Service Forwarding Label for Network Function Virtualization and
Application-centric Traffic Steering
Changcheng Huang
1
and Jiafeng Zhu
2
1
Department of Systems and Computer Engineering, Carleyton University, Ottawa, Ontario, Canada
2
R&D Centre, Huawei Technologies Inc., Santa Clara, U.S.A.
Keywords: Software Defined Network, Network Function Virtualization, Traffic Steering, Service Chaining, VXLAN.
Abstract: New services such as network virtualization, service chaining, and application-centric traffic steering bring
new opportunities for network providers and service providers. Existing Internet architecture is suffering
from the so-called ossification problem, which makes it difficult to support new applications. Software
Defined Network (SDN) is becoming a new paradigm for both WAN and enterprise networks. By
separating control and data planes, SDN allows network providers to sell new services without changing
their physical switches. While SDN is poised to support new services, it is still in development stage.
Mechanisms supporting new services are still missing. In this paper, we propose a new application driven
mechanism called Service Forwarding Label (SFL) which can be used as a uniform label to differentiate
various services and forward packets based on different service requirements. Compared with existing
solutions such as VXLAN, our SFL approach introduces less overhead and supports more new services.
1 INTRODUCTION
Research on Software Defined Network (SDN)
started nearly ten years ago. Most of earlier SDN
networks were designed for datacenter infrastructure
(ONF, 2015, Banikazemi et al., 2013) where
flexibility and scalability are critical. With tens of
thousands commodity switches and servers,
datacenter presents a serious challenge to network
architecture. SDN solves the problem by separating
control plan from data plane. With relatively an
independent control plane, SDN does not need to be
run on proprietary equipment. Instead, the control
plane can be realized with a large number of regular
servers. Through adopting a centralized structure,
these servers can be efficiently facilitated by
datacenter with on-demand service capacity.
In recent years, SDN has found applications in
other areas. Google, for example, built a large scale
SDN-based WAN network (Jain et al., 2013) that
has attracted attentions worldwide. Large carriers are
looking at the possibility of upgrading their network
infrastructures with SDN architecture. Issues such as
scalability have been widely studied. Early
benchmarks on NOX (Tavakkoli et al., 2009) (the
first SDN controller) showed it could only handle
30,000 flow initiations per second while maintaining
a sub-10 ms flow install time. Recent works have
shown that SDN scalability can be extended by
using multicore systems (Tootoonchian et al., 2012)
or deploying multiple OpenFlow controllers (OFCs)
(Tootoonchian and Ganjali, 2010, Yeganeh et al.,
2012).
With a separated control plane, SDN has the
potential to enable new services. One of the primary
new services that have been envisioned is the
network virtualization service, which allows
physical network provider to sell different virtual
networks to different service network providers.
Each service network provider can use its virtual
network just like the way it uses its own private
network while sharing underlying physical network
resources with other service network providers. The
physical network provider, on the other hand, can
enjoy new revenue growth through selling virtual
networks with different granularities.
Service chaining is another area that SDN can
play an important role. Traditionally a service chain
consists of a set of dedicated network service boxes
such as firewall, load balancers, and application
delivery controllers that are concatenated together to
support a specific application (Jacobs, 2015). With a
new service, new devices must be installed and
interconnected in certain order. This can be a very
27
Huang C. and Zhu J..
Service Forwarding Label for Network Function Virtualization and Application-centric Traffic Steering.
DOI: 10.5220/0005528800270034
In Proceedings of the 6th International Conference on Data Communication Networking (DCNET-2015), pages 27-34
ISBN: 978-989-758-112-0
Copyright
c
2015 SCITEPRESS (Science and Technology Publications, Lda.)
complex, time-consuming, and error-prone process,
requiring careful planning of topology changes and
network outages and incurring high OPEX. This
situation is exacerbated when a tenant requires
different service sequences for different traffic flows
or when multiple tenants share the same datacenter
network. Through network function virtualization
(NFV) service, SDN can dynamically create a
virtual environment for a specific service chain and
eliminate the complex hardware and labor work.
While possessing great potential, SDN is still in
development stage. The actual mechanisms for
supporting new services are still not defined.
Meanwhile competitive technologies are emerging.
A good example is VXLAN (Mahalingam, 2015),
which is designed with two major objectives: a) to
extend Virtual LAN (VLAN) service from a local
area network to a network that can span an IP
network; b) to address the shortage of VLAN ID. It
advocates the architecture that overlays virtualized
Layer 2 networks over Layer 3 networks. A 24-bit
VXLAN network identifier is added to allow
identifications of more than 16M different VLANs.
While VXLAN can be implemented relatively easily
with existing technologies, it has several major
disadvantages that make it difficult to meet the long
term challenges. First, VXLAN is an extension of
Layer 2 VLAN. Its main objective is to extend
VLAN service rather than to support arbitrary new
services. Therefore, it inherits characteristics of
Layer 2 technologies which make it far away from
application-centric concept. It is hard to generalize
the VLANs supported by VXLAN to other services
such as service chaining. Second, it is hard to
support resource allocation and optimization for
virtual networks due to its distributed nature. Third,
the overhead for overlaying Layer 2 networks over
Layer 3 networks is overwhelming. Fourth, It is hard
to imagine how to support recursive services
(discussed more in the next section) with VXLAN.
OpenADN (Paul et al., 2013) is another option
that has been proposed recently. With OpenADN,
two new labels are added, one at Layer 3.5 and
another at Layer 4.5. The label at Layer 3.5 is used
to forward packets and the label at Layer 4.5 is used
to interconnect two Layer 3.5 labels after passing a
middle-box. Although OpenADN can solve some of
the issues with VXLAN, it is still too complex to
implement. Furthermore, issues such as recursive
services still cannot be supported.
In this paper, we will propose a new Service
Forwarding Label (SFL) mechanism at Layer 5 that
can be used to identify a service instance for packet
forwarding so that NFV and application-centric
traffic steering can be realized with very little
overhead. Our label creation process is application
driven allowing close integration of applications and
forwarding functions. Combining with the power of
SDN controller, SFL can support resource allocation
and optimization on a per service instance basis.
The paper is organized as follows. In Section 2,
we will use NFV as a new service example to
discuss more detail about the issues to be resolved.
In Section 3, we will propose our SFL scheme. This
will be followed by several use case scenarios in
Section 4. We will finish the paper with some
concluding remarks in Section 5.
2 CHARACTERISTICS OF NEW
SERVICES
In this section, we will discuss in detail the
characteristics and challenges of new services under
SDN context. We will use NFV as a key example for
the new services SDN will provide. In specific, we
will investigate the issues related to recursive
network virtualization, which illustrates the
characteristics of recursive services in more general
cases. The reason we choose NFV as an example is
because it can represent a significant portion of new
services currently being envisioned by network
providers and service providers.
The initial SDN adopters, both vendors and
network providers, have focused on some
fundamental issues related to separating control and
data planes. The benefits of creating new services
have not been fully explored. A good example is the
Google SDN WAN project mentioned earlier (Jain
et al., 2013). Before the SDN project, the WAN
Google had been using for interconnecting their
datacenters had been managed using traditional
approach which was slow due to manual
provisioning process. On average, the utilization of
Google’s WAN at that time was 20 to 30%. Through
multiple years of efforts, Google has upgraded its
WAN with SDN technology. The initial results are
very encouraging. Utilization has been increased to
around 70 to 90%. In some cases, 100% utilization
has been achieved. The key enablers of this
improvement are the SDN dynamic flow creation
capability and a more sophisticated optimization
approach based on Max-min fairness. SDN allows
Google to do centralized traffic engineering that can
balance load distribution across their entire WAN
with sophisticated optimization algorithms. Large
amount of elastic traffic also helps increase the
DCNET2015-InternationalConferenceonDataCommunicationNetworking
28
utilization to 100%.
However, it should be noted that Google’s WAN
is a special case where Google is the user, service
provider, as well as network provider, i.e., Google
provides service to itself on its own network. This
nature allows Google to do global optimization
easily. For example, because Google can control the
traffic carried by the WAN, operators can decide
when and how long the elastic traffic will be
buffered. Also because Google owns both service
network and underlying WAN network, operators
can have a global view of the network and therefore
optimize network usage globally.
In a real world, it is quite common that users,
service providers, and network providers are
separate entities. They may all have their own
objectives which may conflict with each other. Take
the example of enterprise network. With the fast
growth of datacenters, enterprises are becoming
increasingly interested in outsourcing their
enterprise networks to datacenters. Under this
scenario, the owner of the physical network now is
the owner of the datacenter, such as Amazon. The
service network providers are the enterprises who
outsource their enterprise networks to the
datacenters. Therefore a service provider is
independent from the network provider as well as
independent from other service network providers
who share the same physical network. Recognizing
this need, the pioneers of SDN advocate a layer
called FlowVisor (Sherwood et al., 2009) which
plays the same role as the hypervisor for virtual
machines. FlowVisor allows a physical network
provider to partition its physical network into slices
for various service network providers. Each service
network provider can virtually own one slice which
has its own virtual network topology and related
resources generated through a topology abstraction
process. The service network provider can then
optimize its usage of the slice which it owns. A
physical network provider can sell network
virtualization service to service providers with
different granularities selling at different prices
(Drutskoy et al., 2013). This will change the
situation that physical network providers today can
only sell pipes and equip physical network providers
a new venue for revenue growth.
The SDN architecture for virtual network
service is shown in Fig. 1. As shown in the figure,
each entity has its own OpenFlow Controller (OFC).
The OFC of a service provider is in charge of
routing user flows and optimizing resource usage
within its own virtual topology. The OFC of the
network provider will execute the topology
abstraction process through a topology virtualizer
based on a global topology map of the underlying
physical network.
Figure 1: The architecture of virtual network service.
It is envisioned in SDN architecture that a service
provider can further partition its virtual network and
sell virtual network service to other service
providers. This kind of recursive network
virtualization is illustrated in Fig. 2. As shown in the
figure, a network provider sells a virtual network to
a tier-1 service provider which in turn sells virtual
network to a tier-2 service provider. This process
can continue several iterations making service
providers become network providers in a recursive
manner. This kind of recursive services is becoming
popular today. For example, Amazon sells virtual
server service to Netflix and Netflix in turn sells
video streaming service to its clients.
While NFV is a powerful concept, it also faces
several challenges that need to be addressed. These
include overlapping address space, middle-box
traversal, SDN network migration, multiple entities,
etc. We will briefly discuss these problems in the
following paragraphs.
As we mentioned earlier, each service provider
treats its virtual network as its private property. They
decide how to assign VLAN IDs and IP addresses
within their own virtual networks without consulting
each other. Most likely they will use private IP
address blocks to avoid the shortage of address
space. This creates a high possibility of overlapped
address spaces being used by multiple virtual
networks sharing the same physical network. Using
Fig.1 as an example, if the two virtual networks
owned by Tenant A and Tenant B respectively use
the same address space, switches in the physical
ServiceForwardingLabelforNetworkFunctionVirtualizationandApplication-centricTrafficSteering
29
network will not be able to differentiate packets for
the two virtual networks. Clearly some mechanism
is required to identify packets belonging to different
virtual networks.
Figure 2: Illustration of recursive network virtualization.
In a service chain application, packets need to
traverse multiple middle-boxes before reaching their
destinations. Each middle-box such as firewall or
load balancer provides functions that may require
processing packets up to the transport layer.
Therefore the header information including all layers
up to the transport layer will not be maintained end-
to-end. This will make it difficult for physical
switches to steer traffic to the correct next middle-
box on the path to the destination. This issue
suggests that an end-to-end ID above the transport
layer is required to identify traffic belonging to the
same service chain.
SDN as a new architecture is attracting many
vendors and network providers. Given the size of the
Internet today, the evolution towards SDN networks
will likely be a long and slow process. It is easy to
see that SDN networks will be deployed initially as
islands within Internet. Datacenters, for example, are
the first place where SDN will be adopted early.
When datacenters provide virtual network services
to clients, most of the clients are unlikely to be
attached directly to the datacenters. Therefore,
traffic from these clients has to travel through legacy
IP networks before they can reach SDN based
datacenters. Any techniques used to identify virtual
networks in the lower three layers may be lost when
traffic reaches the datacenters. Inter-datacenter
traffic may also span multiple legacy IP networks
making it difficult to support virtual networks across
multiple datacenters.
Recursive network virtualization raises another
new challenge. As we mentioned earlier, each virtual
network may be owned by a different entity. This
leads to the situation that multiple entities may be
involved in recursive network virtualization. Each
entity has the right to control the resources within its
own virtual network and decide how to route traffic
based on its own policy. Orchestration among
multiple OFCs is calling for simple mechanisms that
help streamline business relationship among
different entities.
In summary, the above mentioned challenges
must be addressed before the benefits of SDN can be
fully realized. In the next section, we will describe in
detail how our SFL works.
3 SERVICE FORWARDING
LABEL
As discussed in the last section, we need an ID that
can be used to identify a service instance (In our
context, a service chain instance is considered as a
single service instance.). This ID needs to sit above
Layer 4 so that it can stay intact while a packet
traverses legacy IP networks and middle-boxes. This
naturally points to an ID at Layer 5.
In OSI model, Layer 5 is called the session layer
which is designed to establish, manage, and
terminate connections between local and remote
applications. A good example is a video conference
session where multiple parties join and leave
dynamically. This bears similarity to a service
instance such as a virtual network which carries a
large number of dynamic traffic flows. This
similarity motivates us to define an ID at Layer 5 for
the identification of a service instance (or service
chain instance). We call this ID Service Forwarding
Label (SFL).
SFLs will be created and maintained by a service
provider and used by its clients and OpenFlow
switches to identify and steer traffic belonging to
different service instances. SFLs can be stacked for
applications such as recursive services where each
level of the stack is administered by the owner of the
service level in a recursive business relationship as
discussed in the last section. This allows easy scale
to multiple levels of services with multiple
ownerships nested in the SFL stack.
A SFL must be unique within the space of the
service provider who administers the SFL. Multiple
service providers at the same level will be
differentiated by their SFLs at the lower level. The
DCNET2015-InternationalConferenceonDataCommunicationNetworking
30
combination of the SFLs across different levels in a
label stack uniquely identifies a service at any layer
in a physical substrate domain.
In the following paragraphs, we will discuss SFL
format and the process to establish and terminate a
SFL.
As shown in Fig.3, each SFL is represented by 4
octets. Starting from bit 0 of the 4 octets, the first 30
bits hold the label, bit 30 is reserved for
experimental use, bit 31 is the top-of-stack bit (S).
The S bit is set to one for the last entry in a label
stack, and zero for all other label entries in the stack.
As the header at Layer 5, SFL can run either over
UDP or TCP making it applicable to all kinds of
traffic belonging to the same service instance.
Figure 3: (a) Format of Service Forwarding Label; (b)
Illustration of SFL stack in a packet header.
Each SFL is associated with a lifetime. When the
lifetime expires, the SFL will be terminated or be
renewed. This dynamic mechanism allows a service
provider to maintain a smaller pool of SFLs.
There are various scenarios that may happen
during the lifetime of a SFL. The procedures for
establishing and terminating SFL depend on the
actual scenario encountered. A typical scenario is
shown in Fig. 4. We describe the procedures step by
step in the following part.
A client sends a service request to the OFC
of a service provider with its user ID and
requested service type using HTTP request
message. Metadata can be sent through
HTTP Post message;
The OFC of the service provider decides
whether it can accept the request by
applying optimization process which
determines how to route traffic and allocate
resources for the requested service;
If the request is admissible, the OFC will
create a new SFL which is unique to the
service provider and send the SFL and
associated lifetime to relevant OpenFlow
switches or middle-boxes that need to steer
or process traffic based on the SFL through
an OpenFlow OFPT_FLOW_MOD
message;
Upon receiving the message from the OFC,
the OpenFlow switches or middle-boxes
will set the SFL and its lifetime into their
flow tables as part of a rule set;
The OFC will send HTTP response message
with the SFL and associated lifetime to the
client confirming the acceptance of the
request;
The client will add the label as Layer 5
header to its packets destined for the
requested service and send them out;
When the packets reach the switches or
middle-boxes within the service provider
network, the service provider will match the
Layer 5 header (and other headers in other
layers if necessary) to its rule set and decide
how to forward or process the packets based
on their service requirement;
The switches or middle-boxes will then
process those packets and steer them to the
next switch or middle-box if necessary;
When the lifetime of the SFL expires, the
client can choose either to renew the service
or leave. If it decides to renew, it will send a
HTTP request message with the SFL to the
OFC, the above procedures will be repeated
except that the original SFL will be used
instead of generating a new SFL.
In the next section, we will discuss some use
cases that will help illustrate how the SFL can be
used in real applications.
Figure 4: Flowchart of labelling process.
4 SAMPLE USE CASES
There are numerous use cases that the proposed SFL
ServiceForwardingLabelforNetworkFunctionVirtualizationandApplication-centricTrafficSteering
31
can be applied to. We will discuss some common
use cases briefly in this section.
The first use case is the virtual network service
we mentioned earlier. Here a physical network
provider will serve as the service provider and
service network providers will serve as clients.
Service network providers request virtual networks
from the physical network provider. Each service
network provider will have full control over its
virtual network. One issue we mentioned earlier is
that the address space used by service providers can
be overlapped. An example is shown in Fig. 5 where
Client Network 1 owns Virtual Network 1 and
Client Network 2 owns Virtual Network 2. Both
Virtual Network 1 and Virtual Network 2 share a
physical network owned by a SDN network
provider. When a packet reaches a switch in the
SDN network, the switch needs to decide which
virtual network the packet belongs to.
Figure 5: Virtual networks with overlapped address space.
Through the procedures discussed in the last section,
each client network will receive an SFL assigned by
the SDN network as an identifier of its virtual
network. The client network will inform its users of
adding the SFL for all packets that need to use the
virtual network it owns. When packets reach the
switches in the SDN network, they can be
differentiated using SFL even though their IP
address spaces may be overlapped. This can be done
by a simple match in the flow table. Without SFL,
multiple header fields may need to be matched in
order to identify packets belong to a virtual network,
which will likely cause flow table fragmented and
bloated.
When recursive network virtualization is
deployed, each service network will serve as client
as well as service provider at the same time. As a
client, it receives a SFL from the service provider
one level below it. As the service provider, it
administers the SFLs that identify the virtual
networks it sells. A physical switch can use multiple
levels of the label stack to steer packets for the
correct virtual networks they belong to.
Now we look at the second use case that
demonstrates how service chain can be supported. It
is easy to see that a service chain instance can be
realized using NFV where each virtual node
represents a specific service such as firewall that can
be dynamically mapped to a physical node in the
lower level. By the virtualization of a service chain,
dynamic sharing of physical resources can be
achieved. Traffic flows for different service chain
instances can be uniquely identified and steered by
the combinations of the multiple SFLs in their label
stacks. This enables great flexibility and leads to
significant cost reduction in OPEX. An example is
illustrated in Fig.6.
Figure 6: Service chain as network function virtualization.
Application service providers such as Google are
increasingly interested in providing different
treatments to different types of customers, e.g.
subscribers vs. casual users. Based on the SFLs they
are carrying, user traffic flows can be steered to
different environments with different networking
and computing resources provisioned. Under this
context, SFL provides a simple and effective handle
that connects applications to physical layer devices
directly and enables application-centric traffic
steering. There are many existing Quality of Service
(QoS) schemes such as VLAN and DiffServ. But
they are Layer 2 or 3 mechanisms which are hard to
scale to end-to-end applications. As mentioned
earlier, it is difficult to maintain any code points in
headers up to Layer 4 for end-to-end services due to
middle boxes and different domains a packet may
traverse. By sitting at Layer 5, our SFL can travel
through networks and middle boxes easily and
therefore provide a very strong support for various
end-to-end applications.
There are many application scenarios that can
demonstrate the usage of SFL. For example, a
DCNET2015-InternationalConferenceonDataCommunicationNetworking
32
service provider may want some of its user traffic be
protected from server or link failures while other
traffic not. When a server or link failure happens, the
traffic that needs protection is steered to a protection
path. In OpenFlow switches, packets that require
protection will be matched at a group table instead
of the regular flow table. Therefore incoming
packets must be de-multiplexed into regular flow
table or group table based on whether they need
protection or not. The proposed SFL provides an
excellent option to achieve this function.
Specifically, we can assign one SFL to identify
traffic requiring protection and another SFL for
traffic not requiring protection. As shown in Fig.7,
when packets arrive at a switch, it first goes to a
regular flow table. If the SFL matching indicates a
packet without protection requirement, other header
fields will be matched as regular case; otherwise, the
packets will be forwarded to a group table for
protection matching.
Figure 8: Forwarding packets with and without protection.
Figure 9: Virtual network service with SDN network as an
island.
For the case that SDN networks form some islands
in an end-to-end network, the proposed SFL also
provides an elegant migration solution. An example
is shown in Fig. 8 where the traffic from a client
network needs to traverse a legacy IP network to
utilize the virtual network service provided by a
SDN network. Because SFL sits at Layer 5, it will
not be affected by the legacy network. Packets
carrying SFL will travel through the legacy network
just like a regular packet. This allows the SDN
network island to provide new services even when it
is surrounded by legacy networks.
5 CONCLUSIONS
By separating control plane from data plane and
centralizing resource allocation, SDN has the
potential to allow network and service providers to
create a variety of new services. Existing SDN
products have been focused on some basic functions
such as flow setup and teardown. The potential to
create new services has not been fully explored.
NFV is one of the most promising services SDN
can provide. Through NFV process, a network
provider can sell virtual slices of its physical
network to different service providers. Different
service providers as tenants have full control of the
virtual topologies within their own slices while the
network provider has the control of its physical
network. However lack of an effective virtual
network identifier makes the potential of SDN
difficult to realize.
Application-centric traffic steering is another
type of services SDN can provide. Application
service providers have been looking for ways to
differentiate user traffic flows so that resources can
be used more effectively on generating revenues for
them. An end-to-end identifier that can be utilized to
differentiate user traffic flows is required so that
user traffic flows can be steered to the right server
and network resources for maximizing the benefits
of the service providers.
In this paper, we proposed SFL as an identifier
for service instance. It is controlled by service
providers and used by clients and OpenFlow
switches to steer traffic to different services. The
format of SFL is simple enough to minimize
overhead. Through SFL stacking, recursive services
such as recursive network virtualization can be
supported easily while allowing different entities to
exercise their controls over their own resources.
With SFL as a Layer 5 mechanism, it can traverse
middle-boxes and legacy networks without any
changes so that the relationship between clients and
service providers can be maintained end-to-end.
We have demonstrated various use cases ranging
from NFV, service chaining, to application-centric
traffic steering. Through these use cases, we can see
that the proposed mechanism is simple to implement
with existing protocols and technologies and can
effectively enable various new services.
ServiceForwardingLabelforNetworkFunctionVirtualizationandApplication-centricTrafficSteering
33
REFERENCES
ONF, 2015. https://www.opennetworking.org/
Banikazemi, et al., 2013. Meridian: An SDN Platform for
Cloud Network Services, IEEE Communications
Magazine, Feb. 2013.
Jain, S., et al., 2013. B4: Experience with a Globally-
Deployed Software Defined WAN, ACM SIGCOMM
2013, August 12-16, 2013, Hong Kong.
Tavakkoli, A., et al., 2009. Applying NOX to the
Datacenter, ACM Workshop on Hot Topics in
Networks (HotNets-VIII), October 22-23, 2009, New
York.
Tootoonchian, A., et al., 2012. On Controller Performance
in Software-Defined Networks, 2
nd
USENIX Hot-
ICE’12, April 24, 2012, San Jose.
Tootoonchian, A. and Ganjali, Y., 2010. Hyperflow: A
Distributed Control Plane for OpenFlow, USENIX
NSDI INM/WREN, April 2010, San Jose.
Yeganeh, S. Hassas and Ganjali, Y., 2012. Kandoo: A
Framework for Efficient and Scalable Offloading of
Control Applications, ACM SIGCOMM HotSDN’12,
August 13, 2012, Helsinki.
Jacobs, D., 2015. How SDN and NFV simplify network
service chain provisioning, http://
searchsdn.techtarget.com/tip/How-SDN-and-NFV-
simplify-network-service-chain-provisioning.
Mahalingam, M., et al., 2015. “VXLAN: A Framework for
Overlaying Virtulized Layer 2 Networks over Layer 3
Networks,” IETF draft, http://datatracker.ietf.org/
doc/draft-mahalingam-dutt-dcops-vxlan/.
Paul, S., et al., 2013. OpenADN: A Case for Open
Application Deliver Network,” Proceedings of ICCCN
2013, July 2013.
Sherwood, R., et al., 2009. FlowVisor: A Network
Virtualization Layer, OPENFLOW-TR-2009-1,
OpenFlow Consortium, October 2009.
Drutskoy, D., et al., 2013. Scalable Network Virtualization
in Software Defined Networks, IEEE Internet
Computing, Vol. 17, Iss. 2, March, 2013.
DCNET2015-InternationalConferenceonDataCommunicationNetworking
34