Allocation and Consolidation of Virtual Machines in Cloud Computing Systems

Introduction

The concept of cloud computing and the related environments have seen the introduction of new
computing models through the shift of the location of computational infrastructure to the
internet. This has also seen the reduction of the costs related with the management of software
and hardware resources. It is critical to note that the cloud model utilizes virtualization
technology to efficiently consolidate virtual machines into physical machines. This has the effect
of enhancing the utilizations of physical machines remotely. However, research has revealed that
the average usage of physical machines in numerous cloud data centers has not yet reached
optimal levels than it would have been accepted. There is a need for the introduction of new
approaches which would improve the level of usage of physical machines through the use of the
cloud computing technologies and their centers. This calls for the introduction of a new approach
that would consolidate virtual machines dynamically for there to be an optimal utilization of
physical machines. For this to work there would need to be dynamic programming algorithm that
would be responsible for the selection of the best virtual machines that would cater for the
migration from an overloaded physical machine. This would also call for the consideration of the
migration overhead of a virtual machine (Asyabi and Shariff 1).

Related Works

Surname 2
Since the introduction of cloud computing systems, data centers in private, public, and
hybrid cloud setting has made it possible to offer virtual machines with unparalleled flexibility.
However, it is important to note that buying, using, and maintaining these virtual machines and
the associated physical resources calls for high costs in terms of capital and environmental
resources. This means that cloud providers must offer an optimized usage of the physical
resources through the watchful allocation of virtual machines to the hosts. Cloud services
providers should also continuously balance between the operational costs and the conflicting
needs on performances of the virtual machines (Ts’epoMofolo 42). In the recent percent, there
have been the introductions of many algorithms that through their proposals suggest that the
primary goal is to optimize utilization of the physical machines. However, most of the proposed
approaches have been much diversified through having subtle differences in the applied problem
models. This indicates that there should be a unified problem formulation and the algorithms
applied for algorithms. In this manner, the proposals should highlight their limitations and
strengths and should also point out areas that need further research in the future (GWDG 3).
It is critical to point out that users for cloud computing services have identified this
technology very attractive because of a number of reasons. One of the reasons is the fact that the
users have viewed this technology from the perspective of usage-based schemes for payment, the
ability to avoid up-front investments, and the unlimited scalability. Cloud computing has been
noted for been very friendly on the basis that there are globally available public cloud computing
services. Besides this easy cloud computing approach to the global community, corporate
institutions and other institutions have also been known to take advantage of similar solutions
that take the form of hybrid cloud computing and private cloud computing (Mann 1).

Surname 3

Virtualized large data centers having been given the responsibility of serving a
continuously growing demand for networking, storage, and general computation through cloud
computing. However, it is important to note that efficiency of the operational of the data centers
has become very critical and at the same has developed into a complex issue. This issue revolves
around the traditional cost factors such as staffing and equipment. It is also important to note that
the complexity of the issue has been accelerated by other factors such as energy consumption.
Because of the two issues and the underlying issue of the environmental impacts of the data
centers, there have been many proposals brought forward on how best to manage the data
centers. Statistics indicate that the consumption of energy by the data centers is the fastest
growing part of energy consumption in the ICT ecosystem. It is also important to put into mind
that the up-front purchasing of data center equipment has been outweighed by the cost of its
ongoing electricity consumption (Lee 5).
Data centers have traditionally applied virtual machines in an effort to enhance isolation
of applications and enabling an optimal utilization of the physical resources. Virtual machines
are offered to the customers directly in the case of Infrastructure-as-a-Service (IaaS) provider.
When the provider is a Software-as-a-Service (SaaS) or Platform-as-a-Service (PaaS), virtual
machines are applied to wrap the provisioned applications (Beloglazov and Rajkumar 1).
Since power consumption is central to the issue, data centers have been trying to come up
with viable means of saving power. One of the most attractive means has been identified as the
consolidation of virtual machines. In this way, data centers consolidate the virtual machine to a
minimal number of physical hosts and the hosts that are not used should be switched to a mode
that does not require a lot of energy for example the sleep mode. Although consolidation of
virtual machines is attractive, there is the problem associated with the fact that a virtual machine

Surname 4
consolidation that is too aggressive may result to overloading the hosts. This would have
negative effects on the delivered quality of service and this has the potentiality of violating the
service level agreements with the customers. This implies that virtual machine allocation must
find an optimal balance between quality of service and energy consumption (Lopez-Pires and
Buyya 5).

System Design

The primary idea in the system design for allocating and consolidating virtual machines
in a cloud computing system is the consideration of the degree in which different virtual
machines are affected. In this respect, it is important to determine the virtual machines that
degrade the lease when they are incorporated together in a system. Then there is the
consolidation of these virtual machines to the degree that performance constraints are not
negatively affected (Minarolli 67).
In this regard, there are two assumptions that are made about the system. The first
assumption is associated wit the virtual machine processor core mapping where the assumption
will be that every virtual is assigned one core. The rationale behind this assumption is that the
design is primarily concerned with resource interference as a result of a common cache hierarchy
in a multi-processor chip. The other assumption is the performance degradation data where the
performance is assumed to degrade in every virtual machine. This is when the virtual machines
are consolidated into a system (Lopez-Pires and Baran 6).
With this overview and the assumptions put into consideration, the design has three
components. One of the components is conservatively packed servers where customers submit
virtual machines by using the right cloud APIs. The other component is the virtual profiling

Surname 5
engine. The last component is the consolidation algorithm (Viswanathan 4). The complete design
is shown in the figure below.

Virtue Machine Consolidation Algorithms

Host Under-Load Detection
The polynomial time algorithm for host under-load uses the following precedent. The
first step is constructing weight but undirected graph on 2n nodes. n in this instance correspond
to every n services and at the same time another n is used to serve as ‘dummy’ nodes. For every
pair of service nodes S = {i,j}, if the performance of the services is acceptable when placed
together (d s t < D and d s j < D). After this formulations, the algorithm is given an edje (i,j) with a
weight that is given as w({i,j}). It is critical to note that w(S) represents the cost to run this set
services performed by a virtual machine. For every service j, the system associates a unique
dummy node j’ with j and there is the adding of an edge (j,j’) which is accompanied by the
weight w({j}) which represents the cost of every service j on a particular virtual machine. In the
end, the graph is designed among the dummy nodes, with the aim of ensuring that every pair of
dummy nodes i’ and j’ is added an edge (i’,j’) with a weight of 0 which represents the cost on

C
U
S
T
O
M
E
R
V
M
C
L
O
U
D
D
E
P
L
O
M
E
N
T
A
P
1

VM Profiling
Engine

Hosting
Racks
Consolidation
Algorithm

Conservatively
Packed Servers

Surname 6
non-powered or powered down virtual machines (Hwang and Pedram 5). The illustration below
is a representation of such a graph:

The next step is observing and identification of minimum-weighted perfect pairing in the
graph precisely corresponds to the optimal solution that is associated with solving the problem of
the minimizing the costs of resources. The minimum weight matched takes care of all the nodes
and contains the minimum cost of edge cost, which represents the precise cost per virtual
machine consolidation (Nema et al 4621).

Host Overload Detection
The algorithm of the host overload detection is going to be informed by near-optimal
computation. This implies that the algorithm will find a placement for virtual machines resulting
with a usage that is optimal. Through the use of the profiling method, it becomes easy to filter
out the virtual machine sets that tend to violate the constraint of degradation. When we identify
the remaining ones, the system will need to be allocated the ones that utilize the least resources
in terms of cost. In this regard, we take the assumption that the allowed virtual machine sets is
denoted by Ƒ.
The next step is that the algorithm assigns a value V(S) = w(S) / |S| for all the allowed
sets S € Ƒ. This metric typifies the cost of a set S of virtue machines. Sets with more virtue
machines |S| and low resource use (w(S) yield low V(S). The algorithm sorts the virtual machine

Surname 7
sets in an ascending order that is defined by V(S). The virtual machines that appear earlier in the
order are identified as the ones that incur lower costs and this means that they are the favored
sets (Nguyen et al 3).
The final step for the algorithm is making a single pass through the list and come up with
set that is designated as S as the virtual machines that should be placed in the consolidation
output on the condition that it is a disjoint from the other sets that were chosen in the previous
step. The algorithm ends after it has come to a single pass from the list of sets. It is important to
note that the algorithm can stop earlier if all the virtual machines are incorporated in the sets that
have been chosen. The first set in the list that has been sorted will always be incorporated into
the solution because there are no sets that have been previously chosen making it a disjoint (Lee
5).

Virtual Machine Selection
Following the algorithm, the virtual machines that should be selected are the ones that are
efficient in terms of delivering quality to the customers. This calls for the efficient utilization of
the resources of the existing infrastructure. Following the algorithm, the virtual machines that
should be allocated and consolidated should be the ones that use minimal energy and the
resources available and at the same time ensure that customer requests are delivered in a quality
manner (Mann 13).

Virtual Machine Placement
The algorithms also indicate that virtual machine replacement should be dynamic with
considerations to the stochastic and deterministic demands. To ensure a quick response to the

Surname 8
virtual machine requests and enhance the energy efficiency. This means that a two phase
optimization model. In this model, the virtual machines are organized into a runtime and
consolidated into the servers sporadically. There is also an algorithm that balances the utilization
of resources and at the same time enhancing a live migration procedure that has been developed
(GWDG 3).

Energy Efficiency

It is imperative to note that the primary objective of consolidation is minimizing the
resources used and at the same time to preserve the performance. This means that the host under-
load detection will have an algorithm that has an upper bound on the acceptable degradation
primarily because of interference that will be subjected to each virtual machine. In this case, the
assumption is that the performance constraint will be equal for all virtual machines. However,
there could be considerations for of various qualities of services class which have been found to
have individual degradation limits although basically they do not alter the problem. The
consolidation algorithm could be informed by the following information:
P-Mode Objective: the minimization of energy consumption in relation to the virtual machine
performance constraint
Given:

1. n VMs
2. Servers have k cores
3. Virtual machine degradation for the entire combinations of virtual machines placements
   with up to k collocated virtual machines

Surname 9

4. A cost w(S) for a set of virtual machines that are placed on a server such as the cost of
   energy
5. There is a degradation constraint D > 1 that represents the maximum degradation that is
   tolerable for any virtual machine.
   Find a consolidation of the n VMs through the use of some number b of servers in order to
   minimize

Ʃ b i – 1 w(S i )

In this equation S i represents the set of virtual machines that are consolidated on the i th server.
The consolidation algorithm is also informed by the cost metric with the assumption that
the resource cost w(S) is the most important element in the system. In this way, we assume that if
we have to minimize the number of servers, then the cost metric will be w(S) = 1 for any set
irrespective of how many virtual machines a set contains. The total cost of the resources would
then ensure that the number of servers applied in the system is simplified (Belaglazov and Buyya
1398).

Benchmark for Evaluating Virtual Machine Consolidation Algorithms

Workload Traces
For the degradation data, there is the use of the SPEC CPU 2006 benchmark application.
These degradations have been set in equal ranges for Google’s data center workloads and this
implies that they are to some significant degree representative of workloads in the real world.
The table below shows the application virtual machines and their degradations.
Application VMs Degradations (%)
soplex, Ibm 2,19,7

Surname 10

soplex, soplex 10,10
Ibm, soplex, sjeng 2,4,4.1
Ibm, povray, Ibm 19.6,5.32,19.6
Ibm, soplex, soplex, sjeng 14.56,36.9,36.9,5.83
Ibm, Ibm, Ibm, Ibm 104.6 (each)

Performance Metrics
For the performance metrics, there is the specification of a degradation constraint and the
cost of resource is optimized. This implies that the evaluation metric of interest is the cost of
resources. For every server, there is a fixed and dynamic energy component resulting in the
energy cost w(s) = c j + Ʃ j€ d s j . In this instance, the extra cost for every virtual machine is
represented by the model d s j . However, it is also important to note that modern server
technologies are built with more energy efficiency. This implies that lower idle power cost will
have an exaggeration on the fraction of the overhead as a result of the interference and this
implies that the system will be cost effective (Belaglazov and Buyaa 1).

Performance Evaluation

Experiment Setup
The experiment will be facilitated through the computational of the optimal using 16
virtual machines. These virtual machines will be equally applied in the four benchmarks that had
been identified earlier in the paper. The variance of the degradation will be done from 10% (D =
1.1) to as high as 50% 1 .

Surname 11

Experimental Results and Analysis
The figure below displays the energy overhead of the consolidation that is determined by
the naïve method and by the new system’s virtual machines. The new system’s virtual machines
is within a range of 10% of the optimal and this indicates that it is better than the naïve system
that is currently used.

The figure below displays the sever utilizations that was achieved by the model. The
proposed system’s virtual machine has an achievement of over 80% in utilization in most of the
cases yielding good resource use. As expected, when small levels of degradation are allowed, the
servers are underutilized in order to avert interference, and even the optimal method is not able
to utilize all cores.

Scalability Remarks and Future Directions

Surname 12
A data center’s total cost of ownership incorporates both capital expenses which are paid
upfront and operating expenses including energy bills that are paid on a utilization basis.
Consolidation of virtual machines affects a number of elements of the total cost of ownership.
With an allocation and consolidation of virtual machines, there are a number of effects to the
total cost ownership (Ngenzi and Nair 4). For instance, the fixed costs are amortized of an entire
data center over a period of 15 months which are divided on a monthly basis.

Conclusions

This report has put into consideration the problem of virtue machines and has identified
the primary mechanisms to enhance efficiency for cloud infrastructures. The primary concern
was efficiency. It has been found that with the right allocation and consolidation of virtue
machines, a data center improves its efficiency in energy consumption and at the same time
ensures that the utilization of the physical machines is enhanced.

Surname 13

Works Cited

Asyabi, Esmail, and Mohsen Sharifi. “A New Approach for Dynamic Virtual Machine
Consolidation in Cloud Data Centers.” (2015), pp 1 – 6.
Beloglazov, Anton, and Rajkumar Buyya. “Energy efficient allocation of virtual machines in
cloud data centers.” Cluster, Cloud and Grid Computing (CCGrid), 2010 10th
IEEE/ACM International Conference on. IEEE, 2010, pp. 1 – 2.
Beloglazov, Anton, and Rajkumar Buyya. “Optimal online deterministic algorithms and adaptive
heuristics for energy and performance efficient dynamic consolidation of virtual
machines in cloud data centers.” Concurrency and Computation: Practice and
Experience 24.13 (2012): 1397-1420.
GWDG. Virtual Machine Allocation in Current Cloud Computing Middleware. GWDG eScience
Group, 2015, pp 1 – 4.
Hwang, Inkwon, and Massoud Pedram. “Hierarchical virtual machine consolidation in a cloud
computing system.” Cloud Computing (CLOUD), 2013 IEEE Sixth International
Conference on. IEEE, 2013, pp. 1 – 8.
Lee, S., et al. Validating heuristics for virtual machines consolidation. Microsoft Research.
MSR-TR-2011-9, 2011, pp. 1 – 14.
Lopez-Pires, Fabio, and Benjamin Baran. “Virtual machine placement literature review.” arXiv
preprint arXiv:1506.01509 (2015), p. 1 – 11.
Mann, Zoltán Ádám. “Allocation of virtual machines in cloud data centers–a survey of problem
models and optimization algorithms.” (2015), pp. 1 – 28.
Minarolli, Dorian. “Utility-based Allocation of Resources to Virtual Machines in Cloud
Computing.” (2014) pp 1 – 133.

Surname 14
Nema, Priyanka, Choudhary Sapna, and Nema Tripiti. VM Consolidation Technique for Green
Cloud Computing. International Journal of Computer Science and Information
Technology, 2015, pp. 4620 – 4624.
Ngenzi, Alexander, and Suchithra R. Nair. “Dynamic resource management in Cloud datacenters
for Server consolidation.” arXiv preprint arXiv:1505.00577 (2015), p. 1 – 8.
Nguyen, Quyet Thang, et al. “Virtual machine allocation in cloud computing for minimizing
total execution time on each machine.” Computing, Management and
Telecommunications (ComManTel), 2013 International Conference on. IEEE, 2013, pp. 1
– 5.
Tighe, Michael, et al. “A distributed approach to dynamic VM management.” Network and
Service Management (CNSM), 2013 9th International Conference on. IEEE, 2013, pp. 1
– 5.
Ts’epoMofolo, R. Suchithra. “Heuristic based resource allocation using virtual machine
migration: a cloud computing perspective.” International Refereed Journal of
Engineering and Science (IRJES) 2.5 (2013): 40-45.
Viswanathan, Hariharasudhan, et al. “Energy-aware application-centric vm allocation for hpc
workloads.” Parallel and Distributed Processing Workshops and Phd Forum (IPDPSW),
2011 IEEE International Symposium on. IEEE, 2011, pp. 1 – 7.