Cloud resource scheduling method and device, equipment and storage medium

Abstract

The invention discloses a cloud resource scheduling method and device, equipment and a storage medium, and relates to the technical field of deep learning, and the method comprises the steps: obtaining an initial system operation index, an initial application business index and initial external event information of a cloud computing environment, performing data processing on the initial system operation index, the initial application service index and the initial external event information to obtain target feature sequence data; using the target deep learning model to generate a target cloud resource demand prediction result corresponding to the target feature sequence data; generating a corresponding target resource scheduling strategy based on the target cloud resource demand prediction result by using a target decision engine; the target resource scheduling strategy comprises a target elastic scaling strategy and / or a target task scheduling strategy; and calling a corresponding target cloud resource based on the target resource scheduling strategy, and executing the target resource scheduling strategy by using the target cloud resource. According to the invention, balance between cost and performance can be realized when cloud resources are scheduled.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of deep learning, and in particular to a cloud resource scheduling method, device and equipment, and a storage medium. BACKGROUND

[0002] Cloud computing has the ability to flexibly deploy resources on demand, however, how to efficiently use this flexibility to achieve a balance between cost and performance is still a key problem to be solved. The current technology has the following shortcomings: the automatic scaling mechanism based on threshold value only sets a static threshold value according to a single or small number of system indicators, which has a slow response problem, is difficult to predict load changes in advance, and thus causes insufficient preparation of resources at peak time to affect performance, or causes cost waste due to delayed release at low tide; the traditional time series prediction method is difficult to accurately capture the complex nonlinear load mode in the real scene, and does not fully consider the business driving factors and external event influences, resulting in poor prediction accuracy; the cost and performance management are mutually separated, most management tools focus on performance or only perform simple cost control, and lack an effective mechanism for intelligent trade-off and dynamic optimization between the two; the elastic scaling and task scheduling are independent of each other, and the resource demand and scheduling strategy of online services and offline tasks are not coordinated, missing the opportunity for overall cost optimization; in addition, the existing automatic strategy and business target lack direct association, and it is difficult to achieve a resource optimization goal oriented to business value.

[0003] In summary, how to achieve a balance between cost and performance when scheduling cloud resources is a technical problem to be solved at present. SUMMARY

[0004] Therefore, the purpose of the present application is to provide a cloud resource scheduling method, device and equipment, and a storage medium, which can achieve a balance between cost and performance when scheduling cloud resources. The specific solutions are as follows:

[0005] In a first aspect, the present application provides a cloud resource scheduling method, comprising:

[0006] obtaining initial system running indicators, initial application business indicators and initial external event information of a cloud computing environment, and performing data processing on the initial system running indicators, the initial application business indicators and the initial external event information to obtain target feature sequence data;

[0007] generating a target cloud resource demand prediction result corresponding to the target feature sequence data by using a target deep learning model;

[0008] generating a corresponding target resource scheduling strategy based on the target cloud resource demand prediction result by using a target decision engine; the target resource scheduling strategy includes a target elastic scaling strategy and / or a target task scheduling strategy;

[0009] Invoke a corresponding target cloud resource based on the target resource scheduling strategy, and execute the target resource scheduling strategy by using the target cloud resource.

[0010] Optionally, the obtaining of the initial system running index and the initial application business index of the cloud computing environment comprises:

[0011] Collecting a first performance index corresponding to the infrastructure and a second performance index corresponding to the operating system, and determining the initial system running index based on the first performance index and the second performance index;

[0012] Collecting business load data and user behavior data, and determining the initial application business index based on the business load data and the user behavior data;

[0013] Correspondingly, the data processing of the initial system running index, the initial application business index and the initial external event information to obtain target feature sequence data comprises:

[0014] Timestamp alignment of the initial system running index, the initial application business index and the initial external event information to determine missing time points;

[0015] Filling data corresponding to the missing time points based on a preset filling method to obtain a to-be-cleaned system running index, a to-be-cleaned initial application business index and to-be-cleaned external event information; the preset filling method comprises a forward filling method, a backward filling method and a linear interpolation method;

[0016] Data processing of abnormal data and missing data in the to-be-cleaned system running index, the to-be-cleaned initial application business index and the to-be-cleaned external event information based on a preset data cleaning method to obtain a target system running index, a target initial application business index and a target external event information;

[0017] Processing of the target system running index, the target initial application business index and the target external event information based on a preset feature engineering method to obtain target multi-dimensional feature data;

[0018] Based on a preset time sequence and the target multi-dimensional feature data, a corresponding target feature sequence data is constructed.

[0019] Optionally, before the generating of the target cloud resource demand prediction result corresponding to the target feature sequence data by using the target deep learning model, the method further comprises:

[0020] Determining an initial deep learning model and determining a prediction target corresponding to the initial deep learning model;

[0021] determine a target loss function and a target optimizer of the initial deep learning model based on the predicted target; the target loss function is a loss function determined based on a mean square error loss function, a mean absolute error loss function and a cross-entropy loss function;

[0022] obtain historical feature sequence data and cloud resource demand corresponding to the historical feature sequence data;

[0023] input the historical feature sequence data into the initial deep learning model to obtain a corresponding output result, and determine a target loss value between the output result and the cloud resource demand corresponding to the historical feature sequence data based on the target loss function;

[0024] update parameters of the initial deep learning model based on the target loss value by using the target optimizer to obtain the target deep learning model.

[0025] Optionally, the generating, by the target deep learning model, a target cloud resource demand prediction result corresponding to the target feature sequence data comprises:

[0026] inputting the target feature sequence data into the target deep learning model, generating a target cloud resource demand point prediction value corresponding to each preset time period by using the target deep learning model based on a sliding window mode or a rolling prediction mode, and generating a target cloud resource demand prediction interval corresponding to the target cloud resource demand point prediction value.

[0027] Optionally, the generating, by the target decision engine, a corresponding target resource scheduling strategy based on the target cloud resource demand prediction result comprises:

[0028] determining an initial total cloud resource demand based on the target cloud resource demand prediction result and a service level agreement requirement corresponding to a preset business optimization target by using the target decision engine, and determining a target total cloud resource demand based on a preset confidence interval and the initial total cloud resource demand;

[0029] evaluating a target cloud computing instance combination satisfying the preset business optimization target and the target total cloud resource demand based on cost data and performance data of each cloud computing instance, and generating the target elastic scaling strategy based on the target cloud computing instance combination.

[0030] Optionally, the generating, by the target decision engine, a corresponding target resource scheduling strategy based on the target cloud resource demand prediction result comprises:

[0031] identify a first task meeting a preset delayable condition, a second task meeting a preset interruptible condition and a third task meeting a preset low priority condition, and determine a target task based on the first task, the second task and the third task;

[0032] determine a target Spot instance meeting a preset low cost condition, and schedule the target task to the target Spot instance;

[0033] or, determine a target time period meeting a preset load valley condition based on the target cloud resource demand prediction result, and invoke a cloud computing instance to execute the target task in the target time period;

[0034] or, determine a cloud computing instance corresponding to a preset offline task, and execute the target task by using the cloud computing instance corresponding to the preset offline task.

[0035] Optionally, the invoking the corresponding target cloud resource based on the target resource scheduling strategy comprises:

[0036] determining a target cloud platform interface based on the target resource scheduling strategy, and invoking the target cloud resource through the target cloud platform interface;

[0037] Correspondingly, the executing the target resource scheduling strategy by using the target cloud resource further comprises:

[0038] monitoring cloud resource usage, actual system running indicators and actual application business indicators, and optimizing the target deep learning model based on the cloud resource usage, the actual system running indicators and the actual application business indicators;

[0039] optimizing the target decision engine based on the cloud resource usage, the actual system running indicators and the actual application business indicators.

[0040] In a second aspect, the present application provides a cloud resource scheduling device, comprising:

[0041] a target feature sequence data determination module configured to acquire initial system running indicators, initial application business indicators and initial external event information of a cloud computing environment, and perform data processing on the initial system running indicators, the initial application business indicators and the initial external event information to obtain target feature sequence data;

[0042] a prediction result generation module configured to generate a target cloud resource demand prediction result corresponding to the target feature sequence data by using a target deep learning model;

[0043] a target resource scheduling policy generation module configured to generate a corresponding target resource scheduling policy based on the target cloud resource demand prediction result by using a target decision engine; the target resource scheduling policy comprises a target elastic scaling policy and / or a target task scheduling policy;

[0044] a target cloud resource calling module configured to call a corresponding target cloud resource based on the target resource scheduling policy, and execute the target resource scheduling policy by using the target cloud resource.

[0045] In a third aspect, the present application provides an electronic device, comprising:

[0046] a memory configured to save a computer program;

[0047] a processor configured to execute the computer program to implement the cloud resource scheduling method.

[0048] In a fourth aspect, the present application provides a computer readable storage medium configured to save a computer program; wherein the computer program is executed by a processor to implement the cloud resource scheduling method.

[0049] In the present application, firstly, initial system running indexes, initial application business indexes and initial external event information of a cloud computing environment are acquired, and the initial system running indexes, the initial application business indexes and the initial external event information are subjected to data processing to obtain target feature sequence data; then, a target deep learning model is utilized to generate a target cloud resource demand prediction result corresponding to the target feature sequence data; subsequently, a target decision engine is utilized to generate a corresponding target resource scheduling strategy based on the target cloud resource demand prediction result; the target resource scheduling strategy comprises a target elastic scaling strategy and / or a target task scheduling strategy; finally, corresponding target cloud resources are invoked based on the target resource scheduling strategy, and the target resource scheduling strategy is executed by utilizing the target cloud resources. As can be seen from the above, in the present application, initial system running indexes, initial application business indexes and initial external event information of a cloud computing environment are first collected and processed to obtain target feature sequence data; then, the target feature sequence data is input into a target deep learning model to obtain a target cloud resource demand prediction result; then, a target decision engine is utilized to generate a target resource scheduling strategy based on the target cloud resource demand prediction result; finally, target cloud resources are invoked based on the target resource scheduling strategy to execute the target resource scheduling strategy. In this way, the present application utilizes a target deep learning model to generate a target cloud resource demand prediction result based on initial system running indexes, initial application business indexes and initial external event information and the like multi-dimensional information, which can realize more accurate and longer-period prediction of cloud resource demand, and lay a foundation for cost and performance optimization. In addition, the present application generates a corresponding target resource scheduling strategy based on the target cloud resource demand prediction result by utilizing a target decision engine, which can select the instance type combination with the highest cost-effectiveness to realize quantitative balance and optimization of cost and performance when scheduling cloud resources. Meanwhile, the present application can reduce operation and maintenance complexity and labor cost, and improve resource management automation level and operation and maintenance efficiency. BRIEF DESCRIPTION OF DRAWINGS

[0050] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings needed to be used in the embodiments or prior art description. Obviously, the drawings in the following description are only embodiments of the present application, and for those skilled in the art, other drawings can also be obtained without creative labor based on the provided drawings.

[0051] Figure 1 A system architecture diagram of a cloud resource scheduling scheme provided by the present application;

[0052] Figure 2 A flowchart of a cloud resource scheduling method provided by the present application;

[0053] Figure 3 A flowchart of a specific target feature sequence data generation provided by the present application;

[0054] Figure 4 A specific target cloud resource demand prediction result generation flowchart provided for the present application is shown in the following figure:

[0055] Figure 5 A specific target resource scheduling strategy generation flowchart provided for the present application is shown in the following figure:

[0056] Figure 6 A cloud resource scheduling device structure schematic diagram provided for the present application is shown in the following figure:

[0057] Figure 7 An electronic device structure diagram provided for the present application is shown in the following figure. DETAILED DESCRIPTION

[0058] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the accompanying drawings in the embodiments of the present application. Obviously, the described embodiments are only some of the embodiments of the present application, but not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative work fall within the scope of protection of the present application.

[0059] Cloud computing has the ability to flexibly deploy resources on demand. However, the automatic scaling mechanism based on threshold value only sets a static threshold value according to a single or small number of system indicators, which has a slow response problem, is difficult to predict load changes in advance, and thus causes insufficient preparation of resources at peak time to affect performance, or causes cost waste due to delayed release at low tide. The traditional time series prediction method is difficult to accurately capture the complex nonlinear load mode in the real scene, and does not fully consider the business driving factors and external event influence, resulting in poor prediction accuracy. The cost and performance management are mutually separated, and most management tools either focus on performance guarantee or only perform simple cost control, lacking an effective mechanism for intelligent trade-off and dynamic optimization between the two. Elastic scaling and task scheduling are independent of each other, and the resource demand and scheduling strategy of online services and offline tasks are not coordinated, missing the opportunity for overall cost optimization. In addition, the existing automated strategies and business objectives lack direct correlation, making it difficult to achieve resource optimization goals oriented to business value. Therefore, the present application provides a cloud resource scheduling scheme, which can balance the cost and performance when scheduling cloud resources.

[0060] In the cloud resource scheduling scheme of the present application, the system framework used can be seen from Figure 1As shown, specifically can include data acquisition fusion, deep learning model prediction, strategy generation and optimization, strategy execution four steps. Specifically, in the data acquisition fusion, systematically collect and fuse the multi-dimensional data of the cloud computing environment, including the underlying system running indicators, such as CPU / memory utilization, network IO (Input / Output), etc., upper application business indicators, such as user activity, transaction volume, API (Application Programming Interface) call frequency, etc., and external environment event information, such as holidays, promotional activities. In the deep learning model prediction, a deep learning model, such as LSTM (Long Short-Term Memory), Transformer or its combination, is used to train the fused data to accurately predict the resource demand and potential patterns in the future time window. In the strategy generation and optimization, based on the prediction results, and combined with the preset business goals, such as the lowest cost, the highest SLA (Service Level Agreement) compliance rate or the weighted balance between the two, the adaptive elastic scaling strategy is dynamically generated through the decision engine, such as dynamically adjusting the instance number, optimizing the instance type selection, adjusting the resource ratio, and the task scheduling strategy, such as scheduling non-real-time tasks to low-cost Spot instances or resource idle periods. In the strategy execution, the generated elastic scaling strategy and task scheduling strategy are executed through the action execution module to call the cloud platform interface. In this way, the present application can significantly improve the accuracy of cloud resource demand prediction, realize fine-grained cloud resource management, and effectively reduce the operating cost of cloud computing under the premise of ensuring application performance and meeting SLA requirements.

[0061] Referring to Figure 2 As shown, the embodiment of the present application discloses a cloud resource scheduling method, which can include:

[0062] Step S11, obtaining initial system running indicators, initial application business indicators and initial external event information of the cloud computing environment, and performing data processing on the initial system running indicators, the initial application business indicators and the initial external event information to obtain target feature sequence data.

[0063] Referring to Figure 3As shown, this embodiment can collect multi-dimensional information such as system, business, and external events, and perform deep integration to obtain target feature sequence data. Specifically, obtaining the initial system operating indicators and initial application business indicators of the cloud computing environment can include: collecting a first performance indicator corresponding to the infrastructure and a second performance indicator corresponding to the operating system, and determining the initial system operating indicators based on the first and second performance indicators; and collecting business load data and user behavior data, and determining the initial application business indicators based on the business load data and the user behavior data.

[0064] On the one hand, for system operation indicators, infrastructure and operating system performance metrics can be continuously collected, for example, every minute or less frequently. Specific data collected may include: average / peak CPU utilization, memory usage / availability, network inflow / outflow bytes / packets, disk read / write IOPS (Input / Output Operations Per Second), disk read / write throughput, the number of instances per service / application, container restart counts, and end-to-end response latency of key services, such as P50, P90, P95, and P99. Data sources can include cloud service provider monitoring APIs and operating system-level monitoring agents. Cloud service provider monitoring APIs include but are not limited to AWS CloudWatch, AzureMonitor, and GCP Monitoring; operating system-level monitoring agents include but are not limited to Prometheus NodeExporter and Kubernetes Metrics Server.

[0065] On the other hand, for application business metrics, data reflecting actual business load and user behavior can be collected. Specifically, this may include: the number of active users of the website / app (Daily Active Users (DAU) and Monthly Active Users (MAU)), the number of concurrent sessions, the number of queries per second (QPS) and transactions per second (TPS) of core business interfaces, the number of shopping cart additions, the order generation rate, the number of concurrent video stream viewers, the queue length of data processing tasks, and the success / failure rate of specific API calls. Data sources can include application logs, business databases, API gateway statistics, and message queue monitoring. Application logs can be collected and analyzed using tools such as ELK (Elasticsearch, Logstash, Kibana) and Loki.

[0066] In a third aspect, for external event information, environmental factors that can cause significant changes in business load can be collected. Specifically, it can include: public holiday calendar, company scheduled marketing activities such as Double Eleven and expected intensity, new product / feature launch plan, industry-specific events such as game opening day, and even more relevant weather forecast. Data sources can be internal calendar systems, project management tools, manual input interfaces, third-party APIs such as weather APIs.

[0067] It should be noted that the above data processing of the initial system running index, the initial application business index and the initial external event information to obtain the target feature sequence data can include: first, time stamp alignment is performed on the initial system running index, the initial application business index and the initial external event information to determine missing time points; then, data corresponding to the missing time points is filled based on a preset filling method to obtain cleaned system running index, cleaned initial application business index and cleaned external event information; the preset filling method includes forward filling method, backward filling method and linear interpolation method; then, abnormal data and missing data in the cleaned system running index, the cleaned initial application business index and the cleaned external event information are processed based on a preset data cleaning method to obtain target system running index, target initial application business index and target external event information; then, the target system running index, the target initial application business index and the target external event information are processed based on a preset feature engineering method to obtain target multi-dimensional feature data; finally, the target multi-dimensional feature data is constructed based on a preset time sequence to obtain the target feature sequence data.

[0068] Specifically, first, data from different data sources and different collection frequencies, i.e., initial system running indicators, initial application business indicators, and initial external event information, can be aligned to a unified time grid based on timestamps, for example, to the minute level, and missing points can be processed using forward filling, backward filling, or linear interpolation methods. After that, the timestamp-aligned data needs to be cleaned. For abnormal values, the 3-sigma rule or IQR (Interquartile Range) can be used for detection and correction / rejection; for missing values, mean, median, or interpolation based on time series models can be used for filling. Then, the target system running indicators, target initial application business indicators, and target external event information can be processed based on preset feature engineering methods to obtain target multi-dimensional feature data, including: (1) derived features: calculate the rate of change, such as QPS growth rate, moving average / window statistics, such as past 1 hour average CPU utilization, time features, such as hour of the day, day of the week, whether it is a weekday / holiday flag. (2) Encoding: encoding the classification features of the target external event information, such as One-Hot encoding (i.e., one-hot encoding), label encoding, and numerical representation of event intensity. (3) Normalization / standardization: scale features of different dimensions to a similar range, such as mapping the numerical range of features to [0, 1] through Min-Max Scaling (i.e., normalization), or making the mean 0 and the standard deviation 1 through Z-Score Standardization (i.e., standardization) to facilitate model training. Finally, based on the target multi-dimensional feature data, input sequences are constructed, i.e., the processed target multi-dimensional features are organized in chronological order into target feature sequence data suitable for deep learning model input, for example, a three-dimensional tensor of the form (Batch Size, Time Steps, Feature Dimension), where Batch Size is the batch size, Time Steps is the time step, and Feature Dimension is the feature dimension.

[0069] Step S12, using the target deep learning model, generates a target cloud resource demand prediction result corresponding to the target feature sequence data.

[0070] Referring to Figure 4As shown, the embodiment can generate a target cloud resource demand prediction result based on target feature sequence data using a target deep learning model. It should be noted that before the target deep learning model is used to generate the target cloud resource demand prediction result corresponding to the target feature sequence data, it can further include: first determining an initial deep learning model and determining a prediction target corresponding to the initial deep learning model; then determining a target loss function and a target optimizer of the initial deep learning model based on the prediction target; the target loss function is a loss function determined based on a mean square error loss function, a mean absolute error loss function and a cross-entropy loss function; then obtaining historical feature sequence data and cloud resource demand corresponding to the historical feature sequence data; then inputting the historical feature sequence data into the initial deep learning model to obtain a corresponding output result, and determining a target loss value between the output result and the cloud resource demand corresponding to the historical feature sequence data based on the target loss function; finally, using the target optimizer to update parameters of the initial deep learning model based on the target loss value to obtain the target deep learning model. Specifically, the embodiment preferentially selects an initial deep learning model that is good at processing time series data and capturing long-term dependencies. As shown below:

[0071] LSTM, GRU (Gated Recurrent Unit, i.e., Gated Recurrent Unit): suitable for capturing time series dependencies.

[0072] Transformer: using self-attention mechanism, can better capture long-term and complex feature associations, especially suitable for fusing multi-source heterogeneous input.

[0073] Hybrid model: such as CNN-LSTM or ConvTransformer architecture combining CNN (Convolutional Neural Network, i.e., Convolutional Neural Network) for extracting local patterns and LSTM / Transformer for time series modeling.

[0074] Consider multi-task learning: if multiple resources such as CPU, memory, network, etc. or different service components need to be predicted simultaneously, a multi-task learning architecture can be designed to share part of the underlying representation, improving overall prediction efficiency and accuracy.

[0075] Then, the target loss function of the initial deep learning model is determined according to the prediction target. For regression tasks, such as predicting resource quantity, MSE (Mean Squared Error) or MAE (Mean Absolute Error) is used, and for classification tasks, such as whether expansion is needed, cross-entropy is used. The target optimizer of the initial deep learning model is also determined, such as Adam (Adaptive Moment Estimation), RMSprop (Root Mean Square Propagation), etc. Then, the initial deep learning model can be trained offline using historical feature sequence data and cloud resource demand corresponding to the historical feature sequence data to obtain the target deep learning model. At the same time, a periodic online updating / incremental learning mechanism can be established to enable the target deep learning model to adapt to changes in business patterns.

[0076] In the embodiment, the above-mentioned generating the target cloud resource demand prediction result corresponding to the target feature sequence data by using the target deep learning model can include: inputting the target feature sequence data into the target deep learning model, and generating target cloud resource demand point prediction values corresponding to each preset time period by using the target deep learning model in a sliding window manner or a rolling prediction manner, and generating a target cloud resource demand prediction interval corresponding to the target cloud resource demand point prediction values. Specifically, the target feature sequence data can be input into the target deep learning model, and the target deep learning model can be used to generate target cloud resource demand prediction results for one or more time steps (prediction range) in the future, such as target cloud resource demand prediction results for 1 hour, 6 hours, or 24 hours in the future. The prediction value can be a specific resource quantity, such as the number of CPU cores required, the number of GB of memory, or a prediction of resource utilization, or a prediction of the number of instances required. The target cloud resource demand prediction result not only includes point prediction values, but also can include a prediction interval, i.e., a confidence interval, to reflect the uncertainty of the prediction for reference when making decisions. In the embodiment, patterns in the prediction result can also be analyzed, such as identifying periodicity, such as daily, weekly, expected peak time points and intensity, and load change trends. In this way, by accurately predicting the target cloud resource demand, over-provisioning can be avoided, and waste of idle resources can be reduced. At the same time, it can be ensured that there are enough resources to cope with the business peak period, and problems such as performance degradation, increased response delay, or service interruption caused by insufficient resources can be avoided.

[0077] In step S13, a target resource scheduling strategy is generated based on the target cloud resource demand prediction result by using a target decision engine. The target resource scheduling strategy includes a target elastic scaling strategy and / or a target task scheduling strategy.

[0078] Referring to Figure 5As shown, the embodiment receives (1) target cloud resource demand prediction results; (2) real-time system status, such as current resource configuration, resource utilization, instance status, Spot instance market price, etc.; (3) optimization objectives and policy constraints input by users through a configuration interface. Optimization objectives: such as "cost minimization, meet SLA: P95 latency < 200ms"; or "maximize throughput within budget $X"; or set the weights of cost and performance indicators for weighted optimization, as shown below:

[0079] ;

[0080] wherein W_cost is the weight of the cost indicator, Cost_Index is the cost indicator, W_perf is the weight of the performance indicator, and Performance_Index is the performance indicator.

[0081] Policy constraints: available instance type list, minimum / maximum instance number limit, upper limit of Spot instance usage ratio, scaling cooling time, budget alarm threshold, etc.

[0082] It should be noted that in order to generate the target elastic scaling strategy, the above uses the target decision engine to generate a corresponding target resource scheduling strategy based on the target cloud resource demand prediction result, which includes: first using the target decision engine to determine the initial total cloud resource demand based on the target cloud resource demand prediction result and the service level agreement requirement corresponding to the preset business optimization target, and determining the target total cloud resource demand based on the preset confidence interval and the initial total cloud resource demand; then based on the cost data and performance data of each cloud computing instance, a target cloud computing instance combination that meets the preset business optimization target and the target total cloud resource demand is evaluated, and the target elastic scaling strategy is generated based on the target cloud computing instance combination. Specifically, first, the initial total cloud resource demand, such as total vCPU (Virtual Central Processing Unit, i.e. virtual central processing unit) and total Memory, can be calculated according to the target cloud resource demand prediction result and the SLA requirement. Considering the uncertainty of the target cloud resource demand prediction result, a buffer based on the confidence interval can be added to the predicted mean value to obtain the target total cloud resource demand. Then, the system needs to evaluate the cost and performance of multiple instance types, such as the cost and performance of different specifications of on-demand instances, reserved instances, instances covered by the savings plan, and Spot instances. Then based on the target cloud resource demand prediction result, a combination optimization problem is solved through linear programming, integer programming or heuristic search algorithm, such as genetic algorithm, simulated annealing, etc.: find a combination of cloud computing instance types and quantities that can minimize the total cost under the premise of meeting the target cloud resource demand prediction result and the SLA constraint. Among them, the target cloud resource demand prediction result not only has the total amount, but also can be distinguished by demand characteristics such as CPU-intensive and memory-intensive. In this way, by selecting the target cloud computing instance combination with the highest cost-effectiveness, the resource procurement cost can be optimized.

[0083] In addition, when generating the target elasticity strategy, a Spot instance decision can be made: combined with the bidding strategy, the real-time price of the Spot instance, the historical interruption rate, and the tolerance of the task to interruption, it is decided whether and to what extent the Spot instance is used to meet part of the predicted demand. Resource allocation adjustment is performed in the containerization scenario: for environments such as Kubernetes, not only the number of Pods HPA (Horizontal Pod Autoscaler, i.e., horizontal Pod autoscaler) is adjusted, but also the CPU, Memory Request and Limit of the Pod are dynamically adjusted according to the predicted application internal resource consumption mode, such as VPA (Vertical Pod Autoscaler, i.e., vertical Pod autoscaler), to improve node resource utilization and reduce resource fragmentation. Finally, the target elasticity strategy is generated, and the specific sequence of scaling actions, i.e., when to increase and decrease which type of instance and how many, is output, and considering the execution time and cloud platform API restrictions, a cooling mechanism can be added to prevent frequent fluctuations. As can be seen, the embodiment can improve resource utilization efficiency and reduce internal resource contention through fine resource allocation adjustment.

[0084] It should be noted that in order to generate the target task scheduling strategy, the above-mentioned utilization of the target decision engine to generate a corresponding target resource scheduling strategy based on the target cloud resource demand prediction result can include: first identifying a first task meeting a preset delayable condition, a second task meeting a preset interruptible condition, and a third task meeting a preset low priority condition, and determining a target task based on the first task, the second task, and the third task; determining a target Spot instance meeting a preset low cost condition, and scheduling the target task to the target Spot instance; or, determining a target time period meeting a preset load valley condition based on the target cloud resource demand prediction result, and calling a cloud computing instance to execute the target task in the target time period; or, determining a cloud computing instance corresponding to a preset offline task, and executing the target task by using the cloud computing instance corresponding to the preset offline task. Specifically, first, identify the target task marked as "delayable", "interruptible", or "low priority" in the system, such as the label workload-type=batch, priority=low. Then, according to the resource pool state prediction after the target elastic scaling strategy is executed, such as which resources may be idle at what time period, and in combination with the price prediction of the Spot instance, it is determined to schedule the target task to a low-cost Spot instance, or to arrange the target task to be executed at a predicted system load valley period, or to schedule the target task to a cloud computing instance type specially reserved for offline tasks and having a lower cost. It should be noted that if the target task involves a large amount of data, the target task is considered to be scheduled to the area or node near the data to reduce the data transmission cost and time. The generated target task scheduling strategy needs to be connected and executed with the underlying task scheduling system, such as Kubernetes Job / CronJob, Slurm (Simple Linux Utility for Resource Management), Celery, etc. In this way, by scheduling non-critical tasks to low-cost resources or time periods, the cost space can be further compressed.

[0085] Step S14, calling a corresponding target cloud resource based on the target resource scheduling strategy, and executing the target resource scheduling strategy by using the target cloud resource.

[0086] In the embodiment, the calling of the target cloud resource based on the target resource scheduling strategy can include: determining a target cloud platform interface based on the target resource scheduling strategy, and calling the target cloud resource through the target cloud platform interface. Specifically, the target resource scheduling strategy generated by the target decision engine is converted into a call to a target cloud platform API. The target cloud resource is called through the target cloud platform interface to perform operations such as creation, termination, type change, container resource configuration update, and task submission of an instance. Meanwhile, the embodiment can monitor the execution state and result of the target cloud platform API call.

[0087] It should be noted that after the target cloud resource is used to execute the target resource scheduling strategy, the method can further include: monitoring cloud resource usage, actual system running indicators, and actual application business indicators, and optimizing the target deep learning model based on the cloud resource usage, the actual system running indicators, and the actual application business indicators; and optimizing the target decision engine based on the cloud resource usage, the actual system running indicators, and the actual application business indicators. Specifically, after the target resource scheduling strategy is executed, key cost indicators such as actual cost and resource usage hours, and performance indicators such as SLA compliance rate, delay, error rate, and throughput are continuously monitored. The actual effect of the target resource scheduling strategy is evaluated periodically or event-drivenly, compared with the expected target, and the cost saving amount and the performance improvement degree are calculated. Then, the actual cloud resource usage, the actual system running indicators, and the actual application business indicators after the execution of the target resource scheduling strategy are fed back to the prediction module for online learning or regular retraining of the target deep learning model, so as to continuously improve the prediction accuracy of the target deep learning model. The actual cloud resource usage, the actual system running indicators, the actual application business indicators, and the actual effect of the target resource scheduling strategy such as cost and performance are fed back to the decision engine to adjust and optimize the parameters of the algorithm such as the weight of cost and performance, the Spot instance bidding strategy, and the cooling time, so as to realize self-adaptive optimization of the target resource scheduling strategy. In this way, the target resource scheduling strategy can be dynamically adjusted and optimized by continuously learning to adapt to changes in business patterns, so that the cloud resource management and business development can keep pace with each other.

[0088] As can be seen, in the embodiment, first, initial system running indexes, initial application business indexes and initial external event information of the cloud computing environment are acquired, and the initial system running indexes, the initial application business indexes and the initial external event information are processed to obtain target feature sequence data; then, a target cloud resource demand prediction result corresponding to the target feature sequence data is generated by using a target deep learning model; subsequently, a corresponding target resource scheduling strategy is generated based on the target cloud resource demand prediction result by using a target decision engine; the target resource scheduling strategy includes a target elastic scaling strategy and / or a target task scheduling strategy; finally, corresponding target cloud resources are invoked based on the target resource scheduling strategy, and the target resource scheduling strategy is executed by using the target cloud resources. As can be seen, in the embodiment, first, initial system running indexes, initial application business indexes and initial external event information of the cloud computing environment are acquired and processed to obtain target feature sequence data; then, a target cloud resource demand prediction result is obtained by inputting the target feature sequence data into a target deep learning model; then, a target resource scheduling strategy is generated based on the target cloud resource demand prediction result by using a target decision engine; finally, target cloud resources are invoked to execute the target resource scheduling strategy based on the target resource scheduling strategy. In this way, in the embodiment, a target cloud resource demand prediction result is generated based on multi-dimensional information such as initial system running indexes, initial application business indexes and initial external event information by using a target deep learning model, which can realize more accurate and longer-period prediction of cloud resource demand, and lay a foundation for cost and performance optimization. In addition, in the embodiment, a corresponding target resource scheduling strategy is generated based on a target cloud resource demand prediction result by using a target decision engine, which can select the instance type combination with the highest cost-effectiveness to realize quantitative balance and optimization of cost and performance when scheduling cloud resources. At the same time, the embodiment can reduce operation and maintenance complexity and labor cost, and improve resource management automation level and operation and maintenance efficiency.

[0089] Correspondingly, referring to Figure 6 The embodiment of the application also provides a cloud resource scheduling device, which can include:

[0090] A target feature sequence data determination module 11 is configured to acquire initial system running indexes, initial application business indexes and initial external event information of a cloud computing environment, and process the initial system running indexes, the initial application business indexes and the initial external event information to obtain target feature sequence data.

[0091] A prediction result generation module 12 is configured to generate a target cloud resource demand prediction result corresponding to the target feature sequence data by using a target deep learning model.

[0092] The target resource scheduling strategy generation module 13 is configured to generate a corresponding target resource scheduling strategy based on the target cloud resource demand prediction result by using a target decision engine; the target resource scheduling strategy includes a target elastic scaling strategy and / or a target task scheduling strategy.

[0093] The target cloud resource calling module 14 is configured to call a corresponding target cloud resource based on the target resource scheduling strategy, and execute the target resource scheduling strategy by using the target cloud resource.

[0094] As can be seen from the above, in the present application, firstly, the initial system running index, the initial application business index and the initial external event information of the cloud computing environment are acquired, and the initial system running index, the initial application business index and the initial external event information are processed to obtain target feature sequence data; then, a target cloud resource demand prediction result corresponding to the target feature sequence data is generated by using a target deep learning model; subsequently, a corresponding target resource scheduling strategy is generated based on the target cloud resource demand prediction result by using a target decision engine; the target resource scheduling strategy includes a target elastic scaling strategy and / or a target task scheduling strategy; finally, a corresponding target cloud resource is called based on the target resource scheduling strategy, and the target resource scheduling strategy is executed by using the target cloud resource. As can be seen from the above, in the present application, firstly, the initial system running index, the initial application business index and the initial external event information of the cloud computing environment are acquired and processed to obtain target feature sequence data; then, the target feature sequence data is input into a target deep learning model to obtain a target cloud resource demand prediction result; then, a target resource scheduling strategy is generated based on the target cloud resource demand prediction result by using a target decision engine; finally, a target cloud resource is called based on the target resource scheduling strategy to execute the target resource scheduling strategy. In this way, in the present application, a target cloud resource demand prediction result is generated based on the initial system running index, the initial application business index and the initial external event information and other multi-dimensional information by using a target deep learning model, which can realize more accurate and longer-period prediction of cloud resource demand, and lay a foundation for cost and performance optimization. In addition, in the present application, a corresponding target resource scheduling strategy is generated based on the target cloud resource demand prediction result by using a target decision engine, which can select the instance type combination with the highest cost-effectiveness to realize quantitative balance and optimization of cost and performance when scheduling cloud resources. At the same time, the present application can reduce operation and maintenance complexity and labor cost, and improve resource management automation level and operation and maintenance efficiency.

[0095] In some embodiments, the target feature sequence data determination module 11 can include:

[0096] An initial system running index determination unit is configured to collect a first performance index corresponding to the infrastructure and a second performance index corresponding to the operating system, and determine the initial system running index based on the first performance index and the second performance index.

[0097] An initial application service index determination unit is configured to collect service load data and user behavior data, and determine the initial application service index based on the service load data and the user behavior data.

[0098] Correspondingly, the target feature sequence data determination module 11 can include:

[0099] A missing time point determination unit is configured to perform timestamp alignment on the initial system running index, the initial application service index, and the initial external event information, to determine a missing time point.

[0100] A data filling unit is configured to fill data corresponding to the missing time point based on a preset filling method, to obtain a to-be-cleaned system running index, a to-be-cleaned initial application service index, and to-be-cleaned external event information; the preset filling method includes a forward filling method, a backward filling method, and a linear interpolation method.

[0101] A data processing unit is configured to perform data processing on abnormal data and missing data in the to-be-cleaned system running index, the to-be-cleaned initial application service index, and the to-be-cleaned external event information based on a preset data cleaning method, to obtain a target system running index, a target initial application service index, and target external event information.

[0102] A target multi-dimensional feature data determination unit is configured to process the target system running index, the target initial application service index, and the target external event information based on a preset feature engineering method, to obtain target multi-dimensional feature data.

[0103] A target feature sequence data determination unit is configured to construct corresponding target feature sequence data based on a preset time sequence and the target multi-dimensional feature data.

[0104] In some embodiments, the cloud resource scheduling apparatus can further include:

[0105] An initial deep learning model determination module is configured to determine an initial deep learning model, and determine a prediction target corresponding to the initial deep learning model.

[0106] A target loss function determination module is configured to determine a target loss function and a target optimizer of the initial deep learning model based on the prediction target; the target loss function is a loss function determined based on a mean square error loss function, a mean absolute error loss function, and a cross-entropy loss function.

[0107] a historical feature sequence data obtaining module, configured to obtain historical feature sequence data and cloud resource demand corresponding to the historical feature sequence data;

[0108] a target loss value determining module, configured to input the historical feature sequence data into the initial deep learning model to obtain a corresponding output result, and determine a target loss value between the output result and the cloud resource demand corresponding to the historical feature sequence data based on the target loss function;

[0109] a target deep learning model determining module, configured to update parameters of the initial deep learning model based on the target loss value by using the target optimizer to obtain the target deep learning model.

[0110] In some embodiments, the prediction result generating module 12 can include:

[0111] a prediction result generating unit, configured to input the target feature sequence data into the target deep learning model, and generate target cloud resource demand point prediction values corresponding to each preset time period and target cloud resource demand prediction intervals corresponding to the target cloud resource demand point prediction values by using the target deep learning model based on a sliding window mode or a rolling prediction mode.

[0112] In some embodiments, the target resource scheduling strategy generating module 13 can include:

[0113] a target cloud resource total demand determining unit, configured to determine an initial cloud resource total demand based on the target cloud resource demand prediction result and a service level agreement requirement corresponding to a preset business optimization target by using the target decision engine, and determine a target cloud resource total demand based on a preset confidence interval and the initial cloud resource total demand;

[0114] a target elasticity scaling strategy generating unit, configured to evaluate a target cloud computing instance combination satisfying the preset business optimization target and the target cloud resource total demand based on cost data and performance data of each cloud computing instance, and generate the target elasticity scaling strategy based on the target cloud computing instance combination.

[0115] In some embodiments, the target resource scheduling strategy generating module 13 can include:

[0116] a target task determining unit, configured to identify a first task satisfying a preset delayable condition, a second task satisfying a preset interruptible condition, and a third task satisfying a preset low priority condition, and determine a target task based on the first task, the second task, and the third task;

[0117] The target task scheduling strategy generation unit is configured to determine a target Spot instance that meets a preset low-cost condition, and schedule the target task to the target Spot instance; or determine a target time period that meets a preset load valley condition based on the target cloud resource demand prediction result, and invoke a cloud computing instance to execute the target task in the target time period; or determine a cloud computing instance corresponding to a preset offline task, and execute the target task by using the cloud computing instance corresponding to the preset offline task.

[0118] In some specific embodiments, the target cloud resource invocation module 14 can include:

[0119] The target cloud resource invocation unit is configured to determine a target cloud platform interface based on the target resource scheduling strategy, and invoke the target cloud resource through the target cloud platform interface.

[0120] Correspondingly, the cloud resource scheduling apparatus can further include:

[0121] The target deep learning model optimization module is configured to monitor cloud resource usage, actual system running indicators and actual application business indicators, and optimize the target deep learning model based on the cloud resource usage, the actual system running indicators and the actual application business indicators.

[0122] The target decision engine optimization module is configured to optimize the target decision engine based on the cloud resource usage, the actual system running indicators and the actual application business indicators.

[0123] Further, the embodiment of the present application further discloses an electronic device, Figure 7 is an electronic device 20 structure diagram shown according to an exemplary embodiment, the contents in the figure cannot be considered as any limitation on the use range of the present application. The electronic device 20, specifically can include: at least one processor 21, at least one memory 22, power supply 23, communication interface 24, input output interface 25 and communication bus 26. Wherein, the memory 22 is used for storing computer program, the computer program is loaded and executed by the processor 21, to realize the related steps in the cloud resource scheduling method disclosed in any preceding embodiment. In addition, the electronic device 20 in the embodiment specifically can be electronic computer.

[0124] In this embodiment, the power supply 23 is configured to provide operating voltage for each hardware device on the electronic device 20; the communication interface 24 is configured to create a data transmission channel between the electronic device 20 and external devices, and the communication protocol followed by the communication interface 24 can be any communication protocol applicable to the technical solution of the present application, which will not be specifically limited herein; the input / output interface 25 is configured to obtain external input data or output data to the outside, and the specific interface type can be selected according to the specific application needs, which will not be specifically limited herein.

[0125] In addition, the memory 22 as a carrier for storing resources can be a read-only memory, a random access memory, a magnetic disk or an optical disk, etc., and the resources stored thereon can include an operating system 221, a computer program 222, etc., and the storage mode can be temporary storage or permanent storage.

[0126] The operating system 221 is configured to manage and control each hardware device on the electronic device 20 and the computer program 222, and can be Windows Server, Netware, Unix, Linux, etc. In addition to the computer program capable of completing the cloud resource scheduling method executed by the electronic device 20 disclosed in any of the preceding embodiments, the computer program 222 can further include a computer program capable of completing other specific work.

[0127] Further, the present application also discloses a computer readable storage medium for storing a computer program; wherein the computer program is executed by a processor to implement the cloud resource scheduling method disclosed above. For the specific steps of the method, please refer to the corresponding content disclosed in the preceding embodiments, which will not be described here.

[0128] The embodiments in the specification are described in a progressive manner, and each embodiment focuses on the differences from other embodiments. For the same or similar parts between the embodiments, please refer to each other. For the device disclosed in the embodiments, since it corresponds to the method disclosed in the embodiments, the description is relatively simple, and please refer to the method part for the relevant part.

[0129] The skilled person can further realize that the units and algorithm steps of the examples described in combination with the embodiments disclosed herein can be realized by electronic hardware, computer software or a combination of the two. In order to clearly show the interchangeability of hardware and software, the components and steps of each example have been described in the above description. Whether the functions are realized in hardware or software depends on the specific application and design constraints of the technical solution. The skilled person can use different methods to realize the described functions for each specific application, but such implementation should not be considered beyond the scope of the present application.

[0130] The steps of a method or algorithm described in connection with the embodiments disclosed herein can be embodied directly in hardware, in a software module executed by a processor, or in a combination of the two. A software module can reside in RAM, flash memory, ROM, electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), registers, hard disk, a removable disk, a CD-ROM, or any other form of storage medium known in the art. The

[0131] Finally, it should be noted that, in the description of the application, relational terms such as first and second, and the like can be used solely to distinguish one entity or action from another entity or action without necessarily requiring or implying any actual such relationship or order between such entities or actions. Moreover, the terms "comprises", "comprising", or any other variations thereof, are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements does not include only those elements but can include other elements not expressly listed or inherent to such process, method, article, or apparatus. Without further limitation, an element preceded by "comprises a" does not, without more constraints, foreclose the existence of additional identical elements in the process, method, article, or apparatus that comprises the recited element.

[0132] The above provides a detailed description of the technical solutions of the present application. The principles and implementation modes of the present application are described by applying specific examples. The above description of the embodiments is only used to help understand the method of the present application and its core idea; at the same time, for those skilled in the art, according to the idea of the present application, the specific implementation mode and application range will be changed; in summary, the content of the specification should not be understood as a limitation of the present application.

Claims

1. A cloud resource scheduling method, characterized in that: include: Acquiring initial system operation indicators, initial application business indicators, and initial external event information of the cloud computing environment, and performing data processing on the initial system operation indicators, the initial application business indicators, and the initial external event information to obtain target feature sequence data; Using the target deep learning model, generating a target cloud resource demand prediction result corresponding to the target feature sequence data; Utilizing a target decision engine to generate a corresponding target resource scheduling strategy based on the target cloud resource demand prediction result; The target resource scheduling strategy includes a target elastic scaling strategy and / or a target task scheduling strategy; The corresponding target cloud resources are called based on the target resource scheduling policy, and the target cloud resources are used to execute the target resource scheduling policy.

2. The cloud resource scheduling method according to claim 1, wherein: The obtaining of initial system operation indicators and initial application business indicators of the cloud computing environment includes: Collecting a first performance indicator corresponding to the infrastructure and a second performance indicator corresponding to the operating system, and determining the initial system operation indicator based on the first performance indicator and the second performance indicator; Collecting service load data and user behavior data, and determining the initial application service indicator based on the service load data and the user behavior data; Accordingly, the data processing of the initial system operation indicators, the initial application business indicators, and the initial external event information to obtain target feature sequence data includes: Performing time stamp alignment on the initial system operation indicator, the initial application service indicator, and the initial external event information to determine missing time points; Filling the data corresponding to the missing time points based on a preset filling method to obtain the system operation indicators to be cleaned, the initial application business indicators to be cleaned, and the external event information to be cleaned; the preset filling method includes a forward filling method, a backward filling method, and a linear interpolation method; Based on a preset data cleaning method, data processing is performed on the abnormal data and missing data in the system operation indicators to be cleaned, the initial application business indicators to be cleaned, and the external event information to be cleaned, to obtain target system operation indicators, target initial application business indicators, and target external event information; Processing the target system operation indicators, the target initial application business indicators, and the target external event information based on a preset feature engineering method to obtain target multidimensional feature data; The corresponding target feature sequence data is constructed based on a preset time sequence and the target multi-dimensional feature data.

3. The cloud resource scheduling method according to claim 1, wherein: Before generating the target cloud resource demand prediction result corresponding to the target feature sequence data using the target deep learning model, the method further includes: Determining an initial deep learning model and determining a prediction target corresponding to the initial deep learning model; Determining a target loss function and a target optimizer of the initial deep learning model based on the prediction target; the target loss function is a loss function determined based on a mean square error loss function, a mean absolute error loss function, and a cross entropy loss function; Obtaining historical feature sequence data and cloud resource requirements corresponding to the historical feature sequence data; Inputting the historical feature sequence data into the initial deep learning model to obtain a corresponding output result, and determining a target loss value between the output result and the cloud resource demand corresponding to the historical feature sequence data based on the target loss function; Utilizing the target optimizer, parameters of the initial deep learning model are updated based on the target loss value to obtain the target deep learning model.

4. The cloud resource scheduling method according to claim 1, wherein: The generating a target cloud resource demand prediction result corresponding to the target feature sequence data by using the target deep learning model includes: The target feature sequence data is input into the target deep learning model, and the target deep learning model is used to generate target cloud resource demand point prediction values ​​corresponding to each preset time period based on a sliding window method or a rolling prediction method, and a target cloud resource demand prediction interval corresponding to the target cloud resource demand point prediction value is generated.

5. The cloud resource scheduling method according to claim 1, wherein: The target decision engine is used to generate a corresponding target resource scheduling strategy based on the target cloud resource demand prediction result, including: Determine, using the target decision engine, an initial total cloud resource demand based on the target cloud resource demand forecast result and a service level agreement requirement corresponding to a preset business optimization goal, and determine a target total cloud resource demand based on a preset confidence interval and the initial total cloud resource demand; Based on the cost data and performance data of each cloud computing instance, a target cloud computing instance combination that meets the preset business optimization goal and the target cloud resource total demand is evaluated, and the target elastic scaling policy is generated based on the target cloud computing instance combination.

6. The cloud resource scheduling method according to claim 1, wherein: The target decision engine is used to generate a corresponding target resource scheduling strategy based on the target cloud resource demand prediction result, including: identifying a first task that meets a preset deferrable condition, a second task that meets a preset interruptible condition, and a third task that meets a preset low priority condition, and determining a target task based on the first task, the second task, and the third task; Determine a target Spot instance that meets a preset low-cost condition, and schedule the target task to the target Spot instance; Alternatively, a target time period that satisfies a preset load valley condition is determined based on the target cloud resource demand prediction result, and a cloud computing instance is called to execute the target task during the target time period; Alternatively, a cloud computing instance corresponding to a preset offline task is determined, and the target task is executed using the cloud computing instance corresponding to the preset offline task.

7. The cloud resource scheduling method according to any one of claims 1 to 6, characterized in that: The calling of corresponding target cloud resources based on the target resource scheduling policy includes: Determine a target cloud platform interface based on the target resource scheduling strategy, and call the target cloud resource through the target cloud platform interface; Correspondingly, after executing the target resource scheduling strategy using the target cloud resource, the method further includes: Monitoring cloud resource usage, actual system operation indicators, and actual application business indicators, and optimizing the target deep learning model based on the cloud resource usage, the actual system operation indicators, and the actual application business indicators; The target decision engine is optimized based on the cloud resource usage, the actual system operation indicators and the actual application business indicators.

8. A cloud resource scheduling device, characterized in that: include: a target feature sequence data determination module, configured to obtain initial system operation indicators, initial application business indicators, and initial external event information of the cloud computing environment, and perform data processing on the initial system operation indicators, the initial application business indicators, and the initial external event information to obtain target feature sequence data; A prediction result generation module is used to generate a target cloud resource demand prediction result corresponding to the target feature sequence data using a target deep learning model; A target resource scheduling strategy generation module is used to generate a corresponding target resource scheduling strategy based on the target cloud resource demand prediction result using a target decision engine; The target resource scheduling strategy includes a target elastic scaling strategy and / or a target task scheduling strategy; The target cloud resource calling module is used to call the corresponding target cloud resources based on the target resource scheduling strategy, and use the target cloud resources to execute the target resource scheduling strategy.

9. An electronic device, characterized in that: The electronic device includes a processor and a memory; wherein the memory is used to store a computer program, and the computer program is loaded and executed by the processor to implement the cloud resource scheduling method according to any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that Used to store a computer program, which, when executed by a processor, implements the cloud resource scheduling method according to any one of claims 1 to 7.

Images