Energy Efficiency

Fiorani et al.

VOL. 6, NO. 4/APRIL 2014/J. OPT. COMMUN. NETW. 421

Energy Efficiency of an Integrated
Intra-Data-Center and Core Network
With Edge Caching
Matteo Fiorani, Slavisa Aleksic, Paolo Monti, Jiajia Chen,
Maurizio Casoni, and Lena Wosinska

Abstract—The expected growth of traffic demand may
lead to a dramatic increase in the network energy consumption, which needs to be handled in order to guarantee
scalability and sustainability of the infrastructure. There
are many efforts to improve energy efficiency in communication networks, ranging from the component technology
to the architectural and service-level approaches. Because
data centers and content delivery networks are responsible for the majority of the energy consumption in the
information and communication technology sector, in this
paper we address network energy efficiency at the architectural and service levels and propose a unified network
architecture that provides both intra-data-center and
inter-data-center connectivity together with interconnection toward legacy IP networks. The architecture is well
suited for the carrier cloud model, where both data-center
and telecom infrastructure are owned and operated by the
same entity. It is based on the hybrid optical switching
(HOS) concept for achieving high network performance
and energy efficiency. Therefore, we refer to it as an integrated HOS network. The main advantage of the integration of core and intra-data-center networks comes from
the possibility to avoid the energy-inefficient electronic
interfaces between data centers and telecom networks.
Our results have verified that the integrated HOS network
introduces a higher number of benefits in terms of energy
efficiency and network delays compared to the conventional nonintegrated solution. At the service level, recent
studies demonstrated that the use of distributed video
cache servers can be beneficial in reducing energy consumption of intra-data-center and core networks. However,
these studies only take into consideration conventional
network solutions based on IP electronic switching, which
are characterized by relatively high energy consumption.
When a more energy-efficient switching technology, such
as HOS, is employed, the advantage of using distributed
video cache servers becomes less obvious. In this paper
we evaluate the impact of video servers employed at the
edge nodes of the integrated HOS network to understand
whether edge caching could have any benefit for carrier
cloud operators utilizing a HOS network architecture.
We have demonstrated that if the distributed video cache
servers are not properly dimensioned they may have a

Manuscript received September 27, 2013; revised February 15, 2014; accepted February 17, 2014; published March 31, 2014 (Doc. ID 198202).
M. Fiorani (e-mail: [email protected]) and M. Casoni are with
the Department of Engineering “Enzo Ferrari,” University of Modena and
Reggio Emilia, Modena, Italy.
S. Aleksic is with the Institute of Telecommunications, Vienna University of Technology, Vienna, Austria.
P. Monti, J. Chen, and L. Wosinska are with the KTH Royal Institute of
Technology, Kista, Sweden.
http://dx.doi.org/10.1364/JOCN.6.000421

1943-0620/14/040421-12$15.00/0

negative impact on the benefit obtained by the integrated
HOS network.
Index Terms—Backbone networks; Edge caching; Energy
consumption; Hybrid optical switching; Intra-data-center
networks; Performance analysis.

I. INTRODUCTION

A

lthough information and communication technology
(ICT) can play a fundamental role in enabling a
low-carbon economy, the energy and carbon impact of
the ICT sector itself is already significant, and it is expected to grow rapidly with the proliferation of the number
of connected devices and with the emergence of new services. The energy consumption of the ICT sector can be
divided into (i) energy consumed by the user devices,
(ii) energy consumed by the telecommunication network infrastructure, and (iii) energy consumed by the data centers.
While end user devices are the major contributors, the sum
of the energy consumed by the telecommunication networks and data centers amounts to 51% [1] of the total
ICT consumption. With the expected growth in the Internet and data center traffic [2,3] the energy consumption of
telecommunication networks and data centers is destined
to drastically increase if the network energy efficiency is
not improved. In addition to low-power device technologies,
this problem can be addressed on architectural and service
levels.
Taking into consideration the architectural level, we observe that generally telecommunication networks can be
divided into three areas: access, metro, and core. Several
research papers address the energy consumption of the
different network areas [4,5]. It was shown that although
access networks are currently the major contributor, the
energy consumption of core networks is expected to grow
rapidly to be able to support very high capacities in the
range of several hundreds of terabits per second or even
petabits per second per node [2]. As for data centers, their
energy consumption is divided into energy consumed by
the information technology (IT) equipment, energy consumed by the cooling system, and energy consumed by
the power supply chain. According to the latest specifications, data centers are designed in such a way that the
ICT equipment consumes nearly all the energy within
© 2014 Optical Society of America

422 J. OPT. COMMUN. NETW./VOL. 6, NO. 4/APRIL 2014

the data center. This also means that in modern data centers major energy savings can be achieved by reducing the
power consumption of the IT equipment. According to [6],
the intra-data-center network, which handles the traffic inside a data center as well as that destined to the external
networks, currently represents 23% of the IT equipment
energy consumption. This number is expected to grow in
the future due to the forecasted increase in the data center
traffic [3]. It is, therefore, of the utmost importance to
define new energy-efficient intra-data-center network technologies. In [5,7] it has been shown that the switching
infrastructure consumes the major part of the energy in
core and intra-data-center networks, and it was pointed
out that future research needs to focus on improving the
energy efficiency of switching devices. Today, core and
intra-data-center networks are based on electronic switching; that is, data transmission is performed in the optical
domain, whereas switching and control is done by
electronic equipment. Consequently, electrical-to-optical
(E/O) and optical-to-electrical (O/E) conversions are performed at each hop, which leads to high energy consumption. To solve this problem, several optical switching
solutions have been proposed in core [8,9] and intradata-center [10,11] networks. In particular, [9,11] proposed
two architectures based on hybrid optical switching (HOS)
for achieving high performance and energy efficiency in
core and intra-data-center networks, respectively. The
term hybrid is used to describe the coexistence of different
optical switching paradigms, namely packet, burst, and
circuit switching.
Meanwhile, the latest Cisco Visual Networking Index
forecast [2] reports that consumer Internet video traffic
will increase from 57% to 69% of total Internet traffic in
the period between 2012 and 2017. As a consequence,
energy-efficient video distribution systems are an important tool to maintain a sustainable Internet growth. Video
content can be either stored and distributed from a few centralized servers located in large data centers (referred to as
the centralized approach) or the most popular video contents can be replicated in cache servers located close to
the end users (referred to as the distributed approach).
From the energy consumption perspective it is not obvious
which approach (i.e., centralized or distributed) is most
beneficial. In fact, storing content only in a centralized
server decreases the energy consumption for storage while
increasing transport energy requirements. On the other
hand, replicating some content closer to the users in
distributed cache servers decreases transport energy while
increasing the storage energy requirements. A few recent
studies [12–15] address this trade-off and conclude that the
highest energy efficiency is achieved by storing popular
content in cache servers close to the end users.
Recently, communication service providers are looking
for cloud solutions to reduce costs and create a new level
of efficiency. In this context, one of the most promising solutions is the carrier cloud model, where both data centers
and the core network are owned by the same entity and the
resources are virtualized and shared by multiple tenants.
Several large telecom operators are considering a move to
this novel business model [16,17]. Carrier clouds could

Fiorani et al.

overcome several problems that occur in the existing cloud
solutions, such as unpredictable and nondeterministic
network performance and insufficient availability and
security, which severely complicate or even preclude carrier-grade service level agreements. In order to increase
both the adaptability to different traffic types and the
energy efficiency at the architectural level, this paper
proposes a unified network architecture for carrier cloud
operators. The architecture is based on HOS and provides
both intra-data-center and inter-data-center connectivity
as well as interconnection capabilities toward legacy IP
networks. This architecture is referred to as an integrated
HOS network, in which the traffic is carried in the optical
domain along the entire path from an aggregation switch
inside a data center up to another aggregation switch (in
the same or a different data center) or to an edge node serving as an interface to legacy IP networks. In our study, we
analyze the structure of such an integrated architecture
and evaluate the benefits compared to a nonintegrated
HOS architecture as well as a conventional IP network
based on electronic switching.
Regarding the service level, we observe that the studies
in [12–15] take into consideration only traditional core and
intra-data-center networks based on electronic switches.
Since these networks are characterized by a low energy efficiency, the reduction in the transport energy introduced
by the distributed storing approach generally overcomes
by far the energy consumption of the cache servers. If
we consider instead a carrier cloud operator that relies
on the integrated HOS network, which is able to achieve
high energy efficiency, the advantage of the distributed approach on energy consumption might become less obvious.
For this reason, in this paper we deploy distributed cache
servers at the edge nodes of the integrated HOS network
and evaluate their impact on the network performance and
energy consumption. To the best of our knowledge, the
performance of edge caching in combination with an
energy-efficient network concept based on optical switching has not been evaluated so far.
To summarize, the contribution of this paper is twofold,
namely, (i) we propose and evaluate an integrated intradata-center and core network architecture based on the
HOS concept for carrier cloud operators along with a study
of its benefits and (ii) we assess the impact of distributed
video cache servers on the proposed integrated HOS
network architecture.
The remainder of the paper is organized as follows. In
Section II we describe the proposed integrated core and
intra-data-center HOS network. Section III introduces
the approach used to model the video cache servers. In
Section IV the reference network used for the simulations
and the energy consumption model are described. Section V
presents the simulation results, while Section VI contains
some concluding remarks.

II. INTEGRATED INTRA-DATA-CENTER

AND

CORE NETWORK

Figure 1 shows a high-level representation of the proposed integrated core and intra-data-center network based

Fiorani et al.

VOL. 6, NO. 4/APRIL 2014/J. OPT. COMMUN. NETW. 423

Fig. 1. Representation of the proposed integrated intra-data-center and core network based on hybrid optical switching.

on HOS. The integrated network provides three different
types of interconnections using a unified all-optical infrastructure and a common control plane. The first type of
interconnection is between servers inside the same data
center, referred to as an intra-data-center interconnection.
In Fig. 1 it is represented by a red dotted line to highlight
the path over which data are sent using the HOS paradigm.
The second type of interconnection is between servers
located in different data centers. We refer to it as an
inter-data-center interconnection, and we use a blue dashed
line in Fig. 1 to indicate the path performed in the HOS
domain. The third type of interconnection is between servers inside a data center and HOS edge nodes; that is, it provides the server-to-edge interconnections. An example is
indicated in Fig. 1 by a green solid line for the HOS path.
It should be noted that in the proposed integrated network, core and data centers employ the same unified control plane. A first attempt of an integrated control plane for
intra-data-center and core networks for a carrier cloud has
been recently proposed in [17]. The authors create a proof
of concept for an integrated control plane based on the software-defined network mechanism. However, the proposed
solution is still based on traditional electronic interfaces
between data centers and core networks, and hence it is
not optimized from the energy-efficiency point of view. In
order to minimize energy consumption, in this paper we
propose a novel control plane for intra-data-center and core
networks based on the HOS network model described in
[9,11]. It consists of two layers, the generalized multiprotocol label switching (GMPLS) control layer and the
HOS forwarding layer. The GMPLS control layer is in
charge of configuring and managing the network virtual
topology. It consists of three building blocks: routing,
signaling, and link management. The HOS forwarding
layer performs data aggregation, data scheduling, and
resource reservation. It supports three different optical
transport mechanisms, namely circuits, bursts, and packets. The HOS forwarding layer has the unique feature of
employing a common control packet for managing all three

switching paradigms, enabling circuits, bursts, and packets
to dynamically share the optical resources. The use of optical bursts in combination with packets and circuits allows
the dynamic implementation of different service classes,
leading to an efficient quality-of-service differentiation.

A. HOS Core Network
The HOS core network provides connectivity among
different data centers as well as between data centers
and legacy IP networks. As shown in Fig. 1, each node
in the HOS core network includes a HOS core switch. If
the node is located at the edge of the HOS core network,
it is equipped with an electronic switch for interdomain
connectivity.
An electronic switch in the HOS edge node ensures interoperability between the core network and the legacy IP networks. In the direction toward the HOS core network, the
HOS edge node performs traffic classification and traffic
aggregation. In other words, each incoming IP packet is
classified based on the value of the differentiated service
code point field in the IP header and mapped over the
best-suited optical transport mechanism, as described in
[9]. In the direction toward the legacy IP networks, the
HOS edge node extracts IP packets and performs IP routing. The HOS edge node is divided into two logical building
blocks, one of which consists of an electronic switch to perform IP routing and the second one of which includes all the
electronic components required to (i) perform traffic aggregation and classification in the direction toward the HOS
core network and (ii) perform IP packet extraction in the
direction toward the legacy IP networks. For simplicity,
we will refer to this block as the traffic aggregation block.
High-capacity optical switches provide connectivity
inside the core network. A HOS core switch can be logically
divided into two building blocks, namely the electronic
control logic and the optical switching fabric. The electronic control logic consists of three electronic blocks for

424 J. OPT. COMMUN. NETW./VOL. 6, NO. 4/APRIL 2014

Fiorani et al.

implementing the GMPLS control layer, the HOS forwarding layer, and the switch control unit. The optical switching
fabric is composed of two large optical switches. A fast
optical switch, based on semiconductor optical amplifiers
(SOAs), takes care of the transmission of packets and short
bursts. A slow optical switch, based on microelectromechanical systems (MEMS), handles the transmission of
circuits and long bursts. In the optical switching fabric
block we also include the following active optical components: optical amplifiers (OAs), tunable wavelength converters (TWCs), and control information extraction/
reinsertion (CIE/R) blocks in order to compensate for signal
losses in components, reduce blocking probability, and
encode the control information together with the data
payload on the same optical carrier, respectively.

traffic aggregation, while in the direction toward the data
center servers, the HOS aggregation nodes extract the IP
packets and perform IP routing. The HOS aggregation
nodes consist of the same logical building blocks as the
HOS edge nodes. The main difference between HOS edge
and HOS aggregation nodes is that the HOS edge nodes
could also include the video cache servers, which will be
further elaborated in Section III. The third tier of the
intra-data-center network is represented by a single large
HOS core node. This node has exactly the same architecture as the HOS core switch used in the core network.
For more details we refer to [11].

For a detailed description of the HOS core network we
refer to [9].

To evaluate the impact of distributed video cache servers
on the proposed integrated intra-data-center and core HOS
network, we extend the HOS edge node architecture described in Section II in order to include the video cache
servers. The extended architecture of HOS edge nodes with
cache servers is shown in Fig. 2. It can be logically divided
into three building blocks. Two of them have already been
mentioned in Section II, namely the electronic switch block
and the traffic aggregation block. The former one includes
the switch, the GMPLS module, and the input electronic
line cards, while the latter block comprises the classifier,
the conditioner, the assembler, the resource allocator,
and the packet extractor. The last block, which represents
an extension to the architecture previously presented in
[9,11], is related to the caching operations and consists
of the content tracker, the ToR switch, and the video cache
servers. The content tracker interacts with the HOS control
plane in order to keep track of all the video content inside
the cache servers, process the incoming video requests, and
update the cache servers.

B. HOS Intra-Data-Center Network
The HOS intra-data-center network provides connectivity among the servers inside a data center and connects the
data centers to the HOS core network. It is organized in a
three-tier fat-tree topology. The first tier consists of electronic top-of-rack (ToR) switches. In a conventional highend data center, servers are organized in racks, with each
rack hosting typically 48 blade servers. The ToR switches
interconnect the servers inside a rack and connect the
racks to the second tier of the intra-data-center network,
which is composed of the HOS aggregation nodes. The
HOS aggregation nodes perform the same functions inside
a data center as HOS edge nodes in the HOS core network.
In particular, in the direction toward the network core, the
HOS aggregation nodes perform traffic classification and

III. EDGE CACHE

Fig. 2. Architecture of the HOS edge node with video cache servers.

Fiorani et al.

VOL. 6, NO. 4/APRIL 2014/J. OPT. COMMUN. NETW. 425

As already mentioned in Section I, the impact of distributed cache servers on the network energy efficiency has
already been addressed in previous studies [12–15], mainly
focused on electronic switched networks. The rationale
behind these works is that distributed cache servers reduce
the traffic load, leading to a lower number of electronic
switch ports used in the core and intra-data-center
networks. However, the electronic switching devices commercially available today do not implement dynamic
switching-off of the ports, and thus their energy consumption is almost independent of the traffic load. Techniques
for dynamically switching off the line cards (LCs) have
been proposed in [18,19], but their efficiency in real network scenarios has still to be proven. In fact, scheduling
the switching-off of the LCs in a packet switching network
is a very challenging task because of the stochastic nature
of the traffic, and usually the interarrival time between
two successive packets is very small. The novelty of our approach consists in applying the caching concept to a HOS
network, where we assume that all the optical components
(in the optical switching fabric of the HOS core nodes) are
turned off when they are inactive. This is not as challenging as turning off electronic switch ports [9]. In fact, with
two parallel optical switches, only one needs to be active to
serve traffic from a particular port at a specified time. In
addition, in a HOS network, circuits and bursts are scheduled a priori; thus the incoming traffic is more predictable
than in a traditional packet switched network, that is, one
where the traffic is processed on a packet-by-packet basis.

IV. MODELING APPROACH
In this section, we describe assumptions used to model
and evaluate performance of the proposed integrated intradata-center and core HOS network with edge caching.
First, we present the power consumption model followed
by the description of the reference network scenario, and
finally we introduce the performance metrics.

A. Power Consumption Model

NX
Node
i
1

PiNode 

N
DC
X
j
1

PjDC ;

(1)

where N Node is the number of nodes in the core network and
N DC is the number of data centers. Each node in the HOS
core network performs both edge and core functions. The
power consumption of the ith node in the network is determined by
PiNode




PiEdge



· N W · PES  PA   PiCache ;
PiEdge
N Edge;i
F

PiCore ;

(2)

where PiEdge is the power consumption of the ith HOS edge
part and PiCore is the power consumption of the ith HOS

(3)

where N Edge;i
is the total number of fibers connected to the
F
HOS edge node i and N W is the number of wavelengths per
fiber, which are assumed to be the same for all nodes. In the
formula, PES is the power consumption of the electronic
switch block per port and PA is the power consumption
of the traffic aggregation block per port. The number of
ports of the switch is given by the product of the number
of wavelength channels per fiber and the number of fibers
· N W ); that is, it represents the total number of
(N Edge;i
F
wavelength channels at a HOS edge node. Finally, PiCache
is the power consumption of the cache block. The power
consumption of the cache block of the ith HOS edge node
is obtained through the following formula:
PiCache
PCT  PToR  N iCS · PCS ;

(4)

where PCT is the power consumption of the content tracker,
PToR is the power consumption of the ToR switch, and PCS is
the power consumption of a cache server. Finally, N iCS represents the number of cache servers hosted in the ith HOS
edge node. The cache servers are assumed to have a fixed
storage capacity of 1 TByte. Also the power consumption of
the ith HOS core switch is computed by summing up the
power consumption of its building blocks, as defined by
Eq. (5):
PiCore
PiECL  PiOSF ;

(5)

where PiECL is the power consumption of the electronic control logic and PiOSF is the power consumption of the optical
switching fabric of the ith HOS core switch. The power consumption of the control logic of the ith HOS core switch is
given by Eq. (6):
PiECL
N Core;i
· N W · PGMPLS  PHOS  PSC ;
F

The total power consumption of the integrated core and
intra-data-center HOS network is given by the sum of the
power consumed by each node in the core network (PiNode )
and the power consumed by data centers (PjDC ):
PNetwork


core switch. The power consumption of the ith HOS edge
part is given by the sum of the power consumption of its
building blocks (Section II):

(6)

is the total number of fibers connected to the
where N Core;i
F
HOS core node i. In Eq. (6), PGMPLS is the power consumption of the GMPLS block per port, PHOS is the power consumption of the HOS forwarding layer, and PSC is the
power consumption of the switch control unit. The power
consumption of the optical switching fabric of HOS core
nodes depends on the traffic because we assume that optical switch ports can be turned off when they are inactive. To
compute the power consumption of the optical switching
fabric of the ith HOS core switch, we use Eq. (7):
active;i
· PSOA  N active;i
PiOSF
N active;i
SOA
MEMS · PMEMS  N TWC · PTWC

 N core;i
· N W · PCIE∕R  2 · PEDFA :
F

(7)

active;i
active;i
Here, N active;i
SOA , N MEMS , and N TWC represent the number of
active SOA-switch ports, MEMS-switch ports, and TWCs of
the ith HOS core node, respectively. These values depend

426 J. OPT. COMMUN. NETW./VOL. 6, NO. 4/APRIL 2014

Fiorani et al.

on the traffic load and are computed through simulations.
In Eq. (7), PSOA , PMEMS , and PTWC are the power consumption of the SOA-switch per port, the MEMS-switch per port,
and the TWC, respectively. Finally, PCIE∕R and PEDFA are
the power consumption of the CIE/R block and the OAs.
When computing the power consumption of the HOS
intra-data-center networks we exclude from our analysis
the power consumed by servers and consider only the
power consumed by the network equipment, that is, by
the intra-data-center network. The power consumption
of the jth intra-data-center network is computed using
Eq. (8):
PjDC
N jToR · PToR  N jAggr · PAgrr  PjCore ;

(8)

where N jToR and N jAggr are the numbers of ToR switches and
HOS aggregation switches in the jth data center, respectively. Here, PAggr represents the power consumption of
an HOS aggregation switch and PjCore represents the power
consumption of the HOS core switch inside the jth data
center. We assume that each HOS aggregation switch is
connected to the corresponding HOS core switch in the data
center using one fiber and that the number of ToR switches
connected to the corresponding aggregation node is equal
to the number of wavelength channels per fiber (N W ). To
calculate the power consumption of a HOS aggregation
switch we use Eq. (9):
PAggr
N W · PES  PA :

(9)

Finally, the power consumptions of the HOS core switch
inside the jth data center is computed using Eq. (5)
and replacing the index i with the index j. The energy
consumptions of all the considered network components
are reported in Table I and have been obtained by collecting data from data sheets as well as from research
papers [9,11].

TABLE I
POWER CONSUMPTION OF THE NETWORK
COMPONENTS [9,11]
Components
Electronic switching block per port (PES )
Traffic aggregation block per port (PA )
Content tracker (PCT )
Top-of-rack switch (PToR )
Cache server (PCS )
GMPLS control layer per port (PGMPLS )
HOS forwarding layer (PHOS )
Switch control unit (PSC )
SOA switch per port (PSOA )
MEMS switch per port (PMEMS )
Tunable WC (PTWC )
Control information E./R. (PCIE∕R )
Optical amplifiers (PEDFA )

Power [W]
320
159
330
650
450
6.75
570
300
20
0.1
1.69
17
14

B. Reference Network Scenario
To assess the performance of the proposed integrated
intra-data-center and core HOS network with edge caching, we developed a custom event-driven C++ simulator.
In the following we report the main parameters that
we used in our simulations and present the model that
is applied to generate the network traffic.
We denote N Node as the number of nodes in the network
and N DC as the number of nodes connected to the data
center. We consider the Pan-European network [20] composed of 28 nodes (i.e., N Node
28) and 41 links as the
reference network topology. We assume that 25% of the network nodes are connected to a data center, that is, N DC
7.
In each simulation we randomly connect the data centers to
different nodes of the network. We assume that all the data
centers have the same size and are equipped with 76,800
servers organized in racks. In each rack, 48 servers are connected to a ToR switch using dedicated 1 Gbps links [21].
The number of ToR switches per data center is given by the
ratio between the number of servers and the number of
racks, that is, N jToR
N ToR
1600 ∀ j ∈ N DC . As many
as 64 ToR switches are connected to a HOS aggregation
switch using 40 Gbps links. We obtain that each data
center is equipped with N jAggr
N Aggr
25 ∀ j ∈ N DC
HOS aggregation nodes. Each HOS aggregation node is
connected to the HOS core node inside the data center
using one fiber. The HOS core switch inside a data center
is equipped with 25 fiber ports for interconnecting all the
HOS aggregation switches. In addition it employs 7 fiber
ports for the interconnection toward the Pan-European
network. Thus, it has in total 32 fiber ports. The number
of fiber ports for the interconnection between a data center
and the Pan-European network has been chosen according
to [3], where it is reported that currently 76% of the traffic
generated inside a data center is directed to a server within
the same data center (internal traffic). We assume that
each core node in the Pan-European network also provides
edge functionality. As described before, each data center is
connected to a network node of the Pan-European network
using seven fibers. To ensure that the network nodes have
enough capacity to support the connection toward a data
center without becoming a bottleneck, we assume that each
link in the network is composed of four fibers. We also
assume that each HOS core node is connected to the corresponding HOS edge node using a number of fibers that is
equal to the node degree. As a result, the number of fibers
attached to the ith HOS edge node (N Edge;i
) is equal to the
F
node degree. The number of fibers connected to the ith HOS
core node (N Core;i
) is equal to five times the node degree
F
(four times the node degree for the interconnection toward
other HOS core nodes and one time the node degree for the
interconnection toward the HOS edge node), plus seven
fibers in the case that the HOS core node is directly connected to a data center. Each fiber carries 64 wavelength
channels (N W
64), each of which is operated at 40 Gbps.
As for the edge caching, we assume that the network
nodes that are not directly connected to a data center
are equipped with the same number of cache servers
(N iCS
N CS ∀ i ∈ N Node ). The network nodes that connect

Fiorani et al.

VOL. 6, NO. 4/APRIL 2014/J. OPT. COMMUN. NETW. 427

data centers with the HOS core network do not comprise
any cache servers.

C. Performance Metrics

The cache size of a HOS edge node is defined as the sum
of the storage capacities, expressed in bytes, of all the video
servers hosted in the node. Furthermore, we define the
video content hit rate as the probability that a video request arriving at a HOS edge node determines a cache
hit and thus is served using the local cache servers. The
cache hit rate depends mainly on the cache size. Several
studies report the cache hit rate as a function of the cache
size in real networks based on the YouTube video distribution infrastructure [22,23]. The results of these studies
show that high video hit rates can be achieved even with
small cache sizes and that the cache hit rate exhibits a
logarithmic growth as a function of the cache size. As a consequence, increasing the cache size over a certain value has
a limited impact on the video content hit rate. In our simulation model, we assume that the video content popularity
belongs to a Zipf distribution with a library of 2 million objects and a skew parameter equal to 0.6. These are typical
assumptions for simulating a YouTube-like video content
delivery service [14,15] which lead to cache hit rates consistent with those presented in [22,23]. We also assume
that the size of the video contents is uniformly distributed
between 100 and 500 MByte [15], with an average video
size of 300 MByte, and, consequently, a library that
amounts on average to 600 TByte.

The performance of the proposed integrated intra-datacenter and core network architecture based on HOS is
assessed in terms of energy consumption, average delay,
and average data loss. The energy consumption is measured in terms of joules per bit (J/b) and is computed as
the ratio between the total network power consumption
in watts and the total network throughput in bits per
second.

The IP traffic arriving at the HOS edge nodes from legacy IP networks is modeled using a Poisson distribution.
We assumed that 57% of this traffic consists of requests
for video content [2]. A request for video content can be either served locally by the video cache servers in the HOS
edge node if the required content is available in the cache,
or can be forwarded to the original server located in one of
the data centers. In our simulations we also take into account the possibility that some of the traffic that arrives at
a HOS edge node is destined to another network node, that
is, not to the data center. We refer to this traffic as edge-toedge traffic. Even if it is not directly related to our analysis,
the edge-to-edge traffic is important because it has an
impact on the data losses and the delays as well as on
the energy consumption. For the traffic generated by the
servers, we implemented a more complex traffic model.
According to [24], the interarrival rate distribution of the
packets generated inside a data center can be modeled using a lognormal distribution. We then model the servers as
finite-state machines with two states, namely the lognormal state and the video-transfer state. In the lognormal
state, the servers generate IP packets with a lognormal-distributed interarrival time. The IP packets generated by the
servers in the lognormal state can be addressed either to a
server in the same data center, to a server in a different data
center, or to a specific legacy IP network connected to a HOS
edge node. When a server receives a request for video content from an edge node, it switches to the video-transfer
state. In the video-transfer state the server transmits IP
packets at a constant bit-rate to the requesting HOS edge
node. When all the video content has been transmitted,
the server automatically switches back to the state with
the longnormal interarrival rate distribution.

The delay is defined as the time difference between when
an IP packet is generated (i.e., either by a server in a data
center or a cache server in a HOS edge node or a user of a
network connected to edge HOS nodes) and the time the IP
packet is received (i.e., either by the destination server or
the destination HOS edge node). The global average network delay is defined as the mean value of the delays over
all IP packets measured during a simulation run. The IP
packets that traverse the HOS network can be carried over
different transport mechanisms. We refer then to the
packet delay as the delay experienced by IP packets that
are transmitted as optical packets through the HOS network. Similarly, the short burst delay, long burst delay,
and circuit delay are the delays experienced by IP packets
that are transmitted through the HOS network over a
short burst, a long burst, or a circuit, respectively.
While computing the data loss rates, we assume that all
the electronic switches introduce negligible losses. As a
consequence, the losses in the core and intra-data-center
networks may happen only in the HOS core switches.
We define the packet loss rate as the ratio between the
number of optical packets that are lost along a path
through the HOS network and the total number of generated packets. Similarly, the short burst and the long burst
loss rates are defined as the ratio between the number of
lost and the number of generated short and long bursts,
respectively. Circuits are established using a two-way
reservation mechanism, and consequently the data transmitted over circuits do not experience any losses. However,
in heavily loaded networks a circuit establishment request
could be refused (i.e., blocked) by a core node. As a consequence, we define the circuit establishment failure probability as the ratio between the number of blocked and the
number of generated circuits.
We evaluate the above-mentioned performance for different values of the network load. We define the load as
the ratio between the total amount of traffic offered to
the network by external sources (servers and legacy IP networks) and the maximum amount of traffic that can be
handled by the network, that is, the network capacity.

V. NUMERICAL RESULTS
This section presents a performance analysis of the proposed integrated intra-data-center and core HOS network
architecture with edge caching. First we comment on the
benefits that a carrier cloud operator can achieve by employing the integrated HOS network instead of either a

428 J. OPT. COMMUN. NETW./VOL. 6, NO. 4/APRIL 2014

Fiorani et al.

nonintegrated HOS network or a conventional IP network.
Then we present and discuss the impact of using distributed cache servers on an integrated HOS network.

A. Integrated HOS Network
To better understand the results presented in this
section, in the following, we first explain the difference
between the integrated and nonintegrated HOS architectures. In the nonintegrated HOS architecture, to interconnect data centers and core networks, we employ (i) HOS
edge nodes (one per data center) at the core network side
and (ii) HOS data-center-to-core interfaces at the data
center side. These components are shown in Fig. 3. The
HOS data-center-to-core interfaces perform traffic classification, conditioning, and assembling according to the data
center policies, while the HOS edge nodes perform the
same functions according to the policies used in the core
network. The internal architecture of the HOS datacenter-to-core interfaces is the same as the one of the

HOS aggregation switches used inside the data center.
In the integrated HOS network, there is no need for using
HOS edge nodes and HOS data-center-to-core interfaces to
connect data centers to the core network because both data
center and core network policies are considered when
processing the data center traffic in the HOS aggregation
nodes. Here, the HOS edge nodes are only needed at the
customer (cloud consumer) end.
In Fig. 4 we compare the energy consumption per bit as a
function of the network load for the integrated HOS network, the nonintegrated HOS network, and a conventional
IP network. The conventional IP network has a core and an
intra-data-center network based on electronic switching.
For comparative purposes, we also consider an IP core network able to put the LCs into sleep mode dynamically during idle times [18,19]. In our simulations, we assumed that
all the network nodes which are not directly connected to a
data center are equipped with N CS
10 cache servers, resulting in a total cache size of 10 TByte, which corresponds
to 1/60 of the library.

Fig. 3. Interconnection between data center and core network in the nonintegrated HOS architecture.
80
Integrated HOS network
Nonintegrated HOS network
Electronic IP network w sleeping
Electronic IP network w/o sleeping

45

Energy consumption per bit (nJ/bit)

Energy consumption per bit (nJ/bit)

50

40

35

30

25

20

15

10

5
30 %

35 %

40 %

45 %

50 %

55 %

60 %

65 %

70 %

75 %

80 %

85 %

90 %

Integrated HOS network
Nonintegrated HOS network
Electronic IP network w sleeping
Electronic IP network w/o sleeping

70

60

50

Intra-data-center
network

40

30

Core
network

20

10

0
30 %

35 %

40 %

45 %

50 %

55 %

60 %

Load

Load

(a)

(b)

65 %

70 %

75 %

80 %

85 %

90 %

Fig. 4. Energy consumption per bit as a function of the input load. The energy consumption per bit is the ratio between the network
power consumption and the network throughput. (a) Overall for core and intra-data-center networks and (b) core and intra-data-center
networks shown separately.

Fiorani et al.

VOL. 6, NO. 4/APRIL 2014/J. OPT. COMMUN. NETW. 429

In Fig. 4(a) we show the overall energy consumption per
bit. The figure shows that by employing a sleep-based technique it is possible to achieve large energy savings with
respect to current electronic IP networks, which leave all
LCs in the active mode during the idle times. It is also evident that, even if a sleep-based technique is employed in an
IP network, a HOS network is still able to achieve significantly lower energy consumption values. This is because
the HOS networks are based on an energy-efficient optical
switching technology that benefits from transmitting circuits and long bursts using an optical switch with low power
consumption and relatively slow switching time while using
a small number of fast optical switches for the transmission
of packets and short bursts. The benefit of using HOS becomes more evident for high loads, where sleeping is not
able to provide a significant improvement. Figure 4(a) also
shows the improvement in energy efficiency offered by the
integrated HOS network with respect to the nonintegrated
HOS network. This increment in energy efficiency may
seem small, but it is worth noting that at very high amounts
of network traffic, such as those forecasted in [2,3] and assumed in this paper, even a reduction of a few nanojoules per
bit can result in significant overall energy savings. For instance, at a network load of 35% the integrated HOS network consumes 4 nJ∕b less than the nonintegrated HOS
network. This translates into a total of almost 2 MW saved.
In Fig. 4(b) we show separately the energy consumption
per bit in the core network and the intra-data-center networks. The energy consumption per bit of the core network
is given as the ratio of the core network power consumption
and the core network throughput. In the nonintegrated
HOS network the power consumption of the core includes
the HOS edge nodes dedicated to the interconnection toward the data centers. Similarly, the energy consumption
per bit of the intra-data-center networks is given as the ratio
of the total power consumption of the intra-data-center networks and the total throughput of the intra-data-center
networks. In the nonintegrated HOS architecture the
power consumption of the intra-data-center networks
include the HOS aggregation switches dedicated to the

(a)

interconnection toward the core. From the figure we have
two important observations. First, core networks are more
energy efficient than intra-data-center networks. In fact, according to our calculations the power consumption of the
intra-data-center networks is always much higher than
the power consumption of the core network for similar
amounts of carried traffic. The difference is mainly coming
from the ToR switches introducing an extra level of aggregation in the intra-data-center networks, which is not
present in the core network [see Eqs. (2) and (8)]. The
ToR switches consume a very large amount of power and
dominate the power consumption of the intra-data-center
networks because of the very large number of ToR switches
in current high-capacity data centers based on the three-tier
fat-tree network topology. This fact is more evident for the
integrated HOS network where we observe that at a network load of 35% the energy consumption per bit of the core
network is 5 times lower compared to the intra-data-center
network. Second, the integrated approach has a higher beneficial impact on the energy consumption per bit of the core
network than of the intra-data-center networks. In fact,
when comparing the energy consumption of the integrated
and the nonintegrated HOS networks, we observe that at a
network load of 35%, the integrated approach reduces the
energy consumption per bit by 30.5% of the core network
and by 3.5% in the case of the intra-data-center network.
This is because the additional HOS edge nodes, used in the
nonintegrated HOS network to connect toward the data centers, have a strong impact on the total power consumption
of the core network. This impact is higher than the
impact of the additional HOS aggregation switches used inside the intra-data-center networks.
In Fig. 5 we compare the values of the average network
delays as a function of the network load. Figure 5(a) shows
the average delays in the integrated HOS network, while
Fig. 5(b) presents the average delays in the nonintegrated
HOS network. The figures demonstrate clearly that the integrated approach leads to a better delay performance and
reduces the global average delays of IP packets by always
more than 1 ms. In particular, the integrated approach

(b)

Fig. 5. Average network delays as a function of the input load for the integrated and the nonintegrated HOS networks. (a) Integrated
HOS network and (b) nonintegrated HOS network.

430 J. OPT. COMMUN. NETW./VOL. 6, NO. 4/APRIL 2014

Fiorani et al.

Energy consumption per bit (nJ/bit)

30
Without cache
Cache 1/120 of the library (5 TByte)
Cache 1/60 of the library (10 TByte)
Cache 1/30 of the library (20 TByte)
Cache 1/15 of the library (40 TByte)

25

20

15

10

5
30 %

35 %

40 %

45 %

50 %

55 %

60 %

65 %

70 %

75 %

80 %

85 %

90 %

Load

Fig. 6. Average data loss rate as a function of the input load.

significantly reduces the delays of IP packets transmitted
over short and long bursts. This is due to the fact that
bursts employ a mixed timer-length assembly algorithm
[9] that may take from several hundreds of microseconds
up to a few milliseconds. In the nonintegrated HOS network, the bursts must be disassembled and assembled
again in the electronic interfaces between a data center
and the core network, leading to a strong increase in the
overall network delay.
In this paper we assume that the electronic components
introduce negligible losses. As a consequence, the data loss
rates in the integrated HOS network and in the nonintegrated HOS network are the same. In Fig. 6 we show the
average data loss rates as a function of the network load.
The optical packets are scheduled with the lowest priority,
and thus they experience the highest losses. Optical bursts
are scheduled a priori due to the offset time so that they
receive a sort of prioritized handling in comparison to packets. In particular, long bursts are characterized by long offset times and show loss rates almost three orders of
magnitude lower than packets and almost two orders of
magnitude lower than short bursts. Finally, circuits are
scheduled with the highest priority and achieve a lossless
operation and negligible establishment failure probabilities
in our simulations. To understand where in the network we
observe the highest losses, we plot in Fig. 6 the average loss
rates in the intra-data-center, inter-data-center, and serverto-edge interconnections. We observe that the average loss
rates in inter-data-center interconnections are always the
highest. This is because in the inter-data-center interconnections the data needs to cross on average the highest
number of HOS core switches (in both the HOS core network and intra-data-center network). The lowest average
loss rates are instead achieved by the intra-data-center interconnections where data always have to cross a single
HOS core switch inside the data center.

B. Impact of Edge Caching
In Fig. 7 we show the energy consumption per bit of the
integrated HOS network against the network load and for

Fig. 7. Energy consumption per bit against the input load for
different cache sizes.

different values of the cache size. To vary the cache size we
change the number of cache servers per HOS edge node,
that is, the value of N CS. We always assume that the network nodes connected to a data center are not equipped
with local cache servers. To understand the results
shown in Fig. 7 it should be noted that we do not consider
dynamic switching-off of the electronic LCs, and consequently the energy consumption of the electronic components is independent on the network load. Only the
power consumption of the optical switching fabric of the
HOS core nodes, that is, POSF , changes with the network
load. Furthermore, it should be noted that the energy consumption per bit is defined as the ratio between the network power consumption given in watts and the network
throughput given in bits per second. Figure 7 shows that
at low and moderate loads, the higher the cache size, the
higher the energy consumption per bit. In fact, in our simulations, the increase in the storage energy consumption
introduced by the distributed cache servers (PCache ) is always higher than the reduction of the transport energy
that is obtained by switching off the unused optical switch
ports of the HOS core nodes. When increasing the load, we
observe that the larger the cache size, the faster the decrement of the energy consumption per bit. This is due to
the fact that increasing the number of distributed cache
servers reduces the average data loss rates in the network.
The larger the cache size, the higher the network throughput, especially at high loads. However, the network
throughput does not increase linearly with the cache size.
In fact, as shown in [22,23], the network throughput increases in a log-like way with the increase in the cache
size. This means that the network throughput becomes
saturated when increasing the cache size over a certain
value. On the other hand, increasing the cache size leads
to an almost linear increase of the storage power consumption. As a consequence, at high loads we observe that there
is a trade-off between cache size and energy consumption
per bit. In our simulations, when the load is higher than
50%, the best results in terms of energy consumption are
achieved using 1/60 of the size of the library, that is, setting N CS
10.

Fiorani et al.

VOL. 6, NO. 4/APRIL 2014/J. OPT. COMMUN. NETW. 431
-1

15.5

10

Without cache
Cache 1/120 of the library (5 TByte)
Cache 1/60 of the library (10 TByte)
Cache 1/30 of the library (20 TByte)
Cache 1/15 of the library (40 TByte)

15

10-2

Packets loss rates

14.5

Delays (ms)

14

13.5

13

12.5

12

Without cache
Cache 1/120 of the library (5 TByte)
Cache 1/60 of the library (10 TByte)
Cache 1/30 of the library (20 TByte)
Cache 1/15 of the library (40 TByte)

-3

10

10-4

-5

10

11.5
10-6
25 %

11
30 %

35 %

40 %

45 %

50 %

55 %

60 %

65 %

70 %

75 %

80 %

85 %

90 %

30 %

35 %

40 %

45 %

50 %

Load

(a)

-2

70 %

75 %

80 %

85 %

90 %

65 %

70 %

75 %

80 %

85 %

90 %

(b)
10

Without cache
Cache 1/120 of the library (5 TByte)
Cache 1/60 of the library (10 TByte)
Cache 1/30 of the library (20 TByte)
Cache 1/15 of the library (40 TByte)

-2

10-3

10-4

10-5

10-3

10-4

10-5

-6

25 %

Without cache
Cache 1/120 of the library (5 TByte)
Cache 1/60 of the library (10 TByte)
Cache 1/30 of the library (20 TByte)
Cache 1/15 of the library (40 TByte)

10

Long burst loss rates

Short burst loss rates

65 %

-1

10

10

60 %

Load

-1

10

55 %

-6

30 %

35 %

40 %

45 %

50 %

55 %

60 %

65 %

70 %

75 %

80 %

85 %

90 %

10

25 %

30 %

35 %

40 %

45 %

50 %

55 %

60 %

Load

Load

(c)

(d)

Fig. 8. Average delays and average data loss rates as a function of the input load for different values of the cache size. (a) Average
network delays, (b) packet loss rates, (c) short burst loss rates, and (d) long bursts loss rates.

In Fig. 8(a) we present the global average network delay
as a function of the network load for different values of the
cache size. The figure highlights that the larger the cache
size, the lower the global average network delay. In particular, increasing the cache size from 0 to 1/30 of the size of the
library (i.e., from 0 to 20 TByte) leads to a reduction of the
global average delay in the network by about 2 ms. A further increase of the cache size from 1/30 to 1/15 of the size of
the library (i.e., from 20 to 40 TByte) has a very limited
impact on the global average network delays.
Finally, in Figs. 8(b)–8(d) we show the average loss rates
of packets, short bursts, and long bursts as a function of the
network load for different values of the cache size. The circuit establishment failure probability is always null in the
considered configurations. The figures show that the larger
the cache size, the lower the average loss rates. This is due
to the fact that increasing the cache size keeps the traffic
more local, which corresponds to a higher amount of requests from the end users served by the cache servers. This
leads to a reduction of the traffic in the core and in the
intra-data-center networks and consequently to lower loss
ratios. Figure 8 also shows that by increasing the cache size
from 0 to 1/60 of the size of the library (i.e., from 0 to

10 TByte) we achieve a high reduction in the loss rates,
while increasing the cache size over 1/60 of the size of
the library (i.e., over 10 TByte) has a very limited impact
on the loss rates.

VI. CONCLUSIONS
In this paper we have proposed a unified network
architecture that provides both intra-data-center and
inter-data-center connectivity together with interconnection toward legacy IP networks. This architecture is tailored for the future carrier cloud operators running both
the data centers and the core network. The architecture
is referred to as an integrated core and intra-data-center
network and is based on the HOS technology. The main advantage of the integration of core and intra-data-center
networks in a single infrastructure comes from avoiding
electronic interfaces between the data centers and the core
network. We evaluated the energy consumption along with
the delay and loss performance of the integrated HOS
network and made extensive comparisons with respect
to a nonintegrated HOS solution and a conventional IP

432 J. OPT. COMMUN. NETW./VOL. 6, NO. 4/APRIL 2014

network based on electronic switching. We conclude that
the integrated HOS network achieves by far the highest
energy efficiency. Furthermore, we demonstrated that
the integrated HOS network reduces considerably the
average network delays with respect to a nonintegrated
HOS solution. As a consequence, we conclude that the
integrated HOS network is well suited for application in
carrier clouds.
Furthermore, we studied the impact of distributed video
cache servers on the energy consumption as well as the delay and loss performance of the integrated HOS network.
The existing literature on this topic only takes into account
conventional core and intra-data-center networks based on
IP electronic switching, which are characterized by low energy efficiency. The aim of this study is to identify whether
a carrier cloud operator that relies on the integrated HOS
network concept could increase energy efficiency by employing edge caching. Therefore, we extended HOS edge
node architecture to include cache servers and content
trackers. The content trackers interact with the HOS control plane for updating the servers and processing incoming
video requests. We also developed a novel analytical model
for evaluating the energy consumed by the cache. According to the results we conclude that to achieve both low delay and data loss as well as high energy efficiency in an
integrated HOS network, a careful dimensioning of the
cache size is needed. In particular, at low and moderate
loads we observed the highest energy efficiency is achieved
in the case without any edge caching. Furthermore, our
analysis also leads to the following general conclusion:
when deciding to upgrade the traditional electronic switching-based network to a more-energy efficient one, operators
have to reconsider their edge caching strategy in order to
achieve the best network performance.

REFERENCES
[1] “SMART2020: Enabling the Low Carbon Economy in the
Information Age,” The Climate Group, Global eSustainability
Initiative, Tech. Rep., 2008 [Online]. Available: www
.smart2020.org.
[2] “Cisco Visual Networking Index: Forecast and Methodology,
2012–2017,” Cisco White Paper, May 2013.
[3] “Cisco Global Cloud Index: Forecast and Methodology,
2011–2016,” Cisco White Paper, May 2012.
[4] Y. Zhang, P. Chowdhury, M. Tornatore, and B. Mukherjee,
“Energy efficiency in telecom optical networks,” IEEE
Commun. Surv. Tutorials, vol. 12, no. 4, pp. 441–458, Fourth
Quarter 2010.
[5] R. S. Tucker, “Green optical communications part II: Energy
limitations in networks,” IEEE J. Sel. Top. Quantum Electron., vol. 17, no. 2, pp. 245–260, Mar./Apr. 2011.
[6] “Where does power go?” GreenDataProject, 2008 [Online].
Available: http://www.greendataproject.org.
[7] C. Kachris and I. Tomkos, “A survey on optical interconnects
for data centers,” IEEE Commun. Surv. Tutorials, vol. 14,
no. 4, pp. 1021–1036, Fourth Quarter 2012.
[8] R. Veisllari, S. Bjornstad, and D. Hjelme, “Experimental demonstration of high throughput, ultra-low delay variation

Fiorani et al.
packet/circuit fusion network,” Electron. Lett., vol. 49, no. 2,
pp. 141–143, Jan. 2013.
[9] M. Fiorani, M. Casoni, and S. Aleksic, “Hybrid optical switching for energy-efficiency and QoS differentiation in core
networks,” J. Opt. Commun. Netw., vol. 5, no. 5, pp. 484–
497, May 2013.
[10] O. Liboiron-Ladouceur, I. Cerutti, P. Raponi, N. Andriolli, and
P. Castoldi, “Energy-efficient design of a scalable optical
multiplane interconnection architecture,” IEEE J. Sel.
Top. Quantum Electron., vol. 17, no. 2, pp. 377–383, Mar./
Apr. 2011.
[11] M. Fiorani, S. Aleksic, and M. Casoni, “Hybrid optical switching for data center networks,” J. Electr. Comput. Eng.,
vol. 2014, 139213, 2014.
[12] J. Baliga, R. Ayre, K. Hinton, and R. S. Tucker, “Architectures
for energy-efficient IPTV networks,” in Optical Fiber Communication Conf. (OFC), 2009, paper OThQ5.
[13] C. Jayasundara, A. Nirmalathas, E. Wong, and C. Chan,
“Energy efficient content distribution for VoD services,” in
Optical Fiber Communication Conf. (OFC), 2011, paper OWR3.
[14] C. Chan, E. Wong, A. Nirmalathas, A. Gygax, and C. Leckie,
“Energy efficiency of on-demand video caching systems and
user behavior,” Opt. Express, vol. 19, no. 26, pp. B260–
B269, Dec. 2011.
[15] N. Osman, T. El-Gorashi, and J. Elmirghani, “The impact of
content popularity distribution on energy efficient caching,”
in Proc. Int. Conf. on Transparent Optical Networks
(ICTON), 2013, pp. 1–6.
[16] D. Cai and S. Natarajan, “The evolution of the carrier cloud
networking,” in Proc. IEEE Symp. on Service-Oriented System
Engineering (SOSE), 2012, pp. 286–291.
[17] A. Autenrieth, J. Elbers, P. Kaczmarek, and P. Kostecki,
“Cloud orchestration with SDN/OpenFlow in carrier transport networks,” in Proc. Int. Conf. on Transparent Optical
Networks (ICTON), 2013, pp. 1–4.
[18] F. Idzikowski, S. Orlowski, C. Raack, H. Woesner, and A.
Wolisz, “Saving energy in IP-over-WDM networks by
switching off line cards in low-demand scenarios,” Proc. Conf.
on Optical Network Design and Modeling (ONDM), 2010,
pp. 1–6.
[19] S. Nedevschi, L. Popa, G. Iannaccone, S. Ratnasamy, and D.
Wetherall, “Reducing network energy consumption via
sleeping and rate-adaptation,” in Proc. USENIX Symp. on
Networked Systems Design and Implementation, 2008,
pp. 323–336.
[20] A. Betker, C. Gerlach, R. Hulsermann, M. Jager, M. Barry, S.
Bodamer, J. Spath, C. Gauger, and M. Kohn, “Reference transport network scenarios,” MultiTeraNet Report, July 2003.
[21] “Connectivity solutions for the evolving data center,” Emulex
White Paper, May 2011.
[22] M. Zink, K. Suh, Y. Gu, and J. Kurose, “Characteristics of
YouTube network traffic at a campus network: Measurements, models, and implications,” Comput. Netw., vol. 53,
no. 4, pp. 501–514, Mar. 2009.
[23] L. Braun, A. Klein, G. Carle, H. Reiser, and J. Eisl, “Analyzing
caching benefits for YouTube traffic in edge networks: A
measurement-based evaluation,” in Proc. IEEE Network Operations and Management Symp. (NOMS), 2012, pp. 311–318.
[24] T. Benson, A. Akella, and D. A. Maltz, “Network traffic
characteristics of data centers in the wild,” in Proc. Internet
Measurement Conf. (IMC), 2010, pp. 267–280.