See discussions, stats, and author profiles for this publication at: [Link]

net/publication/273131762

Threshold Based Auto Scaling of Virtual Machines in Cloud Environment

Conference Paper · September 2014

DOI: 10.1007/978-3-662-44917-2_21

CITATIONS READS

33 468

3 authors:

Mohan Murthy M K H. A. Sanjay

Visvesvaraya Technological University Nitte Meenakshi Institute of Technology
12 PUBLICATIONS 67 CITATIONS 88 PUBLICATIONS 614 CITATIONS

SEE PROFILE SEE PROFILE

Anand Jumnal
SECAB Institute of Engineering and Technology
4 PUBLICATIONS 36 CITATIONS

SEE PROFILE

All content following this page was uploaded by Anand Jumnal on 07 May 2016.

The user has requested enhancement of the downloaded file.

 Threshold Based Auto Scaling of Virtual Machines
in Cloud Environment

M.K. Mohan Murthy, H.A. Sanjay, and Jumnal Anand

Nitte Meenakshi Institute of Technology, Bangalore

{maakem,[Link],anandsbj1989}@[Link]

Abstract. Cost effectiveness is one of the reasons behind the popularity of

Cloud. By effective resource utilization cost can be further reduced and
resource wastage can be minimized. The application requirement may vary over
time depending on many factors (for instance load on the application); user may
run different types of application (a simple MS word to complex HPC
application) in a VM. In such cases if the VM instance capacity is fixed there is
a high possibility of mismatch between the VM capacity and application
resource requirement. If the VM capacity is more than the application resource
requirement then resource will be wasted; if the VM capacity is less than the
application resource requirement then the application performance will degrade.
To address these issues we are proposing threshold based auto scaling of virtual
machines in which VMs will be dynamically scaled based on the application
resource utilization (CPU and Memory). Using our approach effective resource
utilization can be achieved.

Keywords: Cloud Computing, Auto Scaling, Virtual Machines.

1 Introduction
In cloud paradigm software, infrastructure, and platform are given as services. In this
work we are considering infrastructure (Virtual Machine). IaaS providers provide
Virtual Machine (VM) to end user. User will use VM instance to host/run his
applications and he will pay some amount as per the SLA (Service Level Agreement).
Many organizations moving towards private cloud; effective resource utilization, cost
reduction, and easy maintenance are some of the reasons behind this. Employees in
the organization will get the VM instances. They have to login to these instances to
use them. Whether it is commercial cloud or private cloud following two scenarios
are possible

1.1 Scenario 1: User Hosts Different Application on the VM

User may use VM to host different applications from a simple MS word to complex
accounting software. If the VM instance is static (normally this will be the case) user
has to select VM instance in such a way to match the application which has the
maximum resource requirement. In this case for instance if the user uses the VM to

C.-H. Hsu et al. (Eds.): NPC 2014, LNCS 8707, pp. 247–256, 2014.
© IFIP International Federation for Information Processing 2014
248 M.K.M. Murthy, H.A. Sanjay, and J. Anand

run his application which has the maximum requirement only for 2 hours in a day
then in remaining 22 hours resource will be wasted. If the application resource
requirement is more than the VM then it leads to application performance
degradation.

1.2 Scenario 2: Application Requirement Vary over Time

Consider a database application which needs more resource when the transactions are
happening. If the transactions are not there it doesn’t need high resources. In case of
the static VM instance this will lead to resource wastage.
Migration of application from one VM to another address’ above mentioned issues
but it is having many disadvantages. It is time consuming, tedious, not cost effective,
and error prone. If VM is dynamically scaled according to the application requirement
resource wastage can be minimized.
To address the above mentioned issues we have developed and tested threshold
based auto scaling of VM mechanism, in which the VM is auto configured according
to the application requirement. In threshold based auto scaling the resource utilization
of the VM is monitored. If they exceed the predefined threshold values then VM
capacity will be increased or decreased dynamically without shutting them down
according to the need, which minimizes resource wastage.

1.3 Up-Scaling

In this work we have considered RAM and CPU utilization of the VM. As the
resource requirement of the application increases, the RAM and CPU utilization of
the VM increases. At some point the application resource requirement will become
more when compare to VMs capacity as a result performance of the application
degrades and ultimately it will hang. To avoid this problem when the CPU and
Memory utilization of the VM crosses the predefined maximum threshold value we
will increase the RAM and CPU capacity of the VM.

1.4 Down-Scaling
As the resource requirement of the application decreases, the RAM and CPU
utilization of the VM decreases. This will lead to resource wastage since the VM
capacity is not fully utilized. To avoid resource wastage when the CPU and Memory
utilization of the VM crosses the predefined minimum threshold value we will
decrease the RAM and CPU capacity of the VM. Monitoring and scaling of RAM and
CPU of the VM are two independent tasks.
Rest of the paper is organized as follows section 2 gives brief description about the
related work; section 3 explains the threshold based auto scaling of VM; section 4
describes the algorithms used to upscale/downscale the VM; section 5 talks about the
experimental setup and results, followed by conclusion.
 Threshold Based Auto Scaling of Virtual Machines in Cloud Environment 249

2 Related Work

There are couple of efforts related to dynamic resource provisioning in cloud

environment. In [1] auto-scaling of the VM instances with respect to the load, where
load is defined as number of jobs submitted and the deadline and budget to complete
the submitted jobs are considered. In the proposed system the VM must be rebooted
after scaling it. So there will be time delay and performance degradation. In some
scenarios rebooting of VM is not acceptable.
In [2] a novel architecture is presented for the dynamic scaling of web applications
based on thresholds in a virtualized cloud computing environment. This work
illustrates the scaling approach with a front-end load balancer for routing and
balancing user requests. Web applications are deployed on web servers installed in
virtual machine instances. In [3] a cloud computing architecture is constructed with a
front-end load balancer, a virtual cluster monitor system and an auto-provisioning
system. The front-end load balancer is utilized to route and balance user requests to
cloud services deployed in a virtual cluster. The virtual cluster monitor system is used
to collect the statistics of the usage of physical resources in each virtual machine in
the virtual cluster. The auto-provisioning system is used to dynamically provision the
virtual machines based on the number of the active sessions or the use of the
resources in the virtual cluster.
In the works [2] [3] front end load balancer is used for load balancing on virtual
machines. In these works a new instance of VM is added to the VM cluster if the
resource utilization crosses the upper threshold. If the utilization is below the lower
threshold then VM instances are removed from the cluster. In these works scaling of
the VM based on the hosted application requirement is not addressed. In [4] a model-
driven engineering approach is presented to optimize the configuration, energy
consumption, and operating cost of cloud auto-scaling infrastructure to create greener
computing. This work concentrates on energy consumption and budget constraints. In
this work pre-configured static VM instances are used. This will lead to resource
wastage as well as application performance degradation.
In all of these works cloud is scaled by adding new VMs. They are not considering
auto scaling of a single VM which is required in the scenario where user will be using
a VM to run his applications in private or public cloud. Our work is about to auto
scale a VM based on the threshold values.

3 Threshold Based Auto Scaling of VM

Application requirement may change over time and also user may host different
applications (which have different resource requirement) on the VM. In these cases
fixed VM capacity may lead to resource wastage or application performance
degradation. This can be addressed by dynamically scaling the VM according to the
hosted application requirement. In threshold based auto scaling the resource
utilization of the VM is monitored. If they exceed the predefined threshold values
then VM capacity will be increased or decreased dynamically according to the need
without shutting down the VMs, which minimizes resource wastage. High level
system overview is shown in figure 1.
250 M.K.M. Murthy, H.A. Sanjay, and J. Anand

Monitor VM VM VM

[Link]

Decision VM
Maker Configuration

Fig. 1. Auto Scaling system overview

Auto-scaling system has the following components

3.1 Monitor
The monitor component monitors the VMs; it reads the CPU and Memory utilization
and passes this data to Decision Maker component. It uses Xen APIs to get the CPU
and Memory utilization of the VMs. It sends request for the CPU and Memory
utilization of VMs to the XCP using Xen APIs. By default it monitors all the active
VMs or we can make it to monitor only specific VMs by configuring the
corresponding values in [Link]. The time interval to send the request to
XCP to get the VMs statistics can be configured in [Link].
When the Monitor module starts it will read all the configuration properties from
the [Link] and monitors the VMs as per the values set to different property
parameters in [Link] file.

3.2 Decision Maker

Decision maker module gets the VM statistics from Monitor module and it also read
the threshold values from the [Link] file, compares against the VM
statistics and decides whether to up/down scale the VM and conveys this decision to
VM configuration module. The information passed to the VM configuration module
includes the VM ID which should be scaled, whether scaling should happen to RAM
or CPU and how much scaling should happen. All the threshold values and scaling
values (helps in to take the decision of how much scaling should happen) are
configurable which are stored in the [Link] file.
There is a possibility that VM’s CPU and RAM utilization may exceed the
threshold value for few seconds and again come back to the normal values. If the
monitor module gets these values it will trigger up/downscaling of the RAM/CPU of
the VM. In the next iteration monitor module gets the normal values again this
triggers down/up-scaling of the RAM/CPU of the VM which results in unnecessary
up/down scaling of VMs. To avoid this problem we have introduced configurable
properties called cpuiteration (min and max) and memoryiteration (min and max).
 Threshold Based Auto Scaling of Virtual Machines in Cloud Environment 251

Any positive integer value from 0 to n can be set to the cpuiteration and
memoryiteration. Both are independent of each other, min memoryiteration and min
cpuiteration are used in case of down-scaling, max memoryiteration and max
cpuiteration are used in case of up-scaling. The Decision Maker initiates
up/downscaling only if the RAM and CPU utilization of the VM exceeds the
threshold values in the successive number of iterations specified in the cpuiteration
(min and max) and memoryiteration (min and max) .

4 Scaling Algorithms

We have written two separate algorithms for memory scaling and CPU scaling.
The working principle of both the algorithms is same.
In case of the memory scaling algorithm each VM’s memory utilization (Mx) is
read and compared with max memory threshold value (Mmx). If Mx is greater than or
equal to Mmx then the max memory utilization counter (MTx) is increased and min
memory utilization counter (MTm) is reset.

4.1 Memory Scaling

Step 1: for each VMx Read Memory utilization Mx

Step 2: if the Mx>=Mmx
- Increment the MTx for the VMx
- Reset MTm for the VMx
else if Mx<=Mmn
- Increment the MTm for the VMx
- Reset MTx for the VMx

Step 3: if MTx>TTMt and if free memory available

- initiate the VM up-scaling for memory
- go to step 5
else if MTm >TTMt
- initiate the VM down-scaling for memory
- go to step 5
Step 4: Go to Step 1
Step 5: Reset the MTx and MTm
Step 6: Go to Step 1

4.2 CPU Scaling

Step 1: for each VMx Read CPU utilization Cx
Step 2: if the Cx>=Cmx
- Increment the CTx for the VMx
- Reset CTm for the VMx
else if Cx<=Cmn
252 M.K.M. Murthy, H.A. Sanjay, and J. Anand

- Increment the CTm for the VMx

- Reset CTx for the VMx
Step 3: if CTx>TTCt and if computing resources available
- initiate the VM up-scaling for CPU
- go to step 5
else if CTm >TTCt
- initiate the VM down-scaling for CPU
- go to step 5
Step 4: Go to Step 1
Step 5: Reset the CTx and CTm
Step 6: Go to Step 1

VMx - Virtual Machine identifier

Mmx - Memory maximum threshold defined
Mmn - Memory minimum threshold defined
Cmx - CPU maximum threshold defined
Cmn - CPU minimum threshold defined
T – Time interval defined to read the Memory, CPU utilization of VMs
TTMt - Time threshold counter defined for Memory.
TTCt - Time threshold counter defined for CPU.
CTx - Max CPU iteration count.
MTx - Max memory iteration count.
CTm - Min CPU iteration count.
MTm - Min memory iteration count.
Cx - CPU utilization of VMx
Mx - Memory utilization of VMx

5 Experimental Setup and Results

A Cloud Environment is set up using Xen Cloud Platform (XCP) [5]. XCP includes
Xen Hypervisor, Xen API tool-stack, vSwitch etc. XCP is an open source enterprise-
ready server virtualization and cloud computing platform. Many of the existing IaaS
providers are using the customized XEN to create the virtualization infrastructure.
XCP delivers the Xen hypervisor with support for a range of guest operating systems
including Windows and Linux network and storage support, management tools in a
single, tested installable image, which is also called XCP appliance. It also supports
the dynamic scaling of virtual machine. XCP APIs are used to get the memory and
CPU utilization statistics of the VMs and to up-scale and down-scale the VMs.

5.1 Memory Scaling

XCP supports two types of memories: static and dynamic. Each will have minimum
and maximum range. The static memory maximum defines the maximum amount of
physical memory that the guest operating system can address from the time the guest
 Threshold Based Auto Scaling of Virtual Machines in Cloud Environment 253

boots up until the time the guest shuts down again. It is not possible to change static
memory when the VM is running. In case of the dynamic memory it is possible to
increase/decrease the range when the VM is running. XCP provides a feature called
dynamic memory controller. Using the following API’s provided by the XCP we
increase/decrease the dynamic memory within the valid range whenever required.

xe vm-param-set uuid=<uuid> memory-dynamic-{min, max};

uuid is the identifier which uniquely identifies a VM.

5.2 CPU Scaling

CPU scaling can be done by either modifying the CPU cap or CPU weight. The CPU
cap optionally fixes the maximum amount of CPU a domain will be able to consume.
The cap is expressed in percentage of one physical CPU: 100 is 1 physical CPU, 50 is
half a CPU, 400 is 4 CPUs, etc... The default, 0, means there is no upper cap [5]. CPU
weight of a VM decides how much CPU is allocated to that VM. A domain with a
weight of 512 will get twice as much CPU as a domain with a weight of 256 on a
contended host. Legal weights range from 1 to 65535 and the default is 256 [5]. In our
work we have used the CPU cap to scale the CPU allocated to VM. The following
API provided by the XCP is used.

xe vm-param-set uuid<uuid> VCPUs-params:cap=<value>

The Cloud Server is setup on a 4 Core Machine with Intel Xeon W3250 processor
with 2.67 GHz, 12 GB of DDR3 RAM, 1 TB Hard disk with 7200 RPM. The desktop
machine which is used to monitor the VMs for memory and CPU utilization of the
application has the Intel Core2 Duo processor with 2.66 GHz clock speed, 1 GB of
RAM, 500 GB hard disk, connected over an Ethernet LAN. CentOS 5.7 Operating
System is used.
To generate load on VMs we have used both computing intensive programs and
memory intensive programs. Once the programs are started the CPU and memory
utilization of the VMs will increase. We have set the scaling factor for memory as
1.25 and for CPU it is 2. The utilization count is set to 3 minutes i.e. if the resource
usage exceeds the upper threshold value for 3 minutes continuously, then the
corresponding VMs will be allocated more resources as specified in the scaling factor
(up-scaling). If the resource usage is below the lower threshold value for 3 minutes
continuously, then the resource will be de-allocated from the VM as per the scaling
factor (down-scaling).
We have setup the threshold values as follows
Upper threshold – 80%, Lower threshold – 25%
Figure 2 and 3 shows memory up-scaling and downscaling respectively. Figure 4
and 5 shows CPU up-scaling and downscaling respectively. Since we have set the
utilization count to 3 minutes, we can observe the scaling at 4th minute in all the cases.
254 M.K.M. Murthy, H.A. Sanjay, and J. Anand

Fig. 2. Memory up-scaling

Fig. 3. Memory down-scaling

 Threshold Based Auto Scaling of Virtual Machines in Cloud Environment 255

Fig. 4. CPU up-scaling

Fig. 5. CPU down-scaling

Following is our observations

 Downscaling minimizes the resource wastage.
 Up-scaling make sure that the application performance is not compromised.
 Choosing the right threshold values is very important for the success of our
approach.
A lower threshold value result in fluctuation of the VM capacity and a higher
threshold value makes our algorithms less responsive to the change in resource
utilization of the VM.
256 M.K.M. Murthy, H.A. Sanjay, and J. Anand

6 Conclusion

By adopting effective resource utilization techniques resource wastage can be

minimized. Our threshold based auto-scaling is one such technique in which VM is
dynamically scaled as per the application resource requirement, thereby minimizing
the resource.
Selecting proper threshold values is very important factor in the success of our
approach. A lower threshold value leads to frequent change in VM configuration and
a higher value reduces the responsiveness of VM to adapt to the new resource
requirement. We can use several techniques to find out optimum threshold values
such as history based, mathematical model etc. In future we are planning to build a
feedback based system to dynamically scale the VM according to the application
requirement. At present our dynamic scaling system is threshold based where
threshold values are static and predefined. In feedback based system we will monitor
the application’s resource utilization and it will be used as feedback to define the
threshold values.

References
1. Mao, M., Li, J., Humphrey, M.: In: Schwarz, T.S.J., Miller, E.L.: Cloud Auto-scaling with
Deadline and Budget Constraints. Department of Computer Science University of Virginia
Charlottesville, VA, USA 22904 {ming, jl3yh, humphrey}@[Link] (2011)
2. Chieu, T.C., Mohindra, A., Karve, A.A., Segal, A.: Dynamic Scaling of Web Applications
in a Virtualized Cloud Computing Environment. In: 2009 IEEE International Conference on
e-Business Engineering (2009)
3. Hung, C.-L., Hu, Y.-C., Li, K.-C.: Auto-Scaling Model for Cloud Computing System. Dept
of Computer Science & Information Engineering, Providence University {clhung, ychu,
kuancli}@[Link]
4. Doughertya, B., Whiteb, J., Schmidta, D.C.: Model-driven Auto-scaling of Green Cloud
Computing. Institute for Software Integrated Systems, Vanderbilt University, Campus Box
1829 Station B, Nashville, TN 37235, Email:{briand,schmidt}@[Link] bECE,
302 Whitemore Hall, Virgnia Tech, Blacksburg, VA 24060, Email:julesw@[Link]
5. [Link]

View publication stats