Network Virtualization - An Overview

Kilian Rausch

Advisor: Michael Herrmann

Seminar Innovative Internet-Technologies and Mobile Communications
Chair for Network Architectures and Services
Department for Computer Science, Technische Universitt Mnchen

Email: [email protected]

ABSTRACT
This paper introduces the basic approach of Network Virtualization, as well as Xen and OpenFlow as two opportunities to realize this approach. A virtual network is an
autonomous, fully isolated network above an existing physical infrastructure. The goal of virtual networks is to run
side-by-side to productive networks without aecting them
and consequently to enable new network innovations to be
tested safely. Xen is a hypervisor running directly on the
hardware and able to host a large number of guest systems.
This paper presents the qualifications of Xen to act as a virtual router platform. OpenFlow itself is an open standard
for network devices to enable the easy deployment of experimental networks over an existing infrastructure. The related
FlowVisor project expands OpenFlow by the opportunity
to create and administrate fully isolated virtual networks.
These two approaches were chosen among others because of
their advanced progress and relevancy to practice.

Keywords
Network Virtualization, Router Virtualization, Software Designed Networking, Virtual Networks, Xen, OpenFlow, FlowVisor

1.

INTRODUCTION

In the past years a high interest in reconsidering the existing network and Internet architecture came up. Some even
describe the todays Internet architecture as ossified [10].
This movement was mainly carried by researchers, wanting
to experiment with new network innovations. But building
a realistic network environment for experimental purposes
would require enormous investments. Therefore the success
of new protocols and network architectures depends the possibility to run and test them coexistent, but isolated to existing network infrastructures. So they can not aect the
productive networks, but can be tested in detail under real
conditions. Productive networks are networks, which reliably carry out the everyday load. Network Virtualization
is an efficient way to overcome this obstacle and to pave
the way for new network ideas and developments. Network
Virtualization is generally achieved through running an additional software on the network devices.
Besides the named above, Network Virtualization oers other
various benefits. For example it allows the network operators to save a fix state of the network (In this paper we will
describe this operation with the term Checkpoint Saving).
This works like saving an image of a virtual operating system
(OS) and could be very useful before changes are deployed.

Seminar FI & IITM SS 2011,

Network Architectures and Services, July 2011

87

Then the previous state of the complete virtual network can

be restored in the case of an error.
Another application scenario would be, that Network Infrastructure Providers have the potential to host multiple
customers on the same hardware. Consequently they could
coordinate the load in a much more efficient way. This leads
to immense cost reductions for both parties and promotes
competition, because then even smaller network operators
have the possibility to use large-scale network infrastructure. Recently the Open Network Foundation (ONF) was
established to support the development of the so called Software Designed Networking (SDN). Among the members are
leading players like Facebook, Google, Microsoft, Deutsche
Telekom and IBM. This non-profit organization especially
supports the development and rollout of OpenFlow.
This paper gives an overview of Network Virtualization and
presents two possible solutions and their relevancy for praxis.
On that account we will especially look at the problems
and challenges, which were discovered in several test cases.
Given, that Virtualization in general brings a performance
loss, we will highlight this topic in the dedicated Sections.
In Section 2 the general approach of Network Virtualization
is illustrated. In Section 3 Xen Hypervisor as solution for
hosting multiple virtual routers is presented and in Section
4 we deal with OpenFlow and FlowVisor as the second possible solution. Finally we conclude in Section 5.

2.

NETWORK VIRTUALIZATION - OVERVIEW AND GENERAL APPROACH

Virtualization in general is a method to concentrate or divide resources of a computer. It abstracts from the physical
hardware, but gives the user the impression of interacting directly with it through the allocation of hardware resources.
Besides the rapid development of virtualization of operating
systems, researchers began to pay more attention to the virtualization of routers. This means running several virtual
routers on the same machine as the basis of administrating
multiple networks. Every router then belongs to a dedicated
network, giving the full administrative power. The approach
of Network Virtualization itself is not really new, as we already know Virtual Private Networks (VPNs) as solution
for connecting dierent networks. The dierence between
VPNs and Virtual Networks (VNs) is illustrated in [1, page
2] as follows:
Although VPNs provide a virtualized channel above a physical network infrastructure, they have a few disadvantages
in comparison with Virtual Networks. So all virtual networks have to be based on the same protocols, topology and

doi: 10.2313/NET-2011-07-2_12

addressing schemes. This mainly penalizes the reasearchers, wanting to deploy many dierent kinds of experimental
networks. Another big issue is the missing isolation of the
VPNs. While a few workarounds exist to deal with that
problem, even these solutions provide no real isolation. Furthermore the infrastructure provider and the VPN service
provider are normally the same entities. This disadvantage
is among other things caused by the missing resource isolation. So multiple providers can not be sure, that another one
is not aecting his network (for example by stressing the infrastructure). But the most important advantage of Virtual
Networks over VPNs is the true isolation of the dierent Virtual Networks [1, page 2]. This mainly targets on hiding of
network infrastructure specifications and even the existence
of another administrative domain. But of course, we must
consider the dierent use cases of VPNs and VNs. VPNs
are often used to safely connect two networks or to establish a single connection remotely, for example to and within
company networks. On the other hand a Virtual Network
is used when researchers want to deploy a complete network
over an existing one to test it under real conditions. So the
Virtual Network provides the feature to be programmed individually with experimental protocols, address schemes et
cetera.

Figure 1: The substrate layer contains the substrate

nodes. On these substrate nodes multiple virtual
nodes are hosted. One physical link contains multiple virtual links, mostly bundled to an aggregate.
[1]
routers.
Mainly the following challenges occurred when developing
Xen [3, page 1]:
1. True Isolation of the Virtual Machines

Another commonly used technique are Virtual Local Area

Networks (VLANs).VLANs are able to set virtual links above
physical ports. The dierence to Virtual Networks is, that
they only can virtualize one specific forwarding algorithm.
So the flexibility of choosing and developing new network
innovations is not supported here [12].

3. Keep the performance overhead as small as possible

3.1

Overview and Design Issues

In a traditional Virtual Machine Monitor (VMM) the hardware is emulated [5]. In contrast to this technique Xen uses
paravirtualization to host a guest OS. This means a software interface is provided for the hardware; similar, but not
equal to the hardware (no hardware emulation). As a consequence of that some slight modifications of the OS are
indispensable. But the relative low costs to port an OS to
Xen are worth to invest. This results in a system, which is
nearly as efficient as a native system [13]. Figure 2 shows
the eorts, measured in lines of codes (LOC), of porting a
convenient OS to Xen.

The basic elements of a virtualized network environment

are the substrate layer, where the physical resources (like
CPU, memory or storage) are located, and the virtual network layer. The substrate layer contains the substrate nodes,
mostly represented through typical network hardware like
routers or switches, which must support virtualization. On
these substrate nodes multiple virtual nodes are hosted, as
well as the physical link contains multiple virtual links. Figure 1 shows the interaction of the virtual and the substrate
layer. Virtual links are bundled to an aggregate. In order
to clearly identify a specific VN in a wide infrastructure, a
Virtual Network ID is used. This ID is globally unique and
consists of the responsible organization ID and the virtual
link ID [1]. When a Virtual Network is set up, a special
VNP/InP Inteface provides all relevant information needed
to build the VN, like the topology of nodes and virtual links,
physical location and traffic characteristics.

3.

2. Support of various operating systems

XEN

In this Section Xen and an approach to use its virtualization technology to build and operate a virtual network is
introduced. Xen is a hypervisor on x86 base which allows
multiple operating systems to run on the same hardware.
It is widely known as virtualization platform for operating
systems. Xen is operating directly on the hardware and is
able to host dierent OSs in the so called domains at the
same time. The primary intention was to keep the performance overhead as small as possible to simultaneously host
up to 100 Virtual Machines [3, page 1] and to isolate the
domains safely. In the following we will introduce how the
virtualization capabilities of Xen can be used to host virtual

Seminar FI & IITM SS 2011,

Network Architectures and Services, July 2011

Figure 2: porting costs, measured in lines of codes

(LOC), of porting Linux or Windows R XP to Xen.
[3]
Important to mention is, that the the guest applications
remain unaected here. The big dierence of Xen to a
convenient operating system, hosting the virtual machines
(normal Virtualization), is the possibility to provide performance isolation [4]. Performance isolation ensures, that

88

doi: 10.2313/NET-2011-07-2_12

one virtual machines performance can not impact the performance of another one (what of course is a desirable goal).
But as in [4, page 19] mentioned, this works only under
certain conditions. We will treat this especially for virtual
router purposes in detail in the Section Evaluation and Performance.

3.2

Xen and the x86 architecture

The Virtual Machine Interface for x86 architecture consists

of three main elements [3]. Memory Management, CPU
Management and I/O Device Management.

3.2.1

Memory Management

The memory management on x86 is quite difficult, because

there is no software-managed Translation Lookaside Buer
(TLB), which translates virtual memory to physical memory. So Xen is designed to hand over the responsibility of
allocating and managing the hardware page tables to the
guest OSs.

3.2.2

Figure 3: Architecture of a machine using Xen with

dierent guest operating systems and the privileged
dom0 control domain [3]

CPU Management

treats the capability of Xen to be used as a system for driving virtual routers. Multiple logically independent software
routers are hosted on one single hardware. Each of them is
responsible for his own isolated Virtual Network. The following statements and evaluations are based on [6] and [2].
Enabled by by powerful hardware and virtualization support, Xens guest domains (called domU) can be equipped
with virtual routers.

In CPU Management x86 can play to its strength by having

four privilege levels (alternatively called rings). Most other
architectures have only two privilege levels. Mostly one privilege level for sensitive and one for non-sensitive commands
(Popek and Goldberg Theorem [7]). Because Xen must urgently be located in level 0 (to have all possible rights), here
the problem occurs, that the guest system has to share the
same level with its applications. The guest system then runs
in a separate address space. This leads to expensive TLB
flushes through permanently involving the hypervisor for access control of the applications [3, page 4]. Running a guest
system in level 0 would completely destroy the idea of isolation, because then this system would be in the position to
change the whole system and consequently the other guest
systems too. So the four rings of x86 are essentially needed
to assign the highest privilege level to Xen, the second one
to the guest OS and the third or fourth one to the applications. This avoids any influence of errors to the hypervisor
and the expensive TLB Flushes.

3.2.3

The fact, that Xen controls every privileged action of the

domains results in two possible scenarios regarding packet
forwarding [6]:
each domU executes the forwarding itself
dom0 bears all the forwarding activities
In both of the scenarios, the packets have to be allocated to
the belonging domU out of the network stream and reverse.
Dom0 sits behind the physical Network Interfaces (Network
Interface Cards) and forwards the packet to the belonging
virtual Interface of the domUs. So there is a lot of packet
traffic between domU (what appears as the real network interface to the outer devices) and dom0, which has to be
coordinated.
This can be done by a bridging (default) or a routing mechanism. In bridging a software bridge within dom0 executes
this task and in routing IP-Adresses are assigned to the
physical interfaces, as well as to the virtual interfaces within
dom0.
In [6] several tests with dierent scenarios and combinations
are carried out in competition with a single native linux.
The authors draw the conclusion, that routing the packets
through dom0 is the better solution (in contrast to bridging),
especially with increasing number of domUs. Furthermore
they found out, that handing over the forwarding function to
the domUs results in a great performance loss. So when the
forwarding task is operated by dom0 with a routing mechanism, then and only then Xen is able to forward packets as
quick as a native linux. Surprisingly, the number of running
virtual routers does not have any impact on the performance
in this scenario. Finally the authors concluded, that when

I/O Device Management

The complete I/O Device Communication of each domain

is processed by Xen with the embedded interfaces of several
device abstractions. As mentioned above, the hardware is
not emulated. The embedded interfaces in combination with
the adapted OS result in much more performance. Additionally this allows Xen to manage and ensure the isolation of
the guest operating systems.
In Figure 3 a sample configuration with Xen and dierent
operating systems is illustrated. Dom0 must be emphasized as privileged control domain of Xen, able to start and
stop other domains (to make this possible, this functionality should be implemented in the operating system of dom0.
Here a XenoLinux is used).

3.3

Evaluation and Performance

As mentioned before in Section 1 there is a high interest

of dedicated network operators and researchers in changing
the existing network and Internet architecture. This Section

Seminar FI & IITM SS 2011,

Network Architectures and Services, July 2011

89

doi: 10.2313/NET-2011-07-2_12

these conditions are taken into account, Xen is suited for the
application as a virtual router platform.

The most popular use case is the experimental test of new

network protocols. This can be pefectly established in campus networks [10], because the researchers here can use their
familiar environment and profit from the OpenFlow Roll-out
at campus networks. The goal is to run experiments in the
existing campus infrastructure without aecting the everyday work of others. Therefore a normal Computer is not
sufficient, because of the low paket forwarding rate and the
small number of ports. Consequently there is a need for an
idealized OpenFlow Switch (Figure 4). This switch meets
the demands of high-performance, low-cost, isolation and
commercial vendors needs.

In [2] another test scenario is evaluated by the use of a

Click Router in combination with Xen. Click is a modular Open-Source Software-Router, implemented as Linux
Kernel Add-On [8] with multi-threading support. In several scenarios it is analyzed, how fast packets are forwarded
by the Click-Xen combination. Here it is experienced, that
the smaller the forwarded packets are, the lower becomes
the forwarding rate (measured in Gb/s). This is caused by
the amount of memory accesses, that small packets are generating. This bottleneck can indirectly be handled by the
CPU. CPU core switching during handling one packet and
the triggered memory accesses can be avoided by allocating
each virtual router to one particular core. Unfortunately this
is the only way to influence the memory access bottleneck.
Another big challenge is sharing a Network Interface Card
(NIC). Hardware-based allocation (every packet to the appropriate virtual router) is not fully supported neither by
Click nor by Xen. But in contrast to the scenario above,
here no software-based allocation is used. In the performed
tests in [2] it is simply assumed, that the NICs are able to
handle this demultiplexing. So the tests here are targeted a
bit dierent. The results show, that core allocation becomes
very difficult, when forwarding paths have dierent forwarding costs. Forwarding costs are for example composed of
table lookups and the packet size. This issue is solved by
a complex extended CPU Scheduler [2, page 5]. But as the
authors appositely presume, sharing a single core might not
be the issue in future, because development heads to increasing numbers of cores in CPUs. So it is concluded, that
even today a virtual router can be realized using Xen and
Click with the given problems to be solved in future. A similar project, just to be mentioned, is Trellis, where the same
conclusion is drawn [9]. Finally, also in [3] the conclusion is
drawn, that Xen is qualified for deploying network-centric
services.

4.

Figure 4: structure of an OpenFlow switch, showing the OpenFlow Software running above the hardware, modifying the FlowTable and communicating
with the Controller [10]
The data path is still a task of the switch hardware, based
on the flow table (provided by the OpenFlow controller).
Here OpenFlow provides an interface to modify the flow tables of commercial router or switch products, exploiting the
fact of many identical functions in the dierent products.
As a result the experimental network traffic can be segregated from production traffic. All in all this enables similar
benefits to the introduced solution with Xen (experiment
with new routing protocols, addressing schemes and even IP
alternatives [10]). The control instance saves a kind of forwarding rules as flow entries in the flowtable, each containing how to handle specific packets. A normal procedure in
an OpenFlow-based hardware is to forward packets, when
a flow entry already exists. The OpenFlow Software then
knows how to forward the packet to which port. Another
case is, that the arriving packet is unknown. Then OpenFlow sends the packet to the Controller via the encrypted
OpenFlow Protocol. The Controller then decides, whether
to create a new entry in the flow tables and to send it back
to the switch, or to drop the packet.
So in an example configuration with many switches and
routers, when a flow is defined at the control unit a new
protocol comes into operation and automatically creates a
new route for the packets by entering automatically all relevant FlowEntries in each switch. This is called a flow.

OPENFLOW AND FLOWVISOR

The OpenFlow standard [10] is another possibility to enhance the possibilities of given networks. This open standard runs as an AddOn on Ethernet switches and routers
and primarily separates the data path from the control path.
The network is programmable and allows administrators to
individually control and channel their data. Therefore only
one single control unit is needed to administrate multiple
routers and switches. This communication is carried out
via a special OpenFlow Protocol, which has the benefit of
beeing independent from hardware vendors. At the moment
major device vendors are implementing OpenFlow in their
hardware. OpenFlow is already be seen as a network virtualization technology by some members of the ONF, because
it provides similar benefits. But this depends on the definition and point of view, of course. But to clarify: no multiple
virtual router instances are running on the devices, like in
Xen. To achieve real Network Virtualization FlowVisor is
used [Section 4.2].

4.1

OpenFlow - Experimental Approach and

Functionality

Seminar FI & IITM SS 2011,

Network Architectures and Services, July 2011

90

doi: 10.2313/NET-2011-07-2_12

In OpenFlow there are two major opportunities to achieve

traffic isolation:

virtual network. The design focuses here on strong isolation.

This involves the controllers of the dierent slices, as well
as the belonging datapath traffic. One slice should not be
able to influence another one. This Isolation is achieved by
following which flow entry to which controller belongs and
assigning a minimum data rate to a slice [10]. FlowVisor
should be transparent (with the meaning of imperceptible)
for the controller and the OpenFlow Switch. This is for
example important when an error or bug occurs in a test
environment. It simplifies the finding and fixing process.
Furthermore FlowVisor should support resource allocation.
This is implemented by a special module, called the Resource
Allocation Policy [12].

forward the packets through the switchs normal processing pipeline

assign VLAN IDs to the groups of packets to allocate
them to dierent VLANs
Many sample applications are given in [10, page 4], just
as creating VLAN similar environments, establishing VoIP
connections, defining new addressing, naming and routing
schemes and even abstracting from flow processing of the
controller to programmable router based processing.

4.2

FlowVisor Network Virtualization Layer

One step ahead goes FlowVisor, a special purpose OpenFlow controller that acts as a transparent proxy between
OpenFlow switches and multiple OpenFlow controllers [11].
So in a FlowVisor application field, when a packet arrives
at an OpenFlow Switch, it is routed by the FlowVisor to
the belonging Controller and reverse forwards the packets
back to the switches. This is performed in an isolated way
by assuring, that no resources, like the FlowTables can affect each other. So FlowVisor supplements OpenFlow by
adding real virtualization possibilities. FlowVisor provides
virtualization of switches to build a divided, fully isolated
and autonomous network above the physical structure. Every network is logically independent and runs in addition to
a productive network on the same hardware. In [12] these
virtual networks are called slices. FlowVisor regards a slice
as any combination of switch ports and layer 2, 3 and 4 of
the OSI-model [14]. Of course, FlowVisor follows the virtual
network approach and supports its advantages, like resource
allocation or Checkpoint Saving [Section 1].

Figure 6: FlowVisor internal structure - FlowVisor

hosts multiple virtual guest controllers, each responsible for one virtual network
The three major design goals of FlowVisor are summed up
in [12] as follows:
1. Transparency
2. Isolation
3. Extensible Slice Definition

4.2.1
Figure 5: FlowVisor layer comparison to Xen and
abstract virtualization - Flow Visor is located betweeen the OpenFlow Switch Software and the Controllers above. Open Roads, PlugN and Open Pipes
are examples of virtual network controllers [12]
The layer location of FlowVisor in Figure 5 shows, that
it is comparable to other virtualization technologies and of
course to Xen, we treated before. Flow Visor is located between the OpenFlow Switch(forwarding path) Software and
the Controllers (control path) above. Open Roads, PlugN
and Open Pipes are examples of virtual network controllers
[12]. To the controllers FlowVisor appears as a set of OpenFlow Switches and to the OpenFlow Switches it appears as
a set of Controllers. FlowVisor itself hosts multiple OpenFlow guest controllers, as shown in Figure 6, one for each

Seminar FI & IITM SS 2011,

Network Architectures and Services, July 2011

Isolation and Challenges

Since isolation is one of the critical issues in Network Virtualization, this Section treats the isolation capabilities of
FlowVisor. Bandwidth Isolation is only provided with the
instruments of VLANs. Each packet has a VLAN Priority
Code Point field, where an entry assigns a packet to a certain priority level. Combined with the traffic class in the
Resource Allocation Policy this enables a kind of bandwith
allocation. But the big disadvantage is, that the exact specifications of a traffic class must be manually implemented in
each switch. This shows, that in this field future research has
to be done to make this issue more practicable. Topology Isolation means, that FlowVisor transmits only information of
the belonging switches to each of its virtual guest controllers
inside. So each virtual network, or each guest controller gets
only information about his own network. Due to the fact
most switch hardware has low-performance CPUs, the risk
of a breakdown of the OpenFlow Software is very high on an
overload. FlowVisor does not support CPU Isolation at the

91

doi: 10.2313/NET-2011-07-2_12

moment. Only a few workarounds exist, which explain how

to deal with the limited CPU. Furthermore the FlowSpace
Isolation assures, that one virtual guest controller can only
aect the own virtual network with its created rule. Even
the connection to the OpenFlow controller is virtualized and
controlled by transaction IDs. This ensures, that no guest
controller can block or even catch a transaction of another
guest controller.

[3]

[4]

4.2.2

Performance

As mentioned in Section 1, adding a additional virtual layer

between two instances results in performance deficits. Of
course this is for FlowVisor as well true as for any other
virtualization technology. The goal rather is to reduce this
performance overhead to a negligible amount. FlowVisor
realizes this through only acting in situations where it is
really necessary. All data and control paths work at full
line rate, without beeing slowed down by FlowVisor. This
also applies for any route selection, carried out by a controller. The only situation when FlowVisor intervenes, is
when a new flow message is send by the switch (a new unknown packet arrived) and port status messages (controller
demands switch to send byte and packet counters for a specific port). The tests in [12] show, that the average overhead
of a port status request is about 0.483ms and for flow messages 16.16ms. So we see the latency for port status request
is quite acceptable, where in contrast the 16.16ms latency is
probably a bit too high for time sensitive applications.

[5]

[6]

[7]

[8]

5.

CONCLUSION

[9]

Our goal was to give an overview over Network Virtualization and to highlight the ability of Xen and OpenFlow/
FlowVisor to realize this approach. So all-in-all the topic
of Network Virtualization is still an experimental field. But
recent developments show, that the industry is interested in
deploying this technique in praxis [Section 1]. What solution has the best chances to become successfully deployed
in large scale networks, depends on many factors and can
not be clearly identified. Here future research is needed.
So Xen provides a relative stable technology yet, while the
quite young OpenFlow Project now receives great support
by the ONF. But a disadvantage of FlowVisor is, that at this
time the software itself is not stable enough to be transferred
into production. Given the quite big performance overhead
of the flow messages, we can summarize, that a lot of work
has to be done in future. With the given constraints, Xen
can even nowadays provide a functional platform for virtual
routers and networks. Due to the increasing interest in programmable, virtual networks, this is going to be a subject
of interest in the future. Especially science will benefit from
this demand in research.

6.

[10]

[11]

[12]

[13]

REFERENCES

[1] Jorge Carapinha and Javier Jimnez: Network

Virtualization a View from the Bottom, VISA 09
Proceedings of the 1st ACM workshop on Virtualized
infrastructure systems and architectures, pages 1-3,
ACM New York, NY, USA c 2009
[2] Norbert Egi, Laurent Mathy, Mickael Hoerdt, Adam
Greenhalgh, Mark Handley and Felipe Huici: Fairness
Issues in Software Virtual Routers, PRESTO 08
Proceedings of the ACM workshop on Programmable

Seminar FI & IITM SS 2011,

Network Architectures and Services, July 2011

[14]

92

routers for extensible services of tomorrow, ACM New

York, NY, USA c 2008
Paul Barham, Boris Dragovic, Keir Fraser, Steven
Hand, Tim Harris, Alex Ho, Rolf Neugebauer, Ian
Pratt and Andrew Warfield: Xen and the Art of
Virtualization, SOSP 03 Proceedings of the
nineteenth ACM symposium on Operating Systems
principles, ACM New York, NY, USA c 2003
Diwaker Gupta, Ludmila Cherkasova, Rob Gardner
and Amin Vahdat: Enforcing Performance Isolation
Across Virtual Machines in Xen, PROCEEDING Middleware 06 Proceedings of the
ACM/IFIP/USENIX 2006 International Conference
on Middleware , Springer-Verlag New York, Inc. New
York, NY, USA c 2006
L. Seawright and R. MacKinnon: VM/370 - A Study
of Multiplicity and Usefulness, IBM Systems Journal,
pages 4-17, 1979
Norbert Egi, Adam Greenhalgh, Mark Handley,
Mickael Hoerdt, Laurent Mathy and Tim Schooley:
Evaluating Xen for Router Virtualization, Proceedings
of 16th International Conference on Computer
Communications and Networks, Computer
Communications and Networks, 2007. ICCCN 2007,
Honolulu, HI c 2007
Gerald J. Popek and Robert P. Goldberg: Formal
requirements for virtualizable third generation
architectures, Commun. ACM, 17(7):412421, 1974.
Daniel Schwencke: Click ein modularer Router,
Seminary paper, TU Braunschweig, July 2006
Sapan Bhatia, Murtaza Motiwala, Wolfgang
Muhlbauer, Vytautas Valancius, Andy Bavier, Nick
Feamster, Larry Peterson and Jennifer Rexford:
Hosting Virtual Networks on Commodity Hardware,
WORKSHOP ON REAL AND OVERLAY
DISTRIBUTED SYSTEMS (WORLDS), 2008
Nick McKeown, Tom Anderson, Hari Balakrishnan,
Guru Parulkar, Larry Peterson, Jennifer Rexford,
Scott Shenker and Jonathan Turner: OpenFlow:
Enabling Innovation in Campus Networks, Newsletter
ACM SIGCOMM Computer Communication Review
archive, Volume 38 Issue 2, April 2008, ACM New
York, NY, USA
FlowVisor: Home,
https://openflow.stanford.edu/display/flowvisor/Home
accessed on May 24th, 2011
Rob Sherwood, Glen Gibb, Kok-Kiong Yap, Guido
Appenzeller, Martin Casado, Nick McKeown and
Guru Parulkar: FlowVisor: A Network Virtualization
Layer, OPENFLOW-TR-2009-1
Christian Kern: Paravirtualisierung, Vanderpool,
Haupseminar Virtualisierungstechnologien, Technische
Universit
at M
unchen, Institut f
ur Informatik,
Lehrstuhl f
ur Rechnertechnik und
Rechnerorganisation, Prof. Dr. Arndt Bode,
22.07.2005
Hubert Zimmermann: OSI Reference Model - The ISO
Model of Architecture for Open Systems
Interconnection, IEEE TRANSACTIONS .ON
COMMUNICATIONS, VOL. COM-28, NO. 4, APRIL
1980

doi: 10.2313/NET-2011-07-2_12