Virtual resource scheduling method and device, electronic equipment, storage medium and product

Abstract

The invention provides a virtual resource scheduling method and apparatus, an electronic device, a storage medium and a product. The method comprises the steps of obtaining current performance index data of a virtual machine; inputting the current performance index data of the virtual machine into a virtual machine resource load prediction model to obtain first prediction performance index data of the virtual machine; wherein the virtual machine resource load prediction model is obtained by performing fine tuning training on a pre-trained large language model based on historical performance index data of a virtual machine; and performing virtual resource scheduling based on the first prediction performance index data of the virtual machine. According to the method, virtual resource intelligent dynamic scheduling based on a large language model fine tuning technology can be realized, the problem that a traditional prediction model is obviously insufficient in perception capability of virtual machine environment characteristics and business scene characteristics is solved, and the identification and prediction capability of a virtual machine resource working mode is obviously improved; moreover, the method can adapt to a dynamic environment, and achieves the periodical and real-time adjustment and optimization of virtual machine resources, thereby guaranteeing the efficient utilization of the resources.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and in particular to a virtual resource scheduling method, device, electronic device, storage medium and product. Background Art

[0002] As a cutting-edge computing service model today, cloud computing can provide elastic computing resources in a convenient, pay-as-you-go manner through the Internet. These computing resources come from a highly integrated and flexibly configurable resource pool, and tenants can obtain and release resources in an efficient and non-human-intervention manner. Virtual resources, as a typical manifestation of cloud resources, provide a solid foundation for the rapid deployment and stable operation of various applications. Specifically, cloud computing aggregates or subdivides cloud resources through the use of virtualization technologies such as the Openstack virtualization platform to achieve on-demand allocation and efficient use of resources.

[0003] However, existing virtual resource management methods are clearly insufficient in their ability to perceive virtual machine environment characteristics and business scenario characteristics, resulting in a decrease in the ability to predict virtual resource usage patterns, which in turn leads to the inefficient use of resources. Summary of the invention

[0004] The present invention provides a virtual resource scheduling method, device, electronic device, storage medium and product, which are used to solve the defect that the prior art has an obviously insufficient perception ability of virtual machine environment characteristics and business scenario characteristics, resulting in a decrease in the prediction ability of virtual resource usage patterns, and realizes intelligent dynamic scheduling of virtual resources based on large language model fine-tuning technology, and periodic and real-time adjustment and optimization of virtual machine resources, thereby ensuring efficient use of resources.

[0005] The present invention provides a virtual resource scheduling method, comprising: Get the current performance indicator data of the virtual machine; Inputting the current performance indicator data of the virtual machine into the virtual machine resource load prediction model to obtain first predicted performance indicator data of the virtual machine; wherein the virtual machine resource load prediction model is obtained by fine-tuning the pre-trained large language model based on the historical performance indicator data of the virtual machine; Virtual resource scheduling is performed based on the first predicted performance indicator data of the virtual machine.

[0006] According to a virtual resource scheduling method provided by the present invention, the virtual machine resource load prediction model is trained in the following manner: Based on a predefined data structure, feature construction is performed on the historical performance indicator data of the virtual machine to obtain a first data set; the data features of the first data set at least include a timestamp and a virtual central processing unit vCPU utilization rate corresponding to each timestamp; Perform feature decomposition on the historical time series data of vCPU utilization to obtain multi-dimensional business trend features; Adding the multi-dimensional business trend feature to the first data set to obtain a second data set; Performing feature engineering and standardization processing on the second data set to obtain a fine-tuned data model set; Based on the fine-tuning data model set, the pre-trained large language model is fine-tuned to obtain the virtual machine resource load prediction model.

[0007] According to a virtual resource scheduling method provided by the present invention, the multi-dimensional business trend characteristics include trend components, period components, residual components, nonlinear trend components and integrated trend indexes; the historical time series data of vCPU utilization is subjected to feature decomposition to obtain multi-dimensional business trend characteristics, including: Using a local weighted regression smoothing algorithm, the original data points in the first data set are weighted with the corresponding weight function to obtain a trend component; Performing weighted calculation on a first difference between the original data point and the corresponding trend component and a weight function corresponding to the first difference to obtain a period component; Subtract the trend component and the period component from the first data set to obtain a residual component; A polynomial regression fitting algorithm is used to calculate the nonlinear trend component based on the target data points; A weighted calculation is performed based on the trend component, the period component, the residual component and the nonlinear trend component to obtain an integrated trend index.

[0008] According to a virtual resource scheduling method provided by the present invention, the pre-trained large language model is fine-tuned based on the fine-tuning data model set to obtain the virtual machine resource load prediction model, including: Dividing the fine-tuning data model set into a training set and a validation set; Inputting the training set into a pre-trained large language model, iteratively performing forward propagation and back propagation on the training set, updating weight parameters of the pre-trained large language model by gradient descent to minimize a loss function, and evaluating model performance based on the validation set; The model parameters of the trained large language model are saved to obtain the virtual machine resource load prediction model.

[0009] According to a virtual resource scheduling method provided by the present invention, the first predicted performance indicator data of the virtual machine is a first predicted sequence value of vCPU utilization; the virtual resource scheduling based on the first predicted performance indicator data of the virtual machine includes: The aggregation strategy algorithm is used to process the first predicted sequence value of the vCPU utilization to obtain the future resource load index of the virtual machine; Calculating the future resource load of the host machine based on the future resource load index of the virtual machine; Determine the virtual machines to be migrated and the target host machine based on the future resource load index of each virtual machine and the future resource load of each host machine; Migrate the virtual machine to be migrated to the target host machine.

[0010] According to a virtual resource scheduling method provided by the present invention, the aggregation strategy algorithm is used to process the first predicted sequence value of the vCPU utilization to obtain the future resource load index of the virtual machine, including: Calculate the mean and standard deviation of the first predicted series of values ​​of vCPU utilization; Based on the mean and the standard deviation, determine a prediction value screening interval, and screen out a plurality of first prediction values ​​within the prediction value screening interval from the first prediction sequence values; Filtering a predetermined number of target first prediction values ​​from the plurality of first prediction values ​​in descending order of the first prediction values; Based on the predetermined number of target first prediction values, a future resource load index of the virtual machine is calculated.

[0011] According to a virtual resource scheduling method provided by the present invention, the virtual machines to be migrated and the target host machines are determined based on the future resource load index of each virtual machine and the future resource load of each host machine, including: Compare the preset host load threshold interval with the future resource load of each host to determine a list of hosts with excessive load and a list of hosts with insufficient load in the business system; The host machine with the largest load is taken out from the list of overloaded host machines, and the virtual machine with the largest load is taken out from the list of virtual machines of the host machine with the largest load to be determined as the virtual machine to be migrated; A host machine with the smallest load is taken from the list of hosts with small loads as a target host machine.

[0012] According to a virtual resource scheduling method provided by the present invention, migrating the virtual machine to be migrated to the target host machine includes: Determine a target host machine prediction data view and a to-be-migrated virtual machine prediction data view; The target host machine prediction data view and the to-be-migrated virtual machine prediction data view are superimposed according to the time series data dimension to obtain a future load prediction set of the target host machine after the virtual machine migration; Comparing the future load prediction set of the target host machine with the preset host machine load threshold interval, and determining whether the virtual machine to be migrated can be successfully migrated according to the comparison result; If the virtual machine to be migrated can be successfully migrated, the virtual machine to be migrated is migrated to the target host machine.

[0013] According to a virtual resource scheduling method provided by the present invention, the step of determining a target host machine prediction data view and a to-be-migrated virtual machine prediction data view includes: Determine, by using the virtual machine resource load prediction model, a second prediction sequence value of the vCPU utilization of the virtual machine to be migrated based on the current time series data of the vCPU utilization of the virtual machine to be migrated, and obtain a prediction data view of the virtual machine to be migrated; Through the virtual machine resource load prediction model, based on the current time series data of the vCPU utilization of the virtual machine on the target host, the predicted sequence value of the vCPU utilization of the target host is determined to obtain a predicted data view of the virtual machine to be migrated.

[0014] According to a virtual resource scheduling method provided by the present invention, the prediction data view of the target host machine and the prediction data view of the virtual machine to be migrated are superimposed according to the time series data dimension to obtain a future load prediction set of the target host machine after the virtual machine migration, including: Calculate the historical data averages of the target host machine prediction data view and the to-be-migrated virtual machine prediction data view respectively; Calculating historical data variances of the target host machine prediction data view and the to-be-migrated virtual machine prediction data view based on the historical data averages respectively; Calculating the confidence rate of the target host machine prediction data view and the confidence rate of the virtual machine prediction data view to be migrated based on the historical data variance of the target host machine prediction data view and the virtual machine prediction data view to be migrated; The target host machine prediction data view and the to-be-migrated virtual machine prediction data view are superimposed and fused based on the confidence rate to obtain a future load prediction set of the target host machine after the virtual machine migration.

[0015] The present invention also provides a virtual resource scheduling device, comprising: The acquisition module is used to obtain the current performance indicator data of the virtual machine; A prediction module, used for inputting the current performance indicator data of the virtual machine into a virtual machine resource load prediction model to obtain first predicted performance indicator data of the virtual machine; wherein the virtual machine resource load prediction model is obtained by fine-tuning a pre-trained large language model based on the historical performance indicator data of the virtual machine; The scheduling module is used to schedule virtual resources based on the first predicted performance indicator data of the virtual machine.

[0016] The present invention also provides an electronic device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements any of the above-mentioned virtual resource scheduling methods when executing the computer program.

[0017] The present invention also provides a non-transitory computer-readable storage medium on which a computer program is stored. When the computer program is executed by a processor, the virtual resource scheduling method described in any one of the above is implemented.

[0018] The present invention also provides a computer program product, comprising a computer program, wherein when the computer program is executed by a processor, the computer program implements any of the above-mentioned virtual resource scheduling methods.

[0019] The virtual resource scheduling method, device, electronic device, storage medium and product provided by the present invention can fine-tune the pre-trained large language model through the historical performance index data of the virtual machine, and obtain a virtual machine resource load prediction model suitable for the characteristics of the virtual machine environment; by inputting the current performance index data of the virtual machine into the virtual machine resource load prediction model, the first predicted performance index data of the virtual machine is obtained, and virtual resource scheduling is performed based on the first predicted performance index data of the virtual machine, which can realize the intelligent dynamic scheduling of virtual resources based on the large language model fine-tuning technology, solve the problem that the traditional prediction model has obviously insufficient perception ability of the virtual machine environment characteristics and business scenario characteristics, and significantly improve the recognition and prediction ability of the virtual machine resource working mode. The present invention can highly adapt to dynamic environments, realize periodic and real-time adjustment and optimization of virtual machine resources, thereby ensuring efficient use of resources. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] In order to more clearly illustrate the technical solutions in the present invention or the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0021] Figure 1 It is a flowchart of a virtual resource scheduling method provided by an embodiment of the present invention.

[0022] Figure 2 It is a schematic diagram of the training process of the virtual machine resource load prediction model provided by an embodiment of the present invention.

[0023] Figure 3 It is a schematic diagram of the overall process provided by an embodiment of the present invention.

[0024] Figure 4 It is a structural diagram of a virtual resource scheduling device provided by an embodiment of the present invention.

[0025] Figure 5 It is a schematic diagram of the structure of an electronic device provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0026] In order to make the purpose, technical solution and advantages of the present invention clearer, the technical solution of the present invention will be clearly and completely described below in conjunction with the drawings of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0027] In the description of the embodiments of the present invention, the terms "include", "comprising" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also includes other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. For those of ordinary skill in the art, the specific meanings of the above terms in the present invention can be understood according to specific circumstances.

[0028] Currently, virtual resource management is usually performed in the following two ways: (1) Static configuration based on the hardware resource specifications specified by the user when applying for a virtual machine. However, this static configuration mode causes the resource specifications of virtual resources to remain constant during their operation cycle, and fails to flexibly adapt to fluctuations in resource load, thus causing resource waste or resource shortage problems. For example, resource preemption often occurs among virtual machine resources on the same host machine, causing a significant drop in the performance of the entire host machine.

[0029] (2) Dynamic scheduling methods for virtual resources: including load balancing algorithms, rule-based scheduling algorithms, and prediction-based scheduling algorithms. Load balancing algorithms aim to ensure that the load is evenly distributed throughout the cloud environment in order to maximize resource utilization and improve performance; rule-based scheduling algorithms use predefined rules and policies to schedule resources, such as priority scheduling and static allocation; prediction-based scheduling algorithms use historical data and prediction models to predict future resource requirements and allocate and schedule resources accordingly. Common methods include time series analysis and machine learning algorithms. However, the existing dynamic scheduling load balancing algorithms for virtual resources are imperfect and cannot effectively consider the real-time changes of resources and future trend predictions, making it difficult to fully optimize resource allocation; rule-based scheduling algorithms cannot adapt to complex dynamic environments and real-time demand changes and lack flexibility. Prediction-based scheduling algorithms require a large amount of historical data and complex prediction models, and it is difficult to take into account the environmental characteristics and business scenario characteristics that affect virtual resources.

[0030] In view of the above problems, the embodiment of the present invention fine-tunes the pre-trained large language model through the historical performance index data of the virtual machine, and a virtual machine resource load prediction model suitable for the characteristics of the virtual machine environment can be obtained; by inputting the current performance index data of the virtual machine into the virtual machine resource load prediction model, the first predicted performance index data of the virtual machine is obtained, and virtual resource scheduling is performed based on the first predicted performance index data of the virtual machine, which can realize the intelligent dynamic scheduling of virtual resources based on the large language model fine-tuning technology, solve the problem that the traditional prediction model has obviously insufficient perception ability of the virtual machine environment characteristics and business scenario characteristics, and significantly improve the recognition and prediction ability of the virtual machine resource working mode. The present invention can highly adapt to the dynamic environment, realize the periodic and real-time adjustment and optimization of virtual machine resources, thereby ensuring the efficient use of resources.

[0031] Figure 1 Schematic diagram of the virtual resource scheduling method provided by the embodiment of the present invention. Figure 1 The embodiment of the present invention provides a virtual resource scheduling method, which may specifically include the following steps: Step 101, obtaining current performance indicator data of the virtual machine.

[0032] It should be noted that the execution subject of the virtual resource scheduling method provided in the embodiment of the present invention may be an electronic device, a component in an electronic device, an integrated circuit or a chip. The electronic device may be a mobile electronic device or a non-mobile electronic device. Exemplarily, the mobile electronic device may be a mobile phone, a tablet computer, a laptop computer, a PDA, a wearable device, an ultra-mobile personal computer (UMPC), a netbook or a personal digital assistant (PDA), etc., and the non-mobile electronic device may be a server, a network attached storage (NAS), a personal computer (PC), a television (TV), a teller machine or a self-service machine, etc., which is not specifically limited in the embodiment of the present invention.

[0033] The performance indicator data refers to the performance indicator related to the virtual machine resource load, which can be used to monitor and evaluate the operation status and performance of the virtual machine. In the embodiment of the present invention, the performance indicator data can refer to the utilization rate of the virtual central processing unit (vCPU).

[0034] In an embodiment of the present invention, a periodic prediction scheduling strategy based on time slices can be defined to build an intelligent dynamic scheduling mechanism that can handle daily periodic changes and flexibly respond to emergencies, thereby improving resource utilization efficiency and business continuity in a cloud computing environment.

[0035] The current performance indicator data may refer to the performance indicator data within the current preset time segment (which may be recorded as H segment). Specifically, the H segment may be composed of the latest N time slices. The performance indicator data of the virtual machine is periodically collected based on the time slices, and the periodic prediction of the future performance indicator data of the virtual machine can be realized, thereby realizing the periodic scheduling of the resource load of the virtual machine.

[0036] Optionally, N may be set to 96. Optionally, 15 minutes may be defined as a time slice.

[0037] The embodiment of the present invention performs periodic prediction scheduling based on time slices. The size of the time slices can be flexibly adjusted based on the current situation to recalculate resource allocation. It can be used in scenarios of regular daily resource detection and dynamic balancing.

[0038] In an embodiment of the present invention, the current performance indicator data of all virtual machines can be obtained in real time, the future performance indicator data of all virtual machines can be predicted based on the current performance indicator data, and virtual resource scheduling can be performed based on the predicted performance indicator data, thereby achieving real-time adjustment and optimization of virtual machine resources, which is conducive to ensuring efficient use of resources.

[0039] Step 102, input the current performance indicator data of the virtual machine into the virtual machine resource load prediction model to obtain the first predicted performance indicator data of the virtual machine; wherein the virtual machine resource load prediction model is obtained by fine-tuning the pre-trained large language model based on the historical performance indicator data of the virtual machine.

[0040] The historical performance indicator data may refer to the performance indicator data of the most recent M days. Optionally, M may be set to 30.

[0041] The first predicted performance indicator data may refer to performance indicator data within a future preset time segment (which may be recorded as P segment). The P segment may be composed of the most recent Q time segments. Optionally, Q may be set to 4.

[0042] Specifically, the historical performance indicator data of the virtual machine can be obtained, and data cleaning and feature construction, time series feature extraction and decomposition, and specialized feature extraction for feature engineering and pre-data standardization can be performed on the historical performance indicator data to obtain a fine-tuned data model set. Based on the fine-tuned data model set, the pre-trained large language model is fine-tuned. After the training is completed, a virtual machine resource load prediction model suitable for the characteristics of the virtual machine environment can be obtained.

[0043] Next, the virtual machine resource load prediction model can be used to predict future performance indicator data based on the current performance indicator data of all virtual machines obtained in real time, thereby achieving real-time adjustment and optimization of virtual machine resources.

[0044] In an optional embodiment, the data view can be composed of H segments (the most recent 96 time slices) and P segments (the predicted 4 time slices in the future), totaling 100 time slices, so as to predict the virtual machine resource load performance indicator data for the next hour based on the virtual machine resource load performance indicator data for the most recent 24 hours.

[0045] The embodiment of the present invention uses a virtual machine resource load prediction model suitable for virtual machine environment characteristics to perform virtual resource scheduling, which can be highly adaptable to dynamic environments and achieve periodic and real-time adjustment and optimization of virtual machine resources, thereby ensuring efficient use of resources.

[0046] The embodiment of the present invention applies a large language model to virtual machine resource load prediction, and uses its technical advantages in deep understanding of natural language and processing complex scenario time series data to migrate the general knowledge of the pre-trained model to the field of virtual machine performance prediction through a fine-tuning strategy. By specifically adjusting the model parameters to make it more suitable for the virtual machine environment and business scenario characteristics, the large model's ability to identify and predict the working mode of virtual machine resources is improved.

[0047] Step 103: Perform virtual resource scheduling based on the first predicted performance indicator data of the virtual machine.

[0048] Specifically, the first predicted performance indicator data of the virtual machine can be integrated and processed to obtain the virtual machine resource load index, thereby providing data support for optimizing the startup host machine to achieve virtual resource scheduling.

[0049] The embodiment of the present invention fine-tunes the pre-trained large language model through the historical performance index data of the virtual machine, and can obtain a virtual machine resource load prediction model suitable for the characteristics of the virtual machine environment; by inputting the current performance index data of the virtual machine into the virtual machine resource load prediction model, the first predicted performance index data of the virtual machine is obtained, and virtual resource scheduling is performed based on the first predicted performance index data of the virtual machine, which can realize the intelligent dynamic scheduling of virtual resources based on the large language model fine-tuning technology, solve the problem that the traditional prediction model has obviously insufficient perception ability of the virtual machine environment characteristics and business scenario characteristics, and significantly improve the recognition and prediction ability of the virtual machine resource working mode. The present invention can highly adapt to the dynamic environment, realize the periodic and real-time adjustment and optimization of virtual machine resources, thereby ensuring the efficient use of resources.

[0050] Figure 2 Schematic diagram of the training process of the virtual machine resource load prediction model provided by the embodiment of the present invention. Figure 2 Based on any of the above embodiments, the virtual machine resource load prediction model can be trained by the following steps: Step 201, based on a predefined data structure, feature construction is performed on the historical performance indicator data of the virtual machine to obtain a first data set; the data features of the first data set at least include a timestamp and a virtual central processing unit vCPU utilization rate corresponding to each timestamp.

[0051] In some embodiments, the historical performance indicator data of the virtual machine is first cleaned, and then feature extraction and construction are performed based on the cleaned historical performance indicator data. The data cleaning may include missing value processing and abnormal point processing.

[0052] In an embodiment of the present invention, the fields of the historical performance indicator data of the virtual machine can be expanded and enriched through the multi-dimensional feature thinking tree technology. Specifically, based on the predefined data structure, the business perception element and the time perception element can be integrated, and the period is one piece of data every 5 minutes, and the historical performance indicator data of the virtual machine can be feature constructed to obtain the first data set.

[0053] The data structure can be seen in Table 1:

[0054] Table 1

[0055] Step 202 , feature decomposition is performed on the historical time series data of vCPU utilization to obtain multi-dimensional business trend features.

[0056] The historical time series data of vCPU utilization may refer to time series data consisting of vCPU utilization collected at various timestamps.

[0057] Specifically, the historical time series data of vCPU utilization can be decomposed into business trend characteristics in five dimensions, namely seasonality, linear trend, residual characteristics, nonlinear trend and integrated trend index, which is conducive to more accurate analysis of the cyclical changes in resource load.

[0058] Step 203: Add the multi-dimensional business trend features to the first data set to obtain a second data set.

[0059] Specifically, after obtaining the five business trend features of vCPU utilization: trend component, period component, residual component, nonlinear trend component and integrated trend index, these business trend features can be added as additional features to the first data set. A new second data set is formed .

[0060] These five business trend features help to gain a deeper understanding of the hidden trends, periodicity, and random fluctuations in the time series data of vCPU utilization, enabling the large language model to more clearly distinguish different components in the data, reduce the impact of noise, and improve generalization ability, thereby better improving the prediction accuracy based on the large language model.

[0061] Step 204: perform feature engineering and standardization processing on the second data set to obtain a fine-tuning data model set.

[0062] In an embodiment of the present invention, the data field type can be converted into a form that can be directly understood by the large language model, so as to facilitate docking with the input of the fine-tuning of the large language model. Specifically, by performing feature engineering and standardization processing on the data fields in the second data set, a fine-tuning data model set can be obtained.

[0063] The data structure conversion rules can be seen in Table 2:

[0064] Table 2

[0065] Step 205: Fine-tune the pre-trained large language model based on the fine-tuning data model set to obtain the virtual machine resource load prediction model.

[0066] After processing the historical time series data of the original virtual machine vCPU utilization, a fine-tuning data model set can be obtained, and then the fine-tuning data model set is input into the pre-trained large language model for fine-tuning training to obtain a virtual machine resource load prediction model. Optionally, the pre-trained large language model can be Qwen-14B-Chat.

[0067] The embodiments of the present invention can realize the cleaning and feature construction of original historical data, the extraction and decomposition of time series features, and specialized feature extraction of feature engineering and data standardization, so as to construct a fine-tuning data model set that integrates business awareness and time awareness specifically for large language models, thereby not only significantly improving the quality of fine-tuning data and improving the accuracy of the model, but also enhancing the interpretability and robustness of the model, providing high-quality data support for building an efficient and reliable virtual machine resource load prediction model.

[0068] Based on any of the above embodiments, the multidimensional business trend features include a trend component, a period component, a residual component, a nonlinear trend component and an integrated trend index; the feature decomposition of the historical time series data of the vCPU utilization to obtain the multidimensional business trend features may specifically include the following sub-steps S11-S15: Sub-step S11, using a local weighted regression smoothing algorithm, performs weighted calculation on the original data points in the first data set and the corresponding weight function to obtain a trend component.

[0069] In the embodiment of the present invention, initialization can be performed first, and the first data set obtained after feature construction of the historical performance indicator data of the virtual machine is set as The first data set can be calculated using a local weighted regression smoothing algorithm The trend component Trend Component The calculation formula can be specifically as follows: (1); in, The first data set can be represented Data point i in Represents a weight function based on Gaussian kernel distance, which can be used to measure the contribution of data point i to target point t.

[0070] Optionally, the local weighted regression smoothing algorithm may adopt the loess smoothing algorithm.

[0071] Sub-step S12, performing weighted calculation on the first difference between the original data point and the corresponding trend component and the weight function corresponding to the first difference to obtain the periodic component.

[0072] In the embodiment of the present invention, the period component can be calculated by subtracting the trend component from the original data points in the first data set. . Periodic component The calculation formula can be specifically as follows: (2); in, The first data set can be represented Data point i in Represents a weight function based on Gaussian kernel distance, which can be used to measure the contribution of data point i to target point t. It can represent the difference between the data point and the trend component.

[0073] Sub-step S13, subtracting the trend component and the period component from the first data set to obtain a residual component.

[0074] In the embodiment of the present invention, after subtracting the trend component and the period component from the original data points in the first data set, the residual component can be obtained: The residual component The calculation formula can be specifically as follows: (3); in, represents the first data set, represents the trend component, Represents the periodic component.

[0075] Sub-step S14, using a polynomial regression fitting algorithm to calculate the nonlinear trend component based on the target data points.

[0076] In the embodiment of the present invention, a polynomial regression fitting algorithm can be used to calculate the nonlinear trend component . Nonlinear trend component The calculation formula can be specifically as follows: (4); in, is the regression coefficient and n is the order.

[0077] Sub-step S15, performing weighted calculation based on the trend component, the periodic component, the residual component and the nonlinear trend component to obtain an integrated trend index.

[0078] In the embodiment of the present invention, weighted calculation can be performed based on the trend component, the period component, the residual component and the nonlinear trend component to define the integrated trend index: , can better capture the various patterns hidden in the data. Integrated trend index The calculation formula can be specifically as follows: (5); in, is the weight of the trend component, is the weight of the periodic component, is the weight of the residual component, is the weight of the nonlinear trend component.

[0079] The embodiment of the present invention uses five business trend characteristics, namely trend component, periodic component, residual component, nonlinear trend component and integrated trend index, to facilitate a deeper understanding of hidden trends, periodicity and random fluctuations in time series data, so that the model can more clearly distinguish different components in the data, reduce the impact of noise, improve generalization ability, and thus better improve the prediction accuracy of the large language model.

[0080] Based on any of the above embodiments, the fine-tuning training of the pre-trained large language model based on the fine-tuning data model set to obtain the virtual machine resource load prediction model may specifically include the following sub-steps S21-S23: Sub-step S21, dividing the fine-tuning data model set into a training set and a validation set.

[0081] Specifically, the fine-tuning data model set can be imported into the training environment, and divided into a training set and a validation set according to the order of the time series, thereby ensuring that the large language model can learn based on past data and then verify its generalization ability on unseen data.

[0082] Sub-step S22, inputting the training set into the pre-trained large language model, iteratively performing forward propagation and back propagation on the training set, updating the weight parameters of the pre-trained large language model by gradient descent to minimize the loss function, and evaluating the model performance based on the validation set.

[0083] Specifically, the pre-trained large language model Qwen-14B-Chat can be used to initialize the model parameters. Since the Qwen-14B-Chat model has a large parameter scale and a wide range of training data sets, the model has fully absorbed and understood a large number of language structures and patterns, and has excellent semantic understanding and sequence generation capabilities. The embodiment of the present invention fine-tunes the large language model so that when processing virtual machine performance indicator sequence data, the improved resource load prediction model can quickly understand its internal logic and trends, providing a solid foundation for subsequent accurate predictions.

[0084] Specifically, according to the characteristics of the virtual machine resource load prediction task, the mean squared error (MSE) and mean absolute error (MAE) can be selected to quantify the gap between the prediction result and the true value. At the same time, the Adam optimizer can be selected to efficiently update the model parameters and accelerate the convergence process.

[0085] Specifically, after the training set is input into the pre-trained large language model, forward propagation and back propagation can be iteratively performed on the training set, and the model weight parameters can be updated through the gradient descent method to minimize the loss function. At the same time, the model performance of the large language model after parameter update can be regularly evaluated on the validation set to monitor the risk of overfitting, so as to adjust the learning rate or terminate the training early to ensure the generalization ability of the model.

[0086] Sub-step S23, saving the model parameters of the large language model after training, and obtaining the virtual machine resource load prediction model.

[0087] After training, the model's predictive capabilities can be fully evaluated on an independent test set, including accuracy, robustness, and real-time responsiveness. The model parameters of the optimized large language model can be saved to generate a virtual machine resource load prediction model for a specific professional field for subsequent deployment and application.

[0088] The virtual machine resource load prediction model obtained through the above fine-tuning training process has been deeply optimized for the virtual machine resource load prediction task. This model can not only accurately capture the dynamic changes in the virtual machine resource working mode, but also has the ability to predict future trends, which can provide strong data support for the intelligent dynamic scheduling of the host machine.

[0089] Based on any of the above embodiments, the first predicted performance indicator data of the virtual machine is a first predicted sequence value of vCPU utilization; and the virtual resource scheduling based on the first predicted performance indicator data of the virtual machine may specifically include the following sub-steps S31-S34: Sub-step S31, using an aggregation strategy algorithm to process the first predicted sequence value of the vCPU utilization to obtain a future resource load index of the virtual machine.

[0090] When scheduling virtual resources, 15 minutes can be defined as a time slice, and the data view can be composed of 100 time slices, including H slices (the latest 96 time slices) and P slices (the predicted 4 time slices in the future), so as to predict the virtual machine resource load performance index data for the next hour based on the virtual machine resource load performance index data for the latest 24 hours. Then, the prediction results are integrated to obtain the virtual machine future resource load index, so as to provide data support for the optimization of the startup host.

[0091] Specifically, the resource load performance indicator data of the virtual machine in the H-segment time slice can be cleaned and constructed with features, time series features can be extracted and decomposed, and specialized feature extraction of feature engineering and data standardization can be performed to obtain the data set to be predicted.

[0092] Specifically, the virtual machine resource load prediction model can be used to predict the resource load of the P segment time slice based on the data set to be predicted, and the first prediction sequence value of the vCPU utilization rate of 12 collection cycles in the next hour can be obtained.

[0093] Specifically, the first predicted sequence value of the vCPU utilization rate may be processed using an aggregation strategy algorithm to obtain a future resource load index of the virtual machine.

[0094] The embodiment of the present invention adopts an aggregation strategy algorithm to calculate the resource load index, which can reduce the impact of abnormal prediction values ​​and better adapt to the distribution characteristics of the prediction values, thereby improving the accuracy and robustness of the prediction; the method of aggregating the prediction results of multiple data points facilitates subsequent host optimization decisions.

[0095] Sub-step S32, calculating the future resource load of the host machine based on the future resource load index of the virtual machine.

[0096] Specifically, the resource load of the host machine in the next hour can be calculated based on the resource load index of all virtual machines on the host machine in the next hour. The calculation formula of the future resource load of the host machine can be as follows: (6); Where N is the number of virtual machines on the host machine. is the vCPU load index of the ith virtual machine, is the vCPU load of the host.

[0097] Sub-step S33, based on the future resource load index of each virtual machine and the future resource load of each host machine, determine the virtual machine to be migrated and the target host machine.

[0098] Specifically, the host with the heaviest load can be determined, and the virtual machine with the heaviest load on the host with the heaviest load can be determined as the virtual machine to be migrated. Specifically, the host with the lightest load can be determined as the target host.

[0099] Sub-step S34, migrating the virtual machine to be migrated to the target host machine.

[0100] Specifically, when it is determined that the virtual machine migration meets the resource requirements of the host machine, the virtual machine to be migrated may be migrated to the target host machine.

[0101] In the embodiment of the present invention, by defining a host resource load matching optimization algorithm, the virtual machines that need to be migrated can be automatically and intelligently identified, and the target host machine that is most suitable for the migration of these virtual machines can be recommended to solve the problem of host resource preemption and achieve load balancing between host machines, which is beneficial to improving resource utilization and business system stability.

[0102] Based on any of the above embodiments, the method of using an aggregation strategy algorithm to process the first predicted sequence value of the vCPU utilization to obtain the future resource load index of the virtual machine may specifically include the following sub-steps S311-S314: Sub-step S311, calculating the mean and standard deviation of the first predicted sequence values ​​of vCPU utilization.

[0103] In an embodiment of the present invention, a prediction sequence data set of vCPU utilization can be defined as , thereby calculating the distribution characteristics of the first predicted sequence value. Where i is the i-th time series data period point, is the predicted value of vCPU utilization at the i-th time series data period point.

[0104] Specifically, the mean of the first predicted sequence values ​​of vCPU utilization The calculation formula can be as follows: (7); in, is the predicted value of vCPU utilization at the i-th time series data cycle point, and n is the number of time series data cycle points.

[0105] Specifically, the standard deviation of the first predicted sequence value of vCPU utilization is The calculation formula is as follows: (8); in, is the predicted value of vCPU utilization at the i-th time series data period point, is the mean.

[0106] Sub-step S312, based on the mean and the standard deviation, determining a prediction value screening interval, and screening out a plurality of first prediction values ​​within the prediction value screening interval from the first prediction sequence values.

[0107] Specifically, the predicted value fluctuation threshold can be determined based on the standard deviation and the preset coefficient; the predicted value screening interval can be determined based on the mean and the predicted value fluctuation threshold.

[0108] Specifically, the predicted value fluctuation threshold can be , is the standard deviation Optionally, It can be 1.5.

[0109] Specifically, you can define the prediction value screening interval .

[0110] Specifically, a plurality of first prediction values ​​within the prediction value screening interval can be screened out from the first prediction sequence value to obtain a prediction value set: .

[0111] Sub-step S313, selecting a predetermined number of target first prediction values ​​from the multiple first prediction values ​​in descending order of the first prediction values.

[0112] Specifically, N largest target first prediction values ​​may be selected from the prediction value set S in descending order of the first prediction values. Optionally, N may be set to 3.

[0113] Sub-step S314, calculating the future resource load index of the virtual machine based on the predetermined number of target first prediction values.

[0114] Specifically, the future resource load index of the virtual machine The calculation formula can be as follows: (9); in, is the predicted value of vCPU utilization at the i-th time series data period point, and N is the number of target first predicted values.

[0115] Through the above steps, the future resource load index of each virtual machine can be calculated respectively.

[0116] Based on any of the above embodiments, the step of determining the virtual machines to be migrated and the target host machine based on the future resource load index of each virtual machine and the future resource load of each host machine may specifically include the following sub-steps S331-S333: Sub-step S331 , comparing the preset host load threshold interval with the future resource load of each host, and determining a list of hosts with excessive load and a list of hosts with insufficient load in the business system.

[0117] Specifically, the optimal threshold interval [min_load, max_load] of the host load can be defined, and the future resource load of each host can be compared and analyzed with the optimal threshold interval. If the future resource load of the host is less than the lower limit min_load of the threshold interval, the host is determined to be a host with too little load in the business system; if the future resource load of the host is greater than the upper limit max_load of the threshold interval, the host is determined to be a host with too much load in the business system.

[0118] After comparing and analyzing the future resource loads of all hosts with the optimal threshold range, a list of hosts with excessive load and hosts with insufficient load in the business system can be obtained, and the lists are sorted by load size and stored in and In the heap data structure, the characteristic of the heap data structure is that each time the largest or smallest item in the heap is taken out.

[0119] Sub-step S332, taking out the host machine with the largest load from the list of overloaded hosts, and taking out the virtual machine with the largest load from the list of virtual machines of the host machine with the largest load to determine as the virtual machine to be migrated.

[0120] When computing and identifying the host to be migrated, you can traverse the overloaded heap data. , take out the host with the largest load ,from Find the virtual machine with the largest vCPU load in the virtual machine list and determine it as the virtual machine to be migrated.

[0121] Sub-step S333, taking the host machine with the smallest load from the list of hosts with too little load as the target host machine.

[0122] When optimizing the most suitable target host for migration, you can select the heap data with too little load. The host with the lightest load is selected and determined as the target host.

[0123] Based on any of the above embodiments, migrating the virtual machine to be migrated to the target host machine may specifically include the following sub-steps S341-S344: Sub-step S341, determining the target host machine prediction data view and the to-be-migrated virtual machine prediction data view.

[0124] Specifically, the target host prediction data view can be determined based on the prediction sequence data of the vCPU utilization of all virtual machines on the target host in the next Q time slices. The to-be-migrated virtual machine prediction data view can be determined based on the prediction sequence data of the vCPU utilization of the to-be-migrated virtual machines in the next Q time slices.

[0125] Sub-step S342, superimposing the target host machine prediction data view and the to-be-migrated virtual machine prediction data view according to the time series data dimension, to obtain a future load prediction set of the target host machine after the virtual machine migration.

[0126] Specifically, the two data views can be superimposed according to the time series data dimension to obtain the future load prediction set of the target host after the virtual machine migration. .

[0127] Sub-step S343, comparing the future load prediction set of the target host machine with the preset host machine load threshold interval, and determining whether the virtual machine to be migrated can be successfully migrated according to the comparison result.

[0128] Specifically, the future load prediction set of the host can be Compare with the upper and lower limits of the host load optimal threshold interval [min_load, max_load] to determine the future load forecast set The value of all data points in the threshold range is determined to determine whether the migration is successful. Optionally, the preset probability threshold can be 95%.

[0129] if If the probability that all data points in the threshold interval are greater than or equal to the preset probability threshold, it can be determined that the virtual machine can be successfully migrated. If the probability that the values ​​of all data points in the threshold interval are less than the preset probability threshold, it can be determined that the virtual machine cannot be successfully migrated.

[0130] Sub-step S344: if the virtual machine to be migrated can be successfully migrated, then the virtual machine to be migrated is migrated to the target host machine.

[0131] After determining that the virtual machine to be migrated can be successfully migrated, the virtual machine to be migrated can be migrated to the target host machine.

[0132] The embodiment of the present invention allocates the degree of contribution based on the reliability and confidence of each predicted data view, which helps to reflect the business scenario characteristics hidden behind the data and significantly improves the accuracy and stability of the view overlay fusion results.

[0133] Based on any of the above embodiments, the step of determining the target host machine prediction data view and the to-be-migrated virtual machine prediction data view may specifically include the following sub-steps S3411-S3412: Sub-step S3411, through the virtual machine resource load prediction model, based on the current time series data of the vCPU utilization of the virtual machine to be migrated, determine the second predicted sequence value of the vCPU utilization of the virtual machine to be migrated, and obtain the predicted data view of the virtual machine to be migrated.

[0134] Specifically, after determining the virtual machine to be migrated After that, the fine-tuned virtual machine resource load prediction model can be used to predict the virtual machine resource load vCPU utilization time series data of 288 collection cycles in the future P segment (96 time slices) based on the virtual machine resource load performance indicator vCPU utilization time series data of the H segment (the latest 96 time slices), and obtain the prediction data view of the virtual machine to be migrated. .

[0135] Sub-step S3412, through the virtual machine resource load prediction model, based on the current time series data of the vCPU utilization of the virtual machine on the target host, determines the predicted sequence value of the vCPU utilization of the target host, and obtains the predicted data view of the virtual machine to be migrated.

[0136] Specifically, after determining the target host After that, the fine-tuned virtual machine resource load prediction model can be used to calculate the vCPU utilization time series data of all virtual machines on the target host in the H segment (the latest 96 time slices) based on the resource load performance indicator vCPU utilization time series data of all virtual machines on the host in the future P segment (96 time slices). After aggregation and integration, the vcpu prediction time series data view of the host in the future P segment (96 time slices) is obtained, and the host prediction data view is obtained. The calculation formula can be as follows: (10); in, Represents the predicted time series data view of each virtual machine, and N represents the number of virtual machines on the host.

[0137] Based on any of the above embodiments, the target host machine prediction data view and the to-be-migrated virtual machine prediction data view are superimposed according to the time series data dimension to obtain a future load prediction set of the target host machine after the virtual machine migration, which may specifically include the following sub-steps S3421-S3424: Sub-step S3421, respectively calculating the historical data averages of the target host machine prediction data view and the to-be-migrated virtual machine prediction data view.

[0138] In the embodiment of the present invention, the two data views may be superimposed according to the time series data dimension to obtain a load prediction set of future P slices (96 time slices) of the target host machine after the virtual machine migration.

[0139] Specifically, you can set the prediction data view of the virtual machine to be migrated The historical data set for the last 30 days is , then the average value of historical data is: (11); Among them, i represents day, j represents data point time, Represents the data point value.

[0140] Specifically, you can set the target host prediction data view The historical data set for the last 30 days is , then the average value of historical data is: (12); Among them, i represents day, j represents data point time, Represents the data point value.

[0141] Sub-step S3422, calculating the historical data variance of the target host machine prediction data view and the to-be-migrated virtual machine prediction data view based on the historical data average value, respectively.

[0142] Specifically, the calculation formula for the historical data variance of the prediction data view of the virtual machine to be migrated is as follows: (13); in, The historical average value of the data view for the virtual machine to be migrated is predicted. The historical data point value of the VM to be migrated.

[0143] Specifically, the calculation formula for the historical data variance of the target host prediction data view can be as follows: (14); in, Predict the historical data average of the data view for the target host, The historical data point value of the target host.

[0144] Sub-step S3423, based on the historical data variance of the target host machine prediction data view and the prediction data view of the virtual machine to be migrated, calculate the confidence rate of the target host machine prediction data view and the confidence rate of the prediction data view of the virtual machine to be migrated.

[0145] Specifically, the calculation formula of the confidence rate of the prediction data view of the virtual machine to be migrated can be as follows: (15); in, The historical data variance of the data view predicted for the virtual machine to be migrated, Predict the historical data variance of the data view for the target host.

[0146] Specifically, the calculation formula for the confidence rate of the target host prediction data view can be as follows: (16); in, The historical data variance of the data view predicted for the virtual machine to be migrated, Predict the historical data variance of the data view for the target host.

[0147] Sub-step S3424, based on the confidence rate, superimposes and fuses the target host machine prediction data view and the to-be-migrated virtual machine prediction data view to obtain a future load prediction set of the target host machine after the virtual machine migration.

[0148] Specifically, the data views based on the confidence rate can be superimposed and fused to obtain the load prediction set of the future P segments (96 time slices) of the target host after migration. , the calculation formula can be as follows: + (17); in, The confidence rate of the data view predicted for the virtual machine to be migrated, The confidence rate of the predicted data view for the target host, Predict data views for virtual machines to be migrated, Predict data views for the target host.

[0149] The embodiment of the present invention is based on a virtual machine migration matching algorithm based on time slice view prediction. By superimposing and fusing the virtual machine prediction data view with the host machine's load prediction data view, and comparing and analyzing with the host machine's resource load optimal threshold range, it is determined whether the virtual machine migration meets the host machine's resource requirements, thereby intelligently identifying the most matching host machine for migration.

[0150] To better understand the host resource load matching optimization algorithm, the following is an explanation using a pseudo code example: / / Initialize variables

[0151] =... / / Host load prediction set

[0152] max_load=... / / The upper limit of the optimal threshold range for host load

[0153] min_load=... / / Lower limit of the optimal threshold range for host load

[0154] =... / / The host with the smallest load

[0155] =... / / Host heap list with excessive load

[0156] =... / / Host heap list with the smallest load

[0157] =... / / From the load of large heap data List and take out the host with the largest load

[0158] =... / / From Find the virtual machine with the largest vCPU load in the virtual machine list and select it as the virtual machine that needs to be migrated

[0159] / / judge Are the values ​​of all data points greater than 95% within the optimal threshold interval [min_load, max_load]?

[0160] if all(data>= min_load and data<= max_load for data in host1_min_dataset)>0.95 then

[0161] / / Virtual machine Can be successfully migrated to the host with the least load superior

[0162] migrate( , )

[0163] / / Recalculate the host Resource load

[0164] new_load = recalculate_load()

[0165] / / Check if the new load new_load is within the optimal threshold range

[0166] if new_load>= min_load and new_load<= max_load then

[0167] / / Then from Take the second overloaded host to identify the virtual machine to be migrated

[0168] next_host_max = pop_second_from_heap( )

[0169] / / Identify the virtual machine to be migrated and execute the steps of "calculating and identifying the most suitable host machine to be migrated; optimizing the host machine and recommending the most suitable host machine for migration" (S4.3-S4.4)

[0170] vm_to_migrate = identify_vm_to_migrate(next_host_max)

[0171] repeat_steps_S4_3_to_S4_4(....)

[0172] else

[0173] / / For the host The second load virtual machine in the system is migrated, and the steps of "calculating and identifying the most suitable host machine to be migrated; optimizing the host machine and recommending the most suitable host machine for migration" are repeated iteratively (S4.3-S4.4).

[0174] second_vm_to_migrate = identify_second_vm_to_migrate( )

[0175] migrate(second_vm_to_migrate, )

[0176] repeat_steps_S4_3_to_S4_4(.....)

[0177] else / / Virtual machine Migrate to the VM with the least load Failure

[0178] / / from Take out the host with the second lightest load Perform migration calculation verification

[0179] host2_min = pop_second_from_heap()

[0180] migrate(, host2_min)

[0181] / / Iterate and repeat the "host optimization and recommend the most suitable host for migration" step (S4.4)

[0182] repeat_steps_S4_4(...)

[0183] In some optional embodiments, multiple virtual machine migration scheduling algorithms may also be verified and compared and analyzed.

[0184] Step 1. Set the implementation environment and provide the virtual machine migration implementation verification results of different migration scheduling algorithms within one month to verify the technical advancement and performance improvement of the embodiment of the present invention. The implementation environment configuration can be shown in Table 3:

[0185] Table 3

[0186] Calculation indicators may include: number of migrations (number of migrations that occurred in a month), migration success rate (number of successful migrations divided by the total number of migration attempts), resource utilization (average utilization of host resources in a month), and simulated business load (using script tools to simulate different business load patterns).

[0187] Step 2: End-to-end closed-loop computing verification of the virtual machine intelligent dynamic matching algorithm based on large model fine-tuning.

[0188] Step 2-1, define the current implementation environment data. The current CPU utilization of the host can be referred to as shown in Table 4:

[0189] Table 4

[0190] The current CPU utilization of virtual machine 1 on host 2 is 77%, and the optimal threshold range of host CPU is [60,75].

[0191] Step 2-2: Build a fine-tuning data model

[0192] The mind tree feature extraction method is used to extract the historical CPU utilization characteristics of all virtual machines, and a fine-tuned data model that integrates business awareness and time awareness is constructed.

[0193] Step 2-3: Fine-tuning of large language model

[0194] Load the Qwen-14B-Chat model, use the fine-tuning data model for fine-tuning training, and save the fine-tuned virtual machine resource load prediction model.

[0195] Step 2-4: Implement virtual machine resource load prediction based on virtual machine resource load prediction model

[0196] Load the fine-tuned model and predict the future CPU utilization of VM 1: 83%.

[0197] Step 2-5, Host optimization

[0198] Referring to Table 5, the future CPU utilization of the host machine is calculated based on the predicted data of all virtual machines as follows:

[0199] Table 5

[0200] At this time, it can be intelligently identified that virtual machine 1 on host machine 2 with the largest load needs to be migrated.

[0201] Through the virtual machine migration matching algorithm based on the time slice view prediction, the host machine 1 with the smallest predicted load is intelligently selected as the most matching host machine, and the step of migrating the virtual machine 1 to the host machine 1 is executed.

[0202] The result after the virtual machine migration is: CPU utilization of host 1: 72%

[0203] CPU utilization of host 2: 74%

[0204] By migrating virtual machine 1 on host 2 to host 1, the CPU utilization of host 2 is significantly reduced, thus avoiding serious performance degradation of host 2. At the same time, the CPU utilization of host 1 also increases steadily, and both fall within the optimal threshold range [60,75], thus achieving the goal of efficient resource utilization and load balancing.

[0205] Step 3. Comprehensive analysis of implementation results

[0206] The implementation results can be seen in Table 6. It can be seen that the method proposed in this proposal can achieve significant performance improvements in indicators such as resource accuracy and migration success rate:

[0207] Table 6

[0208] Figure 3 This is a schematic diagram of the overall process provided by an embodiment of the present invention. Figure 3When constructing a fine-tuning data model, a multi-layer feature extraction method of a virtual machine intelligent time series data thinking tree for a large language model can be used to construct a fine-tuning data model that integrates business awareness and time awareness for large model fine-tuning training. The fine-tuning data model can provide high-quality data support for large language model fine-tuning training, which is the key to the success of large language model fine-tuning training. When fine-tuning a large language model, the large language model can be fine-tuned based on historical performance data to create a virtual machine resource load prediction model that is suitable for virtual machine environment characteristics and business scenario characteristics. When implementing virtual machine resource load prediction based on a large model for virtual machine resource load prediction, an intelligent dynamic scheduling mechanism that can handle daily periodic changes and flexibly respond to emergencies can be constructed based on a periodic prediction scheduling strategy based on time slices. When optimizing the host, the virtual machine migration matching algorithm based on the time slice view prediction can be used to automatically identify the virtual machines that need to be migrated, and recommend the most suitable target host for these virtual machine migrations to solve the problem of host resource preemption.

[0209] In summary, the embodiment of the present invention performs intelligent dynamic scheduling of virtual resources based on large-scale language model fine-tuning technology. By making full use of the technical advantages of large-scale language models in virtual machine resource load prediction, a virtual machine historical performance indicator fine-tuning data model that integrates business awareness and time awareness is constructed, and a pre-trained large-scale language model Qwen-14B-Chat is used for fine-tuning training to obtain a virtual machine resource load prediction model, and resource load prediction is performed based on this model. In addition, a virtual machine migration resource intelligent matching algorithm based on a time slice view is introduced, which aims to migrate virtual machines to the most matching available host machine in time when resource bottlenecks occur, so as to avoid serious performance degradation of the host machine and affect the business. This method can effectively avoid resource preemption problems caused by uneven resource allocation during busy and idle business periods, and ensure that virtual machines are migrated to the most matching host machine, thereby achieving efficient resource utilization and load balancing.

[0210] The embodiment of the present invention provides three main technical implementation solutions involved in the implementation process: First, the embodiment of the present invention proposes a method for extracting features from a virtual machine intelligent time series data mind tree for a large language model, and constructs a fine-tuning data model for the historical performance indicators of a virtual machine that integrates business awareness and time awareness. The fine-tuning data model provides high-quality data support for fine-tuning training of a large language model, which is the key to the success of fine-tuning training of a large language model. The following core processing strategies are given: Strategy 1: Customized data cleaning rules: Improve data quality through customized missing value processing and outlier processing rules, and define specific data structures. For example, expand and enrich field information through multi-dimensional feature thinking tree technology, such as adding business-aware and time-aware cycle multi-dimensional elements, which can significantly improve the understanding and prediction capabilities of large models for complex scenarios.

[0211] Strategy 2: Provides intelligent mining and deconstruction technology for time series data: (1) For time series indicators, multi-period sequence data is used instead of just a single data point. Instead, a time series data set consisting of the most recent 12 periods is used. This can better capture the business trend characteristics of the historical performance indicators of the virtual machine, making the data easier to process and analyze, and improving the ability to accurately perceive the subtle connections between external factors and resource loads.

[0212] (2) Use the mind tree structure to perform five-dimensional expanded decomposition of time series data: By decomposing time series data into five dimensional features, namely seasonality, trend, residual characteristics, nonlinear trend and integrated trend index, it can help to better and more comprehensively understand the trends and cyclical changes in the data, thereby improving the understanding and prediction capabilities of cyclical changes.

[0213] The fine-tuning data model built based on the above core processing strategy can significantly enhance the robustness of the large model in dealing with dynamic changes and uncertainties. It has deep insight into complex scenarios, higher prediction accuracy and stronger generalization ability, and provides high-quality data support for fine-tuning training of large models of virtual machine resource load.

[0214] Second, the embodiment of the present invention provides an aggregation strategy algorithm based on prediction sequence data and provides a method for calculating the resource load index. This method can better adapt to the distribution characteristics of the predicted value, thereby improving the accuracy and robustness of the prediction and facilitating subsequent host optimization decisions.

[0215] Third, an embodiment of the present invention provides a virtual machine migration matching algorithm based on time slice view prediction. The algorithm superimposes and fuses the predicted data view of the virtual machine with the load predicted data view of its host machine based on the confidence rate, and then compares and analyzes it with the optimal threshold interval of the resource load of the host machine to determine whether the virtual machine migration meets the future resource load requirements of the host machine. The advantage of using this algorithm is that it can fully consider the trend changes of the host machine load, discover hidden business scenario characteristics by deeply mining the historical data of the host machine, and superimpose and fuse the predicted results with the actual load trend of the host machine to comprehensively evaluate the load change trend. Doing so can effectively avoid resource preemption problems such as uneven resource allocation of the host machine during busy business hours and idle hours, and ensure that the virtual machine is migrated to the most matching host machine, thereby achieving efficient resource utilization and load balancing.

[0216] The virtual resource scheduling device provided by the present invention is described below. The virtual resource scheduling device described below and the virtual resource scheduling method described above can be referred to each other.

[0217] Figure 4 Schematic diagram of the structure of the virtual resource scheduling device provided by the embodiment of the present invention. Figure 4 The embodiment of the present invention provides a virtual resource scheduling device, which may specifically include the following modules: The acquisition module 410 is used to obtain the current performance indicator data of the virtual machine; Prediction module 420, used for inputting the current performance indicator data of the virtual machine into the virtual machine resource load prediction model to obtain first predicted performance indicator data of the virtual machine; wherein the virtual machine resource load prediction model is obtained by fine-tuning the pre-trained large language model based on the historical performance indicator data of the virtual machine; The scheduling module 430 is used to schedule virtual resources based on the first predicted performance indicator data of the virtual machine.

[0218] The embodiment of the present invention fine-tunes the pre-trained large language model through the historical performance index data of the virtual machine, and can obtain a virtual machine resource load prediction model suitable for the characteristics of the virtual machine environment; by inputting the current performance index data of the virtual machine into the virtual machine resource load prediction model, the first predicted performance index data of the virtual machine is obtained, and virtual resource scheduling is performed based on the first predicted performance index data of the virtual machine, which can realize the intelligent dynamic scheduling of virtual resources based on the large language model fine-tuning technology, solve the problem that the traditional prediction model has obviously insufficient perception ability of the virtual machine environment characteristics and business scenario characteristics, and significantly improve the recognition and prediction ability of the virtual machine resource working mode. The present invention can highly adapt to the dynamic environment, realize the periodic and real-time adjustment and optimization of virtual machine resources, thereby ensuring the efficient use of resources.

[0219] Figure 5 An example of a physical structure diagram of an electronic device is shown in FIG. Figure 5As shown, the electronic device may include: a processor 510, a communication interface 520, a memory 530 and a communication bus 540, wherein the processor 510, the communication interface 520 and the memory 530 communicate with each other through the communication bus 540. The processor 510 may call the logic instructions in the memory 530 to execute the virtual resource scheduling method, which includes: obtaining the current performance indicator data of the virtual machine; inputting the current performance indicator data of the virtual machine into the virtual machine resource load prediction model to obtain the first predicted performance indicator data of the virtual machine; wherein the virtual machine resource load prediction model is obtained by fine-tuning the pre-trained large language model based on the historical performance indicator data of the virtual machine; and scheduling virtual resources based on the first predicted performance indicator data of the virtual machine.

[0220] In addition, the logic instructions in the above-mentioned memory 530 can be implemented in the form of a software functional unit and can be stored in a computer-readable storage medium when it is sold or used as an independent product. Based on such an understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art or the part of the technical solution, can be embodied in the form of a software product, and the computer software product is stored in a storage medium, including a number of instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), disk or optical disk, etc. Various media that can store program codes.

[0221] On the other hand, the present invention also provides a computer program product, which includes a computer program. The computer program can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer can execute the virtual resource scheduling method provided by the above methods, which method includes: obtaining current performance indicator data of the virtual machine; inputting the current performance indicator data of the virtual machine into a virtual machine resource load prediction model to obtain first predicted performance indicator data of the virtual machine; wherein the virtual machine resource load prediction model is obtained by fine-tuning a pre-trained large language model based on the historical performance indicator data of the virtual machine; and performing virtual resource scheduling based on the first predicted performance indicator data of the virtual machine.

[0222] On the other hand, the present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon. When the computer program is executed by a processor, it is implemented to execute the virtual resource scheduling method provided by the above methods, the method comprising: obtaining current performance indicator data of the virtual machine; inputting the current performance indicator data of the virtual machine into a virtual machine resource load prediction model to obtain first predicted performance indicator data of the virtual machine; wherein the virtual machine resource load prediction model is obtained by fine-tuning a pre-trained large language model based on the historical performance indicator data of the virtual machine; and performing virtual resource scheduling based on the first predicted performance indicator data of the virtual machine.

[0223] The device embodiments described above are merely illustrative, wherein the units described as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they may be located in one place, or they may be distributed on multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the scheme of this embodiment. Ordinary technicians in this field can understand and implement it without paying creative labor.

[0224] Through the description of the above implementation methods, those skilled in the art can clearly understand that each implementation method can be implemented by means of software plus a necessary general hardware platform, and of course, can also be implemented by hardware. Based on this understanding, the above technical solution is essentially or the part that contributes to the prior art can be embodied in the form of a software product, and the computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, a disk, an optical disk, etc., including a number of instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in each embodiment or some parts of the embodiments.

[0225] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A virtual resource scheduling method, characterized in that: include: Get the current performance indicator data of the virtual machine; Inputting the current performance indicator data of the virtual machine into the virtual machine resource load prediction model to obtain first predicted performance indicator data of the virtual machine; wherein the virtual machine resource load prediction model is obtained by fine-tuning the pre-trained large language model based on the historical performance indicator data of the virtual machine; Virtual resource scheduling is performed based on the first predicted performance indicator data of the virtual machine.

2. The virtual resource scheduling method according to claim 1, characterized in that: The virtual machine resource load prediction model is trained in the following way: Based on a predefined data structure, feature construction is performed on the historical performance indicator data of the virtual machine to obtain a first data set; the data features of the first data set at least include a timestamp and a virtual central processing unit vCPU utilization rate corresponding to each timestamp; Perform feature decomposition on the historical time series data of vCPU utilization to obtain multi-dimensional business trend features; Adding the multi-dimensional business trend feature to the first data set to obtain a second data set; Performing feature engineering and standardization processing on the second data set to obtain a fine-tuned data model set; Based on the fine-tuning data model set, the pre-trained large language model is fine-tuned to obtain the virtual machine resource load prediction model.

3. The virtual resource scheduling method according to claim 2, characterized in that: The multi-dimensional business trend characteristics include trend component, period component, residual component, nonlinear trend component and integrated trend index; The feature decomposition of the historical time series data of vCPU utilization is performed to obtain multi-dimensional business trend features, including: Using a local weighted regression smoothing algorithm, the original data points in the first data set are weighted with the corresponding weight function to obtain a trend component; Performing weighted calculation on a first difference between the original data point and the corresponding trend component and a weight function corresponding to the first difference to obtain a period component; Subtract the trend component and the period component from the first data set to obtain a residual component; A polynomial regression fitting algorithm is used to calculate the nonlinear trend component based on the target data points; A weighted calculation is performed based on the trend component, the period component, the residual component and the nonlinear trend component to obtain an integrated trend index.

4. The virtual resource scheduling method according to claim 2, characterized in that: The method of fine-tuning the pre-trained large language model based on the fine-tuning data model set to obtain the virtual machine resource load prediction model includes: Dividing the fine-tuning data model set into a training set and a validation set; Inputting the training set into a pre-trained large language model, iteratively performing forward propagation and back propagation on the training set, updating weight parameters of the pre-trained large language model by gradient descent to minimize a loss function, and evaluating model performance based on the validation set; The model parameters of the trained large language model are saved to obtain the virtual machine resource load prediction model.

5. The virtual resource scheduling method according to claim 1, characterized in that: The first predicted performance indicator data of the virtual machine is a first predicted sequence value of vCPU utilization; and the virtual resource scheduling based on the first predicted performance indicator data of the virtual machine includes: The aggregation strategy algorithm is used to process the first predicted sequence value of the vCPU utilization to obtain the future resource load index of the virtual machine; Calculating the future resource load of the host machine based on the future resource load index of the virtual machine; Determine the virtual machines to be migrated and the target host machine based on the future resource load index of each virtual machine and the future resource load of each host machine; Migrate the virtual machine to be migrated to the target host machine.

6. The virtual resource scheduling method according to claim 5, characterized in that: The aggregation strategy algorithm is used to process the first predicted sequence value of the vCPU utilization to obtain the future resource load index of the virtual machine, including: Calculate the mean and standard deviation of the first predicted series of values ​​of vCPU utilization; Based on the mean and the standard deviation, determine a prediction value screening interval, and screen out a plurality of first prediction values ​​within the prediction value screening interval from the first prediction sequence values; Filtering a predetermined number of target first prediction values ​​from the plurality of first prediction values ​​in descending order of the first prediction values; Based on the predetermined number of target first prediction values, a future resource load index of the virtual machine is calculated.

7. The virtual resource scheduling method according to claim 5, characterized in that: The step of determining the virtual machines to be migrated and the target host machines based on the future resource load index of each virtual machine and the future resource load of each host machine includes: Compare the preset host load threshold interval with the future resource load of each host to determine a list of hosts with excessive load and a list of hosts with insufficient load in the business system; The host machine with the largest load is taken out from the list of overloaded host machines, and the virtual machine with the largest load is taken out from the list of virtual machines of the host machine with the largest load to be determined as the virtual machine to be migrated; A host machine with the smallest load is taken from the list of hosts with small loads as a target host machine.

8. The virtual resource scheduling method according to claim 5, characterized in that: The migrating the virtual machine to be migrated to the target host machine includes: Determine a target host machine prediction data view and a to-be-migrated virtual machine prediction data view; The target host machine prediction data view and the to-be-migrated virtual machine prediction data view are superimposed according to the time series data dimension to obtain a future load prediction set of the target host machine after the virtual machine migration; Comparing the future load prediction set of the target host machine with the preset host machine load threshold interval, and determining whether the virtual machine to be migrated can be successfully migrated according to the comparison result; If the virtual machine to be migrated can be successfully migrated, the virtual machine to be migrated is migrated to the target host machine.

9. The virtual resource scheduling method according to claim 8, characterized in that: The step of determining the target host machine prediction data view and the to-be-migrated virtual machine prediction data view includes: Determine, by using the virtual machine resource load prediction model, a second prediction sequence value of the vCPU utilization of the virtual machine to be migrated based on the current time series data of the vCPU utilization of the virtual machine to be migrated, and obtain a prediction data view of the virtual machine to be migrated; Through the virtual machine resource load prediction model, based on the current time series data of the vCPU utilization of the virtual machine on the target host, the predicted sequence value of the vCPU utilization of the target host is determined to obtain a predicted data view of the virtual machine to be migrated.

10. The virtual resource scheduling method according to claim 8, characterized in that: The step of superimposing the target host machine prediction data view and the to-be-migrated virtual machine prediction data view according to the time series data dimension to obtain a future load prediction set of the target host machine after the virtual machine migration includes: Calculate the historical data averages of the target host machine prediction data view and the to-be-migrated virtual machine prediction data view respectively; Calculating historical data variances of the target host machine prediction data view and the to-be-migrated virtual machine prediction data view based on the historical data averages respectively; Calculating the confidence rate of the target host machine prediction data view and the confidence rate of the virtual machine prediction data view to be migrated based on the historical data variance of the target host machine prediction data view and the virtual machine prediction data view to be migrated; The target host machine prediction data view and the to-be-migrated virtual machine prediction data view are superimposed and fused based on the confidence rate to obtain a future load prediction set of the target host machine after the virtual machine migration.

11. A virtual resource scheduling device, characterized in that: include: The acquisition module is used to obtain the current performance indicator data of the virtual machine; A prediction module, used for inputting the current performance indicator data of the virtual machine into a virtual machine resource load prediction model to obtain first predicted performance indicator data of the virtual machine; wherein the virtual machine resource load prediction model is obtained by fine-tuning a pre-trained large language model based on the historical performance indicator data of the virtual machine; The scheduling module is used to schedule virtual resources based on the first predicted performance indicator data of the virtual machine.

12. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the computer program, the virtual resource scheduling method according to any one of claims 1 to 10 is implemented.

13. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the virtual resource scheduling method according to any one of claims 1 to 10 is implemented.

14. A computer program product comprising a computer program, characterized in that When the computer program is executed by a processor, the virtual resource scheduling method according to any one of claims 1 to 10 is implemented.

Images