Professional Documents
Culture Documents
Soft Computing
A Fusion of Foundations,
Methodologies and Applications
ISSN 1432-7643
Volume 23
Number 21
1 23
Your article is protected by copyright and
all rights are held exclusively by Springer-
Verlag GmbH Germany, part of Springer
Nature. This e-offprint is for personal use only
and shall not be self-archived in electronic
repositories. If you wish to self-archive your
article, please use the accepted manuscript
version for posting on your own website. You
may further deposit the accepted manuscript
version in any repository, provided it is only
made publicly available 12 months after
official publication or later and provided
acknowledgement is given to the original
source of publication and a link is inserted
to the published article on Springer's
website. The link must be accompanied by
the following text: "The final publication is
available at link.springer.com”.
1 23
Author's personal copy
Soft Computing (2019) 23:11277–11296
https://doi.org/10.1007/s00500-018-03686-6 (0123456789().,-volV)(0123456789().
,- volV)
Abstract
Virtual machine placement is the concept of hosting the virtual machines to appropriate physical servers so as to meet user
computation requirements. An optimal placement is one of the key concerns in green cloud computing. Virtual machine
placement in distributed cloud environment also imposes propagation time as a key for effective hosting of VM along with
CPU and memory resource constraints. In this paper, MOEA/D a multi-objective evolutionary algorithm is used to find a
non-dominated solution w.r.t. minimal wastage, minimal power consumption and less propagation delay. The proposed
algorithm has been implemented, tested and compared with the existing multi-objective approaches. The statistical analysis
of the simulation results proves that MOEA/D outperforms against the existing algorithms in distributed cloud VM
placement.
123
Author's personal copy
11278 A. Gopu, N. Venkataraman
VM by a physical machine in distributed cloud environ- 2014) is proposed by Alahmadi et al. in the year 2014
ment in a multi-objective problem framework where which is an advanced version of FFD algorithm based on
already minimizing CPU resources and minimum power bin packing concept. CPU utilization is the resource con-
consumption are concerned using multi-objective evolu- sidered for VM host. From conventional FFD algorithm,
tionary algorithm—decomposition approach (Zhang and Li Enhanced FFD differs in terms of VM reuse. Pros: This
2007). Multi-objective optimization has a vast range of approach is more energy efficient and a significant
interests among the researchers in finding single solution improvement in system throughput. Cons: SLA is violated
w.r.t. different aspects of a problem. Some of the appli- due to VM reuse strategy incorporation. This approach
cations of multi-objective optimization methods include outperforms over greedy algorithm, round robin and FFD
(Ahmadi et al. 2015a, b, c, d, e, f; Sadatsakkak et al. algorithms. Singh et al. (2008) in the year 2008 proposed
2015a, b; Ahmadi and Ahmadi 2015a, b; Ahmadi et al. heaviest-first concept for effective server storage virtual-
2016a, b; Ahmadi et al. 2014a, b, c). The detailed view of ization with minimizing CPU utilization as objective
the problem and algorithm is given in the rest of the paper. function. The predominant feature of this algorithm sets a
This paper is organized as follows. Section 2 holds the upper bound value for cost of relocating VM. Pros: A
literature study of existing algorithms addressing VM substantial minimization on relocation cost has been
placement. Section 3 states an example of VM placement achieved. Cons: More number of bins are used for allo-
in distributed cloud environment and multi-objective cating VM. The performance measures outrage the best-
optimization properties. Section 4 holds the problem for- and first-fit algorithms. Based on genetic algorithm, Tang
mulation and objective modeling. Section 5 holds the et al. (2007) proposed a variant of GA called GABA which
description of MOEA/D. Section 6 includes the experi- incorporates reconfigured searching module and request
mental results and statistical analysis of simulation results. forecasting module in it. Pros: It minimizes the number of
Section 7 concludes the proposed work. physical machines used for VM host and with improved
CPU utilization. Cons: solution space. This algorithm
outperforms in VM placement against TSSP07 approach
2 Related work (Tang et al. 2007).
Min-Cut hierarchical clustering approach (Jiankang
In the year 2011, Mishra and Sahoo (2011) presented a et al. 2015) proposed by Jiankang et al. considers CPU
paper on the existing methods for effective VM placement utilization and network bandwidth as the resources to be
which in addition discussed the demerits of those existing minimized. In this concept, Maximum Link Allocation
methods. Previously in the year 2010, Buyya et al. (2010) (MLU) has been considered as an optimization phases and
presented a paper which discusses the existing challenges VM are allowed to be reused. Pros: In terms of energy
in cloud in energy-efficient perspective. Recently in the consumption, it shows a considerable level of minimization
year 2015, Chowdhury et al. (2015) addressed the recent and the network traffic is highly reduced. Cons: Since the
techniques proposed for efficient VM placement on VM can be migrated, the cost of migration becomes more.
appropriate physical machines. This literature consists of It outperforms against BFD and random-based algorithms.
proposed method for solving VM placement, the resources Based on stochastic packing, Wang et al. (2011) proposed
considered, its pros and cons and the algorithms with which group packing algorithm which considers CPU utilization
the proposed outperforms. and its bandwidth usage as resources for minimizing
Ghribi et al. (2013) proposed Exact Allocation and physical machines. Pros: reduced physical machines for
Migration algorithm for effective VM placement. CPU hosting VM. Cons: Consideration of propagation delay can
utilization is the resource considered in this proposed be achievable. It outperforms against first fit, FFD and
system. Pros: The number of physical machine utilization harmonic algorithm. Chen et al. (2011) proposed VM
is highly reduced along with migration cost. Cons: Com- Sizing algorithm for effective VM placement which is
putational complexity is high. It outperforms against the based on stochastic integer programming. The resources
best-fit heuristic algorithm. Bobroff et al. (2007) in 2007 considered for VM placement are CPU utilization and
proposed Measure–Forecast–Remap (MFP) which consid- memory. The overflow of server is monitored with a
ered reducing number of physical machines as the objec- probability of p. Pros: The number of physical machines
tive. For depicting real-life application, a time interval s is for hosting VM is highly reduced. Cons: With this aspect,
introduced in this concept. Pros: MFP meets the targets of more number of resources can be considered by this
SLA and reduces the usage of physical machines. Cons: algorithm for better results. On comparing FFD algorithm,
The performance on considering more objectives is not VM Sizing performs better with reduced number of phys-
addressed. The result reflects that the proposed is better ical machine usage. Singh et al. (2008) in 2008 proposed
than static algorithm. Enhanced FFD (Alahmadi et al. Vector Dot algorithm based on bin packing concept. A
123
Author's personal copy
Optimal VM placement in distributed cloud environment using MOEA/D 11279
maximum of four resources are considered simultaneously Providing an efficient and distributed cloud with mini-
for effective VM placement which includes CPU utiliza- mal propagation time gives a beneficial way from the point
tion, primary memory, network and input–output band- of service providers where the less data transmission is
width. Pros: The loads on server are maintained possible. Optimizing VM placement in distributed cloud
dynamically by Vector Dot and management of overloaded environment is an NP hard problem since the complexity of
nodes is the significant features of Vector Dot. Cons: the problem increases exponentially as the size of the
Predictive and statistical models are supposed to be fol- problem increases. For example, let us consider the dis-
lowed which increases the delay while hosting and tributed cloud environment in Fig. 1. Assume that the
migrating each VM. Vector Dot outperforms against the server is capable of hosting two VM. The location of VM
heuristic methods such as best fit, first fit, worst fit and and servers is denoted in the form of coordinates (x, y).
relaxed best fit. Wood et al. (2009) proposed Sandpiper Based on Euclidean distance, the distance between each
method which is based on bin packing method. CPU uti- server and VM is calculated and represented in matrix form
lization, memory and network are the resources considered as
in order to host VM. Pros: This method detects the hot spot
42:43 24:69 44:28
while placing VM, mitigates it and balances the load of Distance matrix ¼
9:43 47:41 34:01
physical machines. Cons: It resizes the VM resources for
adapting it to the available physical machine and the mit- Each row represents a server, and columns represent the
igation process overheads. Recently in 2014, Song et al. VM. If the servers are allocated irrespective of propagation
(2014) developed a bin packing-based algorithm VISBP. delay between servers and VM, the data are to be trans-
The resources considered for hosting a VM are its CPU ferred far away. In order to reduce the data traffic and for
utilization, memory and the network. Pros: As like in quick response time, a VM can be hosted in order to
sandpiper concept, the hot spot is identified and mitigated achieve minimal propagation delay. This process can be
in VISBP and the load is then balanced. Cons: SLA vio- achievable only if the servers have no restrictions on data
lation occurs. It is compared with sandpiper and Vector Dot resources. But in practical hosts all servers have maximum
which results that the VISBP outrages two methods. Gao resource utilization threshold limit. Obtaining a well-de-
et al. (2013) proposed VMPACS for efficient VM hosting. served solution to place the VM leads to multi-objective
Well-known ant colony system is used for solving this optimization.
optimization problem. CPU utilization and memory are the
resources considered for optimization in this paper. Pros: 3.2 Multi-objective optimization
Substantial minimization has been performed in terms of
number of physical machines utilized, power consumption In this section, the predefined definition of multi-objective
and wastage. Cons: Computational complexity is high. optimization problem (Coello et al. 2007; Deb 2001) and
VMPACS outrages MGGA, FFD and standard ACO. definition of Pareto dominance operators which leads a
On concerning the stated issues on each paper, we guidance for our proposed work are given here.
proposed MOEA/D a multi-objective optimization algo- Assume that a problem’s solution set Sol ¼
rithm for addressing the issue of VM hosting in addition to ½s1 ; s2 ; . . .; sn with n is the number of solutions. Each
the agenda of placing VM in distributed cloud environ- solution si ¼ di;1 ; di;2 ; . . .; di;j ; . . .; di;m where d is the
ment. The further section deals with placing VM in dis- number of decision variables. The intention of the algo-
tributed cloud environment along with minimal resource ! h i
rithm is to find a vector si ¼ si;1 ; si;2 ; . . .; si;m that satis-
wastage and minimal CPU consumption.
fies u equality constraints ui ð~
sÞ ¼ 0; i ¼ 1; 2; . . .; u, and v
inequality constraints vi ð~
sÞ ¼ 0; i ¼ 1; 2; . . .; v so as to find
! h ! !
3 Background out an optimal solution f s ¼ f1 s ; f2 s ; . . .;
!
3.1 Example of VM placement in distributed fk s where k holds the number of objectives.
cloud environment From the given search space G, Sol ¼ ½s1 ; s2 ; . . .; sn is a
set of solutions. A solution si ¼ di;1 ; di;2 ; . . .; di;m is said
In distributed cloud environment, VM and servers that host to dominate sj ¼ dj;1 ; dj;2 ; . . .; dj;m if and only if
VM are located in different geographical regions. The task fx !
si f x !sj 8x 2 k and fx ! si \fx ! sj 9x 2 k and its
is to allocate appropriate servers to VM so as to satisfy the
representation will be ! si !sj such that i 6¼ j. Likewise,
user needs with reduced propagation time. Let us consider
solution si and solution sj are said to be non-dominated by
a distributed cloud environment as shown in Fig. 1.
123
Author's personal copy
11280 A. Gopu, N. Venkataraman
each other when neither ! si !sj nor ! sj ! sj and the 4 Problem formulation
! !
representation is si 4 sj .
A mere example is given for better notation of domi- As it is discussed earlier in Sect. 2 to place VM to a server
nated and non-dominated solutions in Fig. 2. Let us assume node, a three-dimensional characterization is denoted,
a multi-objective problem that holds f1 and f2 as objectives. namely, CPU, memory and propagation time. If more than
From the solution set Sol, three solutions !
s1 ; !
s2 and !
s3 are one VM is placed in a single server, then the CPU and
mapped in Fig. 2 with respect to their objective functions. memory utilization are taken as aggregation of allocated
From the definition stated above, it can be inferred that VM required resources. The propagation time for placing a
solutions !
s2 and !
s3 are dominated by solution ! s1 and it can VM to a server will be calculated based on the distance
! !
be denoted as s1 s2 ; s3 . Solutions s2 and !
! ! s3 are non- between them.
dominated to each other, and it can be denoted as ! s2 4s!3.
4.1 Resource wastage
123
Author's personal copy
Optimal VM placement in distributed cloud environment using MOEA/D 11281
turned off when they enter into idle state. Hence, the 4.4 Objective formulation
mathematical representation of power consumption of
server i is defined as In this section, formulation of objectives for efficient VM
active placement in distributed cloud environment has been
Pi Pidle
i UiCPU þ Pidle
i ; Ui [ 0
Poweri ¼ described. Given a set of physical machines (servers) i 2 I
0 otherwise
in which the VM j 2 J to be hosted. In our simulation
ð2Þ
environment, no VM are prepared so as to require two
where Pactive
i and Pidle
i denote the power consumption of ith servers to process the request.
server. The active and idle power of CPU is fixed as 215 1. Minimize CPU wastage
and 162 watts, respectively.
X
M X
M
Min Wi ¼
4.3 Propagation time i¼1 i¼1
2 PN PN 3
hPi j¼1 xi;j: Rp;j hMi j¼1 xi;j: RM;j þ e
In distributed cloud environment, propagation time of a 4yi P PN 5
N
x i;j: R p;j þ xi;j: R M;j
VM is one of the key factors that have to be considered for j¼1 j¼1
123
Author's personal copy
11282 A. Gopu, N. Venkataraman
123
Author's personal copy
Optimal VM placement in distributed cloud environment using MOEA/D 11283
123
Author's personal copy
11284 A. Gopu, N. Venkataraman
123
Author's personal copy
Optimal VM placement in distributed cloud environment using MOEA/D 11285
123
Author's personal copy
11286 A. Gopu, N. Venkataraman
Spacing (Sp) is a unary performance indicator which For performing analysis of results which are tabulated,
denotes the diversity between end population solutions. ANOVA has been used in order to perform validation.
Values near to 0 in Sp state that the algorithm searches the ANOVA test results whether there exists a significant
search space in diversified manner. Sp can be mathemati- difference between the group of samples given. One-way
cally represented as ANOVA test is performed in this paper to determine
vffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi whether the algorithm group results have a significant
u
u 1 XjPFj
effect over other. In this paper, ANOVA test has been
Sp ¼ t
2
ðd di Þ ð10Þ
jPFj 1 i¼1 performed with a significant level of 95%. If Sig. value is
lesser than a (critical) value (0.05), then H0 (null hypoth-
where d is defined as the mean value of d. esis) should be rejected and H1 (alternate hypothesis)
should be accepted. This shows that the given group of
6.3 Performance analysis values has a significant difference between each other, and
hence, post hoc tests can be performed. Duncan’s Multiple
For analyzing the performance of proposed system over Range Test has been used as post hoc test analyzer. In case
existing algorithms, statistical tools such as ANOVA if Sig. value is higher than 0.05, null hypothesis (H0)
(analysis of variance) and Duncan’s Multiple Range Test should be accepted and post hoc tests cannot be performed.
(DMRT) are used. A description on these two tools is given
below. These tests are performed for the results of power
consumption, wastage and propagation delay which are
tabulated in Tables 3, 4, 5 and 6.
123
Author's personal copy
Optimal VM placement in distributed cloud environment using MOEA/D 11287
6.3.2 Duncan’s Multiple Range Test Table 8(a) and (b) shows the statistical analysis of
ANOVA and DMRT on power consumption of 100 VM
Duncan’s Multiple Range Test is a multiple comparison instances for RCPU = RRAM = 45%, respectively. From
procedure which uses studentized range statistics for Table 8(a) Sig. value, it can be inferred that there exists
comparing the mean values of sets. It computes numerical significant difference between the sets denoted. Hence, null
boundaries for classification of differences between any hypothesis H0 has been rejected and alternate hypothesis
two or more sets which are either significant or non- H1 is accepted. Since H0 is rejected, DMRT is performed
significant. DMRT ranks the sets in increasing or for Table 8(a) instance and the result shows that MOEA/D
decreasing order based on user preference. ranks first among the compared algorithms.
Tables 9 and 10 reflect the statistical analysis on power
6.3.3 Performance analysis w.r.t. power consumption consumption for 200 VM instances of RCPU = RRAM =
25% and RCPU = RRAM = 45%, respectively.
Tables 7, 8, 9 and 10 show the statistical analysis of As per the results of ANOVA analysis in Table 9(a)
algorithms simulated for testing. Table 7 holds statistical w.r.t. power consumption, the Sig. value is less than 0.05
analysis of ANOVA on power consumption of 100 VM which rejected H0 and accepted alternate hypothesis H1 and
instances for RCPU = RRAM = 25%. further shows that there exists significant difference among
Table 7 states that the Sig. value is 0.068 which is the mean values of compared algorithms. Since H0 is
higher than critical value 0.05. As per the statistical anal- rejected based on the results in Table 9(a), DMRT has been
ysis property, there is no significant difference between the performed on power consumption of 200 VM instances for
groups, and hence, H0 should be accepted and H1 has to be RCPU = RRAM = 25%. Table 9(b) presents the results of
rejected. Since H0 is accepted, post hoc tests cannot be DMRT for specified set of results. The post hoc analysis
performed. shows that two homogenous groups can be formed w.r.t.
123
Author's personal copy
11288 A. Gopu, N. Venkataraman
mean values of compared algorithms. MOEA/D forms the Hence, null hypothesis H0 is rejected and alternate
best performing subset among the given algorithms with hypothesis H1 is accepted.
respect to power consumption. Table 11(a) results indicate that H0 has been rejected,
Table 10 shows the ANOVA result analysis on 200 VM and thus, post hoc analysis is possible. Table 11(b) depicts
instances power consumption for RCPU = RRAM = 45%. DMRT results for intended results in Table 11(a). DMRT
Sig. value in Table 10 is greater than critical value (0.05); in Table 11(b) shows that four homogenous groups are
hence, null hypothesis H0 is accepted which results that formed w.r.t. CPU wastage in terms of its mean values.
there is no significant difference among the algorithms Table 12(a) shows the ANOVA results on CPU wastage
stated. Hence, DMRT analysis cannot be performed for 200 resources of 100 VM instances for RCPU = RRAM = 45%.
VM power consumption of CPU and memory resources Sig. value in Table 12(a) shows that there exists significant
which equal 45%. difference in mean values w.r.t. CPU wastage. This rejects
null hypothesis H0 and accepts alternate hypothesis H1.
6.3.4 Performance analysis w.r.t. CPU and memory Hence, post hoc analysis is performed and results are tab-
wastage ulated in Table 12(b). The DMRT results state MOEA/D
outperforms when compared to other algorithms. Three
Tables 11, 12, 13 and 14 hold the statistical analysis on homogenous groups are formed w.r.t. CPU wastage in
CPU wastage of 100 VM instances and 200 VM instances. terms of its mean values.
Table 11(a) shows ANOVA statistical results on wastage Table 13(a) depicts the results of ANOVA on CPU
of 100 VM instances for RCPU = RRAM = 25%. The Sig. wastage of 200 VM instances for RCPU = RRAM = 25%.
value states that there exists significant difference in terms Since Table 13(a) Sig. value is less than the critical values
of mean values between algorithms on CPU wastage. of ANOVA (0.05), it shows there is significant difference
in mean values from obtained results. This rejects null
123
Author's personal copy
Optimal VM placement in distributed cloud environment using MOEA/D 11289
hypothesis H0 and accepts alternate hypothesis H1. Hence, 6.3.5 Performance analysis w.r.t. CPU and propagation
post hoc analysis is performed and results are tabulated in delay
Table 13(b). The DMRT results state MOEA/D and ACS
algorithms outperform when compared to other algorithms. Tables 15, 16, 17 and 18 hold the statistical analysis of the
Four homogenous groups are formed w.r.t. CPU wastage in given dataset w.r.t. propagation delay. Table 15(a) holds
terms of its mean values. the ANOVA results on propagation delay of 100 VM
Table 14 shows ANOVA results on CPU wastage of 200 instances for RCPU = RRAM = 25%.
VM instances for RCPU = RRAM = 45%. The Sig. value Sig. value in Table 15(a) is less than the critical value of
(0.458) is higher than the critical value (0.05). Hence, null ANOVA (0.05). It states that there exists a significant
hypothesis H0 is accepted which shows that no significant difference in the mean value of propagation delay between
difference lies in the CPU wastage results in Table 6. each of the algorithms. Hence, the null hypothesis H0 is
DMRT analysis cannot be performed for this result since rejected which made DMRT possible to find out the
H0 is rejected. homogenous groups between algorithms.
123
Author's personal copy
11290 A. Gopu, N. Venkataraman
Table 15(b) shows DMRT results for Table 15(a) dataset. total of four homogenous groups are formed in terms of its
From Table 15(b), it can be inferred MOEA/D outperforms mean values.
and ranks high when compared to other algorithms and a Table 16 holds the statistical analysis on propagation
delay of 100 VM instances for RCPU = RRAM = 45%.
123
Author's personal copy
Optimal VM placement in distributed cloud environment using MOEA/D 11291
Table 16(a) shows the ANOVA results of the mentioned Table 17(a) consists of ANOVA results on propagation
dataset, and from the results (i.e., Sig. value \ critical delay of 200 VM instances for RCPU = RRAM = 25%. Sig.
value) it can be inferred that significant different lies value in Table 17(a) results less than critical value which
between algorithms w.r.t. mean values of propagation states there is significant difference between the mean
delay. Hence, the null hypothesis H0 has been rejected values of algorithms. As per the property of ANOVA since
followed by accepting alternate hypothesis H1. the Sig. value is less than critical value, the null hypothesis
Results in Table 16(a) Sig. value reject H0, and thus, H0 has been rejected.
DMRT has been analyzed for the given results and is Since Table 17(a) Sig. value is less than the critical
tabulated in Table 16(b). The results state that MOEA/D value, DMRT has been analyzed for finding the perfor-
outperforms when compared to other algorithms followed mance of each algorithm individually. Table 17(b) shows
by ACS and VSIBP which lie under same subset and a total the DMRT results of VM instance of prescribed dataset.
of three homogenous groups are formed in this table. The results state that the proposed algorithm MOEA/D
123
Author's personal copy
11292 A. Gopu, N. Venkataraman
outperforms against the compared algorithms. Four Table 18(a) shows the statistical ANOVA results of poin-
homogenous groups are formed on performing DMRT. ted dataset. Since Sig. value in Table 18(a) is less than
Table 18 depicts statistical analyses on propagation ANOVA critical value, it can be stated that there exists a
delay of 200 VM instances for RCPU = RRAM = 45%. significant difference in between the mean values of
123
Author's personal copy
Optimal VM placement in distributed cloud environment using MOEA/D 11293
123
Author's personal copy
11294 A. Gopu, N. Venkataraman
similarity only with ACS. This shows that MOEA/D shows Results from Tables 15, 16, 17 and 18 clearly illustrate
a significant difference when it is compared to existing the positive impact of MOEA/D on propagation delay in
algorithms Min-Cut, VM Sizing and VISBP. For large VM placement. For all the instances of VM with different
instances of VM on CPU and RAM utilization of 25% levels of utilization (Tables 15(b), 16(b), 17(b) and
(Table 9(b)), MOEA/D forms a standalone homogenous Table 18(b)), MOEA/D outperforms all the existing algo-
group which implicates that it has a significant difference rithms by forming a standalone homogenous group with
against all the compared algorithms. However, for 100 VM ranking first with less propagation delay. In this paper, the
instances with CPU, RAM utilization of 25% (Table 7), overall procedure is to include the new objective called
and for 200 VM instances with CPU, RAM utilization of propagation delay on VM placement and this MOEA/D has
45% (Table 10), MOEA/D fails to show the significant been especially tuned to reduce the overall propagation
difference, and hence, the algorithm needs to be impro- delay, and hence, it shows a significant improvement in all
vised by tuning the parameters which support for finding the instances. Thus, the network propagation delay has
the solution to minimize overall power consumption for been efficiently handled along with the existing objectives
different levels of VM instances on different utilization in the literature using MOEA/D.
schema.
Results from Tables 11, 12, 13 and 14 denote the pos-
itive impact of MOEA/D on CPU and RAM wastage for 7 Conclusion
different VM instances with different utilization levels. For
small-scale VM instances with 100 VM on both the uti- Hosting VM in distributed cloud environment is one among
lization levels of CPU, RAM with 25% (Table 11(b)) and the predominant issue in deploying cloud with green
45% (Table 12(b)), the proposed MOEA/D outperforms all computing. In this paper, a new paradigm of distributed
existing algorithms by forming standalone homogenous cloud environment, namely, propagation delay, has been
group when compared with ACS, VISBP, VM Sizing and addressed as a constraint along with CPU and memory
Min-Cut. And for large-scale VM instances on CPU, RAM utilizations. The aim of this proposal is to find a non-
utilization with 25% (Table 13(b)), MOEA/D forms a dominated solution which simultaneously addresses all
homogenous group by sharing it with ACS which shows three stated objectives. Proposed MOEA/D has been tested
that there exists significant difference when MOEA/D with the instances stated in the literature along with an
results are compared with VISBP, VM Sizing and Min-Cut. added resource propagation delay. The results are com-
However, for large VM instances with 45% of RAM, CPU pared with the existing algorithms presented in the litera-
utilization MOEA/D fails to show a significant difference ture. For analyzing the performance of algorithms,
among the existing algorithms. ANOVA and Duncan’s Multiple Range Test statistical
tools were used. The results show that the proposed
123
Author's personal copy
Optimal VM placement in distributed cloud environment using MOEA/D 11295
MOEA/D outperforms over existing algorithms on VM Ahmadi MH, Ahmadi MA, Feidt M (2016b) Performance optimiza-
hosting. tion of a solar-driven multi-step irreversible Brayton cycle based
on a multi-objective genetic algorithm. Oil Gas Sci Technol
1:1–10. https://doi.org/10.2516/ogst/2014028
Alahmadi A, Alnowiser A, Zhu MM, Che D, Ghodous P (2014)
Enhanced first-fit decreasing algorithm for energy-aware job
Compliance with ethical standards scheduling in cloud. In: 2014 international conference on
computational science and computational intelligence (CSCI),
Conflict of interest The authors certify that they have no affiliations vol 2. IEEE, pp 69–74
with or involvement in any organization or entity with any financial Beloglazov A, Buyya R (2010) Energy efficient resource management
interest or non-financial interest in the subject matter or materials in virtualized cloud data centers. In: Proceedings of the 2010
discussed in this manuscript. 10th IEEE/ACM international conference on cluster, cloud and
grid computing. IEEE Computer Society, Washington
Beloglazov A, Buyya R (2013) Managing overloaded hosts for
dynamic consolidation of virtual machines in cloud data centers
References under quality of service constraints. IEEE Trans Parallel Distrib
Syst 24(7):1366–1379
Ahmadi MH, Ahmadi MA (2015a) Thermodynamic analysis and Bobroff N, Kochut A, Beaty K (2007) Dynamic placement of virtual
optimization of an irreversible radiative type heat engine by machines for managing sla violations. In: 10th IFIP/IEEE
using non-dominated sorting genetic algorithm. Int J Ambient international symposium on integrated network management,
Energy 37:403–408 2007. IM’07. IEEE
Ahmadi MH, Ahmadi MA (2015b) Thermodynamic analysis and Buyya R, Beloglazov A, Abawajy J (2010) Energy-efficient manage-
optimization of an irreversible Ericsson cryogenic refrigerator ment of data center resources for cloud computing: a vision,
cycle. Energy Convers Manag 89C:147–155 architectural elements, and open challenges. arXiv preprint
Ahmadi MH, Ahmadi MA, Mehrpooya M, Hosseinzade H, Feidt M arXiv:1006.0308
(2014a) Thermodynamic and thermo-economic analysis and Chen M, Zhang H, Su YY, Wang X, Jiang G, Yoshihira K (2011)
optimization of performance of irreversible four-temperature- Effective VM sizing in virtualized data centers. In: 2011 IFIP/
level absorption refrigeration. Energy Converg Manag IEEE international symposium on integrated network manage-
88C:1051–1059 ment (IM). IEEE, pp 594–601
Ahmadi MH, Ahmadi MA, Mohammadi AH, Feidt M, Pourkiaei SM Chowdhury MR, Mahmud MR, Rahman RM (2015) Study and
(2014b) Multi-objective optimization of an irreversible Stirling performance analysis of various VM placement strategies. In:
cryogenic refrigerator cycle. Energy Convers Manag 82:351–360 2015 16th IEEE/ACIS international conference on software
Ahmadi MH, Ahmadi MA, Mohammadi AH, Mehrpooya M, Feidt M engineering, artificial intelligence, networking and parallel/
(2014c) Thermodynamic optimization of Stirling heat pump distributed computing (SNPD). IEEE
based on multiple. Energy Convers Manag 80:319–328 Coello CAC, Lamont GB, Van Veldhuizen DA (2007) Evolutionary
Ahmadi MH, Ahmadi MA, Sadatsakkak SA (2015a) Thermodynamic algorithms for solving multi-objective problems, vol 5. Springer,
analysis and performance optimization of irreversible Carnot New York
refrigerator by using multi objective evolutionary algorithms Deb K (2001) Multi-objective optimization using evolutionary
(MOEAs). Renew Sustain Energy Rev. https://doi.org/10.1016/j. algorithms. Wiley, New York
rser.2015.07.006 Dupont C, Schulze T, Giuliani G, Somov A, Hermenier F (2012) An
Ahmadi MH, Ahmadi MA, Shafaei A, Ashouri M, Toghyani S energy aware framework for virtual machine placement in cloud
(2015b) Thermodynamic analysis and optimization of the federated data centres. In: 2012 third international conference on
Atkinson engine by using NSGA-II. Int J Low Carbon Technol future energy systems: where energy, computing and commu-
11:317–324 nication meet (e-energy). IEEE, pp 1–10
Ahmadi MH, Ahmadi MA, Bayat R, Ashouri M, Feidt M (2015c) Fan X, Weber W-D, Barroso LA (2007) Power provisioning for a
Thermo-economic optimization of Stirling heat pump by using warehouse-sized computer. In: ACM SIGARCH computer
non-dominated sorting genetic algorithm. Energy Convers architecture news, vol 35, no 2. ACM, New York
Manag 91:315–322 Gao Y, Guan H, Qi Z, Hou Y, Liu L (2013) A multi-objective ant
Ahmadi MH, Ahmadi MA, Mehrpooya M, Sameti M (2015d) colony system algorithm for virtual machine placement in cloud
Thermo-ecological analysis and optimization performance of computing. J Comput Syst Sci 79(8):1230–1242
an irreversible three-heat-source absorption heat pump. Energy Ghribi C, Hadji M, Zeghlache D (2013) Energy efficient vm
Convers Manag. https://doi.org/10.1016/j.enconman.2014.11. scheduling for cloud data centers: exact allocation and migration
021 algorithms. In: 2013 13th IEEE/ACM international symposium
Ahmadi MH, Ahmadi MA, Feidt M (2015e) Thermodynamic analysis on cluster, cloud and grid computing (CCGrid). IEEE
and evolutionary algorithm based on multi-objective optimiza- Goldberg DE, Lingle R (1985) Alleles, loci, and the traveling
tion of performance of irreversible four-temperature-level salesman problem. In: Proceedings of an international confer-
absorption refrigeration. Mech Ind 16:207 ence on genetic algorithms and their applications, vol 154.
Ahmadi MH, Ahmadi MA, Mehrpooya M, Pourkiaei SM, Khalili M Lawrence Erlbaum, Hillsdale
(2015f) Thermodynamic analysis and evolutionary algorithm Jiankang D, Hongbo W, Shiduan C (2015) Energy-performance
based on multi-objective optimization of Rankine cycle heat tradeoffs in IaaS cloud with virtual machine scheduling. China
engine. Int J Ambient Energy 37:363–371 Commun 12(2):155–166
Ahmadi MH, Ahmadi MA, Mehrpooya M (2016a) Investigation of Miettinen K (2012) Nonlinear multiobjective optimization, vol 12.
design parameters effect on power output and thermal efficiency Springer, Berlin
of the Stirling engine thermodynamic analysis. Int J Low Carbon Mishra M, Sahoo A (2011) On theory of vm placement: anomalies in
Technol. https://doi.org/10.1093/ijlct/ctu030 existing methodologies and their mitigation using a novel vector
123
Author's personal copy
11296 A. Gopu, N. Venkataraman
based approach. In: 2011 IEEE international conference on cloud Proceedings of the 16th international conference on World Wide
computing (CLOUD). IEEE Web. ACM, New York, pp 331–340
Sadatsakkak SA, Ahmadi MH, Ahmadi MA (2015a) Optimization Van Veldhuizen DA (1999) Multiobjective evolutionary algorithms:
performance and thermodynamic analysis of an irreversible nano classifications, analyses, and new innovations. No. AFIT/DS/
scale Brayton cycle operating with Maxwell–Boltzmann gas. ENG/99-01. Air Force Inst of Tech Wright–Patterson AFB OH
Energy Convers Manag 101:592–605 School of Engineering
Sadatsakkak SA, Ahmadi MH, Ahmadi MA (2015b) Thermodynamic Wang M, Meng X, Zhang L (2011) Consolidating virtual machines
and thermo-economic analysis and optimization of an irre- with dynamic bandwidth demand in data centers. In: INFOCOM,
versible regenerative closed Brayton cycle. Energy Convers 2011 proceedings IEEE. IEEE
Manag 94:124–129 Wood T, Shenoy P, Venkataramani A, Yousif M (2009) Sandpiper:
Schott JR (1995) Fault tolerant design using single and multicriteria black-box and gray-box resource management for virtual
genetic algorithm optimization. No. AFIT/CI/CIA-95-039. Air machines. Comput Netw 53(17):2923–2938
Force Inst of Tech Wright–Patterson AFB OH Zhang Q, Li H (2007) MOEA/D: a multiobjective evolutionary
Singh A, Korupolu M, Mohapatra D (2008) Server-storage virtual- algorithm based on decomposition. IEEE Trans Evol Comput
ization: integration and load balancing in data centers. In: 11(6):712–731
Proceedings of the 2008 ACM/IEEE conference on supercom- Zhang L, Zhuang Y, Zhu W (2013) Constraint programming based
puting. IEEE Press virtual cloud resources allocation model. Int J Hybrid Inf
Song W, Xiao Z, Chen Q, Luo H (2014) Adaptive resource Technol 6(6):333–344
provisioning for the cloud using online bin packing. IEEE Trans
Comput 63(11):2647–2660
Speitkamp B, Bichler M (2010) A mathematical programming Publisher’s Note Springer Nature remains neutral with regard to
approach for server consolidation problems in virtualized data jurisdictional claims in published maps and institutional affiliations.
centers. IEEE Trans Serv Comput 3(4):266–278
Tang C, Steinder M, Spreitzer M, Pacifici G (2007) A scalable
application placement controller for enterprise data centers. In:
123