Server Consolidation Algorithms for Virtualized Cloud Environment: A Performance Evaluation

Volume 5, No. 3, March-April 2014
International Journal of Advanced Research in Computer Science
RESEARCH PAPER
Available Online at www.ijarcs.info
© 2010-14, IJARCS All Rights Reserved 140
ISSN No. 0976-5697
Server Consolidation Algorithms for Virtualized Cloud Environment with Variable
Workloads: A Performance Evaluation
Susheel Thakur*
Department of Computer Science
Himachal Pradesh University
Shimla, India
Arvind Kalia
Shimla, India
Jawahar Thakur
Shimla, India
Abstract: Server Consolidation is an efficient approach towards the utilization of physical machines, in order to reduce the total number of
servers that an organization requires. To prevent the server sprawl, this practice was developed. Server Consolidation allow large-scale data
centers to improve their resource utilization and energy efficiency using virtualization technologies. To prevent server sprawl, server
consolidation aims at reducing the number of server machines used in the data centers by consolidating load and enhancing resource utilization
of physical systems. Virtualization enables the migration of virtual machines (VMs) between the physical machines using the technique of live
migration mainly for improving the efficiency. Virtual machine migration is promising approach to realize the objectives of efficient, adaptive
and dynamic resource management in virtualized cloud environment. In this paper, a comprehensive study of the server consolidation algorithms
and their usage towards dynamic resource management in the virtualized cloud environment is presented. We try to simulate and investigate the
impacts of different server consolidation algorithms on the performance of the live machine migration in both source and target machine in terms
of response time. Here in this paper, a performance evaluation of the chosen heuristics and fundamental insights obtained when variable load is
generated over the physical machines, aimed at reducing server sprawl, optimizing power consumption and load balancing across the physical
machines in virtualized cloud environment is presented.
Keywords: Cloud computing; live migration; Load balancing; Server Sprawl; Virtual Machine Monitor (VMM), Hotspot mitigation.
I. INTRODUCTION
In 1969, Leonard Kleinrock [15], one of the chief
scientists of the original Advanced Research Projects
Agency Network (ARPANET) which seeded the Internet,
said: “As of now, computer networks are still in their
infancy, but as they grow up and become sophisticated, we
will probably see the spread of computer utilities which, like
present electric and telephone utilities, will service
individual homes and offices across the country.” This
vision of computing utilities based on a service provisioning
model anticipated the massive transformation of the entire
computing industry in the 21st century whereby computing
services will be readily available on demand, like other
utility services available in today’s society. Cloud
Computing is defined by NIST[21] as a model for enabling
convenient, on demand network access to a shared pool of
configurable computing resources that can be rapidly
provisioned and released with minimal management effort
or service provider interaction.
For simplicity, a cloud is a pool of physical computing
resources i.e. a set of hardware, processors, memory, storage,
networks, etc. which can be provisioned on demand into
services that can grow or shrink in real-time scenario[27].
Virtualization plays a vital role for managing and
coordinating the access from the resource pool. A
virtualized environment that enables the configuration of
systems (i.e. compute power, bandwidth and storage) as well
as the creation of individual virtual machines is the key
features of the cloud computing. Virtualization is ideal for
delivering cloud services. Virtualization Technology enables
the decoupling of the application payload from the
underlying physical hardware and provides virtualized
resources for higher-level applications. An important feature
of a virtual machine is that software running inside it is
limited to resources and abstractions provided by the VM.
The software layer that provides the virtualization is called
virtual machine monitor (VMM). VMM virtualizes all of the
resources of physical machine, thereby supporting the
execution of multiple virtual machines. Virtualization can
provide remarkable benefits in cloud computing by enabling
VM migration to balance load across the data centers [13].
In the surge of rapid usage of virtualization, migration
procedure has been enhanced due to the advantages of live
migration say server consolidation and resource isolation.
Live migration of virtual machines [5] [18] is a technique in
which the virtual machine seems to be active and give
responses to end users all time during migration process.
Live migration facilitates energy efficiency, online
maintenance and load balancing [14]. Live migration helps
to optimize the efficient utilization of available CPU
resources.
Server consolidation is an approach to the efficient
usage of computer server resources in order to reduce the
total number of servers or server locations that an
organization requires. This approach was developed in
response to the problem of “server sprawl”. Server sprawl is
a situation in which multiple underutilized servers

Susheel Thakur et al, International Journal of Advanced Research In Computer Science, 5 (3), March- April, 2014,140-148
accommodate more space and consume more resources that
can be justified by their workload. Many organizations are
turning to server consolidation to reduce infrastructure
complexity, improve system availability and save money.
With increasingly powerful computing hardware, including
multi-core servers; organizations can run large workloads
and more applications on few servers. Reducing the
numbers of servers has tangible benefits for the data centers
as well.
Consolidation will result in reduced power consumption
and thus reducing overall operational costs for data center
administrators. Live migrations achieve this. Based on the
load conditions, under-utilized machines having resource
usage above a certain threshold are identified and migrations
are triggered to tightly pack VMs to increase overall
resource usage on all PMs and free up resources/PMs if
possible[3].
The rest of this paper is organized as follows: Section II
describes the survey of existing literature of various server
consolidation algorithms for cloud computing environment.
Section III provides the overview of the chosen server
consolidation algorithms for performance analysis in
virtualized cloud environment. Section IV gives details
about the experimental test bed used in performance
analysis. Section V discusses the experimental evaluation
and results. Section VI provides the conclusion and future
work.
II. RELATED WORK
Several research groups are working on server
consolidation in both academia and industry. This section
presents studies and systems related to server consolidation.
Khanna et al. [11] presented a dynamic management
algorithm, which is triggered when a physical server
machine becomes overloaded or underloaded. The main aim
of their algorithm was to: i) guarantee that SLAs are not
violated (SLAs are specified in terms of response time and
throughput); ii) minimizing the migration cost; iii)
optimizing the residual capacity of the system; and iv)
minimizing the number of physical machines used. Bobroff
et al. [19] proposed a dynamic server consolidation
algorithm to reduce the amount the amount of required
capacity and the rate of SLA violations. This heuristics
make use of historical data to forecast future demand and
relies on periodic executions to minimize the number of
physical machines to support the virtual machines. Mehta
and Neogi [23] developed the ReCon tool, aimed at
recommending dynamic server consolidation in multi-
cluster data centers. ReCon takes into account both static
and dynamic costs of physical machines, the costs of VM
migration and the historical resource consumption data from
the existing environment in order to provide an optimal plan
of VMs to physical machine mapping over time.
Wood et al. [26] presented the sandpiper system for
monitoring and detecting hotspots, and remapping VMs
whenever necessary. In order to choose which VMs to
migrate, Sandpiper sorts them using a volume-to-size (vsr)
metric, which is a based on cpu, network, and memory
loads. Sandpiper migrates the most loaded VMs from an
overloaded physical machine to one with sufficient capacity.
Srikantaiah et al. [24] studied the problem of request
scheduling for multi-tier web applications in virtualized
heterogeneous systems, to minimize the energy
consumption, while trying to meet the performance
requirements. They try to investigate the effects of
performance degradation due to high utilization of different
resources when the workload is consolidated. They have
determined that the energy consumption per transaction
results in a “U”-Shaped curve, and it is possible to find the
optimal utilization point. The authors have developed a
heuristic for the multidimensional bin packing problem as
an algorithm for the consolidation of workload to handle the
optimization over multiple resources. However, this
approach is workload type and application dependent.
Cardosa et al. [17] have developed an approach for the
problem of power-efficient allocation of VMs in virtualized
heterogeneous computing environments. They have used the
min, max and shares parameters of Xen’s VMM, which
represents minimum, maximum and proportion of the CPU
allocated to VMs sharing the same resource. However, the
approach suits only to enterprise environments as it does not
support strict SLAs and requires the knowledge of
application priorities to define the shares parameter.
Speitkamp and Bichler [4] [16] introduced a linear
programming for the static and dynamic server
consolidation problems. They also formulated extension
constraints for limiting the number of virtual machines in a
physical server, mapping virtual machines to a specific set
of physical servers that have some unique attribute and,
limiting the total number of migrations for dynamic
consolidation. In addition, they also designed an LP-
relaxation based heuristic for minimizing the cost of solving
the linear programming formulations. Emmanuel et al. [6]
proposed a dynamic resource allocation framework based on
their load balancing VM migration algorithm on a test-bed
of three ESX servers, a SAN, VMware VC (Virtual Center)
and VIBM Migration Handler. Similarly, Verma et al. [1]
presented the pMapper architecture and a set of server
consolidation algorithms for heterogeneous virtualized
resources. The algorithms take into consideration power and
migration costs and the performance benefit when
consolidating applications into physical servers.
Ameek et al. [2] has developed a HARMONY set-up
with ESX server and SAN (storage area network) controller
for integrating both server and storage virtualization
technologies to design an agile data centers. Keller et al.
[10] design Golondrina multi-resource management for
operating system-level virtualized environment with client
systems, manager server and cluster gate. Starling [12]
introduced affinity based VM placement and migration in a
decentralized approach with an 8-node cluster of 2x dual-
core AMD machines. Jung et al. [8] [9] have investigated
the problem of dynamic consolidation of VMs running a
multi-tier web-application using live migration, while
meeting SLA requirements. The SLA requirements are
modelled as the response time precomputed for each type of
transactions specific to the web-application. A new VM
placement is produced using bin packing and gradient
search techniques. The migration controller decides whether
there is a reconfiguration that is effective according to the
utility function that accounts for the SLA fulfilment.
However, this approach can be applied only to a single web-
application setup and, therefore, cannot be utilized for a
multitenant IaaS environment.

III. SERVER CONSOLIDATION ALGORITHMS
Server consolidation is an approach to the efficient
usage of computer server resources in order to reduce the
total number of servers or server locations that an
organization requires. This approach was developed in
response to the problem of “server sprawl”. In order to
reduce server sprawl in the data centers, server consolidation
algorithms are implemented. These algorithms are VM
packing heuristics which try to pack as many VMs as
possible on the physical machine (PM) so that resource
usage is improved and under-utilized machines can be
turned off.
A. Sandpiper:
Sandpiper is a system that automates the task of
monitoring and detecting hotspots, determining a new
mapping of physical resources to virtual resources, by
resizing or migrating VM’s to eliminate the hotspots.
Sandpiper makes use of automated black-box and gray box
strategies for virtual machine provisioning in cloud data
centers. Specifically the black-box strategy can make
decisions by simply observing each virtual machine from
the outside and without any knowledge of the application
resident within each VM. The authors present a gray-box
approach that assumes access to OS-level statistics in
addition to external observations to better inform the
provisioning algorithm. Sandpiper implements a hotspot
detection algorithm that determines when to resize or
migrate virtual machines, and a hotspot migration algorithm
that determines what and where to migrate and how many
resources to allocate. The hotspot detection component
employs a monitoring and profiling engine that gathers
usage statistics on various virtual and physical servers and
constructs profiles of resource usage. These profiles are used
in conjunction with prediction techniques to detect hotspots
in the system. Upon detection, Sandpiper grants additional
resources to overloaded servers if available. If necessary,
Sandpiper’s migration is invoked for further hotspot
mitigation. The migration manager employs provisioning
techniques to determine the resource needs of overloaded
VMs to underloaded servers.
Sandpiper supports both black-box and gray-box
monitoring techniques that are combined with profile
generation tools to detect hotspots and predict VM Resource
requirements. Hotspots are detected when CPU usage values
are violated with respect to the CPU thresholds set. Physical
machines (PMs) are classified as underloaded or overloaded.
The PMs are sorted based on the descending order of their
volume metric, and VMs are sorted based on the descending
order of their vsr metric, where volume and vsr are
computed as:
(1)
(2)
where cpu, memory and n/w refers to cpu, memory and
n/w usages of the PMs and VMs respectively and size refers
to the memory footprint of the VM.
To mitigate hotspot on an overloaded PM, the highest
vsr VM is migrated to a least loaded PM amongst the
underloaded ones. If the least loaded PM can’t house the PM,
next PM in the sorted order is checked. Similarly, if the VM
cannot be housed in any of the underloaded PMs, next VM
in the sorted order is checked. This way sandpiper tries to
eliminate hotspots by remapping VMs on PMs through
migration. The experimental results showed that migration
overhead is less than that of swapping overhead; however,
swapping increases the chances of mitigating hotspots in
cluster with high average utilization [25] [26].
B. Khanna’s Algorithm:
Khanna et al., in [11] [25], proposed Dynamic
Management Algorithm (DMA) that is based on
Polynomial-Time Approximation Scheme (PTAS) heuristic
algorithm. The algorithm operates by maintaining two types
of ordering lists, which are migration cost list and residual
capacity list. The PMs are sorted according to the increasing
order of their residual capacities across any resource
dimension like CPU. The VMs on each PM are sorted
according to the increasing order of their resource utilization
like CPU usage. Migration costs of the VMs are determined
based on their resource usage i.e. high usage implies high
costly migration. Whenever a hotspot is detected on a PM
due to violation of upper threshold, VM with least resource
usage is chosen for migration to target host which has the
least residual capacity to house it. If a PM cannot
accommodate the VM, next PM in the sorted order is
checked. Similarly, if the VM cannot be accommodated by
any of the candidate target PMs, next least usage VM from
the sorted order is checked.
Whenever coldspots are detected, the least usage VMs
across all the underloaded PMs is chosen and migrated to a
targeted PM, only if addition of the new VM increases the
variance of residual capacities across all the PMs, else we
choose the next VM in order. If there is no residual space
left for the chosen VM, then the heuristic for coldspot
mitigation stops. Variance is defined as follows:
(3)
(4)
rn= (5)
In above equation, mean is defined as the average of
normalized residual capacities across ‘m’ different resources
like cpu, memory, networks, etc. rescpu , resmem, resnet …
stands for residual capacities across different resource
dimensions. rn is the magnitude of the vector which
comprises of the individual variances across ‘n’ physical
machines.
Khanna’s Algorithm packs the VMs as tightly as
possible trying to minimize the number of PMs by
maximizing the variance across all the PMs. Thus, Khanna’s
algorithm minimizes power consumption by detecting
underutilization in the managed using Max-Min thresholds
selection model.
When the resource usage of a running PM violates a
minimum predefined threshold value, the algorithm tries to
pack the running VMs as close as possible thus trying to
minimize the number of running physical machines.

C. Entropy:
Entropy proposes a consolidation algorithm based on
constraint problem solving. The main idea of the constraint
programming based resource manager is to formulate the
VM resource allocation problem as constraint satisfaction
problem, and then applies a constraint solver to solve the
optimization problem. The ability of this solver to find the
global optimum solution is the main motivation to take this
approach. Entropy resource manager utilizes Choco
constraint solver to achieve the objectives of minimizing the
number of the running nodes and minimizing the migration
cost. Entropy iteratively checks optimality constraint i.e. the
current placement uses minimum number of the running
nodes. If Entropy is successful in constructing a new optimal
placement (uses fewer nodes) at VM packing problem
(VMPP) phase, it will activate the re-allocation. Entropy
employs a migration cost model that relates memory and
CPU usage with migration context. High parallelism
migration steps increases the cost. Using constraint
programming techniques facilitates the task of capturing such
context in two phases.
In the first phase, Entropy computes a tentative
placement (mapping of VMs to PMs) based on the current
topology and resource usage of PMs and VMs and
reconfiguration plan needed to achieve the placement using
minimum number of PMs required. In the second phase, it
tries to improve the reconfiguration plan by reducing the
number of migrations required. Since obtaining the
placement and reconfiguration may take a considerable
amount of time, the time given to the CSP solver is defined
by the users, exceeding which whatever immediate value the
solver has computed is considered for dynamic placement of
VMs. VMs are classified as active or inactive based on their
usage of CPU with respect to thresholds set.
The author define a viable configuration as one in which
every active VM present in the cluster has access to
sufficient cpu and memory resources on any PM. There can
be any number of inactive VM on the PM satisfying the
constraint. The CSP solver takes this viable condition into
account in addition to the resource constraints, while
procuring the final placement plan. However, considering
only viable processing nodes and CPU-Memory Resource
model is the limitation of the Entropy model [7] [25].
IV. EXPERIMENTAL TEST BED
Our implementation and evaluation is based on Xen
Hypervisor. Four PMs with Xen 4.1 hypervisor installed
were used to serve as the physical hosts and another machine
was used as NFS [20] Server to house the VMs images.
Physical machines which worked as clients were used
simultaneously to generate load on the virtual machines
hosted on the PMs. Seven VMs were created with Ubuntu
10.04, lucid host operating system with each 256 MB
memory size and Ubuntu 11.10 as PMs Host operating
system. They all have Apache, PHP and MySQL configured
on them to act as web and Database servers. A separate
machine has been configured which acts as the Management
node, which runs the controller and the Decision Engine. The
VIRT-M cluster management tool was implemented on this
management node. Python programming was used to
prototype these heuristics. Apart from these, RRD tool [22]
was installed on the Management node running the Decision
Engine, for storage of resource data. Our Experimentation
takes these heuristics into consideration and implements
these on idle VMs to investigate their impacts on
performance of live migration in both source and target
machine.
Figure 4.1 Experimental Test Bed
V. EVALUATION AND RESULTS
An experiment was performed with five VMs, two of
which were idle and three having variable workloads. This is
an interesting experiment to perform since if the rate is fixed,
the cpu usage levels will not vary drastically within the
duration of experiment and hence after some reconfiguration
in the initial topology to start with, the PMs are expected to
stabilize without further threshold violation. If the rate
changes, the usage levels will vary more and in such a
scenario, it will be of interest to have an idea how reactively
the algorithms do consolidation. “Httperf” with Webcalender
PHP scripts has been used for this experiment to inject
varying workload. The following scenario was created:
PM84 - lucid08 - no load
PM161 - lucid12 and lucid13 - rates varying as 10 req/sec, 20
req/sec and 30 req/sec on both the VMs
PM170 - lucid09 - rates varying as 10 req/sec, 20 req/sec and
30 req/sec
PM180 - lucid14 - no load
The experiment was performed for 10 minutes duration
for all the algorithms. The upper cpu threshold was kept as
60%. From figures 5.1, 5.2 and 5.3, it was observed that
there is too much of cpu usage level variations which was
expected due to variation of rates. Amongst the three,
variation in Entropy has been slightly lesser amongst the
PMs. Sandpiper mitigates 1 hotspot and triggers 1 migration
from PM170 to PM180. Khanna’s Algorithm performs 5
migrations in total and uses 3 PMs same as sandpiper.

Figure 5.1 Khanna’s Algorithm Figure 5.2 Sandpiper

Figure 5.3 Entropy Algorithm
Entropy reduces the number of PMs by 3 and triggers 3
migrations. As it was analyzed that PM84 has comparatively
higher usages than the others because of VM receptions from
other PMs. From figures 5.1, 5.2 and 5.3, it was observed
that Sandpiper detects a hotspot on PM170, migrates lucid09
to PM180 and reduces the number of PMs by 1. Khanna’s
Algorithm detects a coldspot at PM180, does coldspot
mitigation. It triggers a series of migrations, and finally uses
3 PMs at the end of the experimental run. Lucid14 was
migrated from PM 180 to PM 84. Lucid13 was migrated
from PM161 to PM 170. Lucid 14 from PM84 to PM161,
followed by lucid14 from PM161 to PM84, and finally lucid
08 from PM84 to PM161. Entropy in this case as well puts
all the VMs on PM161. As already known in entropy, viable
configuration is to be satisfied, which means number of
active PMs should have access to sufficient available
processing units. Here, even though the cpu usage of PM161
increases, the viability condition might have been satisfied
for which all the VMs were put on the PM. Also, there was a
decrease in cpu usage trend in the other PMs after entropy
migrates all the VMs to PM 161.
The cpu usage of PM161 were expected to increase
when varying workloads, but comparatively the hike is not
that much. It was believed that this happened because when
the rates are high on a VM, due to increase in processing
time of the requests or delay in sending response to the
client, the cpu usage of the VMs reduce even if the rate is
high from the client. The new topologies generated are:
• Sandpiper - PM84 with lucid08, PM161 with lucid12
and lucid13, PM180 with lucid14 and lucid09
• Khanna’s Algorithm - PM84 with lucid14, PM161 with
lucid08 and lucid12, PM170 with lucid09 and lucid13
• Entropy - all on PM161
Table I describes the measured statistics. It was analyzed
that both sandpiper and Khanna’s Algorithm uses 3 PMs
whereas Entropy uses just 1 PM. Sandpiper and Khanna’s
Algorithm reacts to hotspots when they are formed, whereas
Entropy doesn’t wait for hotspots or coldspots to happen,
based on the current topology configuration, it searches for
an optimal reconfiguration plan, once it finds it, the
reconfiguration plan is implemented. That is why in this
experiment entropy took only more than 1 minute to generate
the new plan whereas, Sandpiper used 3 PMs after
approximately 9.34 minutes from the start of the experiment
and Khanna’s Algorithm took more than 8 minutes to reduce
the number of PMs. Also, Sandpiper triggers just 1 migration
over Khanna’s Algorithm and entropy.
Table I. Measured Evaluation Metrics
Algorithm No. of PMs No. of
Migrations
Time Taken(mins)
Sandpiper 3 1 N/A
Khanna’s Algorithm 3 5 8.22
Entropy 1 3 1.8
As mentioned earlier a plan has been designed to
perform evaluation of application performance which is very
crucial for analysis. But due to much unexpected
circumstances valid correct response time values could not
be logged from which some concrete results could be
inferred. Here, the response time variation of 1 VM, lucid12
across all the three algorithms were presented.

Figure 5.4 Response Time Variation of Lucid12 in case of Sandpiper
Figure 5.5 Response Time Variation of Lucid12 in case of Khanna’s
Algorithm
Figure 5.6 Response Time Variation of Lucid12 in case of Entropy
From figures 5.4, 5.5 and 5.6, it can be observed that
response time for lucid12 in case of sandpiper is higher than
Khanna’s Algorithm and entropy. As seen from the graph, in
experimenting with Khanna’s Algorithm, the average
response time was approximately 200 milliseconds with
transient spikes where the response time value increased to
as high as 600 millisecs and 1100 millisecs. The reason
behind this sudden increase is the fact, that PM161 which
hosted lucid12, triggered migrations of lucid13 (co-located
VM). PM161 had been the source as well destination for
three migrations. And it is known that the application
performance of co-located VMs is affected in such cases.
Reception and migration of VMs from the same host could
occur where lucid12 resides must have been the reason of
sudden spikes in the response time.
In case of Entropy, the results are in conformity to the
events triggered by the algorithm. PM161 where lucid12
resides was chosen as the destination PM for all the VMs in
the topology. As a result, PM161 had been the destination
PM for three migrations. This increases the cpu usage of
PM161 slightly affecting the response time at the server on
lucid12. In addition to this, addition of three more VMs in
lucid12’s host and colocation amongst four other VMs must
have affected its request processing, affecting its application
performance. In Khanna’s Algorithm there are infrequent
spikes, but in entropy there are no spikes, the increase in
response times are persistent, the average is very close to
each other.
Unexpectedly, in sandpiper the response time values are
more compared to both Khanna’s Algorithm and entropy
both of which are comparable. Although in Sandpiper,

lucid12 was not migrated neither did its host PM161
undergo VM receptions from other hosts, the response times
are higher. This is because of the fact that, the MySQL
server running at the VM lucid12 got blocked because of too
many connections for some time. When the workload was
generated with varied rates, since database was inaccessible
for some time, the response time values increased and this
had an effect on the overall response time values of the
entire experimental run. However, it was noticed that the
application performance of the VM gets affected if the host
PM triggers too many migrations or undergoes VM
receptions. A migrating VM will undergo degraded
application performance. Thus, in this experiment, entropy
does best in terms of application performance.
VI. CONCLUSION AND FUTURE WORK
With the popularity of cloud computing systems, live
virtual machines migration will be great beneficial tool for
dynamic resource management in the modern day data
centers. To prevent server sprawl, server consolidation aims
at reducing the number of server machines by consolidating
load, enhancing resource utilization of physical systems
along with provision of isolation & security of the
application hosted. In this paper, we presented a performance
evaluation of the chosen server consolidation algorithms in
virtualized cloud computing environment when no workload
is generated on the client machines. Sandpiper and Khanna’s
Algorithm uses a threshold based technique of triggering VM
migrations. Entropy relies on CSP solver 4.4.1 to perform
consolidation by providing a set of constraints, optimizing
the number of PMs needed to house the VMs and the
migration cost to determine the selection of configuration
plan. In sandpiper, the migration cost is in terms of vsr metric
whereas Khanna’s algorithm considers the resource
utilization as the migration cost metric. All of them intend to
reduce migration cost in terms of the memory allocated to the
VMs. Unlike other algorithms, Entropy tries to obtain a
globally optimal solution, which distinguishes itself in its
consolidation approach. Unlike other algorithms does,
Entropy considers all the hosts in the topology and based on
their current resource usages, finds out an optimal solution
which tries to decrease the migration overhead in terms of
memory. The other algorithms try to achieve consolidation
on a per host basis, making sure that resource violations are
prevented every time each host is scanned, and then the VMs
are packed as closely as possible.
In case of varying workloads, when there are changing
cpu utilization levels, entropy performs better than Khanna’s
Algorithm. It uses less number of PMs and triggers less
number of migrations over Khanna’s Algorithm. Moreover,
it didn’t take much time to procure the reconfiguration plan
and initiate the new mapping. It is within acceptable limits
to cross the upper cpu threshold, at the benefit of obtaining
reduced number of PMs with less migration overhead. Too
many migrations in the system always induces some amount
of instability in the system, also, it could be observed that the
relatively higher variation of cpu utilizations across the PMs
with Khanna’s algorithm than the others. The reasoning is
that it is because of frequent VM relocation across PMs
which happens in Khanna’s Algorithm. Khanna’s algorithm
seems to be more befitting in an environment where there is
need of cpu usages within specified thresholds as a high
priority requirement, in the process of performing
consolidation. It can be seen that better consolidation is
possible than what Khanna’s Algorithm does, but the cpu
usage at the destination PM may reach too high, beyond
acceptance level in some cases.
The application performance of lucid12 is best in
Entropy. But unless it was analyzed the application
performance of all the other VMs present in the topology, it
cannot be ascertained concretely about the goodness of the
algorithms. These algorithms try to efficiently prevent server
sprawl and ensure non-disruptive load balancing in data
centers. Efficiency of the algorithm depends on the resource
parameters and metrics considered. Hence, a comparative
performance analysis was carried out to analyse their
applicability, goodness and incurred overhead. In near future,
Evaluation of these algorithms with mixed load where
different types of applications are used together, can facilitate
in figuring out the distinct cases where an algorithm will
behave well and hence can be used in those cases only to
leverage the maximum benefits.
By increasing the scale of experimentation,
investigations can be done in finding out whether behavior of
algorithms changes when the number of PMs and VMs are
increased in term of the metrics defined to compare the
goodness of the algorithms. Moreover, more algorithms
which does similar jobs like consolidation can be chosen in
near future and their relative behavior can be analysed with
the already chosen algorithms.
VII. REFERENCES
[1] A. Verma, P. Ahuja, A. Neogi, “pMapper: Power and
migration cost aware application placement in virtualized
systems”, in Proceedings of the ACM/IFIP/USENIX 9th
International Middleware Conference, 2008.
[2] Aameek Singh, Madhukar Korupolu, Dushmanta Mohapatra.
“Server-Storage Virtualization: Integration and Load
Balancing in Data Centers”, in Proceedings of the SC 2008
ACM/IEEE conference on Supercomputing, 2008.
[3] Anju Mohan and Shine S, “Survey on Live VM Migration
Techniques”, International Journal of Advanced Research in
Computer Engineering & Technology, vol 2, Jan. 2013.
[4] B. Speitkamp, M. Bichler, “A Mathematical Programming
Approach for Server Consolidation Problems in Virtualized
Data Centers”, IEEE Transactions on Services Computing.
[5] C. Clark, K. Fraser, S. Hand, J. Hansen, E. Jul, C. Limpach,
I. Pratt, and A. Warfiled, “Live Migration of Virtual
Machines”, in Proceedings of the 2nd Conference on
Symposium on Networked Systems Design &
Implementation, vol. 2, pp. 286, 2005.
[6] Emmanual Arzuaga and David R. Kaeli, “Quantifying load
imbalance on virtualized enterprise servers”, in Proceedings
of the first joint WOSP/SIPEW International Conference on
Performance Engineering IEEE/ACM Trans. Netw. Pages
235-242, 2010.
[7] Fabein Hermenier, Xavier Lorca, Jean-Marc Menuad, Gilles
Muller and Julia Lawall, “Entropy: A Consolidation
Machine Manager for Clusters”, in Proceedings of the 2009
ACM SIGPLAN/SIGOPS Internal Conference on Virtual
Execution Networks, VEE, 2008, pp.41-50, 2009.

[8] G Jung, KR Joshi, MA Hiltunen, RD Schlichting RD, C Pu,
“Generating adaptation policies for multi-tier applications in
consolidated server environments”, in Proceedings of the 5th
IEEE International Conference on Autonomic Computing
(ICAC 2008), Chicago, IL, USA, 2008; 23–32.
[9] G Jung, KR Joshi, MA Hiltunen, RD Schlichting RD, C Pu,
“A cost-sensitive adaptation engine for server consolidation
of multitier applications”, in Proceedings of the 10th
ACM/IFIP/USENIX International Conference on
Middleware (Middleware 2009), Urbana Champaign, IL,
USA, 2009; 1–20.
[10] G. Keller and H. Lutfiyya, “Replication and migration as
resource management mechanisms for virtualized
environments”, in Proceedings of the Sixth International
Conference on Autonomic and Autonomous Systems
(ICAS), pages 137-143, 7-13, 2010.
[11] Gunjan Khanna, Kirk Beaty, Gautam Kar and Andrzej
Kochut, “Application Performance Management in
Virtualized Server Environments”, 10th IEEE/IFIP
Conference on Network Operations and Management in
Virtualized Server Environments, NOMS, 2006.
[12] James Greensky, Jason Sonnek, Robert Reutiman, and
Abhishek Chandra, “Starling: Minimizing communication
overhead in virtualized computing platforms using
decentralized affinity-aware migration”, 39th International
Conference on Parallel Processing (ICPP), pages 228-237,
September 2010.
[13] Jyothi Sekhar, Getzi Jeba and S. Durga, “A survey on
Energy Efficient Server Consolidation through VM Live
Migration”, International Journal of Advances in
Engineering and Technology, vol 5, pp.515-525, Nov. 2012.
[14] Kejiang Ye, Xiaohong Jiang, Dawei Huang, Jianhai Chen,
and Bei Wang, “Live migration of Multiple Virtual
Machines with Resource Reservation in Cloud Computing
Environments” , in Proceedings of 2011 IEEE 4th
International Conference On Cloud Computing, pp. 267-274,
2011.
[15] L. Kleinrock, “A Vision for the Internet”, ST Journal of
Research, 2(1):4-5, Nov, 2005.
[16] M. Bichler, T. Setzer, B. Speitkamp, “Capacity planning for
virtualized servers”, in Proceedings of the 16th Annual
Workshop on Information Technologies and Systems
(WITS’06), 2006.
[17] M Cardosa, M Korupolu, A Singh, “Shares and utilities
based power consolidation in virtualized server
environments”, in Proceedings of the 11th IFIP/IEEE
Integrated Network Management (IM 2009), Long Island,
NY, USA, 2009.
[18] M. Nelson, B. Lim, and G. Hutchins, “Fast transparent
migration for virtual machines”, in Proceedings of the
Annual Conference on USENIX Annual Technical
Conference, pp. 25, 2005.
[19] N. Bobroff, A. Kochut, K. A. Beaty, “Dynamic placement of
virtual machines for managing sla violations”, in
Proceedings of the 10th IFIP/IEEE International Symposium
on Integrated Network Management (IM’07), 2007.
[20] NFS setup in Ubuntu,
https://help.ubuntu.com/community/NFSv4Howto.
[21] NIST Definition of Cloud Computing v15,
www.nist.gov/itl/cloud/upload/cloud-def-v15.pdf.
[22] RRD Tool, http://oss.octiker.ch/rrdtool/.
[23] S. Mehta, A. Neogi, “ReCon: A Tool to Recommend
Dynamic Server Consolidation in Multi-Cluster Data
Centers”, in Proceedings of the IEEE Network Operations
and Management Symposium (NOMS’08), Salvador, Bahia,
2008.
[24] S Srikantaiah, A Kansal, and F Zhao, “Energy aware
consolidation for cloud computing”, Cluster Computing
2009; 12:1–15.
[25] Susheel Thakur, Arvind Kalia and Jawahar Thakur, “Server
Consolidation Algorithms for Cloud Computing
Environment: A Review”, International Journal of Advanced
Research in Computer Science and Software Engineering,
vol 3(9), September 2013.
[26] Timothy Wood, Prashant Shenoy, Arun Venkataramani and
Mazin Yousif, “Sandpiper: Black-box and Gray-box
Resource Management for Virtual Machines”, Journal of
Computer Networks, vol.53, pp.2923-2938, Dec.2009.
[27] V. Sarathy, P. Narayan, and Rao Mikkilineni, “Next
Generation Cloud Computing Architecture- Enabling Real-
time Dynamism for Shared Distributed Physical
Infrastructure”, 19th IEEE International Workshops on
enabling Technologies: Infrastructure for Collaborative
Enterprises, WETICE, pp.48- 53, June 2010.

Server Consolidation Algorithms for Virtualized Cloud Environment: A Performance Evaluation

Recommended

More Related Content

What's hot (19)

Similar to Server Consolidation Algorithms for Virtualized Cloud Environment: A Performance Evaluation (20)

More from Susheel Thakur (6)

Recently uploaded (20)

Server Consolidation Algorithms for Virtualized Cloud Environment: A Performance Evaluation