A method and system for dynamic scaling up and down computing resources based on load prediction
By using real-time monitoring and improved load prediction algorithms, combined with LSTM models and preset thresholds, resource allocation is dynamically adjusted, solving the problems of low resource utilization and uneven distribution in cloud computing clusters, and achieving efficient and dynamic computing resource management.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- ANHUI NANRUI JIYUAN POWER GRID TECH CO LTD
- Filing Date
- 2026-05-14
- Publication Date
- 2026-07-10
AI Technical Summary
Traditional cloud computing cluster architectures suffer from low resource utilization, uneven allocation, insufficient prediction accuracy, lagging model training, and unreasonable scaling strategies when facing dynamic business loads, making it difficult to meet the needs of efficient, dynamic, and accurate computing resource scheduling.
By monitoring resource utilization metrics in real time, using an improved LSTM model for load prediction, and combining preset thresholds for scaling up and down decisions, the number of CPU cores and virtual machine instances can be dynamically adjusted to achieve on-demand allocation and elastic scheduling of resources.
It improves the utilization rate of computing resources, optimizes resource allocation, enhances system performance and service quality, reduces operating costs, and solves the problems of uneven resource allocation and delayed expansion in traditional methods.
Smart Images
Figure CN122363933A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the fields of cloud computing and information technology, and in particular to a method and system for dynamically scaling up and down computing resources based on load prediction. Background Technology
[0002] With the deep integration of cloud computing, big data, and artificial intelligence technologies, the business load of application systems exhibits significant characteristics of high dynamism, nonlinear fluctuations, and strong bursts. Traditional fixed-size cluster architectures are difficult to adapt to these dynamic demands, often employing static configuration or experience-based resource allocation strategies. This leads to severe resource idleness and waste during off-peak periods, while during peak periods, insufficient computing power causes response delays, service degradation, and even system crashes, severely restricting system stability and operational cost control.
[0003] Despite the progress made in existing load forecasting algorithms, many technical shortcomings still exist in their practical application in cloud computing environments:
[0004] First, the prediction accuracy and generalization ability are insufficient: most prediction models only focus on a single load dimension and do not fully associate multi-source heterogeneous data such as network traffic, latency, packet loss rate, and business request volume, resulting in poor adaptability to complex load scenarios and large prediction errors when facing sudden traffic and business peaks.
[0005] Second, model training and updates are lagging: Traditional models are mostly trained based on offline historical data, making it difficult to perceive the dynamic changes in node load and network status in the cloud computing environment in real time. This makes it impossible to achieve online iterative optimization of the model, and the prediction results are easily decoupled from the actual load.
[0006] Third, the coupling with scaling-up and scaling-down strategies is low: existing load prediction results are mostly used as a single trigger for scaling-up, without taking into account multiple objectives such as real-time resource utilization of computing nodes, scheduling costs, and service quality for comprehensive decision-making. This can easily lead to over-scaling or untimely scaling-up, and cannot achieve optimal configuration of cluster resources.
[0007] Fourth, there is a conflict between real-time performance and computational overhead: some high-precision deep learning models have complex structures and long inference times, making it difficult to meet the low-latency decision-making requirements of cluster scaling, while lightweight models cannot guarantee prediction accuracy, making it impossible to balance real-time performance and accuracy.
[0008] In summary, existing cluster auto-scaling and scaling technologies based on load prediction still have shortcomings in multi-source data fusion, online adaptive optimization, and multi-objective collaborative decision-making. They are insufficient to fully meet the actual needs of efficient, dynamic, and precise scheduling of computing resources in cloud computing scenarios, and further improvements are urgently needed to enhance the resource utilization and performance of the overall system. Summary of the Invention
[0009] To address the issues of low utilization and uneven distribution of computing resources, the primary objective of this invention is to provide a dynamic scaling-up / scaling method for computing resources based on load prediction, which improves the utilization of computing resources and makes their allocation more rational.
[0010] To achieve the above objectives, the present invention adopts the following technical solution: a method for dynamic scaling up and down of computing resources based on load prediction, the method comprising the following sequential steps:
[0011] (1) Real-time monitoring of computing power resource utilization indicators through a resource monitoring platform, including CPU utilization, GPU utilization, memory size, disk size and network bandwidth;
[0012] (2) Based on resource utilization indicators, the improved load forecasting algorithm is used to predict the future load situation and obtain the forecast results;
[0013] (3) Compare the prediction results with the preset thresholds of various resource utilization indicators and make automatic expansion and contraction decisions;
[0014] (4) Based on the expansion and contraction decision, when it is predicted that the resource utilization index of the future load will exceed the preset threshold of each resource utilization index, the corresponding resources will be automatically increased; when it is predicted that the resource utilization index of the future load will be lower than the preset threshold, the corresponding resources will be automatically reduced to meet the load demand.
[0015] Step (1) specifically includes the following steps in sequence:
[0016] (1a) Create a dedicated Datadog account;
[0017] (1b) Install Datadog Agent on the server to be monitored according to the official documentation;
[0018] (1c) Configure the resource utilization indicators of the computing resources to be monitored through the Datadog dashboard;
[0019] (1d) Create one or more dashboards as needed to display and statistically analyze real-time data.
[0020] In step (2), the improved load prediction algorithm specifically includes the following sequential steps:
[0021] (2a) Preprocess the various resource utilization indicators to obtain preprocessed data. The preprocessed data forms a dataset, which is then divided into training sets according to the proportions. and validation set;
[0022] (2b) Improve the LSTM model to obtain the improved LSTM model;
[0023] (2c) Using the training set The improved LSTM model is trained to obtain the trained model, and then the trained model is validated using a validation set.
[0024] (2d) Input the resource utilization indicators of the current load into the trained model to obtain the prediction results.
[0025] In step (2a), the preprocessing of various resource utilization indicators specifically includes the following steps:
[0026] (2a1) Perform data cleaning on various resource utilization indicators to obtain cleaned data. Data cleaning refers to performing missing value processing, outlier detection and removal operations on the original data to ensure data quality.
[0027] (2a2) Normalize the cleaned data, set a benchmark value for each resource utilization indicator, and divide the value of each resource utilization indicator at each time point by the benchmark value to obtain the normalized value.
[0028] Step (2b) specifically includes the following sequential steps:
[0029] (2b1) Introduce stochastic gradient descent into the LSTM model to iteratively update the model parameters until final convergence; in each iteration, randomly select parameters from the training set... Select a training sample 'a' and update the model parameters in the direction of the negative gradient:
[0030] ;
[0031] in, These are the updated model training parameters; These are the model training parameters before the update; It is the learning rate used to control the step size of gradient descent. ; This represents the actual CPU utilization. CPU utilization as predicted by the LSTM model; This represents the actual GPU utilization. This represents the GPU utilization predicted by the LSTM model. For actual memory; The memory size predicted by the LSTM model; This is the actual disk size; The disk size predicted by the LSTM model; This represents the actual network bandwidth. The network bandwidth predicted by the LSTM model;
[0032] (2b2) Introduce batch normalization in the LSTM model, based on each training set... The inputs to the network layers are standardized to a Gaussian distribution with zero mean and unit variance to accelerate the model convergence process.
[0033] ;
[0034] in, and All of these are trainable parameters, used to amplify and shift the standardized values; for The mean, ; For training set Total number of samples; for variance ;
[0035] (2b3) Using the multi-output mean squared error loss function Replace the mean squared error loss function in the LSTM model:
[0036] .
[0037] Another objective of this invention is to provide a system for a dynamic scaling up and down method of computing resources based on load prediction, comprising:
[0038] The dynamic monitoring module monitors and statistically analyzes resource utilization metrics such as CPU utilization, GPU utilization, memory size, disk size, and network bandwidth in real time, so that the load prediction module can use them to make predictions.
[0039] The load forecasting module is used to predict future load conditions and obtain the forecast results of various resource utilization indicators, which are then used by the resource scaling decision module to make resource scaling decisions.
[0040] The resource expansion / contraction decision module is used to compare the predicted results of various resource utilization indicators with the preset thresholds of various resource utilization indicators, and obtain the comparison results to determine whether to expand or shrink the resources, so that the resource expansion / contraction execution module can expand or shrink the resources.
[0041] The resource scaling execution module is used to automatically increase or decrease the number of CPU cores and virtual machine instances based on the resource scaling decision results.
[0042] As can be seen from the above technical solution, the beneficial effects of the present invention are as follows: First, the present invention monitors and statistically analyzes the current load resource utilization indicators in real time, predicts future load conditions, compares the predicted results of various resource utilization indicators with preset thresholds to determine whether to expand or shrink resources, and automatically increases or decreases hardware resources such as the number of CPU cores and virtual machine instances based on the resource expansion or shrinkage decision results, thereby improving the utilization rate of computing resources and making the allocation of computing resources more reasonable; Second, addressing the industry pain points of low resource utilization, uneven allocation, and inability to adapt to dynamic business needs in traditional fixed-scale clusters, the present invention, through a closed-loop load perception-prediction-decision-execution mechanism, solves the core defects of the traditional static resource allocation model, providing a solution for the efficient management of computing resources in a cloud computing environment. The invention leverages the advantages of a novel technological approach. Third, addressing the shortcomings of traditional time series models and basic deep learning models in predicting insufficient accuracy and weak generalization under nonlinear and sudden load scenarios, this invention significantly enhances the ability to capture complex business load fluctuations by optimizing the model structure and loss function. It can accurately predict load change trends over a future period, enabling early prediction of business peaks and providing core support for forward-looking scaling decisions, fundamentally solving the problems of delayed or blind scaling in traditional methods. Fourth, this invention employs an intelligent scaling decision mechanism based on the linkage between prediction results and thresholds. It compares the predicted values of various resource utilization indicators with preset thresholds in multiple dimensions, automatically generating accurate scaling decisions and dynamically adjusting CPU usage based on the decision results. Hardware and virtualization resources, such as the number of cores and virtual machine instances, enable on-demand allocation and elastic scheduling of computing resources. This mechanism can expand capacity in advance before the load increases, ensuring the performance and service quality of business systems and avoiding service degradation during peak periods; and it can shrink capacity in a timely manner when the load decreases, making use of idle computing resources, significantly improving the overall resource utilization of the cluster, reducing operating costs, and achieving globally optimal allocation of computing resources, thus completely solving the problem of uneven resource allocation in traditional methods. Attached Figure Description
[0043] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation
[0044] like Figure 1 As shown, a method for dynamically scaling up and down computing resources based on load prediction is described. This method includes the following steps in sequence:
[0045] (1) Real-time monitoring of computing power resource utilization indicators through a resource monitoring platform, including CPU utilization, GPU utilization, memory size, disk size and network bandwidth;
[0046] (2) Based on resource utilization indicators, the improved load forecasting algorithm is used to predict the future load situation and obtain the forecast results;
[0047] (3) Compare the prediction results with the preset thresholds of various resource utilization indicators and make automatic expansion and contraction decisions;
[0048] (4) Based on the expansion and contraction decision, when it is predicted that the resource utilization index of the future load will exceed the preset threshold of each resource utilization index, the corresponding resources will be automatically increased; when it is predicted that the resource utilization index of the future load will be lower than the preset threshold, the corresponding resources will be automatically reduced to meet the load demand.
[0049] Step (1) specifically includes the following steps in sequence:
[0050] (1a) Create a dedicated Datadog account;
[0051] (1b) Install Datadog Agent on the server to be monitored according to the official documentation;
[0052] (1c) Configure the resource utilization indicators of the computing resources to be monitored through the Datadog dashboard;
[0053] (1d) Create one or more dashboards as needed to display and statistically analyze real-time data.
[0054] In step (2), the improved load prediction algorithm specifically includes the following sequential steps:
[0055] (2a) Preprocess the various resource utilization indicators to obtain preprocessed data. The preprocessed data forms a dataset, which is then divided into training sets according to the proportions. and validation set;
[0056] (2b) Improve the LSTM model to obtain the improved LSTM model;
[0057] (2c) Using the training set The improved LSTM model is trained to obtain the trained model, and then the trained model is validated using a validation set.
[0058] (2d) Input the resource utilization indicators of the current load into the trained model to obtain the prediction results.
[0059] In step (2a), the preprocessing of various resource utilization indicators specifically includes the following steps:
[0060] (2a1) Perform data cleaning on various resource utilization indicators to obtain cleaned data. Data cleaning refers to performing missing value processing, outlier detection and removal operations on the original data to ensure data quality.
[0061] (2a2) Normalize the cleaned data, set a benchmark value for each resource utilization indicator, and divide the value of each resource utilization indicator at each time point by the benchmark value to obtain the normalized value.
[0062] Step (2b) specifically includes the following sequential steps:
[0063] (2b1) Introduce stochastic gradient descent into the LSTM model to iteratively update the model parameters until final convergence; in each iteration, randomly select parameters from the training set... Select a training sample 'a' and update the model parameters in the direction of the negative gradient:
[0064] ;
[0065] in, These are the updated model training parameters; These are the model training parameters before the update; It is the learning rate used to control the step size of gradient descent. ; This represents the actual CPU utilization. CPU utilization as predicted by the LSTM model; This represents the actual GPU utilization. This represents the GPU utilization predicted by the LSTM model. For actual memory; The memory size predicted by the LSTM model; This is the actual disk size; The disk size predicted by the LSTM model; This represents the actual network bandwidth. The network bandwidth predicted by the LSTM model;
[0066] (2b2) Introduce batch normalization in the LSTM model, based on each training set... The inputs to the network layers are standardized to a Gaussian distribution with zero mean and unit variance to accelerate the model convergence process.
[0067] ;
[0068] in, and All of these are trainable parameters, used to amplify and shift the standardized values; for The mean, ; For training set Total number of samples; for variance ;
[0069] (2b3) Using the multi-output mean squared error loss function Replace the mean squared error loss function in the LSTM model:
[0070] .
[0071] This system includes:
[0072] The dynamic monitoring module monitors and statistically analyzes resource utilization metrics such as CPU utilization, GPU utilization, memory size, disk size, and network bandwidth in real time, so that the load prediction module can use them to make predictions.
[0073] The load forecasting module is used to predict future load conditions and obtain the forecast results of various resource utilization indicators, which are then used by the resource scaling decision module to make resource scaling decisions.
[0074] The resource expansion / contraction decision module is used to compare the predicted results of various resource utilization indicators with the preset thresholds of various resource utilization indicators, and obtain the comparison results to determine whether to expand or shrink the resources, so that the resource expansion / contraction execution module can expand or shrink the resources.
[0075] The resource scaling execution module is used to automatically increase or decrease the number of CPU cores and virtual machine instances based on the resource scaling decision results.
[0076] This invention monitors and statistically analyzes current load resource utilization indicators in real time, predicts future load conditions, and compares the predicted results of various resource utilization indicators with preset thresholds to determine whether to expand or shrink resources. Based on the resource expansion / shrinkage decision, it automatically increases or decreases hardware resources such as CPU cores and virtual machine instances, improving computing resource utilization and making computing resource allocation more rational. It offers the advantages of a new technical path for efficient management of computing resources in a cloud computing environment. It can accurately predict load change trends over a period of time, enabling early prediction of business peaks and providing core support for forward-looking expansion / shrinkage decisions, fundamentally solving the problems of delayed or blind expansion in traditional methods. It achieves on-demand allocation and elastic scheduling of computing resources, allowing for expansion before load increases to ensure the performance and service quality of business systems and avoid service degradation during peak periods; and timely shrinkage when load decreases to revitalize idle computing resources, significantly improving the overall resource utilization of the cluster, reducing operating costs, and achieving globally optimal allocation of computing resources, completely solving the problem of uneven resource allocation in traditional methods.
[0077] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the claimed invention. The scope of protection claimed by the appended claims and their equivalents is defined.
Claims
1. A method for dynamically scaling up and down computing resources based on load prediction, characterized in that: The method includes the following steps in sequence: (1) Real-time monitoring of computing power resource utilization indicators through a resource monitoring platform, including CPU utilization, GPU utilization, memory size, disk size and network bandwidth; (2) Based on the resource utilization index, the improved load forecasting algorithm is used to predict the future load situation and obtain the forecast results; (3) Compare the prediction results with the preset thresholds of various resource utilization indicators and automatically make expansion and contraction decisions; (4) Based on the expansion and contraction decision, when it is predicted that the resource utilization index of the future load will exceed the preset threshold of each resource utilization index, the corresponding resources will be automatically increased; when it is predicted that the resource utilization index of the future load will be lower than the preset threshold, the corresponding resources will be automatically reduced to meet the load demand.
2. The method for dynamic scaling up and down computing resources based on load prediction according to claim 1, characterized in that: Step (1) specifically includes the following steps in sequence: (1a) Create a dedicated Datadog account; (1b) Install Datadog Agent on the server to be monitored according to the official documentation; (1c) Configure the resource utilization indicators of the computing resources to be monitored through the Datadog dashboard; (1d) Create one or more dashboards as needed to display and statistically analyze real-time data.
3. The method for dynamic scaling up and down computing resources based on load prediction according to claim 1, characterized in that: In step (2), the improved load prediction algorithm specifically includes the following sequential steps: (2a) Preprocess the various resource utilization indicators to obtain preprocessed data. The preprocessed data forms a dataset, which is then divided into training sets according to the proportions. and validation set; (2b) Improve the LSTM model to obtain the improved LSTM model; (2c) Using the training set The improved LSTM model is trained to obtain the trained model, and then the trained model is validated using a validation set. (2d) Input the resource utilization indicators of the current load into the trained model to obtain the prediction results.
4. The method for dynamic scaling up and down computing resources based on load prediction according to claim 3, characterized in that: In step (2a), the preprocessing of various resource utilization indicators specifically includes the following steps: (2a1) Perform data cleaning on various resource utilization indicators to obtain cleaned data. Data cleaning refers to performing missing value processing, outlier detection and removal operations on the original data to ensure data quality. (2a2) Normalize the cleaned data, set a benchmark value for each resource utilization indicator, and divide the value of each resource utilization indicator at each time point by the benchmark value to obtain the normalized value.
5. The method for dynamic scaling up and down computing resources based on load prediction according to claim 3, characterized in that: Step (2b) specifically includes the following sequential steps: (2b1) Introduce stochastic gradient descent into the LSTM model to iteratively update the model parameters until final convergence; in each iteration, randomly select parameters from the training set... Select a training sample 'a' and update the model parameters in the direction of the negative gradient: ； in, These are the updated model training parameters; These are the model training parameters before the update; It is the learning rate used to control the step size of gradient descent. ; This represents the actual CPU utilization. CPU utilization as predicted by the LSTM model; This represents the actual GPU utilization. This represents the GPU utilization predicted by the LSTM model. For actual memory; The memory size predicted by the LSTM model; This is the actual disk size; The disk size predicted by the LSTM model; This represents the actual network bandwidth. The network bandwidth predicted by the LSTM model; (2b2) Introduce batch normalization in the LSTM model, based on each training set... The inputs to the network layers are standardized to a Gaussian distribution with zero mean and unit variance to accelerate the model convergence process. ； in, and All of these are trainable parameters, used to amplify and shift the standardized values; for The mean, ; For training set Total number of samples; for variance ; (2b3) Using the multi-output mean squared error loss function Replace the mean squared error loss function in the LSTM model: 。 6. A system for implementing the dynamic scaling up and down method for computing resources based on load prediction as described in any one of claims 1 to 5, characterized in that: include: The dynamic monitoring module monitors and statistically analyzes resource utilization metrics such as CPU utilization, GPU utilization, memory size, disk size, and network bandwidth in real time, so that the load prediction module can use them to make predictions. The load forecasting module is used to predict future load conditions and obtain the forecast results of various resource utilization indicators, which are then used by the resource scaling decision module to make resource scaling decisions. The resource expansion / contraction decision module is used to compare the predicted results of various resource utilization indicators with the preset thresholds of various resource utilization indicators, and obtain the comparison results to determine whether to expand or shrink the resources, so that the resource expansion / contraction execution module can expand or shrink the resources. The resource scaling execution module is used to automatically increase or decrease the number of CPU cores and virtual machine instances based on the resource scaling decision results.