Machine Learning Approach for Live Migration Cost Prediction in

VMware Environments

Mohamed Esam Elsaid

1 a

, Hazem M. Abbas

2 b

and Christoph Meinel

1 c

Internet Technologien und Systeme, Hasso-Plattner Institut, Potsdam Uni., Potsdam, Germany

Dept. Computer and Systems Engineering, Ain Shams University, Cairo, Egypt

Keywords:

Cloud Computing, Virtual, Live Migration, VMWare, vMotion, Modeling, Overhead, Cost, Datacenter,

Prediction, Machine Learning.

Abstract:

Virtualization became a commonly used technology in datacenters during the last decade. Live migration is

an essential feature in most of the clusters hypervisors. Live migration process has a cost that includes the

migration time, downtime, IP network overhead, CPU overhead and power consumption. This migration cost

cannot be ignored, however datacenter admins do live migration without expectations about the resultant cost.

Several research papers have discussed this problem, however they could not provide a practical model that

can be easily implemented for cost prediction in VMware environments. In this paper, we propose a machine

learning approach for live migration cost prediction in VMware environments. The proposed approach is

implemented as a VMware PowerCLI script that can be easily implemented and run in any vCenter Server

Cluster to do data collection of previous migrations statistics, train the machine learning models and then

predict live migration cost. Testing results show how the proposed framework can predict live migration time,

network throughput and power consumption cost with accurate results and for different kinds of workloads.

This helps datacenters admins to have better planning for their VMware environments live migrations.

1 INTRODUCTION

Live migration of virtual machines is a key feature in

virtual environments, private and public cloud com-

puting datacenters. Using live migration, virtual ma-

chines can be moved from a physical host to another

while the applications are running online. This is due

to the negligible service interruption during the mi-

gration process. Servers load balance, power saving,

fault tolerance and dynamic virtual machines alloca-

tion are all dependent on live migration (Choudhary

et al., 2017). During the live migration process, the

VM CPU cache, memory pages and IO buffers con-

tents are migrated. However the storage content is

shared between the source and the target servers, so

storage content is not migrated. Live migration is sup-

ported by VMware (vMotion), Xen (XenMotion), Mi-

crosoft Hyper-V and KVM.

Live Migration cost can be classiﬁed into perfor-

mance of migration and performance loss of VM and

https://orcid.org/0000-0003-1577-5290

https://orcid.org/0000-0001-9128-3111

https://orcid.org/0000-0002-3212-6201

energy overhead (Strunk, 2012). Performance of mi-

gration includes the migration time and the downtime

that is consumed during a live migration process. Per-

formance loss of VM includes the overhead of the

migration process on the servers CPU, memory and

network throughput. And ﬁnally the energy overhead

in Joule or the power consumption overhead in Watt

due to live migration. As we will discuss in section

II, the live migration cost is variable with VM mem-

ory size, network utilization, CPU utilization and the

dirty pages rate which depends on the running work-

load. For the best of our knowledge, until now live

migration is done by datacenter admins with no ex-

pectations about the migration cost. So the admins

do the live migrations and then see the impact of it

on the network, CPU, memory and power consump-

tion. This leads sometimes to facing failures in live

migration, bottlenecks in the datacenter infrastructure

resources, and downgrade in the VMs availability due

to longer down time especially for large memory VMs

migrations.

In this paper, we propose a practical machine

learning framework for live migration cost modelling

and prediction in VMware environments. The pro-

456

Elsaid, M., Abbas, H. and Meinel, C.

Machine Learning Approach for Live Migration Cost Prediction in VMware Environments.

DOI: 10.5220/0007749204560463

In Proceedings of the 9th International Conference on Cloud Computing and Services Science (CLOSER 2019), pages 456-463

ISBN: 978-989-758-365-0

posed framework starts with data collection about the

history of live migrations that have run in the cluster

during the past 12 hours, then use the collected statis-

tics for models training. After training the models,

the prediction phase can start to estimate the future

live migration requests cost. This should help the dat-

acenters admins to estimate the cost of single or mul-

tiple VMs live migration before proceeding with it.

This means better live migration tasks planning and

less resources bottlenecks.

The rest of this paper is structured as following. In

section II, we discuss in details the live migration

cost. In section III, the related research work is dis-

cussed. The proposed prediction technique is pre-

sented in section IV. The the modeling results are an-

alyzed in section V. Then we conclude the paper in

section VI.

2 LIVE MIGRATION COST

Live migration consists of mainly six phases; ini-

tialization, reservation, iterative pre-copy, stop-and-

copy, commitment and activation (Elsaid and Meinel,

2014). The cost of live migration is actually a result

of passing through these six phases. In this paper, we

focus on modeling and predicting the below migra-

tion cost parameters as main overhead costs that can

be measured in our test-bed and modeled with regres-

sion techniques:

• Migration Time: This is the time consumed in

seconds from the migration request initiation un-

til having the VM activated at the destination host.

This is the time consumed by the above 6 phases

of the pre-copy migration.

• Network Overhead: This is the increase in the IP

network throughput in kBps due to the VM con-

tent migration from the source host to the target

one.

• Power Consumption: live migration process con-

sumes signiﬁcant energy. This is due to the in-

crease in CPU and network utilization. In this pa-

per, we model and predict the peak power over-

head in Watt.

The CPU and down time cost parameters will be con-

sidered in an extended work to this paper.

3 RELATED WORK

Live Migration cost is analyzed by different mathe-

matical and empirical methods. Analysis and Predic-

tion of live migration cost is presented using machine

learning, or mathematical approaches. In our paper

(Elsaid and Meinel, 2014), we have provided empiri-

cal models for live migration time, network overhead

and power consumption in relation with the VM ac-

tive memory size for VMware vMotion. Cost predic-

tion could not be used in (Elsaid and Meinel, 2014)

due to the constants in the provided models. These

constants depend on the cluster characteristics like

CPU models and network bandwidth. This work is an

extension of the previous work in (Elsaid and Meinel,

2014); where the contribution is providing a machine

learning based approach for live migration cost esti-

mation. Machine learning is used mainly to deﬁne

the constants in the provided models in (Elsaid and

Meinel, 2014) during the training phase, and then the

empirical models are used to predict the VM migra-

tion cost.

In Reference (Hu et al., 2013), an analysis of

live migration time and downtime is provided and

then a comparison between Xen, KVM, Hyper-V and

VMware vSphere hypervisors is presented in terms of

storage migration and live migration time and down-

time. A comparison between Xen and VMware live

migration time and downtime is also presented in

(Salfner et al., 2011) with more investigation on the

parameters that contol the live migration time and

downtime durations. The authors (Bezerra et al.,

2017) show the impact of a VM live migration on the

running applications performance from client side.

The performance degradation of the application from

client side was measured in operations per second.

The impact of live migration on Internet Web 2.0

applications performance is discussed in (Voorsluys

et al., 2009). This is important for environments with

SLA requirements. For this purpose, a test-bed is built

in (Voorsluys et al., 2009) where the running Web

2.0 workload is Olio application, combined with Fa-

ban load generator that access the Apache 2.2.8 Web

server with MySQL database.

The other category of papers focus on live migra-

tion cost prediction. Classiﬁcation of Live migration

cost is provided in (Strunk, 2012) with an explanation

of the parameters that control migration time, down-

time and energy consumption. Also, Mathematical

models are proposed to estimate live migration time,

downtime and energy consumption.Machine learning

is used in (Berral et al., 2013) for VM placement el-

ements predictive modeling like (CPU, memory, net-

work and energy).The authors (Akoush et al., 2010)

analyze the parameters that control the migration time

and the downtime and show the impact of the work-

load on the migration performance. Markov chains

are used in (Melhem et al., 2018) for hosts utilization

prediction after live migration. The proposed Markov

Machine Learning Approach for Live Migration Cost Prediction in VMware Environments

457

Table 1: Summary of Related Work.

Paper

Research

Scope

Migration

Time

Down

Time CPU Network Power Hypervisor

(Elsaid and Meinel, 2014) Analysis X - - X X VMware

(Hu et al., 2013) Analysis X X - - -

VMware/ Xen/

Hyper-v/ KVM

(Salfner et al., 2011) Analysis X X - - - Xen / VMware

(Voorsluys et al., 2009) Analysis X X - - - Xen

(Strunk, 2012) Prediction X X - - X Xen

(Zhao and Figueiredo, 2007) Prediction X - - - -

VMware but

not vMotion

(Berral et al., 2013) Prediction - - X X X

Oracle

Virtual Box

(Jo et al., 2017) Prediction X X X X -

VMware/ Xen/

Hyper-v/ KVM

(Akoush et al., 2010) Prediction X X - - - Xen

(Salfner et al., 2012) Prediction X X - - -

VMware/

Xen/ KVM

(Melhem et al., 2018) Prediction X - - - - -

(Huang et al., 2014) Prediction X - X X X Xen

(Aldossary and Djemame, 2018) Prediction X - X - X KVM

based prediction model is used for power saving al-

gorithm that can achieve lower SLA violations, lower

VM migrations as well as less power consumption

(Melhem et al., 2018). Time series is used in (Huang

et al., 2014) for time varying resources load predic-

tion. The proposed model is used for power saving by

minimizing the number of active physical machines

with less live migration times and with satisfying the

SLA requirements. The proposed technique is tested

in a Xen cluster. A mathematical based prediction

framework is also proposed in (Aldossary and Dje-

mame, 2018). From Table I, the papers that focus

on VMware cost prediction are (Zhao and Figueiredo,

2007), (Jo et al., 2017) and (Salfner et al., 2012). To

the best of our knowledge, using machine learning for

VMware live migration cost prediction is not covered

in a practical way by any of these papers; which is

the point that we cover in this paper by the proposed

machine learning approach. We discuss the shortage

in practicality of papers (Zhao and Figueiredo, 2007),

(Jo et al., 2017) and (Salfner et al., 2012) in the fol-

lowing points:

• In paper (Zhao and Figueiredo, 2007), a VMware

cluster is built for modeling, but vMotion is not

used. The authors could do the suspend, copy and

resume operations manually for migration time

prediction. This limits the proposed modeling to

be used practically in an enterprise environments

that use vMotion by default as the live migration

feature in vSphere; without any manual opera-

tions.

• The machine learning approach proposed in (Jo

et al., 2017) can be used for most of the live migra-

tion algorithms in different hypervisors as stated

in (Jo et al., 2017). However, the proposed ma-

chine learning technique depends on massive em-

pirical tests for 40,000 live migration that were

run in order to have accurate prediction. This

means that the approach used for live migration

cost prediction will generate an intensive cost it-

self; which blocks the ability to practically imple-

ment the proposed approach in (Jo et al., 2017).

• In the proposed mathematical model in (Salfner

et al., 2012), live migration time and downtime

can be predicted but after the start of the live mi-

gration; not before. This means that the proposed

algorithm does not help the cluster admin to know

the live migration cost before proceeding with mi-

gration.

In the next section, we discuss our paper contribu-

tion in more details.

4 PROPOSED COST

PREDICTION FRAMEWORK

In this paper, we solve the challenge of having a prac-

tical live migration cost prediction for VMware en-

vironments. This is achieved by proposing a ma-

chine learning based approach that is implemented

as VMware PowerCLI script and can connect to any

VMware vCenter server to train the model and then

predict live migration cost. Here, we list our contri-

bution in this paper in the following points:

• We propose a machine learning approach for

VMware vMotion that predicts the live migration

time, network overhead and power consumption

given the active memory size of the VM.

CLOSER 2019 - 9th International Conference on Cloud Computing and Services Science

458

• The proposed approach can be practically used.

It is implemented as a VMware PowerCLI script;

that can be bounded with any VMware vCenter

server and show the cost prediction results.

• The proposed script includes data collection that

is used for the models training phase. This makes

the proposed models adaptable to each VMware

cluster automatically.

• The training phase in the proposed machine learn-

ing algorithm is fed by the ongoing live migra-

tion operations that run in the datacenter; which

increases that the prediction accuracy.

4.1 Modelling of Live Migration Cost

This paper is an extension to the proposed models in

(Elsaid and Meinel, 2014) and (Elsaid and Meinel,

2016). From these papers, the following empirical

models could be proposed for live migration time,

data rate and power consumption after applying the

regression techniques:

• The relation between the network rate and the ac-

tive memory size can be modelled as an exponen-

tial relation; as shown in equation (1).

= αe

Mem

+ β

(1)

: is the source host network throughput over-

head in kBps, V

mem

is the source host active mem-

ory size in kB at the time when the live migration

should start. α and β are the equation constants.

From equation (1).

• Migration Time: A linear relationship is obtained

between the migration time and the division of the

memory size over the transmission rate; as repre-

sented in equation (2).

mig

= a.(

mem

) + b

(2)

mig

is the migration time duration in seconds. a

and b are the equation constants.

• Peak power consumption overhead has linear re-

lation with the transmission rate; as represented in

equation (3).

mig

= c

mig

= c R

(3)

mig

is the peak power overhead in Watt, and c is

constant. From equation (3).

In our previous papers, the above models could be

used for cost analysis but not for cost prediction. This

is because of the equations constants. These con-

stants depend on the cluster hardware conﬁguration

like CPU specs, so they change from a cluster envi-

ronment to another. So in order to determine these

constants and achieve higher accuracy in cost predic-

tion, we propose a machine learning framework to

predict the live migration cost.

4.2 Machine Learning based Cost

Prediction

In this paper, machine learning is used because the

proposed models in equations (1 - 3) can not be used

in live migration cost prediction. This is due to the

constants included in the equations. These constants

values depend on the cluster hardware characteris-

tics; like CPU and network conﬁgurations. So, ma-

chine learning is needed to train the models in ref-

erence to equations (1 - 3) until the constants val-

ues are obtained for each cluster. Then, these equa-

tions can be used for cost prediction. In this sec-

tion, we present the proposed machine learning based

framework for live migration time, transfer rate and

power consumption overhead prediction. As shown

in the ﬂow chart of Fig. 1, the proposed framework

consists of two main phases, the training phase and

the prediction phase. The training phase starts when

the VMware PowerCLI script connects to the cluster

vCenter Server Appliance (vCSA). Then data collec-

tion starts with listing all the events happened in the

cluster during the last 12 hours. This 12 hours cy-

cle can be changed based on the cluster admin pref-

erence. From the collected events, vMotion events

are ﬁltered out. These vMotion events details like the

source host, target host and time stamp are captured.

Then the script calculates the complete and start time

differences in order to get the migration time of each

vMotion request. The performance logs of vCSA are

collected at the start and the completion times at the

vMotion events in order to get the active memory size

of the migrated VMs in kB, the network overhead in

kBps and the peak power change in Watt.

From the above data of each vMotion event, we use

the regression models in equations (1 - 3) to calculate

the equations constants after doing several substitu-

tion and considering the minimum Root Mean Square

Error (RMSE); equation (4).

RMSE =

i=1

− f

)

(4)

Where N is the number of sample points collected

during the last 12 hours. d

is the measured perfor-

mance value and f

is the regression equation value.

If the change in all the constants value became

greater than 10% of the last 12 hours cycle, the script

waits for more 12 hours and run again to continue in

Machine Learning Approach for Live Migration Cost Prediction in VMware Environments

459

Figure 1: Proposed Prediction Framework.

the training phase. If these changes became less than

10% of the last 12 hours, so we consider the the train-

ing phase of this cluster is ﬁnished, and the script then

moves to the prediction phase. The time consumed

until reaching this 10% convergence depends on the

changes that happen in the VMs active memory size;

which depends on the running workload. This se-

quence of data collection and models training makes

the algorithm can ﬁt at any vCenter Server cluster and

adapt its models based on the cluster conﬁguration in

order to provide cost prediction.

In the prediction phase when a vMotion request

is sent by the cluster admin, the active memory size

is captured by the script before proceeding with live

migration. Once the active memory size is known,

equation (1) is used to predict the source host net-

work throughput. Then equation (2) is used to predict

the migration time, and ﬁnally equation (3) is used to

predict the peak power consumption. The prediction

data is exported to a .csv ﬁle that the cluster admin

can read, and decide to proceed with this migration or

not.

5 TESTING ENVIRONMENT

The testing environment is shown in Fig. 2; as shown

it has a similar infrastructure to enterprise datacen-

ters. It includes the following hardware setup; Three

Hosts (Hewlett Packard DL980 G7) with 8 x Intel

Xeon (Nehalem EX) X7560, 8GB RAM, 4 NICs, 2

HBA with 2 Fiber ports per card. The three hosts are

connected to a shared storage EMC VNX5800; 1TB

LUN via FC-SAN network.The Ethernet switch is

Cisco with 1Gbps ports. From software prospective,

VMware ESXi 6.5.0 Hypervisor is used with vCenter

Server that manages both hosts and the VMs live mi-

gration. VMware PowerCLI 6.5.1 build 5377412 is

connected to the vCenter Server to run the framework

algorithm script.

In this set up we have created four Linux Ubuntu

12.04 VMs with 4 vCPU, and different RAM sizes

(1GB, 2GB, 4GB and 8GB). The VMs have mainly 3

categories of workload:

• CPU and Memory intensive: This is considered

as the worst case scenario for a running work-

load. The CPU intensive benchmark that we used

is Linpack (Lin, ) and the memory stress is the

Linux Stress Package (Mem, ).

• Network Intensive: The network stress bench-

mark that we have used is Apache Bench (AB).

Apache Bench tool stresses the web servers with

lots of requests through the network to test the

servers response.

• Idle: VM is simply an idle Ubuntu OS VM; with

no running applications.

With this testing setup, we have run 12 testing sce-

narios; as a matrix of 3 workload categories and 4

different VM sizes. For each conﬁguration, we have

run live migration at least 10 times.

6 RESULTS AND ANALYSIS

After testing the proposed approach in Fig. 1 on the

test-bed of Fig. 2, we present in this section the pre-

diction results for almost 144 readings. We start with

the training phase to show how the models are trained

until obtaining equations (1 - 3) constants with at least

90 percent accuracy.

CLOSER 2019 - 9th International Conference on Cloud Computing and Services Science

460

Figure 2: Testing Lab Layout.

6.1 Training Phase

In this phase, the script collects the last 12 hours

live migration events. Then the performance statistics

of these live migrations are gathered including their

time stamps. The migration time is calculated by the

script; given the start and end time of the live migra-

tion event. The other gathered statistics include the

active memory size, the source host transmission rate

and the peak power change. All these details are used

to train the models of equations (1 - 3) and to obtain

the constants of this cluster by solving several linear

equations. For example in order to calculate a and b of

equation (2), we use every two live migration events

statistics to generate two equations in two unknowns.

These unknowns are a and b in this example, because

the migration time, the active memory size and the

transmission rate are given. So, we gather every two

live migration events statistics to solve for the con-

stants of equation (2). The script keeps on solving for

the values of a and b until ﬁnding the changes in the

values of a and b are less than 10 percent compared to

the calculated values of last equations solution. Fig.

3 shows the changes of a and b constants versus the

number of live migration equations that were used

until reaching the 90 percent saturation. As shown

in Fig. 3; the difference in a is changing with the

number of live migrations which represents solving

more equations until the 90 percent saturation at dif-

ference equals 0.22 after 14 live migrations. At this

point a=9.04. For b constant, the script has run 50 live

migrations to reach the 90 percent saturation at differ-

ence equals 9.16. At this point b=21.04. This means

that modeling with equation (2) could be used after 50

live migration runs for this cluster. For equation (1),

we could also solve every two equations of live mi-

Figure 3: A and B Change until Saturation.

Figure 4: Alpha and Beta Change until Saturation.

Figure 5: C Change until Saturation.

grations data as linearly to obtain the values of α and

β. The is because the values of the active memory

size and the migration time are given, so we can sub-

stitute with them and then solve two equations in two

unknowns; α and β. Fig. 4 shows the differences hap-

pen in the values of α and β after each live migration

until reaching the 90 percent saturation. As shown;

the constant α could reach the saturation at difference

equals 1850 after 54 live migrations. At this point

α equals 2.02 ∗ 10

. The value of β reaches the 90

percent saturation at difference equals 2225 also after

54 live migrations. At this point, β equals 2.33 ∗ 10

This means that modeling with equation (1) can be

used after 54 live migration runs. Finally, equation

(3), which has just one unknown; c and so it can be

resolved given just one live migration statistics. So

Machine Learning Approach for Live Migration Cost Prediction in VMware Environments

461

for each live migration run in the past 12 hours, we

could read the transmission rate and the peak power

overhead and then calculate the constant c. Fig. 5

shows the changes happen with each live migration

calculation to the the constant c; as shown it 90 per-

cent saturates after just fours live migrations runs at

difference of c value equals 0.6 ∗ 10

−5

. At this point

c equals 16 ∗ 10

−5

. From the above analysis, we ﬁnd

that it required 54 live migration runs to be able to

train the models provided in equation (1 - 3). In gen-

eral, the required number of live migrations runs to

ﬁnish the training phase depends on the error gap be-

tween the training data and the regression model.

6.2 Prediction Phase

In this subsection, we build on the training phase that

we have discussed above. Now, the regression mod-

els are trained for this cluster and ready to be used for

future live migration cost prediction. The testing re-

sults in Fig. 6 - 8, show the regression models that

are used and the actual measured data after migration.

The measurement points in Fig. 6, Fig. 7 and Fig.

8 are VMs live migrations with different conﬁgura-

tions including memory size of 1GB, 2GB, 4GB and

8GB VMs that utilize three different kinds of work-

loads. As discussed in section V, these workloads are

CPU and memory intensive, network intensive and

idle VMs. This results in 12 different VM conﬁgu-

rations. Each conﬁguration is tested 12 times; which

represents the existing 144 measurement points in the

following ﬁgure. The prediction starts with Fig. 6;

so given the active memory size of the VM to be mi-

grated, the source host transmission rate can be pre-

dicted. The VM active memory size can be measured

before live migration. Fig. 6 shows the exponential

relation as a valid regression model between the active

memory size and the transmission rate. Table II shows

the RMSE of Fig. 6 in reference to equation (4). Af-

ter obtaining the transmission rate from Fig. 6, we

calculate now the active memory size over the trans-

mission rate; which is the horizontal axis of Fig. 7. So

the migration time can be predicted; using the linear

regression model of Fig. 7. The RMSE of the pre-

diction in Fig. 7 is also listed in Table II. Fig. 7 also

shows that the migration time can consume several

minutes in case of large memory and memory inten-

sive VMs. The last model is for the source host peak

power change; which is shown in Fig. 8. So given the

source host transmission rate, the peak power change

can be obtained.

All these predicted live migration cost parameters

are exported to a .csv ﬁle that can be accessed by the

cluster admin to check the estimated cost if he/she

Figure 6: Rate vs Active Memory Size.

Figure 7: Migration Time vs Active Memory Size/Rate.

Figure 8: Peak Power Overhead vs Transmission Rate.

decides to do live migration to a certain VM. This

help the admins to have better planning for live migra-

tions. This proposed framework script can adapt itself

by changing the models constants using the training

phase; which make it ﬂexible with any VMware clus-

ter.

Table 2: RMSE of the Regression Models.

Model Fig. RMSE

Transmission Rate Fig. 6 8187

Migration Time Fig. 7 15.5

Peak Power Fig. 8 1.7

CLOSER 2019 - 9th International Conference on Cloud Computing and Services Science

462

7 CONCLUSION

Live migration cost can not be ignored and might lead

to resources bottlenecks, service availability degrada-

tion and live migration failures. Several related papers

have discussed this problem by applying mathemati-

cal and empirical studies, however to the best of our

knowledge there is no related paper that could pro-

vide a practical approach that can be used and inte-

grated with VMware clusters. In this paper, we pro-

posed a practical machine learning based approach

that helps the datacenter admins to predict the live

migration cost in VMware environments. The pro-

posed framework is implemented as VMware Power-

CLI script and can connect to any vSphere vCenter

Server. We considered simplicity in the proposed ap-

proach to minimize the CPU consumption overhead

due to running the proposed approach and so make

it agile enough to be implemented in enterprise dat-

acenters. In this paper, we predict the live migra-

tion time, network throughput and power consump-

tion overhead. Testing results show that the proposed

regression based models can be used for cost predic-

tion with acceptable error.

REFERENCES

https://people.seas.harvard.edu/ apw/stress/.

www.netlib.org/linpack/.

Akoush, S., Sohan, R., Rice, A., Moore, A. W., and Hop-

per, A. (2010). Predicting the performance of virtual

machine migration. In Proceedings of the 2010 IEEE

International Symposium on Modeling, Analysis and

Simulation of Computer and Telecommunication Sys-

tems, MASCOTS ’10, pages 37–46, Washington, DC,

USA. IEEE Computer Society.

Aldossary, M. and Djemame, K. (2018). Performance and

energy-based cost prediction of virtual machines live

migration in clouds. In Proceedings of the 8th Interna-

tional Conference on Cloud Computing and Services

Science, CLOSER 2018, Funchal, Madeira, Portugal,

March 19-21, 2018., pages 384–391.

Berral, J. L., Gavald

a, R., and Torres, J. (2013). Power-

aware multi-data center management using machine

learning. In Proceedings of the 2013 42Nd Interna-

tional Conference on Parallel Processing, ICPP ’13,

pages 858–867, Washington, DC, USA. IEEE Com-

puter Society.

Bezerra, P., Martins, G., Gomes, R., Cavalcante, F., and

Costa, A. (2017). Evaluating live virtual machine mi-

gration overhead on client’s application perspective.

In 2017 International Conference on Information Net-

working (ICOIN), pages 503–508.

Choudhary, A., Govil, M. C., Singh, G., Awasthi, L. K.,

Pilli, E. S., and Kapil, D. (2017). A critical survey of

live virtual machine migration techniques. J. Cloud

Comput., 6(1):92:1–92:41.

Elsaid, M. E. and Meinel, C. (2014). Live migration impact

on virtual datacenter performance: Vmware vmotion

based study. In 2014 International Conference on Fu-

ture Internet of Things and Cloud, pages 216–221.

Elsaid, M. E. and Meinel, C. (2016). Multiple virtual

machines live migration performance modelling –

vmware vmotion based study. In 2016 IEEE Inter-

national Conference on Cloud Engineering (IC2E),

pages 212–213.

Hu, W., Hicks, A., Zhang, L., Dow, E. M., Soni, V., Jiang,

H., Bull, R., and Matthews, J. N. (2013). A quan-

titative study of virtual machine live migration. In

Proceedings of the 2013 ACM Cloud and Autonomic

Computing Conference, CAC ’13, pages 11:1–11:10,

New York, NY, USA. ACM.

Huang, Q., Shuang, K., Xu, P., Liu, X., and Su, S. (2014).

Prediction-based dynamic resource scheduling for vir-

tualized cloud systems. JNW, 9:375–383.

Jo, C., Cho, Y., and Egger, B. (2017). A machine learning

approach to live migration modeling. In Proceedings

of the 2017 Symposium on Cloud Computing, SoCC

’17, pages 351–364, New York, NY, USA. ACM.

Melhem, S. B., Agarwal, A., Goel, N., and Zaman, M.

(2018). Markov prediction model for host load de-

tection and VM placement in live migration. IEEE

Access, 6:7190–7205.

Salfner, F., Tr oger, P., and Polze, A. (2011). Down-

time Analysis of Virtual Machine Live Migration. In

The Fourth International Conference on Dependabil-

ity, pages 100–105. IARIA, IARIA.

Salfner, F., Tr oger, P., and Richly, M. (2012). Dependable

Estimation of Downtime for Virtual Machine Live Mi-

gration. International Journal On Advances in Sys-

tems and Measurements, 5:70–88.

Strunk, A. (2012). Costs of virtual machine live migration:

A survey. In Proceedings of the 2012 IEEE Eighth

World Congress on Services, SERVICES ’12, pages

323–329, Washington, DC, USA. IEEE Computer So-

ciety.

Voorsluys, W., Broberg, J., Venugopal, S., and Buyya, R.

(2009). Cost of virtual machine live migration in

clouds: A performance evaluation. In Proceedings of

the 1st International Conference on Cloud Comput-

ing, CloudCom ’09, pages 254–265, Berlin, Heidel-

berg. Springer-Verlag.

Zhao, M. and Figueiredo, R. J. (2007). Experimental study

of virtual machine migration in support of reserva-

tion of cluster resources. In Proceedings of the 2Nd

International Workshop on Virtualization Technology

in Distributed Computing, VTDC ’07, pages 5:1–5:8,

New York, NY, USA. ACM.

Machine Learning Approach for Live Migration Cost Prediction in VMware Environments

463