A serverless user workload prediction method, apparatus and medium
By collecting and extracting multiple features in a serverless architecture and training them with a CNN-LSTM model, the problem of weak data feature extraction and large prediction bias caused by single feature prediction in existing technologies is solved, and more accurate workload prediction and resource scheduling are achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- CHINA UNITED NETWORK COMM GRP CO LTD
- Filing Date
- 2023-06-19
- Publication Date
- 2026-06-30
AI Technical Summary
Existing serverless workload prediction methods use single features for prediction, resulting in weak data feature extraction, large prediction bias, and difficulty in achieving optimal resource scheduling.
Historical data from the serverless architecture is collected, and root cause features, time-series features, difference features, and interaction features are extracted. A convolutional long short-term memory neural network (CNN-LSTM) model is used for training to build a workload model and perform multi-feature integrated prediction.
By using a multi-feature ensemble approach, the stability and accuracy of workload prediction are improved, significantly enhancing prediction performance and enabling better resource scheduling.
Smart Images
Figure CN116781535B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of cloud computing technology, and in particular to a serverless user workload prediction method, apparatus and medium. Background Technology
[0002] In the era of cloud computing, serverless architecture can scale according to increasing business needs, meaning workloads can be elastically adjusted. Serverless allows code to run only in response to events, eliminating the need for containers or servers to run for extended periods. Serverless effectively reduces resource waste during idle periods and business overload during peak periods. Therefore, achieving optimal scheduling of elastic resources requires accurate workload prediction behind the scenes. A crucial issue for a practical serverless service is how to quickly anticipate and predict business peaks and troughs.
[0003] Traditional resource scheduling and control systems generally use maximum thresholds as the basis for decision-making, focusing only on currently monitored values. Valuable historical data is not utilized, and data assets are not leveraged. Furthermore, the lack of historical data makes it difficult to accurately grasp workload trends, and relying solely on human experience makes it difficult to adjust resources in advance, resulting in significant time lags. In recent years, with enterprises increasingly migrating to the cloud, workload forecasting in cloud environments has become a classic and extremely challenging problem. Traditional workload forecasting methods use single features for prediction, such as raw features or historical data features, resulting in weak data feature extraction and large prediction biases. Summary of the Invention
[0004] The technical problem to be solved by the present invention is to address the above-mentioned shortcomings of the prior art by providing a serverless user workload prediction method, apparatus and medium, in order to solve the problem that the existing workload prediction methods are prone to weak data feature extraction and large prediction deviation due to the use of single features for prediction.
[0005] In a first aspect, the present invention provides a serverless user workload prediction method, the method comprising:
[0006] Collect historical data from the serverless architecture, including performance monitoring data as independent variables and the number of jobs waiting to be executed as dependent variables;
[0007] Root cause features, time-division features, differential features, and interaction features are extracted based on the performance monitoring data and the number of jobs waiting to be executed.
[0008] The performance monitoring data, root cause features, time-division features, difference features, and interaction features are used as training datasets to train a preset workload model, resulting in a trained workload model.
[0009] Predict serverless user workloads using a trained workload model.
[0010] Furthermore, the performance monitoring data includes queue ID, queue status, memory usage, CPU usage, collection time, number of canceled tasks, and number of failed tasks.
[0011] Furthermore, before extracting root cause features, time-division features, differential features, and interaction features based on the performance monitoring data and the number of jobs waiting to be executed, the method further includes:
[0012] The historical data is preprocessed, including deleting duplicate data, deleting missing data, and converting the collection time into a timestamp.
[0013] Furthermore, the root cause features include features that affect CPU utilization;
[0014] The time-division feature includes hour and minute information;
[0015] The differential feature is a feature obtained by performing an N-order difference on the number of jobs waiting to be executed, where N is greater than or equal to 1;
[0016] The interactive features include the actual number of canceled tasks obtained by subtracting the number of failed tasks from the number of canceled tasks.
[0017] Furthermore, the preset workload model is trained using five-fold cross-validation.
[0018] Furthermore, the step of using the trained workload model to predict serverless user workload specifically includes:
[0019] Input the data to be predicted into the trained workload model to obtain the differential number of jobs waiting to be executed;
[0020] The predicted number of jobs waiting to be executed is obtained based on the differential number of jobs waiting to be executed.
[0021] Furthermore, the workload model is a convolutional long short-term memory neural network (CNN-LSTM).
[0022] Secondly, the present invention provides a serverless user workload prediction device, comprising:
[0023] The data acquisition module is used to collect historical data in the serverless architecture. The historical data includes performance monitoring data as independent variables and the number of jobs waiting to be executed as dependent variables.
[0024] The feature extraction module, connected to the data acquisition module, is used to extract root cause features, time-division features, differential features, and interaction features based on the performance monitoring data and the number of jobs waiting to be executed.
[0025] The model training module, connected to the feature extraction module, is used to train a preset workload model using the performance monitoring data, root cause features, time-division features, difference features, and interaction features as training datasets to obtain a trained workload model.
[0026] The workload prediction module is connected to the model training module and is used to predict the workload of Serverless users using the trained workload model.
[0027] Thirdly, the present invention provides a serverless user workload prediction device, including a memory and a processor, wherein the memory stores a computer program, and the processor is configured to run the computer program to implement the serverless user workload prediction method described in the first aspect above.
[0028] Fourthly, the present invention provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the serverless user workload prediction method described in the first aspect.
[0029] This invention provides a serverless user workload prediction method, apparatus, and medium. First, historical data from the serverless architecture is collected, including performance monitoring data as the independent variable and the number of jobs waiting to be executed as the dependent variable. Then, root cause features, time-division features, differential features, and interaction features are extracted based on the performance monitoring data and the number of jobs waiting to be executed. Next, the performance monitoring data, root cause features, time-division features, differential features, and interaction features are used as a training dataset to train a pre-defined workload model, resulting in a trained workload model. Finally, the trained workload model is used to predict serverless user workload. Because this application uses multi-feature ensemble for feature extraction, which encompasses root cause features, time-division features, differential features, and interaction features, the trained workload model exhibits strong stability, significantly improving the actual prediction performance. This solves the problem of existing workload prediction methods, which often use single features for prediction, leading to weak data feature extraction and large prediction deviations. Attached Figure Description
[0030] Figure 1 This is a flowchart of a serverless user workload prediction method according to Embodiment 1 of the present invention;
[0031] Figure 2 This is a schematic diagram of the structure of a serverless user workload prediction device according to Embodiment 2 of the present invention;
[0032] Figure 3 This is a schematic diagram of the structure of a serverless user workload prediction device according to Embodiment 3 of the present invention. Detailed Implementation
[0033] To enable those skilled in the art to better understand the technical solution of the present invention, the embodiments of the present invention will be further described in detail below with reference to the accompanying drawings.
[0034] It is understood that the specific embodiments and accompanying drawings described herein are merely for explaining the invention and are not intended to limit the invention.
[0035] It is understood that, without conflict, the various embodiments and features in the embodiments of the present invention can be combined with each other.
[0036] It is understood that, for ease of description, only the parts related to the present invention are shown in the accompanying drawings, while the parts unrelated to the present invention are not shown in the drawings.
[0037] It is understood that each unit or module involved in the embodiments of the present invention may correspond to only one entity structure, or may be composed of multiple entity structures, or multiple units or modules may be integrated into one entity structure.
[0038] It is understood that, without conflict, the functions and steps marked in the flowcharts and block diagrams of this invention may occur in a different order than that marked in the accompanying drawings.
[0039] It is understood that the flowcharts and block diagrams of this invention illustrate the possible architecture, functions, and operations of systems, apparatuses, devices, and methods according to various embodiments of this invention. Each block in the flowchart or block diagram may represent a unit, module, program segment, or code, containing executable instructions for implementing the specified function. Furthermore, each block or combination of blocks in the block diagram and flowchart can be implemented using a hardware-based system to achieve the specified function, or using a combination of hardware and computer instructions.
[0040] It is understood that the units and modules involved in the embodiments of the present invention can be implemented by software or by hardware. For example, the units and modules can be located in a processor.
[0041] Example 1:
[0042] This embodiment provides a serverless user workload prediction method, such as... Figure 1 As shown, the method includes:
[0043] Step S101: Collect historical data from the Serverless architecture, including performance monitoring data as independent variables and the number of jobs waiting to be executed as dependent variables.
[0044] In this embodiment, the performance monitoring data includes queue ID, queue status, memory usage, CPU (central processing unit) usage, collection time, number of canceled tasks, and number of failed tasks. The queue status includes whether the queue is currently available or unavailable. The collection frequency can be once every 5 minutes, or once every 1 minute, or once every 3 minutes; the specific collection frequency can be set according to actual needs. The number of jobs waiting to be executed corresponds to the number of jobs waiting to be executed based on the performance monitoring data.
[0045] Step S102: Extract root cause features, time-division features, differential features, and interaction features based on the performance monitoring data and the number of jobs waiting to be executed.
[0046] Optionally, the root cause features include features that affect CPU utilization;
[0047] The time-division feature includes hour and minute information;
[0048] The differential feature is a feature obtained by performing an N-order difference on the number of jobs waiting to be executed, where N is greater than or equal to 1;
[0049] The interactive features include the actual number of canceled tasks obtained by subtracting the number of failed tasks from the number of canceled tasks.
[0050] Step S103: Use the performance monitoring data, root cause features, time-division features, difference features, and interaction features as training datasets to train a preset workload model, and obtain a trained workload model;
[0051] Specifically, when a normal serverless architecture is in operation, CPU utilization is not only related to task scheduling, but also to many other factors. Therefore, starting from known features, we need to find the root causes of CPU utilization, explore the actual factors affecting CPU utilization, obtain root cause features, and indirectly use them as model features input to the model.
[0052] Specifically, in actual data acquisition, data is usually stored in the basic structure of year, month, day, hour, minute, and second. General prediction tasks only use it directly as a time feature without affecting the prediction. This invention extracts the hour and minute information as time and minute features.
[0053] Specifically, in actual data acquisition, the intervals between data collections are usually very short, typically measured in seconds. This leads to very strong correlations between data points, resulting in ill-conditioned matrices, which cause significant bias in prediction tasks and render the results unusable for real-world business applications. Therefore, to avoid ill-conditioned matrices, this invention employs elastic differencing, performing first-order and second-order differencing operations on the data. The choice of differencing order can be determined based on the similarity matrix of the coefficients of the differrated data; lower similarity is better.
[0054] Specifically, the original independent variables included the number of canceled tasks and the number of failed tasks. We added an interactive feature: actual canceled tasks = number of canceled tasks - number of failed tasks.
[0055] Specifically, by using root cause features as dependent variables and other features and data as independent variables to train a pre-defined workload model, a well-trained workload model can be obtained.
[0056] It should be noted that, during the research and practice of existing technologies, the inventors discovered that traditional workload forecasting methods generally use the actual time of the raw data directly without processing the raw time, thus failing to effectively extract time features. Furthermore, traditional workload forecasting data features typically include CPU utilization, but the processing method simply uses this indicator without multivariate feature extraction, resulting in loss of feature information. Additionally, traditional workload forecasting methods only use historical data features such as difference values for statistical analysis, which cannot be directly used as model features. Moreover, existing methods using model stacking struggle to leverage the model's versatility. Therefore, to overcome the above problems of traditional workload forecasting methods, this application proposes a multi-feature ensemble method. This multi-feature ensemble encompasses root cause features, time-division features, difference features, and interaction features. Using these features, along with performance monitoring data from historical data, to train the workload model results in a highly stable workload model and improved accuracy of the prediction results.
[0057] Optionally, the preset workload model can be trained using five-fold cross-validation.
[0058] In this embodiment, in order to effectively avoid overfitting and underfitting and to effectively evaluate the quality of the model, the present invention uses five-fold cross-validation to train the preset workload model. Five-fold cross-validation means dividing the training dataset into five equal parts, using one part for testing in each training session, and using the rest for training. The experiment is repeated five times and the average value is calculated.
[0059] Optionally, the workload model is a convolutional long short-term memory neural network (CNN-LSTM).
[0060] In this embodiment, to improve the efficiency of model learning, CNN-LSTM can be used as the workload model. CNN-LSTM is a deep learning model that combines Convolutional Neural Networks (CNN) and Long Short-Term Memory (LSTM). It has strong feature extraction and memory capabilities and can be applied to different types of data.
[0061] Optionally, before extracting root cause features, time-division features, difference features, and interaction features based on the performance monitoring data and the number of jobs waiting to be executed, the method further includes:
[0062] The historical data is preprocessed, including deleting duplicate data, deleting missing data, and converting the collection time into a timestamp.
[0063] In this embodiment, deleting duplicate data refers to deleting data where the independent and dependent variables are exactly the same, and deleting missing data refers to deleting missing data at the beginning and end of the task. Since the task has not yet started or stopped at this time, the data has no practical meaning.
[0064] Step S104: Use the trained workload model to predict the workload of Serverless users.
[0065] Optionally, predicting serverless user workload using a trained workload model specifically includes:
[0066] Input the data to be predicted into the trained workload model to obtain the differential number of jobs waiting to be executed;
[0067] The predicted number of jobs waiting to be executed is obtained based on the differential number of jobs waiting to be executed.
[0068] In this embodiment, by using the difference value as the prediction target instead of the original prediction target, the problem of poor model stability caused by the strong correlation between adjacent points of scheduling data in the prior art can be solved. Specifically, the data to be predicted is input into the trained workload model to obtain the number of jobs waiting to be executed in the difference, and then the number of jobs waiting to be executed in the difference is converted into the predicted number of jobs waiting to be executed through the difference formula.
[0069] In this embodiment, after obtaining the predicted number of jobs waiting to be executed, the necessary resources can be prepared in advance according to the predicted number of jobs, and the service can be flexibly scheduled to avoid potential resource waste or shortage.
[0070] In one specific embodiment, the serverless user workload prediction method may include the following steps:
[0071] (1) Data preparation
[0072] Independent variables include queue ID, queue status, memory usage, CPU usage, collection time, number of canceled tasks, and number of failed tasks.
[0073] Dependent variable: The number of jobs waiting to be executed.
[0074] The target of the prediction is the dependent variable, which is the number of jobs waiting to be executed, in order to make a trend judgment on the scheduling of work tasks.
[0075] (2) Data preprocessing
[0076] Delete duplicate data;
[0077] Delete missing data at the beginning and end of the task: At this point, the task has not yet started or closed, and the data has no practical meaning.
[0078] Data collection time is converted into timestamps, etc.
[0079] (3) Feature extraction and integration
[0080] Root cause characteristics: When a normal serverless architecture is in operation, CPU utilization is not only related to task scheduling, but also to many other factors. Therefore, it is necessary to start from known characteristics, find the root cause of CPU utilization, and explore the actual factors affecting CPU utilization, which can be indirectly used as input features into the model. The indirect factors in this prediction task include factors such as the number of running jobs.
[0081] Time-division features: In actual data acquisition, data is stored in the basic structure of year, month, day, hour, minute, and second. Typical prediction tasks only use this as a time feature without affecting the prediction itself. This invention extracts the hour and minute information as time-division features. Simultaneously, it smooths out abnormal data such as holidays.
[0082] Difference Characteristics: In actual data acquisition, the data acquisition intervals are very short, typically measured in seconds, leading to very strong correlations between data points. This can result in ill-conditioned matrices, causing significant bias in prediction tasks and rendering the results unusable for real-world business applications. Therefore, to avoid ill-conditioned matrices, this invention employs elastic differencing, performing first-order and second-order differencing operations on the data. The choice of differencing order can be determined based on the similarity matrix of the coefficients of the differrated data; lower similarity is better.
[0083] Interactive Feature: The original independent variables include the number of canceled tasks and the number of failed tasks. We add an interactive feature: actual canceled tasks = number of canceled tasks - number of failed tasks.
[0084] Finally, the above features are integrated, that is, the root cause features, time-division features, and interaction features are transformed into a part of the independent variable features, and the difference features are transformed into dependent variables.
[0085] (4) Model building and training
[0086] The workload model is built using CNN-LSTM as the basic model architecture.
[0087] Five-fold cross-validation was used for model training.
[0088] (5) Model Prediction
[0089] The trained workload model is used for actual Serverless user workload prediction.
[0090] (6) Scheduling
[0091] Based on model predictions, the service is elastically scheduled.
[0092] The serverless user workload prediction method provided in this invention first collects historical data from the serverless architecture, including performance monitoring data as the independent variable and the number of jobs waiting to be executed as the dependent variable. Then, root cause features, time-division features, differential features, and interaction features are extracted based on the performance monitoring data and the number of jobs waiting to be executed. Next, the performance monitoring data, root cause features, time-division features, differential features, and interaction features are used as a training dataset to train a preset workload model, resulting in a trained workload model. Finally, the trained workload model is used to predict the serverless user workload. Because this application uses multi-feature ensemble for feature extraction, which encompasses root cause features, time-division features, differential features, and interaction features, the trained workload model exhibits strong stability, significantly improving the actual prediction performance. This solves the problem of existing workload prediction methods, which often use single features for prediction, leading to weak data feature extraction and large prediction deviations.
[0093] Example 2:
[0094] like Figure 2 As shown, this embodiment provides a serverless user workload prediction device for performing the above-described serverless user workload prediction method, including:
[0095] The data acquisition module 11 is used to collect historical data in the serverless architecture. The historical data includes performance monitoring data as an independent variable and the number of jobs waiting to be executed as a dependent variable.
[0096] The feature extraction module 12, connected to the data acquisition module 11, is used to extract root cause features, time-division features, differential features, and interaction features based on the performance monitoring data and the number of jobs waiting to be executed.
[0097] The model training module 13 is connected to the feature extraction module 12 and is used to train a preset workload model using the performance monitoring data, root cause features, time-division features, difference features and interaction features as training datasets to obtain a trained workload model.
[0098] The workload prediction module 14 is connected to the model training module 13 and is used to predict the workload of Serverless users using the trained workload model.
[0099] Optionally, the performance monitoring data includes queue ID, queue status, memory usage, CPU usage, collection time, number of canceled tasks, and number of failed tasks.
[0100] Optionally, the device further includes:
[0101] The preprocessing module is used to preprocess the historical data, and the preprocessing includes: deleting duplicate data, deleting missing data, and converting the collection time into a timestamp.
[0102] Optionally, the root cause features include features that affect CPU utilization;
[0103] The time-division feature includes hour and minute information;
[0104] The differential feature is a feature obtained by performing an N-order difference on the number of jobs waiting to be executed, where N is greater than or equal to 1;
[0105] The interactive features include the actual number of canceled tasks obtained by subtracting the number of failed tasks from the number of canceled tasks.
[0106] Optionally, the model training module 13 is further configured to train the preset workload model using five-fold cross-validation.
[0107] Optionally, the workload prediction module 14 specifically includes:
[0108] The input unit is used to input the data to be predicted into the trained workload model to obtain the differential number of jobs waiting to be executed;
[0109] The prediction unit is used to obtain the predicted number of jobs waiting to be executed based on the differential number of jobs waiting to be executed.
[0110] Optionally, the workload model is a convolutional long short-term memory neural network (CNN-LSTM).
[0111] Example 3:
[0112] refer to Figure 3 This embodiment provides a serverless user workload prediction device, including a memory 21 and a processor 22. The memory 21 stores a computer program, and the processor 22 is configured to run the computer program to execute the serverless user workload prediction method in Embodiment 1.
[0113] The memory 21 is connected to the processor 22. The memory 21 can be a flash memory, a read-only memory or other memory, and the processor 22 can be a central processing unit or a microcontroller.
[0114] Example 4:
[0115] This embodiment provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the serverless user workload prediction method in Embodiment 1 above.
[0116] The computer-readable storage medium includes volatile or non-volatile, removable or non-removable media implemented in any method or technology for storing information (such as computer-readable instructions, data structures, computer program modules, or other data). Computer-readable storage media include, but are not limited to, RAM (Random Access Memory), ROM (Read-Only Memory), EEPROM (Electrically Erasable Programmable Read-Only Memory), flash memory or other memory technologies, CD-ROM (Compact Disc Read-Only Memory), DVD or other optical disc storage, cartridges, magnetic tapes, disk storage or other magnetic storage devices, or any other medium that can be used to store desired information and is accessible to a computer.
[0117] In summary, the serverless user workload prediction method, apparatus, and medium provided in this embodiment of the invention first collect historical data from the serverless architecture, including performance monitoring data as the independent variable and the number of jobs waiting to be executed as the dependent variable. Then, root cause features, time-division features, differential features, and interaction features are extracted based on the performance monitoring data and the number of jobs waiting to be executed. Next, the performance monitoring data, root cause features, time-division features, differential features, and interaction features are used as a training dataset to train a preset workload model, resulting in a trained workload model. Finally, the trained workload model is used to predict serverless user workload. Because this application uses multi-feature ensemble for feature extraction, which encompasses root cause features, time-division features, differential features, and interaction features, the trained workload model exhibits strong stability, significantly improving the actual prediction performance. This solves the problem that existing workload prediction methods, which use single features for prediction, are prone to weak data feature extraction and large prediction deviations.
[0118] It is understood that the above embodiments are merely exemplary implementations used to illustrate the principles of the present invention, and the present invention is not limited thereto. For those skilled in the art, various modifications and improvements can be made without departing from the spirit and essence of the present invention, and these modifications and improvements are also considered to be within the scope of protection of the present invention.
Claims
1. A serverless user workload prediction method, characterized in that, include: Collect historical data from the serverless architecture, including performance monitoring data as independent variables and the number of jobs waiting to be executed as dependent variables; Root cause features, time-division features, differential features, and interaction features are extracted based on the performance monitoring data and the number of jobs waiting to be executed. The performance monitoring data, root cause features, time-division features, difference features, and interaction features are used as training datasets to train a preset workload model, resulting in a trained workload model. Predict serverless user workloads using a trained workload model; The root cause features include features that affect CPU utilization, including the number of running jobs in the Serverless architecture. The time-division feature includes hour and minute information; The difference feature is the feature obtained by performing an N-order difference on the number of jobs waiting to be executed. The N is greater than or equal to 1, and N is selected based on the coefficient similarity matrix of the differenced data. The order that minimizes the similarity represented by the coefficient similarity matrix is selected as N. The interactive features include the actual number of canceled tasks obtained by subtracting the number of failed tasks from the number of canceled tasks.
2. The method according to claim 1, characterized in that, The performance monitoring data includes queue ID, queue status, memory usage, CPU usage, collection time, number of canceled tasks, and number of failed tasks.
3. The method according to claim 2, characterized in that, Before extracting root cause features, time-division features, differential features, and interaction features based on the performance monitoring data and the number of jobs waiting to be executed, the method further includes: The historical data is preprocessed, including deleting duplicate data, deleting missing data, and converting the collection time into a timestamp.
4. The method according to claim 1, characterized in that, The preset workload model is trained using five-fold cross-validation.
5. The method according to claim 1, characterized in that, The method of using a trained workload model to predict serverless user workloads specifically includes: Input the data to be predicted into the trained workload model to obtain the differential number of jobs waiting to be executed; The predicted number of jobs waiting to be executed is obtained based on the differential number of jobs waiting to be executed.
6. The method according to claim 2, characterized in that, The workload model is a convolutional long short-term memory neural network (CNN-LSTM).
7. A serverless user workload prediction device, characterized in that, include: The data acquisition module is used to collect historical data in the serverless architecture. The historical data includes performance monitoring data as independent variables and the number of jobs waiting to be executed as dependent variables. The feature extraction module, connected to the data acquisition module, is used to extract root cause features, time-division features, differential features, and interaction features based on the performance monitoring data and the number of jobs waiting to be executed. The model training module, connected to the feature extraction module, is used to train a preset workload model using the performance monitoring data, root cause features, time-division features, difference features, and interaction features as training datasets to obtain a trained workload model. The workload prediction module is connected to the model training module and is used to predict the workload of Serverless users using the trained workload model. The root cause features include features that affect CPU utilization, including the number of running jobs in the Serverless architecture. The time-division feature includes hour and minute information; The difference feature is the feature obtained by performing an N-order difference on the number of jobs waiting to be executed. The N is greater than or equal to 1, and N is selected based on the coefficient similarity matrix of the differenced data. The order that minimizes the similarity represented by the coefficient similarity matrix is selected as N. The interactive features include the actual number of canceled tasks obtained by subtracting the number of failed tasks from the number of canceled tasks.
8. A serverless user workload prediction device, characterized in that, It includes a memory and a processor, wherein the memory stores a computer program and the processor is configured to run the computer program to implement the serverless user workload prediction method as described in any one of claims 1-6.
9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the serverless user workload prediction method as described in any one of claims 1-6.