Implementation and Operation of User Defined Network on IaaS Clouds
using Layer 3 Overlay
Ryo Nakamura, Yuji Sekiya and Hiroshi Esaki
The University of Tokyo, Bunkyoku, Tokyo, Japan
Keywords:
IaaS, Cloud Network, Layer 3 Overlay Network, LISP, VXLAN.
Abstract:
Server virtualization technology achieves ”Infrastructure as a Service (IaaS)” model services. Anyone can use
virtual server resource easily without preparing computer hardware, thus IaaS environments provides multiple
operational benefits to users. Therefore, today various types of users, both of personal users and enterprises
users, deploy virtual resources on IaaS cloud. On the other hand, many enterprise users still have their own
environments and operate them by themselves (i.e, on-premise environments). When they migrate the services
from physical facilities to IaaS clouds, they have two problems. The first problem is that services are bound
to IP addresses. The second problem is lack of flexibility in network design and management on IaaS clouds.
To solve these problems, we propose an operation model to migrate on-premise services to IaaS clouds using
LISP and VXLAN. Using our proposed method, users can migrate IP addresses and services among IaaS
clouds. In this paper, we proposed new migration method of IP address among IaaS clouds and implemented
the method using LISP and VXLAN software on Linux systems. Then we evaluate performance of our system
and present a operation on actual IaaS clouds.
1 INTRODUCTION
Many virtual machines (VM) can work on sin-
gle physical machine by virtualization technologies.
Then, there are some service providers that provide
virtual machine environments that are called Infras-
tructure as a Service (IaaS) to users. IaaS providers
provide VM to users, and users pay fee depending on
resources of VM. Today, these IaaS services are used
by many consumers and enterprises. IaaS service of-
fers many operational benefits such as reducing the
costs of facilities and fault tolerance. Hence, enter-
prises tend to adopt IaaS clouds for benefits.
However, migrating existing services that are al-
ready operated in on-premise environment to IaaS
clouds. The first problem is flexibility of IP ad-
dresses. Almost services are bound to IP address,
e.g., www and IP address based access control (ACL).
IP addresses that are used for service servers can
be changed easily using the DNS system. However,
changing IP addresses that are embedded into config-
uration such as ACL requires expensive operational
costs. An IP address of IaaS provider is assigned to a
VM. Thus, if a user want to migrate their services to
IaaS cloud, IP addresses must be changed.
Besides, the layer 2 segment design and its con-
struction on IaaS cloud should have also flexibility.
In on-premise environments, layer 2 network seg-
ments are separated by policies and usages. This
sort of network separations should be possible in IaaS
cloud. Several IaaS clouds and researches provide
functions to design flexible layer 2 network (Amazon,
2012)(Benson et al., 2011)(Cabuk et al., 2007). How-
ever, this flexibility is sometimes restrected by con-
tracts and basically design of IaaS clouds.
The second problem is redundancy. If a user mi-
grate their systems to IaaS cloud, the services are
not so redundant. A service provider of IaaS cloud
may construct their system highly redundant, how-
ever, there are some possibility of failures that de-
ploying services on single IaaS cloud rather than on
multiple IaaS clouds. Thus, the service which is mi-
grated from on-premise environment to IaaS clouds
must be constructed among several IaaS clouds.
In this paper, we present an approach that can iso-
late user networks from underlay IaaS clouds using
layer 3 overlay networks. By using layer 3 overlay,
user can define networks among several IaaS clouds
with their own IP address prefixes. In order to achieve
this isolation, we implemented Locator ID Separa-
tion Protocol and Virtual eXtensible LAN on Linux
system. Finally, we evaluated implementations and
present the operational model of our system.
366
Nakamura R., Sekiya Y. and Esaki H..
Implementation and Operation of User Defined Network on IaaS Clouds using Layer 3 Overlay.
DOI: 10.5220/0004373803660369
In Proceedings of the 3rd International Conference on Cloud Computing and Services Science (CLOSER-2013), pages 366-369
ISBN: 978-989-8565-52-5
Copyright
c
2013 SCITEPRESS (Science and Technology Publications, Lda.)
2 APPROACH
In this section, we describe the requirements to mi-
grate systems from on-premise environments to IaaS
clouds, and the overview of proposed approach.
2.1 Requirements
Requirements to achieve migration from on-premise
environments to IaaS clouds are shown below.
Migrating IP address prefix from on-premise net-
works to IaaS clouds.
Achieving flexible layer 2 network design on IaaS
cloud.
Constructing user networks among several IaaS
clouds.
In this research, to realize these requirements, we
introduce layer 3 overlay networks. If using specific
functions that depend on each IaaS cloud, migration
can be realized. However, on assumption that a user
network is constructed among several IaaS clouds,
the network can not depend on these specific func-
tions. Then, in our proposed method, layer 3 forms
boundary between user network and IaaS clouds. All
of functions to construct network are provided by
over layer 3. Figure 1 shows the overview of pro-
posed model. A user defined network is isolated from
IaaS clouds through integration of LISP (Locator/ID
Separation Protocol) and VXLAN (Virtual eXtensible
LAN). The layer 3 topology in this system is defined
by LISP that is IP over IP overlay protocol, and the
layer 2 segment is defined by VXLAN that is ethernet
over IP overlay protocol.
LISP (Farinacci et al., 2012) manages two aspects
of the IP address, addressing and locator, in isola-
tion. LISP manages pair of the prefix (EID), and
the Routing Locator (RLOC) which accommodates
the EID prefix. Then, LISP overlay routing table
is constructed of this pair. By this approach, LISP
achieves multipoint to multipoint layer 3 overlay IP
routing. VXLAN (M.Mahalingam et al., 2011) is eth-
ernet over IP overlay architecture. VXLAN Tunnel
End Point (VTEP), is a gateway between VXLAN
network and non-VXLAN network. VTEPs encap-
sulate broadcast, unknown unicast and multicast eth-
ernet frame in IP multicast, and well-known unicast
in IP unicast. Thus, VTEPs construct overlay for-
warding database like ethernet switching. Therefore,
VXLAN constructs the Forwarding Information Base
for multipoint Ethernet over IP overlay network.
User Defined Network
IaaS Clouds
IP layer
LISP
VXLAN
Virtual
LISP Router
Virtual
LISP Router
VM
Figure 1: Isolation of user defined network and underlay
IaaS clouds using layer 3 overlay networks.
2.2 Design Overview of proposed
Approach
The proposed method achieves to isolate user net-
works from IaaS clouds. By this method, users can
design and construct networks using own address pre-
fixes freely without changes to underlay IaaS clouds,
then users can migrate on-premise systems. LISP is
used to achieve address prefix migration. LISP real-
izes network mobility through separation of EID and
RLOC. Moreover, LISP avoids triangular routing be-
tween LISP sites by exchanging overlay routing table,
then, applying a LISP site to a IaaS cloud enables IP
address prefix migration and route optimization be-
tween IaaS clouds. Simultaneously, VXLAN is used
to define layer 2 segments on IaaS clouds. VXLAN
realize multi tenancy without changes to outer net-
works. VXLAN is originally assumed that it is used
into underlay IaaS systems to separate user networks.
However, By VMs directly terminate VXLAN, VM
users can design and construct layer 2 segments freely
without specific functions provided by IaaS clouds.
Figure 2 shows the overview of constructed sys-
tem. User defined network is constructed among sev-
eral IaaS clouds. Layer 2 segments in each IaaS
cloud are constructed by VXLAN. Moreover, each
layer 2 segment is accommodated by LISP router. An
IP address prefix owned by user is associated with
a VXLAN segment. In IaaS clouds, an IP address
of IaaS provider is assigned to a VM. Then, A LISP
router uses IaaS provider’s address that is assigned as
a locator address, and uses the prefix that is assigned
to VXLAN interface as EID prefix. A LISP router ad-
vertises this prefix as EID prefix to the LISP overlay
network. Thus, user defined network is constructed
among several IaaS clouds over layer 3 networks.
ImplementationandOperationofUserDefinedNetworkonIaaSCloudsusingLayer3Overlay
367
IaaS Cloud #2IaaS Cloud #1
Internet
on-premises
LISP Router (XTR)
Virtual Machine
VXLAN Segment
VNI 11
VNI 12
VNI 21
VNI 22
PXTR
Map Server
XTR XTR XTR XTR
Figure 2: The overview of proposed system.
2.3 Implementation and Operation
We implemented LISP and VXLAN in Linux user
space, and they can be configured through Vyatta.
In proposed system, LISP router including VXLAN
function is deployed as a virtual router into IaaS
clouds. Thus, we implemented them as vyatta exten-
sion. Vyatta, is open source routing suite, provides
CLI to configure many open source software for net-
working. Furthermore, vyatta is provided VM image.
By implementing LISP and VXLAN as vyatta exten-
sion, anyone can use proposed system in cloud envi-
ronments with some networking functions that vyatta
provides. These implementations, that are called lixy
and vxlan-vyatta, are published.
We deployed proposed system into the two actual
IaaS clouds. The first IaaS cloud is WIDE Cloud
(WIDE Project, 2012) that is academic IaaS environ-
ment. WIDE Cloud is constructed from over 20 HVs
around Japan, and some HVs are located in foreign
countries. It is controlled by WIDE Cloud Controller
that is developed by WIDE project. All of HVs con-
nect to a large scale layer 2 segment that accommo-
dates VMs using vlan and vpn tunnels. And another
cloud is constructed independently at Keio University
using WIDE Cloud Controller too.
We deployed Vyatta with LISP and VXLAN ex-
tensions to these IaaS clouds. We extend LISP beta
network (Lisp, 2012) to these clouds using proposed
system. The wide-xtr that is located in WIDE back-
bone registers assigned prefix to the map server of
beta network periodically. Moreover, two XTRs that
are located in two IaaS clouds, register a part of as-
signed prefix and own address as a locator address
a map server that is located in WIDE backbone. In
this way, we separate assigned prefix into two small
prefixes that are used on two clouds. Moreover,
layer 2 segments in each cloud are constructed from
VXLAN. Thus, we use LISP beta network prefix
without any changes to underlay IaaS clouds using
proposed method.
3 PERFORMANCE EVALUATION
In our proposed method, all of traffics from VMs go
through layer 3 overlay. Layer 3 overlay causes de-
grading bandwidth because MTU is decreased due to
layer 3 encapsulation. Thus, this point is obvious bot-
tleneck of this system. This section shows the perfor-
mance evaluation result of LISP and VXLAN that we
implemented.
3.1 Performance Degradation due to
Fragmentation
First of all, we evaluated implementations to clarify
the performance degradation due to fragmentation.
Therefore, we tested the throughput of implementa-
tions changing message size of test traffic that is UDP
payload length from 50 bytes to 1500 bytes with 50
bytes step on a physical node environment.
Figure 3 shows throughput of LISP and VXLAN
on physical nodes. With the result of LISP test,
when message size 1450 bytes message, throughput
was 952Mbps, and when the message size was 1500
bytes, throughput was 917Mbps. LISP encapsulates
IP packets in LISP header, so that, when message size
is over 1422 bytes, packets are fragmented by LISP
router. With this result, performance degradation due
to fragmentation is about 3.7%. With the result of
VXLAN test, when message size was 1400 bytes,
throughput was 924Mbps, and when message size
was over 1400 bytes, throughput degraded. VXLAN
encapsulates ether frames in VXLAN header, so that,
when message size is over 1408 bytes, packets are
fragmented by VXLAN node. With this result, per-
formance degradation due to fragmentation is about
25%.
3.2 Performance on Virtual
Environment
Implementations are assumed to be used as VM on
IaaS clouds, and so we evaluated performance on a
virtual environment. With the result of above experi-
ments, if the message size is over the threshold, per-
formance is degraded. Therefore, we evaluated the
performance on a virtual environment with the largest
message size that packets were not fragmented.
Figure 4 shows performances of LISP and
VXLAN on a virtual environment using kernel vir-
tual machine (KVM). The x axis means the direction
of test traffic. The performance of LISP implementa-
tion is degraded to less than a half of the performance
of the physical node environment. In addition, the
performance of virtual node to physical node is 25%
CLOSER2013-3rdInternationalConferenceonCloudComputingandServicesScience
368
0
100
200
300
400
500
600
700
800
900
1000
0 200 400 600 800 1000 1200 1400
throughput [Mbps]
message size [Byte]
Throughput of LISP and VXLAN
LISP
VXLAN
Figure 3: UDP throughput measurement of LISP and
VXLAN implementations.
lower than physical to virtual. When packets come
from edge network of LISP router, router has to look
up LISP map table and encapsulates packets. There-
fore, this result shows that looking up LISP map table
and encapsulation overhead causes of 25% degrada-
tion of performance on a virtual environment. The
performance of VXLAN implementation is also de-
graded to less than a half of the performance of phys-
ical. In addition, virtual to physical case performance
is half of physical to virtual case. When the test traffic
was transmitted from VM, vCPU usage on VM was
full.
0
200
400
600
800
1000
Physical to Physical Physical to Vritual Virtual to Physical
throughput [Mbps]
Throughput of LISP and VXLAN
LISP
VXLAN
Figure 4: UDP throughput of LISP with 1422 bytes mes-
sage, and UDP throughput of VXLAN with 1408bytes mes-
sage.
3.3 Consideration
With these results, performances of each implemen-
tation are over 900Mbps on a physical environment.
However, on the virtual environment, performances
are degraded to less than a half of the performance of
physical. Thus, using this implementation, enterprise
systems without large scale services like office branch
systems can be migrated to IaaS cloud. However, per-
formancedegradation in a virtual environmentcan not
be ignored.
4 CONCLUSIONS
Also we deployed them in our actual IaaS cloud and
evaluated the implementations.The evaluation results
show that the implementations have some perfor-
mance degradation of network communications, how-
ever, it is acceptable degree of degradation for small
enterprise systems. Although current performance
of our systems is not feasible to accommodate large
scale systems that require high performance. More-
over, flexible layer 2 design in IaaS clouds is achieved
by VXLAN. VXLAN requires IP multicast to under-
lay IaaS system and IP multicast does not work glob-
ally today. As a result, migratable layer 2 domain is
limited in a IaaS cloud. Therefore, our future works
are improving packet transfer performance on vir-
tual environments, and design a more flexible method
for constructing layer 2 network among several IaaS
clouds.
REFERENCES
Amazon (2012). Amazon EC2. http://aws.amazon.com/
ec2.
Benson, T., Akella, A., Shaikh, A., and Sahu, S. (2011).
CloudNaaS: a cloud networking platform for enter-
prise applications. In Proceedings of the 2nd ACM
Symposium on Cloud Computing, SOCC ’11, pages
8:1–8:13, New York, NY, USA. ACM.
Cabuk, S., Dalton, C. I., Ramasamy, H., and Schunter, M.
(2007). Towards automated provisioning of secure
virtualized networks. In Proceedings of the 14th ACM
conference on Computer and communications secu-
rity, CCS ’07, pages 235–245, New York, NY, USA.
ACM.
Farinacci, D., Fuller, V., Meyer, D., and Lewis, D. (2012).
Locator/ID separation protocol (LISP) draft-ietf-lisp-
23. Internet Draft, IETF Network Working Group.
Lisp (2012). LISP beta network. http://www.lisp4.net/beta-
network.
M.Mahalingam, D.Dutt, K.Duda, P.Agarwal, Kreeger, L.,
Sridhar, T., M.Bursell, and C.Wright (2011). draft-
mahalingam-dutt-dcops-vxlan-00.txt. Internet Draft,
IETF.
WIDE Project (2012). WIDE Cloud. https://
wcc.wide.ad.jp/.
ImplementationandOperationofUserDefinedNetworkonIaaSCloudsusingLayer3Overlay
369