A Survey of Performance Comparison between Virtual Machines and Containers

© 2016, IJCSE All Rights Reserved 55
International Journal of Computer Sciences and EngineeringInternational Journal of Computer Sciences and EngineeringInternational Journal of Computer Sciences and EngineeringInternational Journal of Computer Sciences and Engineering Open Access
Survey Paper Volume-4, Issue-7 E-ISSN: 2347-2693
A Survey of Performance Comparison between Virtual Machines and
Containers
Prashant Ramchandra Desai
Department of Computer Engineering, Bharati Vidyapeeth University, Pune, India
Available online at: www.ijcseonline.org
Received:18/Jun/2016 Revised: 26/Jun/2016 Accepted: 16/Jul/2016 Published: 31/Jul/2016
Abstract—: Since the onset of Cloud computing and its inroads into infrastructure as a service, Virtualization has become peak
of importance in the field of abstraction and resource management. However, these additional layers of abstraction provided by
virtualization come at a trade-off between performance and cost in a cloud environment where everything is on a pay-per-use
basis. Containers which are perceived to be the future of virtualization are developed to address these issues. This study paper
scrutinizes the performance of a conventional virtual machine and contrasts them with the containers. We cover the critical
assessment of each parameter and its behavior when its subjected to various stress tests. We discuss the implementations and
their performance metrics to help us draw conclusions on which one is ideal to use for desired needs. After assessment of the
result and discussion of the limitations, we conclude with prospects for future research.
Keywords—Performance, Virtual Machines(VMs), Containers, Virtualization, Kernel Virtual machines (KVM), Docker, Hypervisor.
I. INTRODUCTION
The abundance of cloud computing over the past decade has
lead to significant growth in virtualization techniques used.
The services provided by Cloud Computing indirectly
depend on the level of virtualization provided by the
underlying virtual machines. Hardware virtualization
consists of abstracting complete hardware resources so that
entire software like an operating system can run on it.
Hardware virtualization is categorized further depending
upon the levels of abstraction it offers into full virtualization,
partial virtualization, and Para virtualization. In Para
virtualization, the guest operating system is recompiled
before being installed in a virtual machine. The hypervisor
serves as a host for the guest OS and adds a layer of
virtualization. It also acts as an interface between guest OS
and the hardware. In full virtualization, the hypervisor is
directly installed on the hardware. Each guest operating gets
all the features of the underlying hardware because of the
layer of abstraction provided it doesn't feel the hypervisor's
presence. It is a commonly implemented form of
virtualization in virtual machines [1][2]. Since the
virtualization software is a bundle of hardware drivers and
interfaces which needs to be pre-installed before creating the
virtual machines. It is been observed, in some cases [1]some
of the interfaces and drivers do not function as expected thus
enabling the usage of Dockers where a small size
Unix/Linux operating system is installed on the bare-metal
and the virtual machines are hosted from the Docker or the
container-based virtualization. Container-based
virtualization is an alternate technology to virtual machines
and is quickly replacing them in the cloud environment. The
virtualization layer runs as an application, where the
operating system's kernel runs on bare metal with guest OS
installed on top of it. These are known as containers.
II. BACKGROUND
A. KVM
Kernel-based Virtual Machine (KVM) is a virtualization
technology for Linux systems that turns it into a hypervisor
and is designed for x86 processor architecture. The
hypervisor is built on Linux kernel, running an open source
operating system. It provides support for multiple guest OS
include windows, BSD, and Solaris. KVM is a full
virtualization technique where you can provide virtualized
hardware: that allocates the number of CPUs, memory, hard
disk space to the guest operating system.
There are a number of management tools as libvirt, oVirt
and Virtual Machine Manager available to easily manage
KVM through graphical user interface . One of the features
supported by KVM is the live migration in which entire data
centers can be backed up for maintenance without affecting
the functioning of the guest operating system.KVM supports
"hot plugging" that means you can resize or add additional
resources to the virtual machine while it is in operation
without disrupting its services.
B. Containers
Containers are similar to virtual machines in the services
they provide except they don't come with an overhead of
running a separate kernel and virtualizing all hardware
components. It modifies the host operating system to provide
abstraction. They have a container ID and group permission
system. Its main concept is built on the use of kernel

International Journal of Computer Sciences and Engineering Vol.-4(7), PP(55-59) Jul 2016, E-ISSN: 2347-2693
namespaces where isolated containers are created. These
have no access to objects outside the container. A container
which runs a full operating system is a system container
whereas the one which runs an application is an application
container. Because a containerized system has only one
kernel, there is only one level of resource scheduling and
allocation. The process that run inside a container is not
aware of the limit on the resources it can utilize hence there
always remains a risk where an application may over-
allocate itself resources.[3]
Security in containers is achieved by managing the group
permissions and by creating namespace awareness, where
the users are not treated with the same privileges inside the
container as those outside them. Docker is the most widely
used container today that deploys applications inside
software containers along with its code, runtime, system
tools, system libraries, etc . Docker has layered file system
images, supported by AUFS (Another UnionFS). Containers
have a faster boot time than virtual machines.
III. ASSESSMENT
A virtual machine can be benchmarked using various
performance measurements tools that integrate with your
hypervisor accordingly. Some of the tools used are VMware
ESXi, VSphere, vCenter. These tools monitor the resource
utilization of an active virtual machine and later aggregate
data into excel format for analysis. These same tools can be
used to benchmark application or system containers.
The performance can be measured on the basis of various
parameters such as throughput, latency, bandwidth, etc of
the system. We are mainly concerned with the overhead
generated by native, virtualized and non-virtualized
environments and how it can be optimized to improve
system performance and prove which one is suitable to use
under preferred conditions. [7]-[10]
The behavior of various parameters under the defined
workload is observed carefully and recorded, to study its
characteristics and get a consolidated comparison between
the three implementations.
A. CPU Performance
Throughput is the parameter used to measure the output
of the workload the CPU is subjected to a compression, High
Performance Computing (HPC) test. As observed the native
and Docker performance is similar in compression while
KVM is slower as compared to them.
HPC performance is similar on native and Docker but is
reasonably slow on KVM because of abstraction that works
as a disadvantage in this situation. The CPU schedulers do
not influence the processor in the native and Docker setup
hence there is no difference in performance.
B. Memory Performance
Bandwidth is the parameter used to measure the speed of
memory access and operations. Under various benchmarks
developed to stress test the memory in sequential and
random access methods, it is observed that the
performance of native, Docker and KVM environment is
almost the same for various operations with very little
deviation. The testing was carried out on a single node on
large datasets. Container based systems have the ability to
return unused memory to the host resulting in better usage.
Para virtualization systems suffer from double cache since
the same memory blocks are used by the host and the
virtual machine.
C. Network Performance
Bandwidth is used to measure the performance of the
network communication. The communication scenario is
bulk data transfer over a single TCP connection like the
client -server model.
The data transfer rate is measured in both directions since
TCP/IP stack has different policies for sending and
receiving data. The major component that causes a
bottleneck in performance is the NIC in which case we use
the CPU cycles to measure the overhead. As for the
performance goes Docker uses bridging and NAT which
increases the route length. Dockers which do not use NAT
have same performance to native systems.KVM
performance can be improved if the VM can directly
communicate with the host bypassing the in-between
layers.
Latency can also be another network parameter used to
determine performance. In the above setup of the
experiment, the client sends 200 bytes of request to the
server and the server response with a 400 bytes reply in
this case the client has to wait to process the complete
request. If we consider wireless and wired communication
protocols Docker with NAT doubles latency time KVM
also adds to latency time compared to its native system
implementations. The difference in network performance
is due to the virtualized network devices implemented by
virtual machines and Docker.
D. Disk Performance
Throughput is used to measure the efficiency of the disk
operations. As duly noted for sequential read and
sequential write operations, the Docker and KVM add
very little overhead when compared to native, but there is
a lot of performance variance in KVM's case cause of
suspected bottleneck in fiber channel.
For random read and random write operations, Docker has
no overhead, but KVM's performance decreases

significantly. Disk performance is affected by the I/O
scheduler used by the system.
E. Application Performance-Redis
Redis (REmote Dictionary server) is the most popular
NoSQL database used in containers and is open source. It is
an in-memory data store and key-valued database.
Generalized operations between client and server where
performance issues are noticed only in the form of network
latency. This arises due to the large number of concurrent
requests the client makes simultaneously to the server. Redis
is widely used in the cloud environment. In this test scenario,
the number of requests made simultaneously is ten times the
number of clients. Redis is extensively used to test the
networking and memory subsystem.
The native environment can handle a large number of clients
connected to the networking subsystem, so we can easily
scale the number of clients connected. The bottleneck in
performance is visible at the CPU because Redis is a single-
threaded application; it increases the latency of the CPU.
Docker's performance is virtually similar to that of native
except with NAT enabled, latency increases as new
connections are added. KVM's performance is lower initially
but approaches that of native systems as concurrency
increases. Hence we can draw the conclusion that Redis
application needs to be concurrent to be fully utilized.
F. Application Performance-MySQL
MySQL is a relational database that can be used to stress the
memory, file system, networking and inter-process
communication subsystems. Various open source
benchmarks can be run against the database to test the
transaction throughput of the system. Various configurations
were used to test MySQL, running under the native system,
MySQL under Docker that uses the host's networking,
MySQL under Docker using NAT for networking and KVM
running MySQL.
Throughput is the number of transactions per second
increase until the peak point where it becomes stable.
Docker has similar performance to that of native systems;
KVM has higher overhead compared to all other
configurations used to demonstrate. The Docker's layered
file system also introduces overhead because of the I/O
request getting redirected through various layers. The
latency of the system increases with the load, but Docker
does not show this observation when compared to other to
other setups because of lower throughput at lower
concurrency rates. When it comes to CPU peak utilization
native system is able to achieve a higher rate as compared to
Dockers which show that Dockers have a small but
significant impact. [4]-[16]
CPU utilization measured in throughput is minimal for the
same amount of CPU used for Docker and Docker with
NAT, but the latency for Docker is higher for lower values
of concurrency. This is due to mutex contention which
prevents MySQL from fully utilizing the CPU for
configurations. [17]-[26]
G. Discussion
We can analyze the results, as expected for the given
environment containers and VMs generate a negligible
amount of overhead when it comes to CPU and memory
usage. When it comes to network usage both Docker and
KVM add to the latency and hence need to tune and are of
no good with their standard implementations. Most of the
overhead added are because of the bottlenecks in end
communication devices. When it comes to disk performance
Docker and native are similar to each other but KVM sees a
considerable downfall in performance since each operation
has to pass through AUFS.
The major limitation of virtualization technology is
bottlenecks in I/O intensive applications. The idea is to get
very close to native systems when it comes to I/O processing
and decrease as many layers as possible in virtual machines
and containers. Container-based virtualization system has
better CPU and I/O performance because of its ability to
release used unused and resources and work in isolation.
CONCLUSIONS AND FUTURE WORK
This study analyzes the behavior of various well-known
parameters, under different workload conditions to give us a
comprehensive performance comparison between Containers
and Virtual Machines. As the survey clearly states, that no
one, technology can be substituted by another right away,
each having its own merits and demerits and each
technology can be used efficiently with proper configuration
suiting your particular needs.
Some of the areas which are not researched to its full extent
are:
• Performance analysis in isolation when more than
more than one workload is executing on the server.
• The performance of different kinds of workloads such
as data intensive workloads which are more I/O bound
in nature.
• Performance trade-off between live migration and
restarting virtual machines and containers.
REFERENCES
[1] Xavier, M.G., Neves, M.V. and De Rose, C.A.F., 2014,
February. A performance comparison of container-
based virtualization systems for mapreduce clusters.
In 2014 22nd Euromicro International Conference on
Parallel, Distributed, and Network-Based
Processing (pp. 299-306). IEEE.

[2] Soltesz, S., Pötzl, H., Fiuczynski, M.E., Bavier, A. and
Peterson, L., 2007, March. Container-based operating
system virtualization: a scalable, high-performance
alternative to hypervisors. In ACM SIGOPS Operating
Systems Review (Vol. 41, No. 3, pp. 275-287). ACM.
[3] Vaughan-Nichols, S.J., 2006. New approach to
virtualization is a lightweight.Computer, 39(11), pp.12-
14.
[4] Raval, K.S., Suryawanshi, R.S., Naveenkumar, J. and
Thakore, D.M., 2011. The Anatomy of a Small-Scale
Document Search Engine Tool: Incorporating a new
Ranking Algorithm. International Journal of
Engineering Science and Technology, 1(3), pp.5802-
5808.
[5] Archana, R.C., Naveenkumar, J. and Patil, S.H., 2011.
Iris Image Pre-Processing And Minutiae Points
Extraction. International Journal of Computer Science
and Information Security, 9(6), p.171.
[6] Jayakumar, M.N., Zaeimfar, M.F., Joshi, M.M. and
Joshi, S.D., 2014. INTERNATIONAL JOURNAL OF
COMPUTER ENGINEERING & TECHNOLOGY
(IJCET). Journal Impact Factor, 5(1), pp.46-51.
[7] Naveenkumar, J. and Joshi, S.D., 2015. Evaluation of
Active Storage System Realized through MobilityRPC.
[8] Jayakumar, D.T. and Naveenkumar, R., 2012.
SDjoshi,“. International Journal of Advanced Research
in Computer Science and Software Engineering,” Int.
J, 2(9), pp.62-70.
[9] Jayakumar, N., Singh, S., Patil, S.H. and Joshi, S.D.,
Evaluation Parameters of Infrastructure Resources
Required for Integrating Parallel Computing Algorithm
and Distributed File System.
[10]Jayakumar, N., Bhardwaj, T., Pant, K., Joshi, S.D. and
Patil, S.H., A Holistic Approach for Performance
Analysis of Embedded Storage Array.
[11]Naveenkumar, J., Makwana, R., Joshi, S.D. and
Thakore, D.M., 2015. OFFLOADING
COMPRESSION AND DECOMPRESSION LOGIC
CLOSER TO VIDEO FILES USING REMOTE
PROCEDURE CALL. Journal Impact Factor, 6(3),
pp.37-45.
[12]Naveenkumar, J., Makwana, R., Joshi, S.D. and
Thakore, D.M., 2015. Performance Impact Analysis of
Application Implemented on Active Storage
Framework. International Journal, 5(2).
[13]Salunkhe, R., Kadam, A.D., Jayakumar, N. and
Thakore, D., In Search of a Scalable File System State-
of-the-art File Systems Review and Map view of new
Scalable File system.
[14]Salunkhe, R., Kadam, A.D., Jayakumar, N. and Joshi,
S., Luster A Scalable Architecture File System: A
Research Implementation on Active Storage Array
Framework with Luster file System.
[15]Jayakumar, N., Reducts and Discretization Concepts,
tools for Predicting Student’s Performance.
[16]Jayakumar, M.N., Zaeimfar, M.F., Joshi, M.M. and
Joshi, S.D., 2014. INTERNATIONAL JOURNAL OF
COMPUTER ENGINEERING & TECHNOLOGY
(IJCET). Journal Impact Factor, 5(1), pp.46-51.
[17]Kumar, N., Angral, S. and Sharma, R., 2014.
Integrating Intrusion Detection System with Network
Monitoring. International Journal of Scientific and
Research Publications, 4, pp.1-4.
[18]Namdeo, J. and Jayakumar, N., 2014. Predicting
Students Performance Using Data Mining Technique
with Rough Set Theory Concepts. International
Journal, 2(2).
[19]Naveenkumar, J., Keyword Extraction through
Applying Rules of Association and Threshold
Values. International Journal of Advanced Research in
Computer and Communication Engineering
(IJARCCE), ISSN, pp.2278-1021.
[20]Kakamanshadi, G., Naveenkumar, J. and Patil, S.H.,
2011. A Method to Find Shortest Reliable Path by
Hardware Testing and Software
Implementation. International Journal of Engineering
Science and Technology (IJEST), ISSN, pp.0975-5462.
[21]Naveenkumar, J. and Raval, K.S., Clouds Explained
Using Use-Case Scenarios.
[22]Naveenkumar J, S.D.J., 2015. Evaluation of Active
Storage System Realized Through Hadoop.
International Journal of Computer Science and Mobile
Computing, 4(12), pp.67–73.
[23]Rishikesh Salunkhe, N.J., 2016. Query Bound
Application Offloading: Approach Towards Increase
Performance of Big Data Computing. Journal of
Emerging Technologies and Innovative Research, 3(6),
pp.188–191.
[24]Sagar S lad s d joshi, N.J., 2015. Comparison study on
Hadoop’s HDFS with Lustre File System. International
Journal of Scientific Engineering and Applied Science,
1(8), pp.491–494.
[25]Salunkhe, R. et al., 2015. In Search of a Scalable File
System State-of-the-art File Systems Review and Map
view of new Scalable File system. International
Conference on Electrical, Electronics, and Optimization
Techniques (ICEEOT) - 2016. pp. 1–8.
[26]Naveenkumar, J. and Raval, K.S., Clouds Explained
Using Use-Case Scenarios. INDIACom-2011
Computing for Nation Development, 2011/3.
[27]N. Jayakumar, “Reducts and Discretization Concepts,
tools for Predicting Student’s Performance,” Int. J.
Eng. Sci. Innov. Technol., vol. 3, no. 2, pp. 7–15, 2014.
[28]BVDUCOE, B.B., 2011. Iris Image Pre-Processing and
Minutiae Points Extraction. International Journal of
Computer Science & Information Security.
[29]P. D. S. D. J. Naveenkumar J, “Evaluation of Active
Storage System Realized through MobilityRPC,” Int. J.
Innov. Res. Comput. Commun. Eng., vol. 3, no. 11, pp.
11329–11335, 2015.

[30]Anjali Nigam and Vineet Singh, “Securing Data
Transmission in Cloud using Encryption Algorithms”,
International Journal of Computer Sciences and
Engineering Research Volume 4, Issue 6, June 2016.
[31]Pritika Goel, “An Improved Load Balancing Technique
in Weighted Clustering Algorithm”, International
Journal of Computer Sciences and Engineering
Research Volume 4, Issue 6, June 2016.
[32]Anish Babu S , Hareesh M J , John Paul Martin , Sijo
Cherian and Yedhu Sastri.”System Performance
evaluation of Paravirtualization, Container
virtualization and Full virtualization using Xen,
OpenVZ and XenServer” Fourth International
Conference on Advances in Computing and
Communications.pp 247-250 IEEE 2014
AUTHORS PROFILE
Prashant Ramchandra Desai has received his B.E(2014)
in computer engineering from Karmeer Bhaurao Patil
College of Engineering , Satara. He is currently pursuing
M.Tech in computer engineering from Bharati Vidyapeeth
College Of Engineering ,Pune. His area of interest include
virtualization,Big Data and storage area network.

A Survey of Performance Comparison between Virtual Machines and Containers

More Related Content

What's hot (20)

Similar to A Survey of Performance Comparison between Virtual Machines and Containers (20)

Recently uploaded (20)

A Survey of Performance Comparison between Virtual Machines and Containers