CN116594855A - A Virtual Machine Load Prediction Method Based on Missing Value Filling - Google Patents

Abstract

The invention designs a virtual machine load prediction method based on deficiency value filling, and belongs to the technical field of cloud computing; the concurrent access quantity and the resource quantity of the virtual machine to be subjected to load prediction in a certain period are taken as known values, the virtual machine load to be predicted is taken as a missing value, and the virtual machine load is filled by using a missing value filling method, so that a task of predicting the virtual machine load is realized, and the virtual machine load prediction problem is converted into a missing virtual machine load value filling problem; then constructing and training a GAIN-VMLP (virtual machine load prediction model) based on GAIN; predicting the load of the virtual machine by using the trained GAIN-VMLP model; the virtual machine load prediction method and device can effectively solve the virtual machine load prediction problem in cloud computing, and therefore effective support is provided for elastic expansion adjustment of virtual machine resources in cloud computing.

Description

A Virtual Machine Load Prediction Method Based on Missing Value Filling

technical field

The invention belongs to the technical field of cloud computing, in particular to a virtual machine load prediction method based on missing value filling.

Background technique

With the development of cloud computing, more and more software systems are deployed in the cloud environment as SaaS to provide services. A SaaS layer application (that is, a cloud application) provides services to users by signing a service level agreement (SLA, Service Level Agreement) with the users.

Generally, cloud application providers need to lease resources (such as virtual machines) provided by cloud resource providers, deploy cloud applications, and provide users with services that meet SLA requirements. The performance of a cloud application deployed on a virtual machine (VM, Virtual Machine) is related to the amount of resources (such as CPU type and number of cores, memory capacity, network bandwidth, system disk type and capacity, etc.) and load (such as CPU Utilization, memory occupancy, etc.) are closely related, if the virtual machines under the same configuration have different loads, the performance of the cloud applications running on them will also be different. The virtual machine load is affected by factors such as the amount of virtual machine resources and the number of concurrent user visits. Under the same amount of concurrent user visits, if the amount of virtual machine resources is large, the load may be low and the performance may be good; if the amount of virtual machine resources is small, the load may be high and performance may be poor.

Cloud application providers always hope to rent resources to deploy cloud applications with the minimum resource usage cost, and provide users with services that meet SLA requirements. In order to ensure the performance of cloud applications with the minimum cost of resource usage, it is necessary to limit the load of the virtual machine within a certain range, and if it exceeds the range, it is necessary to dynamically adjust the resources of the virtual machine. In order to satisfy the initial application as much as possible or the adjusted virtual machine load is within the limited range, it is necessary to establish a model for predicting the virtual machine load based on the amount of virtual machine resources and user concurrent visits, and use this model to calculate the current concurrent visits How to adjust the amount of virtual machine resources under certain conditions, so how to predict the virtual machine load based on the amount of virtual machine resources and the number of concurrent user accesses becomes a key issue in the elastic adjustment of virtual machine resources.

Virtual machine load prediction based on virtual machine resources and user concurrent visits is a regression problem. Its input is the virtual machine's resources and concurrent visits, and its output is the virtual machine's load.

Contents of the invention

Aiming at the deficiencies in the prior art, the present invention proposes a virtual machine load prediction method based on missing value filling; the virtual machine configuration information and concurrency amount are regarded as known values, and the virtual machine load to be predicted is regarded as a missing value. GAIN (Generative Adversarial Imputation Nets, Generative Adversarial Imputation Nets) fills in missing values.

A virtual machine load prediction method based on missing value filling, specifically comprising the following steps:

Step 1: Transform the virtual machine load prediction problem into a missing virtual machine load filling problem;

The virtual machine load prediction is to predict the load of the virtual machine in a certain period of time according to the concurrent visits and resources of the virtual machine; wherein the concurrent visits, resources and load indicators are set according to actual application requirements;

The amount of concurrent visits and resources of the virtual machine for load prediction in a certain period of time is regarded as a known value, and the load of the virtual machine to be predicted is regarded as a missing value, and the virtual machine load is filled by using the missing value filling method to realize the virtual Machine load prediction task, so that the virtual machine load prediction problem is transformed into a missing virtual machine load value filling problem; specifically:

According to the actual application requirements, the concurrent visits and n virtual machine resource indicators are used to predict the load of m virtual machines, and the transformation method of transforming the problem of virtual machine load prediction into the problem of filling the missing virtual machine load value is as follows:

The virtual machine's n items of virtual machine resource quantity indicators, virtual machine concurrent visits and m items of virtual machine load indicators together form a virtual machine state described by k items (k=n+1+m) indicators; wherein, the first n items is the amount of virtual machine resources, the n+1th item is the amount of concurrent access, and the last m items are the virtual machine load to be predicted;

The specific values of each state index of the virtual machine in a certain period of time form a k-dimensional virtual machine state vector; for a virtual machine whose load is to be predicted, the virtual machine load is calculated based on the amount of virtual machine resources and the amount of concurrent user visits. Prediction, that is, using the first n+1 items of the virtual machine state vector to predict the value of the next m items; the first n+1 items of the virtual machine state vector are regarded as known values, and the last m items are regarded as missing values. The missing value is filled to complete the prediction of the virtual machine load, thus converting the virtual machine load prediction problem into a missing virtual machine load value filling problem;

Step 2: Build a GAIN-based virtual machine load prediction model GAIN-VMLP;

The GAIN-VMLP regards the amount of virtual machine resources and the amount of concurrent access (i.e., the first n+1 items of the virtual machine state vector) as known values, and takes the virtual machine load to be predicted (i.e., the last m of the virtual machine state vector) item) as the missing value, transform the virtual machine load prediction problem into the problem of filling the missing virtual machine load value, and use GAIN to fill the missing virtual machine load value, so as to complete the prediction of the virtual machine load, thus constructing a virtual machine GAIN-VMLP model for load forecasting;

The use of GAIN to fill in the missing virtual machine load value is specifically: generating fitting data for the missing virtual machine load value through the generator, and then using the discriminator to judge whether the data is real to achieve the purpose of confrontation;

The input of the GAIN-VMLP model is a k-dimensional virtual machine state vector with a missing virtual machine load value (the latter m items are the virtual machine load to be predicted), and the output is a k-dimensional virtual machine with a virtual machine load prediction value State vector (the last m items are virtual machine loads that have been predicted);

The construction of the GAIN-based virtual machine load prediction model GAIN-VMLP, the specific process is as follows:

Step S1: designing the input data vector;

The missing item in the virtual machine state vector is marked with a special value, and the special value is not within the value range of any index, forming the input data vector X; for the k-dimensional virtual machine state vector, the number of concurrent visits and the virtual machine The first n+1 items of the resource amount are known, while the last m items representing the load of the virtual machine are missing;

Step S2: designing the mask vector;

Mark the position of the missing data through the mask vector; the k-dimensional vector M=(M ₁ ,...,M _k ), the value of each component is 0 or 1, where 1 means that the value of the corresponding component in X is not missing , 0 means that the value of the corresponding component in X is not missing; each X corresponds to an M, and the value of each component in M is set according to X, when the component _Xi (i∈[1,k]) in X is 0 , the corresponding component _Mi in M is also 0; when the component _Xi (i∈[1,k]) in X is not 0, the corresponding component Mi in _M is 1, thus forming a mask vector M;

Step S3: designing random noise vectors;

Use random noise to fill in the missing data initially; randomly generate a k-dimensional random noise vector Z=(Z ₁ ,…,Z _k ), and the value range of each component is [0,1], thus forming random noise vector Z;

Step S4: designing the prompt vector;

The prompt vector strengthens the confrontation process between the generator and the discriminator, prompts the discriminator with some missing information in the original data, makes the discriminator pay more attention to the part prompted by the prompt vector, and forces the generator to generate more real data at the same time; the prompt vector The generation method of H is as follows:

Step S4.1: first generate a random vector B; the value of each component of the k-dimensional vector B=(B ₁ ,...,B _k ) is 0 or 1; the generation method of each component value is: from {1, ...,k} Randomly generate a number p, set the value of the pth component in B to 0, and set the value of the other components to 1;

Step S4.2: Generate a hint vector H according to B; the value of each component of the k-dimensional vector H=(H ₁ ,...,H _k ) is 0, 0.5 or 1, and the generation method of H is shown in formula 1:

H＝B⊙M+0.5(1-B) (1)

Among them, ⊙ represents the point product calculation of the vector;

Step S5: Design the generator G;

In the GAIN-VMLP, a generator is used to generate prediction data; the input of the generator is a k-dimensional input data vector X, a mask vector M and a random noise vector Z with missing values, and after three fully connected layers, the output is a band with output data vector with virtual machine load predictions Since the generator not only predicts the value of the position to be predicted, but also predicts the value of the original input data, the output vector should not only make the predicted value deceive the discriminator, but also make the value output of the original input data close to the real value; Therefore, the two loss functions of the generator are shown in formula 2 and formula 3:

Among them, m represents the mask vector; b represents the random vector; Indicates the discrimination result of the discriminator, indicating the probability of whether the data generated by the generator is missing data; x represents the input data vector; /> Represents the data vector generated by the generator; k represents the dimension of the data vector x;

The goal of the generator G is to minimize the weighted sum of the two loss functions, as shown in Equation 4:

Among them, K _G represents the number of training samples per batch when the generator G is trained using the gradient descent method, and α is a hyperparameter;

Step S6: Design the discriminator D;

In the GAIN-VMLP, a discriminator is used to determine whether the data is real data in the data set or false data generated by the generator, and its input is the output of the generator And the hint vector H, through the three layers of fully connected layers, output the judgment result expressed in the form of probability />

The loss function of the discriminator is shown in Equation 5:

Among them, m represents the mask vector; b represents the random vector; Indicates the discrimination result of the discriminator, indicating the probability of whether the data generated by the generator is missing data;

For Equation 5, it is expected that the discriminator can accurately judge the authenticity of the predicted virtual machine load, so the training criterion of the discriminator is shown in Equation 6:

Among them, K _D represents the number of training samples per batch when the discriminator D is trained using the gradient descent method;

Step S7: standardize the design of the input data;

Each dimension of the GAIN-VMLP input data is standardized, and the data is mapped to the interval [0, 1]. The normalization formula is shown in formula 7:

Step 3: Train the GAIN-VMLP model;

The training process of the GAIN-VMLP model is as follows:

Step 3.1: Process the training data set; the training data set is composed of several pieces of data obtained by running a virtual machine or performing a benchmark test, and each piece of data consists of n virtual machine resource index values, concurrent visits, m Composition of virtual machine load index values; randomly select several pieces of data in the data set according to the ratio, and clear the virtual machine load index values to indicate missing values;

Step 3.2: Generate the input data vector; according to the standardized design of the input data, use Equation 7 to standardize each piece of data in the data set, and mark the items with empty load values in each piece of data with -1, thus forming the A data set X composed of normalized input data vectors of several k-dimensional virtual machine state vectors;

Step 3.3: Set the batch size s; use the mini-batch gradient descent method to train the GAIN-VMLP model; the number of virtual machine state vectors input into GAIN-VMLP in each batch will be calculated according to the batch size (batch size) parameter s to control;

Step 3.4: Calculate the mask vector; according to the mask vector design method, generate a mask vector m for each data vector x in the data set X, thereby forming the mask set M of the data set X;

Step 3.5: Perform discriminator optimization training: The discriminator optimization training process is as follows:

Step 3.5.1: Select s data vectors from the dataset X And at the same time select the mask vector corresponding to the s data vector from the mask set M

Step 3.5.2: Generate s independent and identically distributed random noise Z;

Step 3.5.3: Generate s independent and identically distributed random vectors B;

Step 3.5.4: Use the generator for s data vectors generate s data vectors />

Step 3.5.5: Use the gradient descent method to update the training discriminator D based on Equation 5;

Step 3.6: Perform generator optimization training; the generator optimization training process is as follows:

Step 3.6.1: Select s data vectors from the dataset X And at the same time select the mask vector corresponding to the s data from the mask set M />

Step 3.6.2: Generate s independent and identically distributed random noise Z;

Step 3.6.3: Generate s independent and identically distributed random vectors B;

Step 3.6.4: Generate s hint vectors based on Equation 1

Step 3.6.5: Use the gradient descent method to update the training generator G based on Equation 4;

Step 4: Use the trained GAIN-VMLP model to predict the load of the virtual machine;

Set the virtual machine resource quantity index and virtual machine load index according to the actual application requirements, combine the given concurrency, form the virtual machine state vector, input it into the GAIN-VMLP model, and the model output result is the virtual machine load prediction result, so as to realize the The load of the virtual machine is predicted.

Beneficial technical effect of the present invention:

1. The present invention builds a virtual machine load prediction method based on missing value filling, discusses the problem of model establishment in the prediction method, and predicts the virtual machine load in cloud computing. The selected model and related parameters are determined according to the experimental results and actual application requirements.

2. The present invention is based on a virtual machine load prediction method filled with missing values, and introduces the prediction process and prediction algorithm of the model. After comparative analysis of experiments, this method has achieved good results in prediction accuracy and stability.

3. The present invention regards the amount of virtual machine resources and the amount of concurrent visits as known values, and regards the virtual machine load to be predicted as a missing value, thereby transforming the problem of virtual machine load prediction into a problem of filling in missing virtual machine load values, and then proposes A virtual machine load prediction method based on missing value filling is proposed, and a missing value filling algorithm GAIN is used to solve the problem, which can effectively solve the problem of virtual machine load prediction in cloud computing, and thus provide for the elasticity of virtual machine resources in cloud computing Telescopic adjustment provides effective support.

Description of drawings

Fig. 1 is a schematic diagram of a virtual machine state vector of the GAIN-VMLP model described in the embodiment of the present invention;

The mask vector diagram of the GAIN-VMLP model described in the embodiment of the present invention in Fig. 2;

FIG. 3 is an architecture diagram of the GAIN-VMLP model described in the embodiment of the present invention;

Fig. 4 the prediction result of CPU load according to the embodiment of the present invention;

Fig. 5 is the prediction result of RAM load according to the embodiment of the present invention;

Fig. 6 The experimental results of the influence of different sample sizes on the algorithm in the embodiment of the present invention;

Fig. 7 The experimental results of the influence of different distribution forms of the data set of the embodiment of the present invention on the accuracy of the algorithm;

Fig. 8 is the experimental result of the impact of different iterations on the accuracy of the algorithm in the algorithm of the embodiment of the present invention;

Fig. 9 Experimental results of test sets with different proportions at 10,000 rounds of iterations for different fully connected layers in the embodiment of the present invention;

Fig. 10 Experimental results of different proportions of test sets in different fully connected layers of the embodiment of the present invention at 20,000 rounds of iterations;

Fig. 11 Experimental results of test sets with different ratios at 30,000 rounds of iterations for different fully connected layers in the embodiment of the present invention;

Fig. 12 The experimental results of the influence of different hyperparameters α on the accuracy of the algorithm in the embodiment of the present invention;

Fig. 13 is the experimental result of the influence of different predicted positions on the accuracy of the algorithm according to the embodiment of the present invention.

Detailed ways

The present invention will be further described below in conjunction with accompanying drawing and embodiment;

A virtual machine load prediction method based on missing value filling, specifically comprising the following steps:

Step 1: Transform the virtual machine load prediction problem into a missing virtual machine load filling problem;

The load prediction of the virtual machine is to predict the load of the virtual machine in a certain period of time according to the amount of concurrent visits and the amount of resources of the virtual machine; wherein the amount of resources of the virtual machine includes CPU main frequency, number of CPU cores, memory capacity, and intranet bandwidth , external network bandwidth, system disk type, system disk capacity and many other indicators, which indicators to use depends on the virtual machine resource description provided by the cloud server provider and the actual application requirements; the load of the virtual machine includes CPU utilization, memory utilization Rate, IO consumption, internal network bandwidth usage, external network bandwidth usage and many other indicators, which indicators will be predicted will depend on actual application requirements;

The amount of concurrent visits and resources of the virtual machine for load prediction in a certain period of time is regarded as a known value, and the load of the virtual machine to be predicted is regarded as a missing value, and the virtual machine load is filled by using the missing value filling method to realize the virtual Machine load prediction task, so that the virtual machine load prediction problem is transformed into a missing virtual machine load value filling problem; specifically:

According to the actual application requirements, the concurrent visits and n virtual machine resource indicators are used to predict the load of m virtual machines, and the transformation method of transforming the problem of virtual machine load prediction into the problem of filling the missing virtual machine load value is as follows:

The virtual machine's n items of virtual machine resource quantity indicators, virtual machine concurrent visits and m items of virtual machine load indicators together form a virtual machine state described by k items (k=n+1+m) indicators; wherein, the first n items is the amount of virtual machine resources, the n+1th item is the amount of concurrent access, and the last m items are the virtual machine load to be predicted;

The specific values of each state index of the virtual machine in a certain period of time form a k-dimensional virtual machine state vector, as shown in Figure 1; The amount of visits predicts the virtual machine load, that is, uses the first n+1 items of the virtual machine state vector to predict the value of the next m items; the first n+1 items of the virtual machine state vector are regarded as known values, and the last m items As a missing value, the missing value is filled to complete the prediction of the virtual machine load, thereby converting the virtual machine load prediction problem into a missing virtual machine load value filling problem;

In the embodiment of the present invention, two virtual machine resource indicators, such as the number of CPU cores and memory capacity, and concurrent visits are used to predict two virtual machine loads, such as the CPU utilization rate and memory utilization rate of the virtual machine, for illustration; the two virtual machine resources Volume index, concurrent visits, and two virtual machine load indexes together form a virtual machine state described by five indexes; the first two items of the virtual machine state are the number of CPU cores and memory capacity, and the third is the number of concurrent visits , the last two items are CPU utilization rate and memory utilization rate; for a virtual machine whose load is to be predicted, the first three items of its virtual machine state vector are all known values, while the last two items are missing values, and the last two items are missing values Filling can complete the virtual machine load prediction work, thus transforming the virtual machine load prediction problem into a missing virtual machine load value filling problem.

Step 2: Build a GAIN-based virtual machine load prediction model GAIN-VMLP; (GAIN for VM LoadPrediction); as shown in Figure 3;

The GAIN-VMLP regards the amount of virtual machine resources and the amount of concurrent access (i.e., the first n+1 items of the virtual machine state vector) as known values, and takes the virtual machine load to be predicted (i.e., the last m of the virtual machine state vector) item) as the missing value, transform the virtual machine load prediction problem into the problem of filling the missing virtual machine load value, and use GAIN to fill the missing virtual machine load value, so as to complete the prediction of the virtual machine load, thus constructing a virtual machine GAIN-VMLP model for load forecasting;

The use of GAIN to fill in the missing virtual machine load value is specifically: generating fitting data for the missing virtual machine load value through the generator, and then using the discriminator to judge whether the data is real to achieve the purpose of confrontation;

The input of the GAIN-VMLP model is a k-dimensional virtual machine state vector with a missing virtual machine load value (the latter m items are the virtual machine load to be predicted), and the output is a k-dimensional virtual machine with a virtual machine load prediction value State vector (the last m items are virtual machine loads that have been predicted);

The construction of the GAIN-based virtual machine load prediction model GAIN-VMLP, the specific process is as follows:

Step S1: designing the input data vector;

The missing item in the virtual machine state vector is marked with a special value (this special value is not within the value range of any index) to form the input data vector X; for the k-dimensional virtual machine state vector, it represents the amount of concurrent access and The first n+1 items of the virtual machine resource amount are known, while the last m items representing the virtual machine load are missing;

Step S2: designing the mask vector; the mask vector is shown in Figure 2;

Mark the position of the missing data through the mask vector; the k-dimensional vector M=(M ₁ ,...,M _k ), the value of each component is 0 or 1, where 1 means that the value of the corresponding component in X is not missing , 0 means that the value of the corresponding component in X is not missing; each X corresponds to an M, and the value of each component in M is set according to X, when the component _Xi (i∈[1,k]) in X is 0 , the corresponding component _Mi in M is also 0; when the component _Xi (i∈[1,k]) in X is not 0, the corresponding component Mi in _M is 1, thus forming a mask vector M;

Step S3: designing random noise vectors;

Use random noise to fill in the missing data initially; randomly generate a k-dimensional random noise vector Z=(Z ₁ ,…,Z _k ), and the value range of each component is [0,1], thus forming random noise vector Z;

Step S4: designing the prompt vector;

The prompt vector strengthens the confrontation process between the generator and the discriminator, prompts the discriminator with some missing information in the original data, makes the discriminator pay more attention to the part prompted by the prompt vector, and forces the generator to generate more real data at the same time; the prompt vector The generation method of H is as follows:

Step S4.1: First generate a random vector B; the value of each component of the k-dimensional vector B=(B ₁ ,...,B _k ) is 0 or 1; the generation method of each component value is: from {1, ...,k} Randomly generate a number p, set the value of the pth component in B to 0, and set the value of the other components to 1;

Step S4.2: Generate a hint vector H according to B; the value of each component of the k-dimensional vector H=(H ₁ ,...,H _k ) is 0, 0.5 or 1, and the generation method of H is shown in formula 1:

H＝B⊙M+0.5(1-B) (1)

Among them, ⊙ represents the point product calculation of the vector;

Step S5: Design the generator G;

In the GAIN-VMLP, a generator is used to generate prediction data; the input of the generator is a k-dimensional input data vector X, a mask vector M and a random noise vector Z with missing values, and after three fully connected layers, the output is a band with output data vector with virtual machine load predictions Because the generator not only predicts the value of the position to be predicted, but also predicts the value of the original input data, so the output vector not only makes the predicted value deceive the discriminator, but also makes the value output of the original input data close to the real value; So the two loss functions of the generator are shown in Equation 2 and Equation 3:

Among them, m represents the mask vector; b represents the random vector; Indicates the discrimination result of the discriminator, indicating the probability of whether the data generated by the generator is missing data; x represents the input data vector; /> Represents the data vector generated by the generator; k represents the dimension of the data vector x;

The goal of the generator G is to minimize the weighted sum of the two loss functions, as shown in Equation 4:

Among them, K _G represents the number of training samples per batch when the generator G is trained using the gradient descent method, and α is a hyperparameter;

Step S6: Design the discriminator D;

In the GAIN-VMLP, a discriminator is used to determine whether the data is real data in the data set or false data generated by the generator, and its input is the output of the generator And the hint vector H, through the three layers of fully connected layers, output the judgment result expressed in the form of probability />

The loss function of the discriminator is shown in Equation 5:

Among them, m represents the mask vector; b represents the random vector; Indicates the discrimination result of the discriminator, indicating the probability of whether the data generated by the generator is missing data;

For Equation 5, it is expected that the discriminator can accurately judge the authenticity of the predicted virtual machine load, so the training criterion of the discriminator is shown in Equation 6:

Among them, K _D represents the number of training samples per batch when the discriminator D is trained using the gradient descent method;

Step S7: standardize the design of the input data;

Each dimension of the GAIN-VMLP input data is standardized, and the data is mapped to the [0,1] interval, and the standardized formula is shown in formula 7:

Step 3: Train the GAIN-VMLP model;

The training process of the GAIN-VMLP model is as follows:

Step 3.1: Process the training data set; the training data set is composed of several pieces of data obtained by running a virtual machine or performing a benchmark test, and each piece of data consists of n virtual machine resource index values, concurrent visits, m Composition of virtual machine load index values; randomly select several pieces of data in the data set according to the ratio, and clear the virtual machine load index values to indicate missing values;

Step 3.2: Generate the input data vector; according to the standardized design of the input data, use Equation 7 to standardize each piece of data in the data set, and mark the items with empty load values in each piece of data with -1, thus forming the A data set X composed of normalized input data vectors of several k-dimensional virtual machine state vectors;

Step 3.3: Set the batch size s; use the mini-batch gradient descent method to train the GAIN-VMLP model; the number of virtual machine state vectors input into GAIN-VMLP in each batch will be calculated according to the batch size (batch size) parameter s to control; if s is set to 4, 4 virtual machine state vectors will be selected for each batch and input to GAIN-VMLP for processing.

Step 3.4: Calculate the mask vector; according to the mask vector design method, generate a mask vector m for each data vector x in the data set X, thereby forming the mask set M of the data set X;

Step 3.5: Perform discriminator optimization training: The discriminator optimization training process is as follows:

Step 3.5.1: Select s data vectors from the dataset X And at the same time select the mask vector corresponding to the s data vector from the mask set M />

Step 3.5.2: Generate s independent and identically distributed random noise Z;

Step 3.5.3: Generate s independent and identically distributed random vectors B;

Step 3.5.4: Use the generator for s data vectors generate s data vectors />

Step 3.5.5: Use the gradient descent method to update the training discriminator D based on Equation 5;

Step 3.6: Perform generator optimization training; the generator optimization training process is as follows:

Step 3.6.1: Select s data vectors from the dataset X And at the same time select the mask vector corresponding to the s data from the mask set M />

Step 3.6.2: Generate s independent and identically distributed random noise Z;

Step 3.6.3: Generate s independent and identically distributed random vectors B;

Step 3.6.4: Generate s hint vectors based on Equation 1

Step 3.6.5: Use the gradient descent method to update the training generator G based on Equation 4;

Step 4: Use the trained GAIN-VMLP model to predict the load of the virtual machine;

Set the virtual machine resource quantity index and virtual machine load index according to the actual application requirements, combine the given concurrency, form the virtual machine state vector, input it into the GAIN-VMLP model, and the model output result is the virtual machine load prediction result, so as to realize the The load of the virtual machine is predicted.

The present invention includes a virtual machine load prediction method based on missing value filling, the prediction process of the model and a prediction algorithm; the virtual machine load prediction problem is transformed into a missing value filling problem, and the CPU load and RAM load of the virtual machine are proposed using the GAIN algorithm method for making predictions.

The virtual machine load prediction example based on missing value filling described in the present invention is as follows:

Experimental protocol design;

1. Experimental dataset. Fetch data from the cloud and assume only one service is running on each virtual machine. In this paper, a total of 16 static attribute allocation schemes for virtual machine resources are preset, and simulate Poisson distribution, normal distribution, uniform distribution, chi-square distribution, mutation distribution, exponential distribution, increasing distribution, decreasing distribution, random distribution, etc. 9 In order to form different virtual machine resource load conditions (that is, different virtual machine resource states), a total of 100 pieces of data are generated for each distributed user access frequency under each allocation method, and a total of 14400 experimental data are obtained. The data is divided into two parts, training data and test data.

2. Experimental protocol. Based on the above data sets, considering the influence of different parameters on the operation of the algorithm and the results, seven experimental schemes were designed, and the root mean square error (RMSE) was used as the evaluation basis. The smaller the RMSE, the smaller the gap between the predicted result and the actual value. The definition of RMSE is shown in Equation 9.

Experiment 1: It is used to explore the difference between the accuracy of predicting CPU load and RAM load alone and the accuracy of predicting two loads at the same time when GAIN algorithm predicts virtual machine load. In this experiment, eight groups of comparative experiments were designed, and the proportion of the test set was increased from 0.1 to 0.8 with a step size of 0.1. In each group of experiments, the GAIN algorithm is used to predict the CPU load alone, the RAM load alone, and the CPU and RAM load at the same time, and record the RMSE.

Experiment 2: It is used to explore the influence of the different number of data set samples in the GAIN algorithm on the algorithm. Set the test set ratio to 0.2, gradually increase the sample size from 2000 to 18000, and the step size is 2000, and record the RMSE of the algorithm.

Experiment 3: It is used to explore the influence of different distribution forms of data sets in the GAIN algorithm on the accuracy of the algorithm. Experiments were carried out on 9 distribution types, including Poisson distribution, normal distribution, uniform distribution, chi-square distribution, mutation distribution, exponential distribution, increasing distribution, decreasing distribution, and random distribution.

Experiment 4: It is used to explore the influence of different iterations in the GAIN algorithm on the accuracy of the algorithm. Gradually increase the number of iterations of the GAIN algorithm from 1000 to 10000, with a step size of 1000, and record its RMSE.

Experiment 5: It is used to explore the influence of different numbers of fully connected layers in the GAIN algorithm on the accuracy of the algorithm. The default GAIN is three fully connected layers. In this experiment, the number of fully connected layers is increased from 3 to 5, and the step size is 1, and the RMSE under different numbers of fully connected layers is recorded.

Experiment 6: It is used to explore the influence of different hyperparameters alpha in the GAIN algorithm on the accuracy of the algorithm. The hyperparameter alpha affects the loss function of the generator G. In this group, the hyperparameter alpha is increased from 1 to 512.

Experiment 7: It is used to explore the influence of different predicted positions in the GAIN algorithm on the accuracy of the algorithm. The position to be predicted by the GAIN-based virtual machine prediction algorithm is the last two columns, that is, the CPU load and RAM load are placed at the end. In this experiment, three sets of control experiments were set up, and the data to be predicted were placed in the first two columns, the middle two columns, and the last two columns of the data set, respectively, and the RMSE was recorded respectively to explore the influence of different prediction positions on the accuracy of the algorithm.

Experimental results and analysis;

Experiment 1: The results of GAIN algorithm predicting RAM load alone and the results of simultaneously predicting CPU load and RAM load in CPU load prediction are shown in Figure 4. The comparison of load forecasting results is shown in Figure 5.

It can be seen from Figure 4 and Figure 5 that when the GAIN algorithm predicts the CPU load and RAM load at the same time, the accuracy is almost the same as when the two loads are predicted separately. Predicting two workloads at the same time can save half the training time, and hardly reduce the prediction accuracy, which greatly reduces the cost of training the model.

Experiment 2: Experiment with the parameters of Experiment 3 in the previous section, and only predict the CPU load. The experimental results are shown in Figure 6. It can be seen from Figure 6 that with the increase of the data set, the RMSE gradually decreases, and the algorithm solution accuracy gradually increases.

Experiment 3: Forecasting only for CPU load, the experimental results are shown in Figure 7. It can be seen from Figure 7 that under different concurrency distributions, the prediction errors of the GAIN algorithm and the ELM algorithm are relatively stable, while the prediction errors of the BP algorithm and the DBN algorithm fluctuate greatly, and the prediction error of the GAIN algorithm is the smallest and the effect Best, it shows that the method proposed in this paper has high accuracy and stability when predicting CPU load.

Experiment 4: Forecasting only for CPU load, the experimental results are shown in Figure 8. It can be seen from Figure 8 that with the gradual increase in the number of iterations, the solution accuracy of the GAIN algorithm is also gradually improved.

Experiment 5: Forecasting only for CPU load, the results are shown in Figures 9, 10 and 11. The GAIN algorithm of the five-layer fully connected layer has a relatively large fluctuation and poor stability at 10,000 iterations, and is relatively stable at 20,000 and 30,000 iterations. When the number of iterations exceeds 20,000 rounds, the accuracy of the five-layer fully-connected layer is slightly higher than that of the three-layer fully-connected layer, especially when the number of iterations exceeds 20,000 rounds, the accuracy improvement is relatively large.

Experiment 6: Forecasting only for CPU load, the results are shown in Figure 12. It can be seen from Figure 12 that when alpha is lower than 128, the accuracy of the algorithm is relatively high, and when the alpha exceeds 128, the RMSE of the algorithm increases a lot, and the accuracy of the algorithm drops significantly.

Experiment 7: Forecasting only for CPU load, the results are shown in Figure 13. It can be seen from Figure 13 that under different test set ratios, different prediction positions have little effect on the accuracy of the algorithm. When the proportion of the test set exceeds 0.5, the error of the predicted position in the middle will be slightly larger than the other two positions.

Claims (9)

1. The virtual machine load prediction method based on the deficiency value filling is characterized by comprising the following steps of:

step 1: converting the virtual machine load prediction problem into a missing virtual machine load value filling problem;

step 2: constructing a GAIN-VMLP (virtual machine load prediction model) based on GAIN;

step 3: training a GAIN-VMLP model;

step 4: predicting the load of the virtual machine by using the trained GAIN-VMLP model;

setting a virtual machine resource quantity index and a virtual machine load index according to actual application requirements, combining given concurrency quantity to form a virtual machine state vector, inputting the virtual machine state vector into a GAIN-VMLP model, and obtaining a model output result which is a virtual machine load prediction result, thereby realizing the prediction of the load of the virtual machine.

2. The virtual machine load prediction method based on missing value filling according to claim 1, wherein the virtual machine load prediction in step 1 predicts the load of a virtual machine in a certain period according to the concurrent access amount and the resource amount of the virtual machine in the certain period; and setting indexes of the concurrent access quantity, the resource quantity and the load according to actual application requirements.

3. The virtual machine load prediction method based on missing value filling according to claim 1, wherein step 1 specifically comprises:

the concurrent access quantity and the resource quantity of the virtual machine to be subjected to load prediction in a certain period are taken as known values, the virtual machine load to be predicted is taken as a missing value, and the virtual machine load is filled by using a missing value filling method, so that a task of predicting the virtual machine load is realized, and the virtual machine load prediction problem is converted into a missing virtual machine load value filling problem; the method comprises the following steps:

the method for converting the virtual machine load prediction problem into the missing virtual machine load value filling problem by predicting m virtual machine loads according to the actual application requirements by adopting concurrent access quantity and n virtual machine resource quantity indexes comprises the following steps:

n virtual machine resource quantity indexes, virtual machine concurrent access quantity and m virtual machine load indexes of the virtual machine jointly form a virtual machine state described by k (k=n+1+m) indexes; the first n items are virtual machine resource amounts, the n+1th item is concurrent access amount, and the last m items are virtual machine loads to be predicted;

the specific value of each state index of the virtual machine in a certain time period forms a k-dimensional virtual machine state vector; predicting the virtual machine load based on the virtual machine resource amount and the user concurrent access amount aiming at a virtual machine of the load to be predicted, namely predicting the value of the m items by using the first n+1 items of the virtual machine state vector; and taking the first n+1 items of the virtual machine state vector as known values, taking the last m items as missing values, filling the missing values to complete the prediction work of the virtual machine load, and converting the virtual machine load prediction problem into a missing virtual machine load value filling problem.

4. The virtual machine load prediction method based on missing value filling according to claim 1, wherein in step 2, the GAIN-VMLP takes the first n+1 item of the virtual machine resource amount and the concurrent access amount, i.e. the virtual machine state vector, as a known value, takes the virtual machine load to be predicted, i.e. the last m items of the virtual machine state vector, as a missing value, converts the virtual machine load prediction problem into a missing virtual machine load value filling problem, and fills the missing virtual machine load value by adopting GAIN, thereby completing the prediction of the virtual machine load, and thus constructing a GAIN-VMLP model for virtual machine load prediction;

the filling of the missing virtual machine load value by adopting GAIN specifically comprises the following steps: generating fitting data of the missing virtual machine load value through a generator, and judging whether the data are real through a discriminator so as to achieve the aim of countermeasure;

the input of the GAIN-VMLP model is the k-dimensional virtual machine state vector with the missing virtual machine load value, namely the last m items are virtual machine loads to be predicted, and the output is the k-dimensional virtual machine state vector with the virtual machine load predicted value, namely the last m items are virtual machine loads for which the result is predicted.

5. The virtual machine load prediction method based on missing value filling according to claim 1, wherein in the step 2, the virtual machine load prediction model GAIN-VMLP based on GAIN is constructed, and the specific process is as follows:

step S1: designing an input data vector;

the missing items in the virtual machine state vector are marked by a special value which is not in the value range of any index to form an input data vector X; for a k-dimensional virtual machine state vector, the first n+1 terms of concurrent access amount and virtual machine resource amount are known, while the last m terms representing virtual machine load are missing;

step S2: designing a mask vector;

marking the missing data position through a mask vector; k-dimensional vector m= (M ₁ ,...,M _k ) The value of each component is 0 or 1, wherein 1 indicates that the value of the corresponding component in X is not missing, and 0 indicates that the value of the corresponding component in X is not missing; each X corresponds to one M, the values of the components in M are set according to X, when the components in X are _i (i∈[1,k]) When 0, the component M in the corresponding M _i Also 0; when component X in X _i (i∈[1,k]) When the value is not 0, the component M in the corresponding M _i 1, thereby forming a mask vector M;

step S3: designing a random noise vector;

initially padding the missing data using random noise; randomly generating a k-dimensional random noise vector z= (Z) ₁ ,…,Z _k ) The value range of each component is 0,1]Thereby forming a random noise vector Z;

step S4: designing a prompt vector;

the prompt vector intensifies the countermeasure process of the generator and the discriminator, prompts the discriminator for partial missing information of the original data, enables the discriminator to pay more attention to the prompt part of the prompt vector, and simultaneously forces the generator to generate more real data;

step S5: designing a generator G;

a generator used in the GAIN-VMLP to generate prediction data; the input of the generator is k-dimensional input data vector X with missing value, mask vector M and random noise vector Z, and the output data vector with virtual machine load predicted value is output through three full-connection layersThe output vector is not only required to make the predicted value deception discriminator, but also to make the value output of the original input data approach to the true value; the two loss functions of the generator are therefore shown in equations 2 and 3:

where m represents a mask vector; b represents a random vector;a discrimination result of the discriminator, which indicates a probability of whether or not the data generated by the generator is missing data; x represents an input data vector; />Representing the data vector generated by the generator; k represents the dimension of the data vector x;

the goal of generator G is to minimize the weighted sum of the two loss functions, as shown in equation 4:

wherein K is _G The number of training samples in each batch when the generator G adopts a gradient descent method for training is represented, and alpha is a super parameter;

step S6: designing a discriminator D;

the GAIN-VMLP uses a discriminator to determine whether the data is real data in the dataset or false data generated by the generator, the input of which is the output of the generatorAnd a prompt vector H for outputting a judgment result expressed in a probability form through three full connection layers>

The loss function of the discriminator is shown in fig. 5:

where m represents a mask vector; b represents a random vector;a discrimination result of the discriminator, which indicates a probability of whether or not the data generated by the generator is missing data;

for equation 5, the training criteria for the expected arbiter for the authenticity determination of the predicted virtual machine load is shown in equation 6:

wherein K is _D The number of training samples of each batch when the discriminator D adopts a gradient descent method for training is represented;

step S7: carrying out standardized design on input data;

carrying out standardization processing on each dimension of the GAIN-VMLP input data, mapping the data into a [0,1] interval, wherein a standardization formula is shown in a formula 7:

6. the virtual machine load prediction method based on missing value padding of claim 5, wherein the generating method of the hint vector H in step S4 is as follows:

step S4.1: firstly, generating a random vector B; k-dimensional vector b= (B) ₁ ,…,B _k ) The value of each component of (2) is 0 or 1; the generation method of each component value comprises the following steps: randomly generating a number p from {1, …, k }, setting the value of the p-th component in B to 0, and setting the values of the rest components to 1;

step S4.2: generating a prompt vector H according to the B; k-dimensional vector h= (H) ₁ ,…,H _k ) The generation method of H with the value of 0, 0.5 or 1 of each component is shown in the formula 1:

H＝B⊙M+0.5(1-B) (1)

wherein, as indicated by the letter, "-represents the dot product calculation of the vector.

7. The virtual machine load prediction method based on missing value filling according to claim 1, wherein step 3 specifically comprises:

step 3.1: processing the training data set; the training data set consists of a plurality of pieces of data obtained by running or performing benchmark test of the virtual machine, wherein each piece of data consists of n virtual machine resource quantity index values, concurrent access quantity and m virtual machine load index values; randomly selecting a plurality of data in the data set according to a proportion, emptying the load index value of the virtual machine to represent a missing value;

step 3.2: generating an input data vector; according to the standardized design of input data, carrying out standardized processing on each piece of data in a data set by adopting a 7, and marking items with empty load values in each piece of data by using a-1, so as to form a data set X consisting of a plurality of input data vectors standardized by the state vectors of the k-dimensional virtual machine;

step 3.3: setting a batch processing size s; training the GAIN-VMLP model using a small batch gradient descent method; the number of virtual machine state vectors input into the GAIN-VMLP per batch will be controlled according to a batch size (batch size) parameter s;

step 3.4: calculating a mask vector; generating a mask vector M for each data vector X in the data set X according to a mask vector design method, thereby forming a mask set M of the data set X;

step 3.5: and (3) performing discriminant optimization training:

step 3.6: and performing generator optimization training.

8. The virtual machine load prediction method based on missing value padding of claim 7, wherein the optimized training process of the step 3.5 arbiter is as follows:

step 3.5.1: selecting s data vectors from data set XAnd simultaneously selecting mask vectors corresponding to the s data vectors from the mask set M +.>

Step 3.5.2: generating s independent random noise Z distributed in the same way;

step 3.5.3: generating s independent random vectors B distributed in the same way;

step 3.5.4: using a generator for s data vectorsGenerating s data vectors->

Step 3.5.5: the training discriminant D is updated based on equation 5 using a gradient descent method.

9. The virtual machine load prediction method based on missing value padding of claim 7, wherein the step 3.6 generator optimizes the training process as follows:

step 3.6.1: selecting s data vectors from data set XAnd simultaneously selecting mask vectors corresponding to the s data from the mask set M>

Step 3.6.2: generating s independent random noise Z distributed in the same way;

step 3.6.3: generating s independent random vectors B distributed in the same way;

step 3.6.4: generating s hint vectors based on 1

Step 3.6.5: the training generator G is updated based on equation 4 using a gradient descent method.