Prediction-based multi-target virtual machine integration method

Abstract

The invention discloses a prediction-based multi-target virtual machine integration method, belongs to the field of Internet cloud computing application, and solves the problems that the future resource use condition of a virtual machine cannot be accurately predicted and a plurality of targets cannot be simultaneously optimized in an existing method. According to the method, the resource use condition of the VM in the future is predicted by using a method of mixing Autoregressive Integrated Moving Average (ARIMA) and Support Vector Regression (SVR), and the energy consumption, the performance reduction, the migration cost and the resource waste are minimized at the same time by using a non-dominant sorting genetic algorithm (NSGA)-II (Non-dominant sorting genetic algorithm II). And based on the predicted multi-objective optimization result, a decision is made for integration of the virtual machine, so that the integration of the virtual machine has good timeliness and practical application value.

Description

Technical Field

[0001] The present invention relates to the application in the field of Internet cloud computing, and in particular to a prediction-based multi-objective virtual machine integration method. Background Art

[0002] Virtual machine (VM) consolidation aims to run VMs on the minimum number of physical machines (PMs). VM consolidation can significantly reduce energy consumption in applications, improve the quality of service (QoS) of applications, and improve the resource utilization of the system. However, the load on the cloud is unstable and highly variable. If VM consolidation is performed based on the current state of the load, the consolidation results may soon become invalid. VM consolidation based on utilization prediction makes consolidation decisions timely, thereby better optimizing the performance goals of the system and applications. On the other hand, the service provider (SP) and different users in the data center may have different interests. Therefore, when performing VM consolidation, multiple different goals need to be optimized simultaneously. Some research methods attempt to predict the resource requirements of cloud environments. [1] Gray models are used to predict the CPU and memory usage of hosts. Gray models do not require a large amount of training data. However, it cannot guarantee accurate predictions for frequently fluctuating workloads. [2] Gray-Markov (GM)-based models are used to predict future resource usage. [3] used a discrete-time Markov chain (DTMC) model to predict future resource usage. It proposed a multi-objective VM placement method by using the ε-dominance-based multi-objective artificial bee colony (ε-MOABC) algorithm to search for the optimal solution to the VM placement problem. Some studies also used machine learning (ML) techniques to predict resource usage. [4] proposed a host states naive Bayesian prediction (HSNBP) model in cloud data centers. The HSNBP model can predict overloaded hosts using a naive Bayesian (NB) classifier. [5] proposed an ARIMA model for workload prediction. Due to its simple implementation, LR has been used to estimate future resource usage in many studies. [6-9] used the LR method to predict CPU usage. However, the LR technique only considers linear features in time series data. The relationship between resource demand data and time is more in line with a curvilinear relationship. In

[10] , the authors replaced LR with K-nearest neighbor regression (KNNR). To avoid unnecessary VM migration,

[11] proposed a real-time VM migration estimation model based on Bayesian network.

[12] proposed a self-guided load forecasting method (SDWF).SDWF uses a multilayer neural network (MNN) to learn historical data and predict the next workload value.

[13] A multi-resource feed-forward neural network (MR-FNN) is proposed to simultaneously predict multiple resource requirements. MR-FNN uses a differential evolution algorithm to enhance its learning and prediction capabilities.

[14] SVR (Support vector regression) is used to predict the future usage of multi-attribute host resources. It uses the sequential minimal optimization (SMO) algorithm to train the SVR model.

[15] A cloud resource allocation framework based on reinforcement learning (RL) mechanism and fuzzy logic (FL) is proposed;.

[0003] The present invention proposes a multi-objective VM integration method based on prediction, which uses a hybrid method of Auto-Regressive Integrated Moving Average (ARIMA) and SVR to predict the future resource usage of VMs, and then performs VM integration based on the predicted values. This method can accurately predict the future resource usage of VMs, and can also minimize energy consumption, performance degradation, migration costs, and resource waste at the same time, which makes the VM integration method proposed in the present invention have good timeliness and practical application value;

[0004] References

[0005] [1]J.Jheng, F.Tseng, H.Chao, Li-Der Chou, "A novel VM workload prediction using greyforecasting model in cloud data center", presented at the2014Int.Conf.Inf.Net., Phuket,

[0006] TH, Feb. 10-12, 2014.

[0007] [2]Sun-Yuan Hsieh et al.,“Utilization-prediction-aware virtualmachine consolidation approachfor energy-efficient cloud data centers”,JPARALLEL DISTR COM,vol.139,pp.99-109,

[0008] May.2020.DOI:10.1016 / j.jpdc.2019.12.014.

[0009] [3]Monireh H.Sayadnavard et al.,“A multi-objective approach forenergy-efficientand reliabledynamic VM consolidation in cloud data centers”,ENG SCI TECHNOL,vol.26,Feb.2021.

[0010] DOI:10.1016 / j.jestch.2021.04.014.

[0011] [4]Li L,Dong J,Zuo D,Liu J,“SLA-Aware and Energy-Efficient VMConsolidation in CloudData Centers Using Host States Naive BayesianPrediction Model”,presented at the 2018

[0012] IEEE.Int.Conf.Ubiq.Comp.,Melbourne,VC,AUS,Dec.11-13,2018.

[0013] [5]Calheiros R N et al.,“Workload Prediction Using ARIMA Model andIts Impact on CloudApplications'QoS”,IEEE T CLOUD COMPUT,vol.3,no.4,pp.449-458 Aug.2015.

[0014] DOI:10.1109 / TCC.2014.2350475.

[0015] [6]F.Farahnakian,P.Liljeberg,J.Plosila,“Lircup:Linear regressionbased cpu usage predictionalgorithm for live migration ofvirtual machines indata centers”,presented at the 2013 Euro.

[0016] Conf.Soft.Eng.Adv.Appl.,Santander,Cantabria,ES,Sep.4-6,2013.

[0017] [7]Farahnakian,F.et al.,“Using ant colony system to consolidate vmsfor green cloudcomputing”,IEEE T SERVICE COMPUT,pp.187-198,Oct.Pt.2.2015.DOI:

[0018] 10.1109 / TSC.2014.2382555.

[0019] [8]Hieu et al.,“Virtual machine consolidation with usage predictionfor energy-efficient clouddata centers”,presented at the 2015 IEEE 8thInt.Conf.Clou.Comp.,,New York,USA,

[0020] Jun.27-Jul.2,2015.

[0021] [9]S.B.Shaw,J.P.Kumar,A.K.Singh,“Energy-performance trade-offthroughrestricted virtualmachine consolidation in cloud data center”,Presented atthe

[0022] 2017 Int.Conf.Inte.Comp.Cont.,Coimbatore,IND,Jun.23-24,2017.

[0023]

[10] F.Farahnakian et al.,“Energy aware consolidation algorithm basedon k-nearest neighborregression for cloud data centers”,presented at the2013IEEE / ACM Int.Conf.Util.Clou.

[0024] Comp.,Dresden,GRE,Dec.9-12,2013.

[0025]

[11] Li,Z.,Yan,C.,Yu,X.,Yu,N.,“Bayesian network-based virtual machinesconsolidationmethod”,FUTURE GENER COMP SY,vol.69,pp.75-87,Apr.2017.DOI:

[0026] 10.1016 / j.future.2016.12.008.

[0027]

[12] Jitendra Kumar,Ashutosh Kumar Singh,Rajkumar Buyya,“Self directedlearning basedworkload forecasting model for cloud resource management”,INFORM SCIENCES,vol.

[0028] 543,pp.345-366,Jan.2021.DOI:10.1016 / j.ins.2020.07.012.

[0029]

[13] Deepika Saxena, Ashutosh Kumar Singh, "A proactive autoscaling and energy-efficientVM allocation framework using online multi-resource neural network for cloud data center",

[0030] Neurocomputing,Vol.426,pp.248-264,Feb.2021.DOI:10.1016 / j.neucom.2020.08.076.

[0031]

[14] L.Abdullah et al., "Predicting Multi-Attribute Host ResourceUtilization Using SupportVector Regression Technique", IEEE ACCESS, vol.8, pp.66048-66067, Mar.2020.DOI:

[0032] 10.1109 / ACCESS.2020.2984056.

[0033]

[15] Thandar Thein et al., "Reinforcement learning based methodology for energy-efficient resource allocation in cloud data centers", J KING SAUDUNIV-COM, vol.32, no.10, pp.

[0034] 1127-1139,Dec.2020.DOI:10.1016 / j.jksuci.2018.11.005. Summary of the invention

[0035] The present invention can accurately predict the future resource usage of VMs and use the second-generation non-dominated sorting genetic II algorithm (NSGA-II) to optimize four objectives simultaneously. Based on the predicted multi-objective optimization results, decisions are made on virtual machine integration, making virtual machine integration have good timeliness and practical application value;

[0036] In order to achieve the above object, the present invention is implemented by the following technical solutions:

[0037] A prediction-based multi-objective virtual machine integration method comprises the following steps:

[0038] Step 1) establishing energy consumption model, performance model, migration model and resource waste model;

[0039] Step 2) predicting the resource usage of the VM at the next moment, including the following sub-steps:

[0040] 2.1. Divide the dataset of VM resource usage into training set and test set;

[0041] 2.2. Determine p, d and q of the ARIMA model based on the training set;

[0042] 2.3. Use the planning method to determine the α and β parameters of the ARIMA model;

[0043] 2.4. Preprocess data for SVR model;

[0044] 2.5. Apply the k-CV strategy to determine the optimal parameters for the kernel function of SVR;

[0045] 2.6. Use the SMO algorithm to train the SVR, that is, determine the parameters λ and b of the SVR model;

[0046] 2.7. Test the SVR model using the test set;

[0047] 2.8. Use the ARIMA model obtained in steps 2.2 and 2.3 to predict the VM resource usage value at the next moment;

[0048] 2.9. Use the ARIMA model to predict several VM resource usage values ​​in history, and then calculate a set of error histories;

[0049] 2.10. Based on the error history obtained in step 2.9, use the SVR model obtained in step 2.6 to predict the error at the next moment.

[0050] 2.11. Use the error prediction value obtained in step 2.10 to correct the prediction value obtained in step 2.8 to obtain the final prediction value of VM resource usage at the next moment;

[0051] Step 3) using the resource usage prediction value of the VM obtained in step 2) to detect the overload / underload of the host, and select the VM that needs to be migrated;

[0052] Step 4) using the predicted value of VM resource usage obtained in step 2) to calculate the value of the objective function corresponding to the model established in step 1), using the second generation non-dominated sorting genetic algorithm (Non-dominated sorting genetic algorithm 2, Nsga2) to optimize the four objectives of energy consumption, performance, migration cost and resource waste, and obtain the solution to the VM integration problem, that is, the optimal mapping of VM to PM;

[0053] Step 5) using the optimal solution obtained in step 4), implement VM integration;

[0054] Further, the ARIMA model in steps 2.2, 2.3 and 2.9 is expressed as: t =α0+β1y t-1 +β2y t-2 +…+β p y t-p +∈ t -α1∈ t-1 -α2∈ t-2 -…-α q ∈ t-q , where β i (i=1,2,…,p) is the AR coefficient at lag p, α i (i=0,1,…,q) is the MA coefficient at lag q,∈ t is the prediction error for period t, y t is a real value, and the integers p and q are the orders of the model;

[0055] Furthermore, the SVR model in steps 2.4-2.7 is expressed as: in, and α i is the Lagrange multiplier, k(x,x i ) is the kernel function, b is a deviation;

[0056] Furthermore, the SMO in step 2.6 is expressed as:

[0057]

[0058] Where the input T is the training data, and the outputs λ and b are the parameters of the SVR model function. Eq. (1) in line 6 of the algorithm is expressed as: Where the input T is the training data, and the outputs λ and b are the parameters of the SVR model function. Eq. (1) in line 6 of the algorithm is expressed as: in, represents the old value of λ2, E1 and E2 are errors, expressed as: E i =f(x i )-y i,i=1,2,…,mk 11 , k 22 and k 12 Using the formula k ij =φ(x i ) T φ(x j )=k(x i ,x j ) indicates that here, It is used to map the n-dimensional vector x to an m-dimensional vector. The 7th line of Eq. (2) is expressed as: in, is the original solution of λ2 without shearing, H and L are two constants. The 8th line of Eq. (3) is expressed as: in and are the old values ​​of λ1 and λ2, is the new value of λ2. Eq.(4) in line 9 is expressed as: Where ∈ is the tolerable error. Eq. (5) is expressed as: It is expressed as: It is expressed as: It is expressed as: It is expressed as:

[0059] Compared with the prior art, the present invention has the following beneficial effects:

[0060] 1. The present invention adopts a hybrid model of ARIMA and SVR as a prediction model, uses historical data of VM resource demand to train ARIMA, and uses historical error data to train SVR. After SVR corrects ARIMA, the prediction model can accurately predict the future demand for VM resources, thereby providing accurate prediction values ​​for host overload / overload detection and VM integration;

[0061] 2. The present invention uses the NSGA-II algorithm to generate offspring using the two operations of non-dominant sorting and crowding calculation based on the selection operator, crossover operator and mutation operator. This makes the present invention solve the problem of being unable to compare the quality of solutions on multiple objectives compared with other algorithms that can only perform single-objective optimization, and the diversity and convergence of the solutions obtained by the search are better;

[0062] 3. The present invention optimizes four objectives, namely, energy consumption, performance, migration cost and resource waste, based on the predicted value of VM resources. Thus, the present invention can reduce the energy consumption of the data center and improve the performance, and reduce the number of VM migrations and resource waste, which is more valuable in practical applications. BRIEF DESCRIPTION OF THE DRAWINGS

[0063] Figure 1 Virtual Machine Integration Flowchart

[0064] Figure 2 Performance comparison with Nsga2 algorithm

[0065] Figure 3 Comparison with other overload detection algorithms. DETAILED DESCRIPTION

[0066] The present invention is further described in detail below with reference to the accompanying drawings and implementation examples;

[0067] A prediction-based multi-objective virtual machine integration method comprises the following steps:

[0068] Step 1) establishing energy consumption model, performance model, migration model and resource waste model;

[0069] Step 2) predicting the resource usage of the VM at the next moment, including the following sub-steps:

[0070] 2.1. Divide the dataset of VM resource usage into training set and test set;

[0071] 2.2. Determine p, d and q of the ARIMA model based on the training set;

[0072] 2.3. Use the planning method to determine the α and β parameters of the ARIMA model;

[0073] 2.4. Preprocess data for SVR model;

[0074] 2.5. Apply the k-CV strategy to determine the optimal parameters for the kernel function of SVR;

[0075] 2.6. Use the SMO algorithm to train the SVR, that is, determine the parameters λ and b of the SVR model;

[0076] 2.7. Test the SVR model using the test set;

[0077] 2.8. Use the ARIMA model obtained in steps 2.2 and 2.3 to predict the VM resource usage value at the next moment;

[0078] 2.9. Use the ARIMA model to predict several VM resource usage values ​​in history, and then calculate a set of error histories;

[0079] 2.10. Based on the error history obtained in step 2.9, use the SVR model obtained in step 2.6 to predict the error at the next moment.

[0080] 2.11. Use the error prediction value obtained in step 2.10 to correct the prediction value obtained in step 2.8 to obtain the final prediction value of VM resource usage at the next moment;

[0081] Step 3) using the resource usage prediction value of the VM obtained in step 2) to detect the overload / underload of the host, and select the VM that needs to be migrated;

[0082] Step 4) using the predicted value of VM resource usage obtained in step 2) to calculate the value of the objective function corresponding to the model established in step 1), using the second generation non-dominated sorting genetic algorithm (Non-dominated sorting genetic algorithm 2, Nsga2) to optimize the four objectives of energy consumption, performance, migration cost and resource waste, and obtain the solution to the VM integration problem, that is, the optimal mapping of VM to PM;

[0083] Step 5) using the optimal solution obtained in step 4), implement VM integration;

[0084] The present invention is run on a PC computer (Core i7, 3.33GHz, 6G RAM) with Windows 10 operating system and simulated using CloudSim 5.0 simulation platform. The data center consists of 50 heterogeneous physical machines and is tracked using Microsoft Azure data. The types of virtual machines include four types of virtual machine instances of Amazon EC2. Table 1 shows the CPU capabilities and memory capacities of these four virtual machine instances. The configuration of the physical machine is shown in Table 2;

[0085] Table 1 Virtual machine types

[0086]

[0087] Table 2 Physical machine configuration.

[0088]

[0089] We first compare the prediction algorithm used in the present invention with three benchmark algorithms (ARIMA, SVR, BP-NN) to compare their prediction performance. The comparison results are shown in Table 3:

[0090] Table 3 Comparative analysis of different prediction methods

[0091]

[0092] It can be concluded from Table 3 that the performance of the prediction algorithm of the present invention is the best among all the methods;

[0093] Next, we select energy consumption, resource waste, migration cost and SLA violation as performance indicators, and select the Nsga2 algorithm without prediction function as the benchmark method. The comparison between the present invention and the benchmark method in terms of energy consumption, resource waste, etc. is shown in the attached figure of the specification. Figure 2 As shown. The results show that the present invention is superior to the benchmark method in terms of energy consumption, performance, migration cost, and resource waste, which is due to the high predictive ability of the present invention. The present invention provides predictability for overload and underload detection, avoiding the problem of rapid invalidation of detection results, which also provides predictability for VM integration;

[0094] Finally, we compare with other overload / underload detection algorithms (LR, MAD and IQR). The comparison results are as follows Figure 3 As shown. The results show that the present invention outperforms the baseline method in terms of energy consumption, performance, migration cost, and resource waste. This is due to the fact that the present invention can capture the linear and nonlinear characteristics of the data, achieve timely overload and underload detection, accurate VM and target PM selection, and avoid unnecessary VM migration;

[0095] The above is only a specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any changes or substitutions that can be easily thought of by a person skilled in the art within the technical scope disclosed by the present invention should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention shall be based on the protection scope of the claims.

Claims

1. A prediction-based multi-objective virtual machine integration method, comprising the following steps: Step 1) establishing energy consumption model, performance model, migration model and resource waste model; Step 2) predicting the resource usage of the VM at the next moment, including the following sub-steps: 2.

1. Divide the dataset of VM resource usage into training set and test set; 2.

2. Determine p, d and q of the ARIMA model based on the training set; 2.

3. Use the planning method to determine the α and β parameters of the ARIMA model; 2.

4. Preprocess data for SVR model; 2.

5. Apply the k-CV strategy to determine the optimal parameters for the kernel function of SVR; 2.

6. Use the SMO algorithm to train the SVR, that is, determine the parameters λ and b of the SVR model; 2.

7. Test the SVR model using the test set; 2.

8. Use the ARIMA model obtained in steps 2.2 and 2.3 to predict the VM resource usage value at the next moment; 2.

9. Use the ARIMA model to predict several VM resource usage values ​​in history, and then calculate a set of error histories; 2.

10. Based on the error history obtained in step 2.9, use the SVR model obtained in step 2.6 to predict the error at the next moment. 2.

11. Use the error prediction value obtained in step 2.10 to correct the prediction value obtained in step 2.8 to obtain the final prediction value of VM resource usage at the next moment; Step 3) using the resource usage prediction value of the VM obtained in step 2) to detect the overload / underload of the host, and select the VM that needs to be migrated; Step 4) using the predicted value of VM resource usage obtained in step 2) to calculate the value of the objective function corresponding to the model established in step 1), using the second generation non-dominated sorting genetic algorithm (Nsga2) to optimize the four objectives of energy consumption, performance, migration cost and resource waste, to obtain the solution to the VM integration problem, that is, the optimal mapping from VM to PM; Step 5) Using the optimal solution obtained in step 4), VM consolidation is performed.

2. The prediction-based multi-objective virtual machine integration method according to claim 1, characterized in that: The ARIMA model in steps 2.2, 2.3 and 2.9 is represented as: t =α0+β1y t-1 +β2y t-2 +…+β p y t-p +∈ t -α1∈ t-1 -α2∈ t-2 -…-α q ∈ t-q , where β i (i=1,2,…,p) is the AR coefficient at lag p, α i (i=0,1,…,q) is the MA coefficient at lag q,∈ t is the prediction error for period t, y t are real values, and the integers p and q are the orders of the model.

3. The multi-target virtual machine integration method according to claim 1, characterized in that: The SVR model in steps 2.4-2.7 is expressed as: in, and α i is the Lagrange multiplier, k(x,x i ) is the kernel function and b is a deviation.

4. The prediction-based multi-objective virtual machine integration method according to claim 1, characterized in that: The SMO in step 2.6) is expressed as: Where the input T is the training data, and the outputs λ and b are the parameters of the SVR model function. Eq. (1) in line 6 of the algorithm is expressed as: Where the input T is the training data, and the outputs λ and b are the parameters of the SVR model function. Eq. (1) in line 6 of the algorithm is expressed as: in, represents the old value of λ2, E1 and E2 are errors, expressed as: E i =f(x i )-y i ,i=1,2,…,mk 11 , k 22 and k 12 Using the formula k ij =φ(x i ) T φ(x j )=k(x i ,x j ) indicates that here, It is used to map the n-dimensional vector x to an m-dimensional vector. The 7th line of Eq. (2) is expressed as: in, is the original solution of λ2 without shearing, H and L are two constants. The 8th line of Eq. (3) is expressed as: in and are the old values ​​of λ1 and λ2, is the new value of λ2. Eq.(4) in line 9 is expressed as: Where ∈ is the tolerable error. Eq. (5) is expressed as: Eq. (6) is expressed as: Eq. (7) is expressed as: Eq. (8) is expressed as: Eq. (9) is expressed as:

Images