory occupancy, and network bandwidth. The computing task attributes are obtained, and the computing task attributes at least include the computing type, the data size, and the SLA constraint. The environment data is obtained, and the environment data at least includes the regional electricity price, the network delay, and the cooling cost. The cloud computing data is mapped to a preset data interval for normalization processing, and the missing values of the cloud computing data are processed by mean filling. The processed cloud computing data is converted into a structured feature vector. The cloud computing data at least includes the cloud node state data, the computing task attributes, and the environment data.

[0035] Specifically, in the embodiment of the application, when the cloud node state data is acquired, the core indicators such as CPU utilization, GPU utilization, memory occupancy and network bandwidth can be collected. When the computing task attributes are acquired, the computing type determines the preference of the task for resources such as CPU or GPU, the data size affects the demand for memory and network bandwidth, and the SLA constraint explicitly specifies the time, reliability and other hard requirements for task execution. When the environment data is acquired, the regional electricity price, network delay and cooling cost are mainly covered.

[0036] Further, the cloud computing data is normalized, which uniformly maps indicators of different magnitudes to a preset data interval, eliminating the interference caused by the magnitude difference. At the same time, the missing values are processed by the mean filling method, that is, the missing values are supplemented according to the average value of the historical data of the indicator, ensuring the integrity of the data and providing a coherent and comparable data set for subsequent analysis. Secondly, the processed cloud computing data is converted into a structured feature vector, which integrates various types of scattered data into a unified vector form according to the preset dimensions. Each dimension corresponds to a processed indicator, such as node status, task attributes and environment data, which respectively occupy corresponding dimensions to form a vector that not only retains the key information of the original data but also has a regular structure, facilitating subsequent feature extraction and model calculation through algorithms, thereby improving the efficiency and accuracy of data processing.

[0037] S102, through the multi-head attention mechanism, the task resource space-time correlation features corresponding to the structured feature vector are extracted.

[0038] In an implementation manner of the application, based on the feature dimension attribute, the structured feature vector is divided into a plurality of feature subsets, and each feature subset is linearly transformed and mapped to different subspaces to generate corresponding query vectors, key vectors and value vectors. Through the multi-head attention mechanism, a plurality of parallel attention sublayers corresponding to the multi-head attention mechanism are used to associate the query vectors, key vectors and value vectors corresponding to each feature subset respectively to obtain the associated feature components of each feature subset. Based on different time points of the time dimension, the associated feature components are spliced and fused to generate task resource space-time correlation features containing time dimension and space dimension correlation information, and cloud computing task fusion feature matrices are generated according to the task resource space-time correlation features.

[0039] Specifically, according to the meaning represented by the feature, such as node state, task attribute, environment data, etc., the structured feature vector is divided into multiple feature subsets. For example, node resource features such as CPU utilization and GPU utilization are classified into one subset, and task features such as computation type and data size are classified into another subset. For each feature subset, a preset linear transformation matrix is used to operate to convert the feature vectors in the subset into query vectors, key vectors and value vectors respectively. Different subsets use independent transformation matrices to make the generated vectors map to non-overlapping subspaces, and each subspace focuses on capturing the feature correlation pattern of the corresponding attribute category.

[0040] Further, an attention sublayer is assigned to each feature subset, and the sublayer receives the query, key and value vectors of the subset. The similarity score of the query vector and the key vector is calculated, and the attention weight is obtained after normalization. The value vector is weighted and summed with the weight to output the correlation feature component of the subset. Multiple sublayers operate in parallel to simultaneously obtain local correlation features of each subset such as node resources and task requirements.

[0041] Further, the correlation feature components at different times are collected in timestamp order, and all components at the same time are spliced in order of feature categories to form a comprehensive feature vector at that time. Then, the comprehensive feature vectors at different times are arranged in time axis to obtain a task resource space-time correlation feature containing time evolution law and spatial correlation. The task resource space-time correlation feature is reshaped into a matrix form according to the time dimension and the feature dimension, the rows of the matrix correspond to different times, the columns correspond to different feature dimensions, and the matrix elements are feature values corresponding to the time and the dimension, forming a structured cloud computing task fusion feature matrix.

[0042] S103, the task resource space-time correlation feature is input into the weight dynamic prediction module and the resource matching degree scoring module for parallel processing of weight prediction and matching degree scoring calculation.

[0043] In an implementation manner of the present application, the task resource space-time correlation feature is input into the weight dynamic prediction model, and at the same time, the task resource space-time correlation feature is input into the resource matching degree scoring module. Parallel processing is performed in different modules.

[0044] In an implementation manner of the present application, in the weight dynamic prediction module, the historical cloud computing data is input into the preset parameter prediction model to output the initial weight parameter. The initial weight parameter and the cloud computing task fusion feature matrix are input into the preset dynamic adjustment model to output the optimized weight.

[0045] Specifically, historical cloud computing data is collected, including historical cloud node state data, task attribute data, environment data and corresponding historical weight parameters. After sorting the historical cloud computing data in time sequence, the historical cloud computing data is input into the preset parameter prediction model. The model learns the time sequence characteristics in the historical data, captures the change rule of the weight parameters in different scenarios, and outputs the initial weight parameters based on the historical trend prediction through the fitting training of the historical data. The parameters reflect the weight distribution benchmark in similar historical scenarios.

[0046] In the embodiments of the present application, the preset parameter prediction model can be an LSTM model. During training, the input training sample is the time sequence of historical cloud computing data, including cloud node state, task attribute, environment data at multiple consecutive time points; and the output training sample is the historical weight parameter at the subsequent time point of the corresponding time sequence, which is used to learn the mapping relationship between the historical data and the weight parameter by the model.

[0047] Further, the obtained initial weight parameters are input into the preset dynamic adjustment model together with the cloud computing task fusion feature matrix. The model takes the fusion feature matrix as the current environment state input, combines the initial weight parameters, and performs online learning through interaction with the environment. During the learning process, the model continuously adjusts the weight parameters according to the preset reward mechanism, such as positive rewards of resource utilization rate improvement and delay reduction, and finally outputs the optimized weight parameters that adapt to the real-time scenario, so that the weight parameters are more suitable for the actual situation of the current task and resources.

[0048] In the embodiments of the present application, the preset dynamic adjustment model can be an A3C model. The corresponding input training sample of the A3C model includes the historical cloud computing task fusion feature matrix, the corresponding initial weight parameters and the environment feedback data; and the output training sample is the weight parameter optimized by manual or automatic evaluation, which is used to learn how to adjust the initial weight to obtain a better weight according to the current state.

[0049] In one implementation manner of the present application, after the task resource space-time association features are input into the weight dynamic prediction module and the resource matching degree scoring module, the method further includes, in the resource matching degree scoring module, constructing a cloud node topology graph based on the connection relationship between the cloud nodes. The cloud computing task fusion feature matrix and the cloud node topology graph are input into the preset graph neural network. The features corresponding to each node in the cloud node topology graph are aggregated by the graph convolution layer in the preset graph neural network to output a comprehensive feature vector of each cloud node for the current task; wherein the comprehensive feature vector at least includes the current cloud node resource state, the communication cost between other nodes and the task demand association information. Based on the comprehensive feature vector and the preset scoring function, the task resource matching degree score is obtained.

[0050] Specifically, a cloud node topology graph is constructed based on the physical connection relationship between each cloud node, such as network wiring, communication link and logical association. Each node in the graph represents a cloud node, and the node attribute contains its basic resource information; the edge between the nodes represents the connection relationship, and the edge attribute records the initial communication basic parameters, so as to intuitively present the distribution and association structure of the cloud nodes. The cloud computing task fusion feature matrix is combined with the constructed cloud node topology Figure 1 The input preset graph neural network is combined. The fusion feature matrix provides the spatio-temporal association information of the task and the resource, and the topology graph provides the node connection structure, and the combination of the two enables the model to perceive the dynamic task demand and understand the association mode between the nodes.

[0051] Further, the graph neural network aggregates the features of each node in the topology graph through a graph convolution layer. Each node retains its own attributes while fusing the feature information of adjacent nodes, such as the resource load and communication state obtained from adjacent nodes, and dynamically updates the node features in combination with the task association information in the fusion feature matrix. After multi-layer aggregation, the comprehensive feature vector of each cloud node for the current task is output, which covers the node's own resource state, real-time communication cost with other nodes, and matching degree with task demand, etc.

[0052] Further, based on the comprehensive feature vector of each cloud node, a preset scoring function is substituted. The scoring function comprehensively considers the degree of node resource meeting task demand, the influence of node-to-node communication cost on task execution, etc., and calculates the final quantitative task resource matching degree score according to the set weight, and the higher the score, the better the adaptability of the node to the current task.

[0053] In an implementation manner of the present application, index data for calculating the task resource matching degree score is determined; wherein the index data at least includes resource compliance rate, resource redundancy rate, communication delay proportion and data transmission energy consumption. The index data is quantitatively processed to obtain the quantitative value corresponding to each index data. Based on the quantitative value, the index weight corresponding to the index data and the correction coefficient, a preset scoring function is constructed.

[0054] Specifically, the core index data for calculating the matching degree score is selected from the association features of the task and the resource, including the resource compliance rate, the resource redundancy rate, the communication delay proportion and the data transmission energy consumption. Among them, the resource compliance rate reflects the degree of cloud node resource meeting the basic demand of the task, the resource redundancy rate reflects the redundancy level of node resource exceeding the task demand, the communication delay proportion measures the influence of node-to-node communication delay on the total execution period of the task, and the data transmission energy consumption represents the energy consumption of task data transmission between nodes.

[0055] Further, the ratio of the actual capacity of the CPU, GPU, memory and other resources of the cloud node to the demand capacity of the task is calculated, and if the ratio is greater than or equal to 1, it is recorded as 1, otherwise the actual ratio is taken as the quantitative value, which is used as the resource compliance rate. The difference between the actual capacity of the node resources and the demand capacity of the task is determined, and the resource redundancy rate is determined based on the ratio of the difference to the actual capacity of the node resources. If the result is negative, it is recorded as 0, that is, there is no redundancy, otherwise the actual result is taken as the quantitative value. The data transmission delay between nodes is divided by the total execution period of the task to obtain the quantitative value of the communication delay ratio, and the smaller the value, the smaller the communication delay impact. The actual energy consumption is calculated based on the real-time bandwidth between nodes, the amount of transmitted data and the unit energy consumption coefficient, and then divided by the maximum energy consumption threshold allowed by the task to obtain the quantitative value corresponding to the data transmission energy consumption.

[0056] Further, the process of constructing the preset scoring function in the embodiment of the application is as follows: based on the task type, such as computation-intensive and data-intensive, and the SLA constraints, such as delay requirement and cost upper limit, the corresponding index weight is assigned to each index data, for example, the resource compliance rate weight can be increased for a computation-intensive task, and the communication delay ratio weight can be increased for a data-intensive task. A correction coefficient is set, which is 1 when all key indicators meet the SLA requirements; if there is a key indicator that does not meet the requirements, the correction coefficient is adjusted downward according to the proportion of the number of non-compliant indicators.

[0057] Further, the scoring function is constructed: matching degree score=(resource compliance rate×weight 1+(1-resource redundancy rate)×weight 2+(1-communication delay ratio)×weight 3+(1-data transmission energy consumption)×weight 4)×correction coefficient, through which the quantitative indicators are weighted and fused to output the final quantitative matching degree score.

[0058] In an implementation manner of the application, maximizing the resource task matching degree, minimizing the resource allocation cost and minimizing the task execution delay are taken as optimization objectives to construct a multi-objective optimization function. The task execution deadline, the maximum load capacity of the cloud node and the network bandwidth are taken as constraint conditions. The multi-objective optimization function is solved under the constraint conditions by using a heuristic optimization algorithm to generate an optimal solution set. Based on a preset decision rule, an algorithm scheduling strategy is determined in the optimal solution set; wherein the preset decision rule is related to the task priority and the actual demand of the business.

[0059] Specifically, maximizing resource-task matching degree, minimizing resource allocation cost, and minimizing task execution delay are three core optimization objectives. A sub-function is constructed for each objective. The resource-task matching degree sub-function adopts the weighted sum of the aforementioned matching degree scores. The resource allocation cost sub-function calculates the total cost in combination with the node unit time tariff, task execution time length, and node quantity. The task execution delay sub-function comprehensively considers the calculation time of the task on the node and the data transmission delay between nodes. The three sub-functions are combined by weighting according to the weight parameters output by the dynamic weight prediction module to form a multi-objective optimization function.

[0060] Further, based on the task execution deadline, a constraint is set that the total task execution delay should not exceed the deadline. In view of the maximum load capacity of the cloud node, it is stipulated that the resource utilization rate of the node, such as CPU and GPU, should not exceed 100%. According to the network bandwidth, the data transmission rate between nodes is limited to not exceeding the upper limit of the bandwidth, and the total transmission amount of the cross-node task should be adapted to the bandwidth.

[0061] Further, the embodiments of the present application use a heuristic algorithm such as a genetic algorithm or a particle swarm optimization algorithm to take the multi-objective optimization function as a fitness function. In the algorithm initialization stage, a random resource allocation scheme that meets the constraint condition is generated as an initial population, and the scheme includes information such as task allocation node and execution order. The population is iteratively updated through selection, crossover, and mutation operations. In each iteration, the performance of the scheme is evaluated according to the fitness function, and the scheme with better performance is retained. After a preset number of iterations, the algorithm converges and outputs a Pareto optimal solution set.

[0062] Further, the final strategy is selected from the optimal solution set based on a preset decision rule. The decision rule is bound to the task priority and business demand, for example, an urgent task preferentially selects a scheme with the minimum delay, a cost-sensitive business preferentially selects a scheme with the lowest cost, and a balanced business selects a compromise scheme that is relatively optimal in all three objectives. The schemes in the solution set are sorted according to the decision rule, and the scheme ranked first is selected as the final computing power scheduling strategy, which includes specific task-node allocation relationship, resource reservation quantity, and execution timing arrangement.

[0063] S104, based on the parallel processing result, the multi-objective optimization function corresponding to the computing power scheduling is solved, and a computing power scheduling instruction is generated based on the solving result.

[0064] In an implementation manner of the present application, the computing power scheduling instruction is sent to the cloud resource pool to execute the computing power scheduling instruction. The task execution result is obtained, and the weight dynamic prediction module is optimized based on the execution result.

[0065] Specifically, the determined computing power scheduling strategy is converted into specific computing power scheduling instructions, and the instruction content includes the target cloud node of task allocation, the resource allocation amount of each node, the start time of task execution, priority setting, etc. The instructions are sent to the cloud resource pool through the cloud resource management interface to trigger the scheduling execution mechanism of the resource pool. After receiving the instructions, the cloud resource pool allocates resources according to the instruction content. The target cloud node adjusts its state according to the allocated resource amount, such as activating idle resources, adjusting the priority of the task queue, loading the specified task to the corresponding resource for execution. In the execution process, the resource usage and task progress are monitored in real time.

[0066] Further, after the task execution is completed, relevant execution result data is collected, including task actual execution time, resource actual utilization rate, resource allocation cost, communication delay in task execution process, and whether the final SLA constraint is met, etc. Information is formed into a complete execution result report. Compare the task execution result with the predicted value when the scheduling strategy is formulated, calculate the deviation index. According to the deviation index, judge the applicability of the weight parameter output by the weight dynamic prediction module in the current scene. If the deviation exceeds the preset threshold, adjust the parameters in the module, so that the module can more accurately predict the weight parameter in subsequent scheduling, and improve the adaptability of the scheduling strategy.

[0067] Figure 3 A structural schematic diagram of a cloud computing computing power intelligent scheduling device based on deep learning is provided for the embodiments of the present application. As shown in Figure 3 The cloud computing computing power intelligent scheduling device 200 based on deep learning includes at least one processor 201 and a memory 202 communicatively connected to the at least one processor 201. The memory 202 stores instructions executable by the at least one processor 201. The instructions are executed by the at least one processor 201 to enable the at least one processor 201 to: preprocess the obtained cloud computing data to generate a structured feature vector; wherein the cloud computing data at least includes cloud node state data, computing task attributes and environment data; extract task resource space-time correlation features corresponding to the structured feature vector through a multi-head attention mechanism; input the task resource space-time correlation features into a weight dynamic prediction module and a resource matching degree scoring module at the same time to perform parallel processing of weight prediction and matching degree scoring calculation; based on the parallel processing result, solve the multi-objective optimization function corresponding to the computing power scheduling, and generate the computing power scheduling instruction based on the solving result.

[0068] The embodiment of the application provides a kind of nonvolatile computer storage medium, and computer executable instruction is stored, computer executable instruction is set as: the cloud computing data obtained is preprocessed, and structured feature vector is generated;Wherein, cloud computing data at least includes cloud node state data, computing task attribute and environment data;Through multi-head attention mechanism, the task resource space-time correlation characteristics corresponding to structured feature vector are extracted;Task resource space-time correlation characteristics are simultaneously input into weight dynamic prediction module and resource matching degree scoring module, to carry out the parallel processing of weight prediction and matching degree score calculation;Based on the parallel processing result, the multi-objective optimization function corresponding to computing power scheduling is solved, and computing power scheduling instruction is generated based on the solving result.

[0069] Each of the embodiments in the application is described in a progressive manner, and the same or similar parts between each of the embodiments can be referred to each other, and each of the embodiments focuses on the difference from other embodiments. Especially, for the device, equipment, nonvolatile computer storage medium embodiment, since it is basically similar to the method embodiment, it is described simply, and the relevant part is referred to the part of the method embodiment.

[0070] The above only describes the embodiments of the application, and is not used to limit the application. For those skilled in the art, the embodiments of the application can be variously changed and changed. And these modifications or replacements do not make the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the application.

Claims

1. A cloud computing power intelligent scheduling method based on deep learning, characterized in that, The method includes: The acquired cloud computing data is preprocessed to generate structured feature vectors; wherein the cloud computing data includes at least cloud node status data, computing task attributes, and environmental data. The multi-head attention mechanism is used to extract the spatiotemporal correlation features of task resources corresponding to the structured feature vector; The spatiotemporal correlation features of the task resources are simultaneously input into the weight dynamic prediction module and the resource matching degree scoring module to perform parallel processing of weight prediction and matching degree scoring calculation. The multi-objective optimization function corresponding to computing power scheduling is solved based on the parallel processing results, and computing power scheduling instructions are generated based on the solution results.

2. The dynamic computing power scheduling method based on deep learning according to claim 1, characterized in that, The preprocessing of the acquired cloud computing data to generate structured feature vectors specifically includes: Obtain cloud node status data, which includes at least CPU utilization, GPU utilization, memory usage, and network bandwidth. Obtain the computing task attributes, which include at least the computing type, data size, and SLA constraints; The environmental data is acquired, including at least regional electricity prices, network latency, and cooling costs. The cloud computing data is mapped to a preset data range and normalized. The missing values ​​of the cloud computing data are then filled with the mean. The processed cloud computing data is converted into structured feature vectors.

3. The dynamic computing power scheduling method based on deep learning according to claim 1, characterized in that, The step of extracting the task resource spatiotemporal correlation features corresponding to the structured feature vector through a multi-head attention mechanism specifically includes: Based on the feature dimension attribute, the structured feature vector is divided into multiple feature subsets, and each feature subset is linearly transformed and mapped to different subspaces to generate corresponding query vectors, key vectors and value vectors; By using multiple parallel attention sub-layers corresponding to the multi-head attention mechanism, the query vector, key vector, and value vector corresponding to each feature subset are associated to obtain the associated feature components of each feature subset. Based on different moments in the time dimension, the associated feature components are spliced ​​and fused to generate task resource spatiotemporal association features containing time and spatial dimension association information, and a cloud computing task fusion feature matrix is ​​generated based on the task resource spatiotemporal association features.

4. The dynamic computing power scheduling method based on deep learning according to claim 3, characterized in that, After simultaneously inputting the spatiotemporal correlation features of the task resources into the weighted dynamic prediction module and the resource matching degree scoring module, the method further includes: In the weight dynamic prediction module, historical cloud computing data is input into the preset parameter prediction model to output the initial weight parameters; The initial weight parameters are fused with the feature matrix of the cloud computing task and input into a preset dynamic adjustment model to output the optimized weights.

5. The dynamic computing power scheduling method based on deep learning according to claim 3, characterized in that, After simultaneously inputting the spatiotemporal correlation features of the task resources into the weighted dynamic prediction module and the resource matching degree scoring module, the method further includes: In the resource matching score module, a cloud node topology map is constructed based on the connection relationships between each cloud node; The feature matrix of the cloud computing task is fused with the cloud node topology graph and input into a preset graph neural network. The graph convolutional layer in the pre-set graph neural network aggregates the features corresponding to each node in the cloud node topology graph to output a comprehensive feature vector for each cloud node for the current task. The comprehensive feature vector includes at least the current cloud node resource status, communication cost with other nodes, and task requirement association information. Based on the comprehensive feature vector and the preset scoring function, the task resource matching score is obtained.

6. The dynamic computing power scheduling method based on deep learning according to claim 5, characterized in that, Before obtaining the task resource matching score based on the comprehensive feature vector and the preset scoring function, the method further includes: The indicator data for calculating the task resource matching score is determined; wherein, the indicator data includes at least the resource compliance rate, resource redundancy rate, communication latency ratio, and data transmission energy consumption. The data of each indicator are quantified to obtain the quantified value corresponding to each indicator data. Based on the quantified value, the indicator weights corresponding to the indicator data, and the correction coefficients, the preset scoring function is constructed.

7. The dynamic computing power scheduling method based on deep learning according to claim 1, characterized in that, The process of solving the multi-objective optimization function corresponding to computing power scheduling based on parallel processing results specifically includes: The multi-objective optimization function is constructed with maximizing resource-task matching, minimizing resource allocation cost, and minimizing task execution latency as optimization objectives. The constraints are based on the task execution deadline, the maximum load capacity of the cloud node, and the network bandwidth. The multi-objective optimization function is solved under the constraints using a heuristic optimization algorithm to generate an optimal solution set. Based on preset decision rules, a computing power scheduling strategy is determined in the optimal solution set; wherein, the preset decision rules are related to task priority and actual business needs.

8. The dynamic computing power scheduling method based on deep learning according to claim 1, characterized in that, After generating the computing power scheduling instruction based on the solution result, the method further includes: The computing power scheduling instruction is sent to the cloud resource pool for execution; Obtain the task execution results; Based on the execution results, the parameters of the weight dynamic prediction module are optimized.

9. A cloud computing power intelligent scheduling device based on deep learning, characterized in that, The device includes a memory for storing computer program instructions and a processor for executing the program instructions, wherein when the computer program instructions are executed by the processor, the device is triggered to perform the method described in any one of claims 1-8.

10. A non-volatile computer storage medium storing computer-executable instructions, characterized in that, The computer-executable instructions are capable of performing the method described in any one of claims 1-8.

Images