Conference PaperPDF Available

Greening the Internet with Nano Data Centers


Abstract and Figures

Motivated by increased concern over energy consumption in mod- ern data centers, we propose a new, distributed computing platform called Nano Data Centers (NaDa). NaDa uses ISP-controlled home gateways to provide computing and storage services and adopts a managed peer-to-peer model to form a distributed data center in- frastructure. To evaluate the potential for energy savings in NaDa platform we pick Video-on-Demand (VoD) services. We develop an energy consumption model for VoD in traditional and in NaDa data centers and evaluate this model using a large set of empiri- cal VoD access data. We find that even under the most pessimistic scenarios, NaDa saves at least 20% to 30% of the energy com- pared to traditional data centers. These savings stem from energy- preserving properties inherent to NaDa such as the reuse of al- ready committed baseline power on underutilized gateways, the avoidance of cooling costs, and the reduction of network energy consumption as a result of demand and service co-localization in NaDa.
Content may be subject to copyright.
Greening the Internet with Nano Data Centers
Vytautas Valancius
Georgia Tech
Nikolaos Laoutaris
Telefonica Research
Laurent Massoulié
Christophe Diot
Pablo Rodriguez
Telefonica Research
Motivated by increased concern over energy consumption in mod-
ern data centers, we propose a new, distributed computing platform
called Nano Data Centers (NaDa). NaDa uses ISP-controlled home
gateways to provide computing and storage services and adopts a
managed peer-to-peer model to form a distributed data center in-
frastructure. To evaluate the potential for energy savings in NaDa
platform we pick Video-on-Demand (VoD) services. We develop
an energy consumption model for VoD in traditional and in NaDa
data centers and evaluate this model using a large set of empiri-
cal VoD access data. We find that even under the most pessimistic
scenarios, NaDa saves at least 20% to 30% of the energy com-
pared to traditional data centers. These savings stem from energy-
preserving properties inherent to NaDa such as the reuse of al-
ready committed baseline power on underutilized gateways, the
avoidance of cooling costs, and the reduction of network energy
consumption as a result of demand and service co-localization in
Categories and Subject Descriptors
C.2.4 [Computer Communication Networks]: Distributed Sys-
tems—Distributed Applications
General Terms
Design, Management, Measurument
Energy Efficiency, Data Centers, Nano Data Centers, Video
Most current Internet applications are served from a large num-
ber of collocated servers stacked together in one of multiple data
center facilities around the world. This centralized hosting model is
a classic example of the economies of scale: large numbers of sim-
ilar servers yields relatively low manning requirements and eases
Permission to make digital or hard copies of all or part of this work for
personal or classroom use is granted without fee provided that copies are
not made or distributed for profit or commercial advantage and that copies
bear this notice and the full citation on the first page. To copy otherwise, to
republish, to post on servers or to redistribute to lists, requires prior specific
permission and/or a fee.
CoNEXT’09, December 1–4, 2009, Rome, Italy.
Copyright 2009 ACM 978-1-60558-636-6/09/12 ...$10.00.
Figure 1: Growth of video content in the Internet.
procurement procedures. Homogeneous hosting environments al-
low for better resource optimization and managmement, while ex-
tensive use of virtualization technologies provide an abstraction of
dedicated platforms to developers.
Centralization trend is not without its limitations. Data centers
are prone to: 1) over-provisioning, 2) hight cost of heat dissipa-
tion, and 3) increased distance to end-users. All of these issues,
as we will see later, lead to ever increasing energy bills that the
data center and network operators need to foot. The data centers
are over-provisioned because they need to match the peak demand
despite that the average load remains much lower throughout most
of the day; in addition, redundancy requirements might push up
the number of necessary servers significantly. Data centers also
are expensive to cool. Despite significant efforts for improving the
server power-efficiency, an average data center spends as much en-
ergy on cooling as it spends for powering its servers [23]. Even
in the data centers using state-of-the-art cooling technologies heat
dissipation accounts for at least 20% to 50% of the total power con-
sumption [4]. Centralization trend also increases the data center
distance to the users. Not every service can be hosted at well con-
nected central interconnection points. Higher distance from end
users increases bandwidth-mileage requirements and adds to the
energy consumption of the networking equipment. It is not sur-
prising therefore that data centers recently made headlines with re-
ports that they consume 1.5% of total US electricity [3], or that
their carbon-dioxide emissions are projected to surpass those of the
airline industry by the year 2020 [6]. These concerns are indeed
expected to amplify in view of growth projections of data center-
hosted applications like video distribution [13] (Figure 1).
While data centers are an example of centralization trend, an-
other, opposite trend manifests though Peer-to-Peer (P2P) systems.
The decentralized P2P systems free-ride on end-user computers
and, being spatially distributed, incur little or no heat dissipation
costs. In addition, some P2P systems can exploit locality and re-
duce the redundancy requirements, because each P2P user often
is P2P server at the same time. Unfortunately, conventional P2P
system performance typically suffers from free-riding, node churn,
and lack of awareness of underlying network conditions.
Motivated by the problem of energy consumption, in this paper
we propose a new way to deliver Internet services based on our
Nano Data Center (NaDa) platform. The key idea behind NaDa
is to create a distributed service platform based on tiny managed
“servers” located at the edges of the network. In NaDa, both the
nano servers and access bandwidth to those servers are controlled
and managed by a single entity (typically an ISP) similarly to what
is suggested in [20, 27]. Significant opportunities already exist for
hosting such tiny servers on ISP owned devices like Triple-Play
gateways and DSL/cable modems that sit behind standard broad-
band accesses. Such gateways form the core of the NaDa platform
and, in theory, can host many of the Internet services currently
hosted in the data centers. In this paper, however, we will focus
on video streaming services, which, as shown in Figure 1, exhib-
ited large growth in past few years. In addition, video streaming
services allow for easier energy usage modeling and simulation.
Figure 2 shows a high level architecture of such streaming archi-
tecture using home gateways as nano servers.
NaDa follows a P2P philosophy, but contrary to typical P2P,
NaDa is coordinated and managed by an ISP that installs and runs
the gateways that act as nano servers, as well as the network that
interconnects them. Due to it’s managed nature, NaDa avoids most
of the shortcomings of classic unmanaged P2P. ISPs can easily im-
plement NaDa by providing new customers with slightly over di-
mensioned gateways, whose extra storage and bandwidth resources
would be used by NaDa to implement services like video hosting,
all of which will be totally isolated from the end-user via virtual-
ization technologies. Thus, with a rather small investment in higher
capacity devices, ISPs can go beyond their traditional role as com-
munication carriers, and enter the potentially highly profitable ser-
vice and content hosting market.
Obviously designing a system like NaDa spans a magnitude of
issues that cannot all be covered here and thus in this paper we
mainly focus on energy efficiency which, as discussed earlier, is
probably the Achilles heel of monolithic data centers. As we will
demonstrate in coming sections, NaDa can reduce the energy costs
of offering Internet services, particularly content distribution. The
savings come in three ways. First, through improved heat dissi-
pation stemming from the distributed nature of NaDa which min-
imizes the need for cooling. Second, through traffic localization
resulting from the co-location between NaDa and end users. Fi-
nally, NaDa performs efficient energy use by running on gateways
that are already powered up and utilized for content access or other
services. NaDa avoids wasting the high baseline power already
paid for in an online gateway. We elaborate further on these energy
saving aspects in Section 2.
Our main contributions are the following. We develop a model
to evaluate the energy needed to provide services in both central-
ized data centers and distributed NaDa. This model relies on a large
collection of empirical data from various sources, including power
measurements of operational content distribution servers and end-
user devices (Section 3). We then apply this model in the evalua-
1. Push
2. Serve
Figure 2: High level NaDa architecture. Content is served from
home gateways whenever possible.
tion of energy consumption in the context of video-on-demand. We
detail NaDa platform and placement mechanisms for video con-
tent (Section 4), and use trace-driven simulations to quantify en-
ergy savings under various interesting configurations. We find that
even under the most pessimistic conditions NaDa, achieves at least
15% to 30% overall energy savings compared to traditional data
centers, while more optimistic scenarios achieve up to 60% greater
energy efficiency (Section 5). Related work and final remarks are
discussed in Sections 6 and 7 respectively.
In this section we present a more detailed argument on why
NaDa platform can be more energy efficient than conventional data
centers and why it is certainly feasible to build it. We first detail the
energy reduction principles of NaDa. We then quantify the avail-
ability of the end devices to be used as NaDa nano nodes.
2.1 Advantages of NaDa
Four main principles allow NaDa’s to surpass the energy effi-
ciency of conventional data centers:
Heat Dissipation. Recent reports [4, 19] show that data cen-
ters consume large amounts of energy for energy transmis-
sion/conversion, and most importantly, cooling and heat dissipa-
tion. Such overheads are captured by the Power Usage Efficiency
(PUE) metric, which is defined as the ratio between the total power
consumed by a data center and the power actually delivered to its
IT equipment (servers, networking equipment). The PUE factor
ranges from as high as 2.0 in legacy data centers [23] to as low as
1.2 in recent state of the art facilities [4]. Operators are trying to
improve their PUE by leveraging advanced scheduling [26], per-
forming heat reuse [22], adopting free flow systems, and even ex-
ploiting colder geographical locations [12]. All these approaches,
however, have limitations. Scheduling, heat reuse and advanced
cooling systems save only a fraction of energy costs; placing data
centers in remote locations sacrifices proximity to users and speed
of service delivery. Finally, and most important, heat dissipation is
hard to tackle due to the high density of co-located IT equipment,
which is an inherent characteristic of monolithic data centers.
Church et al. [12] have recently proposed de-densifying IT
equipment by spreading then among container sized data centers.
NaDa pushes this idea to its extreme by distributing data center IT
equipment to user premises where heat dissipation costs are neg-
ligible. We show that the additional energy consumed by a single
gateway is negligible and thus easy to dissipate.
Service Proximity. Content providers build data centers in mul-
tiple locations to reduce service delays/communication costs and
improve fault tolerance. Because of high construction costs, only a
small number of data centers are built. They then have to be placed
at major network interconnection points which are often subject to
high real estate costs. Even at such locations, classic data centers
are relatively far from end users in comparison to servers of con-
tent distribution networks, and a fortiori of NaDa, whose PoPs are
inside user residences. Apart from reducing delays, service prox-
imity reduces the distance that information has to travel and thus
also the energy in powering and cooling the networking equipment
that has to carry it.
Self-scalability. Conventional data centers and the networks sup-
porting them are provisioned for peak load, which leads to rather
low average utilization [23]. Utilization is worsen further by the
fact that backup equipment needs to be powered and keep running
so as to be able to receive immediately any load resulting from
other failing equipment. Unlike conventional data centers, NaDa
is largely self adaptive. As the user population grows, so does the
number of available gateways to NaDa. Thus, NaDa reduces over-
provisioning while providing built-in redundancy.
Energy efficiency. Today’s routers and servers spend most of their
energy on the baseline activities such as running the fans, spinning
the disks, powering the backplane, and powering the memory. Even
in an idle state, modern systems can be consuming anything from
50% to 80% of the power consumed under maximum load [7, 10].
Therefore, operators constantly try to maximize the utilization of
their servers, but this becomes increasingly complicated due to the
daily variations of load, as illustrated in the following subsection.
Even though operators have perfected service load balancing, idle
servers are not necessarily switched off due to the need to keep
spare capacity readily available. The problem is even more promi-
nent in networks, where it is hard to shut off a router at off-peak
times [24]. In contrast, no baseline powering is paid by NaDa, as it
runs on gateways that are already on for servicing another purpose
(basic connectivity).
2.2 Gateway availability
Re-using the already committed base-line power of DSL gate-
ways is key to energy savings but depends largely on the number
of the devices which can be found online at a point in time. In
this subsection we quantify the above number through upper and
lower bounds derived from real gateway deployments, and thus
provide an idea regarding the expected capacity of NaDa as defined
by the above availability of powered devices. The upper bound is
computed from gateway response traces produced by Gummadi et
al. [14], while the lower bound is derived from set-top-box activ-
ity traces, exploiting the fact that if a user is accessing content, his
DSL gateway must be up.
Upper bound. While some users keep their gateways online at all
times, there definitely exist energy-conscious ones that regularly
switch them of when not in need, e.g., during work hours or trips.
The study of Gummadi et al. includes reports on the availability of
the residential gateways in 12 major ISPs over one month. The data
consists of gateway responses to active probing performed every 15
minutes. Based on these data, we computed the gateway uptimes
by accumulating all the intervals during which we found a gate-
way to be active. To discount for possible network effects (such as
firewalls) we removed the users that never replied to the probes.
Based on the above analysis, we found that on average, equip-
ment at the end user premises is up 85% of the time1, as seen in
Figure 3. We also observed only insignificant, 5-8% uptime varia-
1Note though that devices probed by Gummadi et. al. belong to
Figure 3: Average user uptime in different provider networks.
tion thought out a day. Therefore, 85% gives a rough upper bound
on the time we can expect to find a gateway available for processing
some workload sent to it by NaDa.
Lower bound. To derive a lower bound we made the observation
that a customer’s gateway has to be up and running for at least as
much as the time that we find a user actively accessing content on
his set-top-box. To test the idea we resorted to the IPTV traces col-
lected by Cha et al. [9]. The IPTV traces consist of 251,368 active
users over 63 days. The traces record each operation the user is
performing, including channel selection and switching. Figure 4
shows IPTV user activity throughout the day. We see a high vari-
ability in usage, depending on the time of the day. In the peak hour
approximately 19% of customers are using their Internet connec-
tion for IPTV services; on average the usage is at around 7%.
Figure 4: Usage of the IPTV service.
This bound is highly pessimistic since customers can use other
services requiring Internet access. For instance, triple play users
would want to keep their DSL gateways on because of the tele-
phony service. In Section 5 we will examine both scenarios and
show that even in such pessimistic case NaDa achieves consider-
able energy savings.
In this section we first quantify the energy consumption charac-
teristics of different devices engaged in NaDa. Then we develop an
energy consumption model that we use for performing back-of-the-
envelope calculations of system-wide energy consumption. This
BitTorrent users, thus biasing the sample.
model is further refined in Section 4 to capture additional content
related details.
3.1 Basic Parameters
We distinguish between the servers, the gateways and the net-
work in assessing the energy needed to deliver cloud services with
NaDa and with monolithic data centers.
Servers. To illustrate typical server performance, we measure a
video streaming server used in production in a Sapphire VOD plat-
form [5], which provides both live IPTV and VoD services. More
precisely, we monitored a server, with 5GB of RAM and 2 Intel
Xeon DualCore 2GHz processors. We varied the load by chang-
ing the rates and numbers of video streams downloaded from the
server, and tracked the resulting power consumption through a spe-
cialized energy measurement device that provides a precision of
±0.01Watt (HIOKI 3333 Power HiTester).
The results, reported in Figure 5, show that power is a function
of aggregate load, and does not depend on individual video rates or
number of concurrent video streams. More importantly, the base-
line power (325W) constitutes more than 80% of the overall power
use at peak utilization, which corresponds to 1.8Gbps aggregate
streaming rate. In addition, we observe that the load-dependent
power consumption (obtained by discounting baseline power) de-
pends linearly on aggregate load.
Thus for a service rate of xMbps, the corresponding power con-
sumption reads bs+asx, where subscript sis for server, bsis the
baseline consumption, here equal to 325W, and asis the slope of
the load-dependent consumption, here approximately equal to 30.6
Joule per Gigabit. In the present case, xvaries between 0 and
MaxLoad, with MaxLoad equal to 1.8Gbps. Taking into ac-
count the baseline power, the energy per bit reads as+bs/x,and
is minimized when the server is fully used. In that case, it equals
as+bs/MaxLoad, which with the current parameters gives us
aservercostofγs:= 211.1Joules per Gbit. Note that 211.1
Joules/Gbit represents the most efficient use of this VoD system. In
fact this metric, which we nevertheless use in our evaluation and
build a case for more efficient NaDa system, implies a data cen-
ter without any stand by resources and with all servers running at
100% capacity.
Figure 5: Power use of a video server.
Gateways We performed similar measurements on a commercial
Thomson Triple Play VDSL2+ gateway used by France Telecom.
We monitored TG787v model which has a Broadcom chipset, with
a 300 MHz MIPS32 processor. We uploaded video from the gate-
way, storage being provided either by an external hard disk, or a
flash memory, connected by USB port.
Figure 6 shows the resulting power consumption for varying
loads. Baseline consumption of the Gateway (in the absence of
external storage) equals 13.5W. To power the USB port requires
some additional power, which depends on the model of flash mem-
ory or HDD used. For example, the Flash 1GB stick increased the
baseline by 0.5W. As we see from Figure 6, the additional, load-
dependent power used is negligible compared to baseline. Also, as
in the server case, it appears to depend linearly on load, and corre-
lates with CPU activity.
Turning on such a gateway to serve load with Flash (254MB)
storage incurs a cost of 14.5W/10Mbps, or 1450 Joules per Giga-
bit, considerably higher than the corresponding metric for servers.
However, if we consider a gateway that is already on, its baseline
consumption of 13.5W is already spent anyway. Thus the energy
per bit becomes 1W/10Mbps = 100 Joules per Gigabit, about half
of the corresponding metric for servers. We shall denote this cost
for serving one bit as γn(subscript nbeing for NaDa).
If we also consider next-generation Gateways with built-in flash
memory, we might speculate that it is possible to eliminate the in-
crease in baseline due to powering the USB port. Data in Figure 6
then yields an improved cost of only 18 Joules per Gigabit. Despite
this opportunity, our evaluation in Section 5 is conducted with a
more conservative 100 Joules/Gbit.
Figure 6: Power use of a DSL modem serving content.
Network To evaluate the network based power consumption, we
pinpoint the cost in terms of Joules per bit incurred at each net-
work hop, and then determine the number of hops that need to be
traversed to deliver a service. We ignore non-routing equipment
such as simple Ethernet switches or optical equipment. We argue
that, for the same load, such equipment energy use is a fraction
of routing equipment energy use, since it does not perform packet
processing and route table lookups.
To evaluate energy cost per hop, we rely on the study of
Chabarek et al. [10]. In this study the most efficient platform was
found to be the Cisco GSR 12000 router chassis, whose power con-
sumption was approximately 375W. A representative line card con-
figuration used in that study achieves peak capacity of 5Gbps. We
focus on this particular configuration for our evaluation. Assuming
(optimistically) that routers are typically loaded at 50%, this gives a
rate of 2.5Gbps. Combined with the power consumption, this gives
a ratio of 375W/2.5Gbps=150 Joules per Gigabit. We shall denote
this cost per bit as γr(subscript rbeing for router).
We now evaluate the typical number of hops involved in data
transfers. We envision two distinct scenarios: 1) NaDa used across
the wide-area Internet, and 2) NaDa used within an ISP. To find
out distances to servers and between clients on the Internet we em-
ployed the DipZoom measurement [28]. DipZoom has about 300
measurement nodes; the clients provided by DipZoom are from
North America, Europe and Asia. We grouped these clients based
on the country, city or autonomous system they belong to. Table 1
shows the resulting data. The “Same DSLAM” line is not from
DipZoom, and gives the hop count between two Gateways under
the same DSLAM, assuming the DSLAM has routing capability.
The same Table 1 shows the typical distance between clients in
the ISP in the same metropolitan area and distance from clients to
VoD servers. We obtained these metrics from private communi-
cation with several large service providers. We recognize that it
is hard to find both meaningful router efficiency metrics and per-
fect distance measurements that would satisfy all network scenar-
ios. Therefore, in our evaluation section we provide NaDa multiple
efficiency results for differing assumptions about the underlying
network properties.
Measurement Distance Standard
WAN: Popular servers 14.01 4.2
WAN: Clients within a country 14.09 4.01
WAN: Clients within a city 13.53 4.29
WAN: Clients within an AS 9.1 2.56
ISP: Accessing server 4 -
ISP: Accessing another client 2 -
Table 1: Number of router hops traversed between clients and
3.2 Energy Consumption Model
Power Server
baseli ne
Figure 7: Energy usage slope.
Previous discussion led us to consider a linear model for the
power consumption/load for all three components, namely servers,
gateways and routers, characterized respectively by slopes γs,γn
and γr. For gateways, such a linear model is motivated by the mea-
surements in Figure 6 and the fact that in NaDa, only gateways that
are already on will be used. Hence the baseline power should not
be accounted, being already wasted no matter whether additional
load is put on the gateway or not.
For servers, assuming that one uses the minimal number of
servers that can handle the load, the actual power/load relation-
ship is given by the “staircase” curve on Figure 3.2. However, if
servers could be put to sleep (and hence not consume energy) when
idle, and process load at maximal speed when awake, then the lin-
ear model would be appropriate. Hence the linear model for servers
can either be thought of as an approximation of the staircase model,
or as an accurate description of servers with efficient sleep mode.
Using the linear model for routers is harder to defend, as a router
needs to remain powered, even when forwarding data at a low rate.
An alternative assumption consists in assuming that router power
consumption is load-independent, and thus savings can only come
from shifting load from servers to gateways. We shall indeed con-
sider energy savings under this assumption. However if network
load is reduced, then network capacity upgrade can be delayed, and
long-term energy consumption of network components can be re-
duced accordingly. We shall thus also evaluate energy savings with
the linear model for routers.
Apart from power consumption of individual components, the
overall system efficiency is also affected by power conversion,
power loss, and cooling costs. Data centers, CDN networks and
routers incur cooling and power conversion overhead; to account
for it, we inflate the corresponding power numbers by the PUE fac-
tor defined earlier. Recent reports [4, 19] indicate that PUE of 2
is common, while PUE of 1.2 is the best to date, and unlikely to
be improved much upon. To be fair to data center architecture, we
assume that conventional equipment is subject to a PUE of 1.2.
Thus for a given load xexpressed in terms of bandwidth, we
propose the following expression for the corresponding power con-
sumption in data centers, that we denote Edc (x):
Edc(x)=xPUE(γs+hdc γr)(1)
In the above, hdc is the average number of hops necessary to reach
the client in a conventional, centralized system. In view of Table 1,
we evaluate both hdc =14and hdc =4. We may alternatively re-
move the term reflecting networking equipment, hdc γrif we want
to account only for energy used in servers and gateways.
In the case of NaDa, gateways, unlike routers and servers, do
not require cooling. Nevertheless they do incur overhead costs for
energy transportation, distribution and conversion. Surveys [1, 2]
indicate that such losses amount to 7.4% and 6.2% of the delivered
energy respectively for the US and the UK. When evaluating energy
consumption at gateways, we thus inflate our initial evaluation by
a “loss” factor that we take equal to 1.07.
Thus for a given load xin bandwidth, the corresponding power
consumption in NaDa, denoted En(x),is:
En(x)=x(γn+PUE hnγr),(2)
where hnis the average number of hops necessary to reach the
client inNaDa, which in view of Table 1, could be taken as hn=9
or hn=2for a WAN or an ISP scenarios respectively.
3.3 Back-of-the-envelope calculations
The expressions (1,2) allow us to assess potential benefits from
NaDa for content access services. Specifically, by using the pro-
posed values γs= 211.1J/Gb, γn= 100J/Gb, γr= 150J/Gb,
PUE=1.2, =1.07, hdc =14,andhn=2, summarized in Table 4
we find:
Edc(·)=1.07 100 + 1.29150
1.2(211.1+14150) 59%.(3)
Thus NaDa potentially reduces power costs by 41%. Similarly,
we may compare NaDa efficiency to that of Content Delivery Net-
works (CDNs). For CDNs we would use the same formula (1) as
for data centers, except that we would set the hop count term to
hcdn =4in view of Table 1. The corresponding expression, that
we may denote Ecdn, would yield the following evaluation:
Ecdn(·)=1.07 100 + 1.22150
1.2(211.1+4150) 38%.(4)
Thus we may expect energy savings of the order of 62% in NaDa
compared to CDNs.
When we consider only gateway and server energy, by removing
the network component in the above ratios, we arrive at a value of
1.07 100/(1.2211.1) 42%, thus savings on data centers only
(i.e. network excluded) of potentially 58%.
The above values are idealized values of course. To quantify
more precisely a concrete application scenario, e.g., NaDa-based
VoD, many more parameters need to be considered including the
memory used at Gateways, the placement strategies, and the popu-
larity profile of content. We perform an in-depth evaluation of these
impacts in the next section.
NaDa platform is ambitious in the sense that it could eventually
host various third party distributed applications (e.g. distributed
games, file sharing, social networks, web hosting, etc). However,
designing applications to run on such highly distributed environ-
ment is non-trivial (security concerns, data consistency and parti-
tioning, etc) and thus it is outside of the scope of this paper. To
highlight the benefits of NaDa in terms of energy savings, instead
we focus on one particular application: a distributed VoD system.
Video content distribution will likely be the main driver of Internet
traffic growth in the next 5 years [13].
Results on the energy consumption of VoD delivered from NaDa,
evaluated via event-driven simulations, are presented in Section 5.
In this Section we set the stage by describing the system architec-
ture, the data sets, the content placement method and the simulation
setup used in the evaluation.
4.1 VoD Service Architecture in NaDa
The fundamental element in NaDa is the gateway. In addition,
atracker manages gateway resources. Finally, content servers take
care of ingesting content in the system, and provide a fall-back op-
tion under shortage of gateway resources. We now detail further
the roles of these three key components:
Gateways provide storage, and bandwidth resources. In the case
of VoD, they store full or partial replicas of video content objects,
and provide uplink bandwidth to deliver these objects to other gate-
ways. NaDa gateways have two separate virtual circuits on their
upstream line with different capacities allocated to each: one vir-
tual circuit is dedicated to conventional Internet use while the other
is allocated for NaDa use. Such setting is challenging to achieve
on cable networks, but readily available with DSL and fiber-to-the-
home technologies. In the case of DSL service providers can sepa-
rate and contain the traffic from gateways using different VPI/VCI
designators, while in the case of FTTH networks ISPs can use
VLANs or GPON compatible technologies. In our evaluation we
assume gateways with 1Mbps to 2Mbps upstream speeds.
Current home gateways today are (almost) always on, making ef-
ficient use of baseline energy an easy target. In the future however,
we envision that gateways will have the ability to enter so called
“sleep mode”. Such gateways operate as usual when they are in ac-
tive use and switch to a sleep mode to preserve power when no user
activity is present. While in practice the gateways can be active
due to normal Internet use, in the most pessimistic case for NaDa
the gateways can serve the content only when the same gateway is
used to retrieve some other content. As we shall see, even in such a
pessimistic scenario we can achieve up to 15-30% energy savings.
It may be necessary to provide incentives for home users to host
NaDa gateways, especially since the required power is paid for by
the users. This could be achieved through “miles” or service dis-
count schemes introduced to reward users. However, as we can see
in the Figure 11, the typical daily load on a gateway is about 4000
megabits of traffic which translates to an additional 2.5 kW/hour a
year. Given the price range of 10-20 cents per kW-hour, this addi-
tional energy cost to a single home user is not significant.
The tracker coordinates all VoD-related activities in NaDa. It
monitors the availability and content possession of gateways, an-
swers requests for content by lists of gateways holding the desired
content. It is then up to the requesting gateway to download movie
parts from other gateways. Such downloads are performed at a
rate equal to the minimum of the spare uplink bandwidth and the
streaming rate; if spare uplink bandwidth is less than streaming
rate, other gateways are contacted in parallel. When the total down-
load rate from gateways is less than the streaming rate, content
servers provide the missing data at the desired speed.
The tracker also has the role to inform gateways of suitable con-
tent updates to perform. Such content updates are performed at low
network utilization periods, possibly by using multicast. A cached
content copy at the gateway can be retrieved numerous times thus
rendering content update costs marginal.
Content servers provide the content from legacy data centers
or caches. Content servers can belong to the entity that manages
the NaDa platform, or to content providers. Their primary role
is to pre-load gateways in offline mode with content that they can
subsequently re-distribute. Content servers can also serve content
requests online if no gateway can treat the request.
We evaluate NaDa platform by performing content placement
and access simulation in a large metropolitan area served by one
PoP and a few thousands of users. As a consequence, each user
is assumed to have identical network distance to every other user
in a network (this is what would happen on a mid-sized metropoli-
tan area network). Similarly every user has the same distance to
the content servers. We describe the data sets we use to drive the
simulations and then proceed to the VoD placement and serving
4.2 Datasets
We use content access traces from three sources (see Table 2):
(1) Netflix movie database, (2) Youtube traces [8], and (3) IPTV
access statistics from a large ISP [9]. Netflix movie database con-
tains the number of rentals for each movie - we interpret a rental as
a content access. For each movie in Netflix database, we assume
the same duration, namely 90 minutes. Youtube traces contain both
view count and the length of each video object. The IPTV statis-
tics represent a service with approximately 250,000 customers and
contains anonymized 60 day-long trace. We take a random sample
of 2,000 users and we treat each TV program as a separate ob-
ject. While today IPTV programs are mostly streamed to users
using multicast, service evolutions such as catch-up TV will make
IPTV look more and more as another VoD service. Finally we use
amixed trace which is an equal mix of all the content in the three
traces mentioned above.
Figures 8 and 9 show length and popularity distributions for the
traces. Netflix is not plotted in the length figure as all movies are
assumed to have the same length. Length of IPTV content is almost
constant, while Youtube content duration is more variable. The sec-
ond dip in the Youtube curve reflects the 10 minute limit imposed
Trace Objects Avg. len (s) Avg. views
IPTV 65,486 2246.2 2127.9
Youtube 1,686,264 166.5 2191.1
Netflix 17,770 5400 6555.5
Table 2: Content trace properties.
on uploaders. Some privileged users, however, can upload longer
Figure 8: IPTV and Youtube content length.
Figure 9: Content view counts.
Figure 9 represents content popularity, measured in view counts.
The heaviest tail is observed in the Youtube content, while IPTV
and Netflix offer slightly more uniform, but still highly skewed
popularity distributions.
We perform two transformations to our content traces before we
feed them into the simulator. First, we normalize the content to
match the number of users in the simulated network. We use the
data collected by Cha et al. [9] to determine the average amount
of time users spend daily watching IPTV, which we find to equal
to approximatelly 100 minutes. We then multiply this average by
the number of users in our population to get the aggregate number
of content-hours. We then use this aggregate number of hours as
a target to scale down our original traces. Once the number of ac-
cesses per item in the trace is normalized, we spread the accesses
throughout the span of a single day to reproduce the daily use pat-
tern observed in Figure 4.
4.3 Content placement strategy
The content must to be split into smaller chunks called data win-
dows to make it more manageable. We use the following conven-
tion for content formatting in NaDa system. Each movie is split
into data windows, whose length is a parameter wthat we let vary
from 10 to 120 seconds. The number of windows constituting a
movie is then the ratio of movie duration by parameter w.
We perform placement by first determining, for each movie f,
the number of movie replicas nfwe target to store overall on gate-
ways. This in turn gives us the number of data windows we store
in the system: if nfis an integer, each window is replicated exactly
nftimes. For fractional nf, a subset of windows is replicated nf
times, while other windows are replicated nf+1times, in or-
der to have approximately nftimes the full movie size replicated
overall. Given the number of replicas of each window of content,
we then place these at random on gateways, under the constraint
imposed by memory resources at each gateway.
Optimization formulation. We now explain the rationale we
used to determine the replication number nfof each movie f.We
assume that information is available on movie popularity, captured
by the average number ρfof simultaneous viewings of movie f
during the peak hour. The latter could be defined by the 95-th per-
centile, say, of the daily distribution of viewings. Based on the
diurnal activity patterns in Figure 4, we would take the peak load
equal to three times the average load over a day.
Let nfdenote the number of copies of movie fstored in total.
We phrase the problem of selecting such parameters nfas an op-
timization problem, whose objective is to minimize the number of
requests that must be served by the infrastructure. To this end, we
list some constraints that are natural to the problem. First, if nfis
allowed to be less than 1 (i.e. only a subset of the original movie
is stored within NaDa), then the number of simultaneous requests,
xf, that can be served from gateways, must satisfy
Indeed, each one of the ρfrequests can’t receive from gateways
more than the amount nfstored there in total.
Second, denote by Lfthe length of movie fin seconds. Then for
an upstream bandwidth of uand a streaming rate of r, the number
xfof simultaneous requests that can be served from gateways must
Indeed, the first term u/r reflects the fact that a window of con-
tent stored on a given box can be streamed at rate rsimultaneously
to at most u/r downloaders. Now, the number of boxes holding
such a window of content is precisely nftimes the number of win-
dows per replica of the movie, that is Lf/w.
Another effect we account for is the following. A fraction nf/N
of the content needed for the ρfviewings is available on the gate-
way of the user requesting the content, where Nis the total num-
ber of gateways in the system. The residual number of requests
that need to be served from either other gateways, or infrastructure
servers, is given by
f=ρf(1 nf/N ).(7)
(a) Imagenio IPTV. (b) Youtube. (c) Netflix.
Figure 10: Representative diurnal patterns of energy use. NaDa energy use is further split into server and gateway components.
Energy is accounted in 1 minute intervals.
Our optimization formulation is then the following:
Minimize X
where ρ
fis given by (7). Minimization is over the non-negative
variables nf,xf,forallfin the whole catalogue F, under con-
straints (5), (6), and the following:
nfLfsN (9)
where sdenotes the storage of each gateway, to which we add a
constraint on total usage of uplink bandwidth:
The “hot-warm-cold” placement method. The optimization
problem (5-10) is a standard linear program. It can be shown us-
ing elementary properties of linear programming that an optimal
solution has the following structure.
Assume that movies fin the catalogue Fare sorted by decreas-
ing popularity normalized by duration, i.e. ρ1/L1ρ2/L2....
Then movies should be partitioned into three groups: the most pop-
ular f1movies, constituting “hot” content, should be replicated on
all gateways. The subsequent f2most popular movies, constituting
“warm” content, are replicated minimally so that all their requests
can be served from gateways, that is:
Finally, the less popular movies, constituting “cold” content, are
not stored within NaDa. The sizes f1and f2of the “hot” and
“warm” groups are determined so that constraints (9-10) are met
with equality, where xfissetequaltoρ
We now comment on general properties of this placement strat-
egy. When storage is scarce, the “hot” group can vanish; at the
other extreme, with massive storage most content ends up being
replicated everywhere.
Concerning the “warm” content, for small window size wrela-
tive to movie length Lf, constraint (6) becomes irrelevant, and as
a result the optimal number of replicas nfin (11) ends up equal to
1. That is, a single copy is stored. While this may seem counter-
intuitive, numerical evaluations to be presented next will confirm
that this suffices to serve most requests.
Parameter Va l u e s Notation
Network size 1,000-30,000 users N
Upstream bandwidth 0.1-2Mbps u
Streaming rate 0.1-8Mbps r
Gateway storage 100-128,000MB s
Window length 10-120sec w
Simulation duration 86400sec d
Peak to average ratio 3p
Table 3: Simulation parameters.
Note finally that in the case of gateways with efficient sleep
mode, the above placement strategy should be adjusted as follows.
If at peak hour a fraction of gateways is expected to be on, then in
Equations (5–7) variable nfshould be replaced by nf.Thecor-
responding optimal placement strategy has the same structure as
before, the expression for nfin (11) being scaled up by the factor
4.4 Simulation environment
We implement event-driven simulator to investigate NaDa per-
formance and dimensioning for VoD services. The simulations
consist of two phases: (1) object placement which is performed of-
fline, according to the mechanisms described in the previous sub-
section; and (2) trace-driven simulation using our IPTV, Netflix,
YouTube and mixed traces, transformed to reflect user video watch-
ing time and diurnal variations as we previously explained.
Each simulation represents 24 hours of user activity. We vary
NaDa system parameters within the range given in Table 3. We
limit the number of users to 30,000 in in our largest simulations
order to make the simulations scalable. All other values have been
chosen to match realistic operational settings. The output of the
simulation is then used to calculate energy consumption, using the
models of Section 3 and the parameters listed in Table 4. We
set hop count to 2 for gateway-to-gateway communication, and 4
for gateway-to-server communication in ISP scenario; and we set
counts to 9 and 14 for WAN scenario. These parameters are sum-
marized in Table 4.
Unless otherwise noted, in this section we run simulations using
video streaming rate of 2Mbps, a network of 6000 Gateways, each
equipped with 8GB storage, 1Mbps upstream and hosting content
with window size of 60 seconds. The default workload, if none
Figure 11: Bandwidth to the servers and to the NaDa system as we increase storage in each Gateway. Left graph represents band-
width using legacy Gateways, right graph represents a pessimistic scenario for Gateways with with sleeping support.
Parameter Va l u e Notation
Hops to servers ISP:4, WAN:14 hcdn
Hops between clients ISP:2, WAN:9 hn
Hop(router) energy/bit 150 109J γr
Server energy/bit 211 109J γs
NaDa energy/bit 100 109J γn
Data center PUE 1.2 PUE
Home energy PUE 1.1
Table 4: Parameters for energy computation.
specified, as achieved by using a mixed trace from our three data
sources. We also use router distances from an ISP network as de-
scribed in Table 1. All the traces in the simulations are normalized
to represent a daily population as described in Section 4.2. We
provide results both for legacy Gateways (Figures 10, 11 and 12),
expected to be always on, and for Gateways with efficient sleep
mode (Figures 11, 13 and 14).
5.1 Effects of daily usage patterns
Figure 10 illustrates energy consumption pattern throughout the
day for three different content traces. The top curve in Figure 10
represents the energy used when relying on a traditional Data Cen-
ter. Going down, the second line shows the energy consumed with
NaDa. The bottom two plots show the fraction of NaDa energy due
to the requests directed to infrastructure and gateways respectively.
We first observe that benefits of NaDa are larger at high activ-
ity periods. At peak time, using default settings described above,
NaDa saves around 30% of the legacy Data Center energy. Fig-
ure 10 also shows how NaDa can smoothen the load on the net-
work core by reducing its peak use. While hard to quantify, such
smoothing enables to save on investments by delaying upgrades of
network capacity and equipment.
A second observation is that the benefits of NaDa change with
the characteristics of the content. Savings are the smallest for the
Youtube trace. This may sound counter-intuitive, as Youtube has
the most skewed popularity distribution of the three traces (i.e.
relative popularity of popular content is higher than for IPTV or
Netflix). Skewness is advantageous to NaDa, as it increases the
predictability of which content is more likely to be accessed. How-
ever, Youtube also has the largest catalogue, and requests for “cold”
content in Youtube more than offset requests for “hot” content.
5.2 Proportion of “on” Gateways
NaDa makes use of storage and uplink bandwidth of active Gate-
ways. As discussed in Section 2, legacy Gateways are (almost)
always on, and thus always available as NaDa resources. If we
transition to Gateways with efficient sleep mode (we shall refer to
these as “sleeping Gateways”), fewer resources will be available to
NaDa. At the very least, Gateways of customers requesting content
must be on, and thus NaDa capacity scales with demand.
Figure 11 shows side to side the efficiency of NaDa system with
legacy and with sleeping Gateways in terms of load served, as we
vary the storage per Gateway. In a legacy scenario, NaDa handles
all demand with as little as 6GB of storage per Gateway. NaDa
with sleeping Gateways has lower bandwidth efficiency and takes
18GB of storage per Gateway to attract approximately 1/3 of the
total load. Additional storage can in principle lead to all demand
being served by NaDa (e.g. when all content ends up replicated
everywhere), but in practice efficiency increases very slowly with
additional storage above 18GB.
This feature stems from the fact that with our default settings,
video streaming rate (2Mbps)is larger than upstream rate (1Mbps).
When the upstream to video streaming rate ratio is larger than one,
far less storage is needed for all load to be diverted from Servers
to Gateways. This can be seen on Figure 14, which shows over-
all energy savings with sleeping Gateways and various choices of
upstream and video streaming rates.
5.3 Storage and window size
While Figure 11 shows bandwidth use in NaDa, Figure 12 trans-
lates such bandwidth to overall energy consumed per user. Zero
storage corresponds to traditional system without NaDa support.
As expected, additional storage decreases energy use, down to the
point where the benefit of additional storage becomes negligible.
This happens at around 4GB for all three traces, in the considered
Further increases in efficiency are driven by availability of local
copies of content. Note that for Youtube content, energy consump-
tion decreases more slowly than for the other traces because of the
relatively huge size of the Youtube catalogue. Nevertheless, 4GB
of memory is small and affordable, and achieves a 55% savings
with legacy Gateways.
We now discuss the impact of window size variation (introduced
in Section 4.3) on the efficiency of NaDa. Figure 12 displays a
“perfect” split curve for the mixed dataset. Perfect split here means
that the content is divided into infinitely small chunks. Such split-
ting results in an idealized system where the collection of Gate-
ways behaves as a single server with aggregate upstream capacity
of uN, equal to the sum of Gateway uplink bandwidths. Indeed,
contention at the individual Gateway level decreases with window
size, since the smaller the window, the less likely it is that distinct
requests compete for the same Gateway’s uplink bandwidth.
As can be seen in Figure 12, the energy consumption for a win-
dow size of 60 seconds follows closely the energy consumption of
the idealized system. From the viewpoint of bandwidth utilization,
window sizes of 60 seconds are already small enough for an effi-
Figure 12: Energy usage with legacy Gateways as a function
of storage. *Mixed trace represents a “perfect splitting” with
infinitely small window size.
Figure 13: Energy savings for NaDa with sleeping Gateway
support using different assumptions about the underlying net-
cient operation. While smaller windows may have other advantages
– for instance reduced start-up delay – they introduce management
overhead at the tracker and at the Gateways.
5.4 Network and population size
We now discuss the impact of network hop count and user pop-
ulation size on savings. Figure 13 shows NaDa efficiency with dif-
ferent assumptions about the underlying network.
WAN scenario assumes 14 router hops to a server and 9 router
hops between two clients. ISP scenario assumes 4 router hops to
the server and 2 router hops between clients. The last scenario in
Figure 13 measures NaDa efficiency only as it applies to servers
and Gateways. We observe that the introduction of the network
into the equation results in reduced overall relative savings. How-
ever, even in the most adverse scenario, we achieve more than 20%
energy savings.
Energy savings also occur as we increase user population partic-
ipating in NaDa. Figure 14 shows how energy efficiency grows as
we increase the number of users. We observe that with 9000 users,
video rate of 0.7Mbps and upstream rate of 1Mbps NaDa achieves
39% energy savings. As we increase the network size to 27000
users the energy savings grow to 44%. As more Gateways partic-
ipate as clients, more Gateways can act as servers, thus increasing
the likelihood of finding needed content.
In the legacy system the savings with larger population size
would grow faster. With limited Gateway storage, NaDa perfor-
mance is memory-bound rather than bandwidth-bound (at least for
small populations). As population grows, the total memory avail-
able to the system increases, and the corresponding memory limi-
tation is relaxed.
5.5 Streaming and uplink rates
Each set of three groups of bars in Figure 14 shows energy ef-
ficiency as a function of video streaming rate and upstream band-
width available to Gateway. Energy savings increase steeply until
we hit the point where the upstream rate is equal to the stream-
ing rate. For example, for 18,000 users switching from 1Mbps to
2Mbps service at video rate of 2Mbps increases savings from 15%
to 25%. On the other hand switching from 1Mbps to 2Mbps at a
video rate of 8Mbps improves savings only insignificantly. With
upstream rate less than streaming rate, individual Gateways serve
windows of content at a rate below the streaming rate. One thus
needs to download content from several Gateways simultaneously
if one wants to obtain it fast enough from NaDa. This causes more
contention for Gateway uplink bandwidth, and hence reduces the
efficiency at which Gateway uplink bandwidth is used.
Figure 14: Energy savings for NaDa with sleeping Gateway
support using different population sizes with varying video and
upstream rates.
Component level energy savings in the IT equipment is an ac-
tive research area, with ideas ranging from dynamic voltage Scal-
ing [29] and memory power management [15], to operating system
scheduling [21]. A holy grail of efficient energy use is a energy
proportional computing [7], where the overall energy consumption
of a system is linearly proportional to the load. Achieving this goal
would eliminate the waste of baseline power in servers that are not
100% utilized. This objective is strongly related to our goal of
leveraging baseline power wasted today in Gateways. Energy pro-
portionality is a challenging objective and it remains to be seen
how quickly it will move forward. Until then, new energy-aware
distributed systems techniques like the one put forth by NaDa will
be needed for keeping energy consumption under control.
Still at the scale of individual components, the authors of [17]
address a networking issue, namely energy consumption in LAN
Ethernet cards. They make interesting proposals, suggesting to use
proxying and splitting to put LAN cards to sleep without loosing
existing TCP connections.
At a larger scale, one of the early works to discuss aggregate
network consumption was a 2004 position paper on “Greening of
the Internet” [18] that coined the idea of energy efficiency of In-
ternet components and protocols. More concrete subsequent works
include [10] and [25]. The first one looks at energy gains by recon-
figuring Internet topologies either dynamically through routing or
earlier at the stage of network design. The second proposes delay-
ing batches of traffic at buffers at the network edges so as to allow
time for core network components to sleep and save energy. The
above works all approach energy savings essentially at the network
layer and below, and thus differ substantially from our application-
level approach in NaDa.
One of the early works on energy-efficient Data Center designs
is [11] whereas more recent results and an overview of the current
state of research efforts can be found in [16]. These works take the
monolithic paradigm of Data Centers for granted and look at how
to improve its energy efficiency. Our work has suggested that there
are fundamental limitations inherent to the paradigm itself and has
proposed the NaDa approach as an alternative or a partial substitute
to monolithic Data Centers.
To the best of our knowledge the most directly related works to
ours are two 2008 position papers [20] and [12]. Both of these
works coin the idea of breaking and distributing monolithic Data
Centers. The first one goes all the way to suggesting using edge
devices like in NaDa, whereas the second one stays at the coarser
granularity of container-sized mini Data Centers. Our work is
aligned with these approaches and extends them by providing a
thorough quantification of gains based on concrete application sce-
narios and with the use of extensive measurement data.
We have introduced NaDa, a new communication architecture
for Internet content and service delivery. In NaDa, content and ser-
vices are stored on home gateways instead of data centers. The
access to these services on gateways is provided by using a man-
aged P2P infrastructure. NaDa greatly reduces the need for content
servers and network resources. We use real-life service load traces
and extensive simulations to show that NaDa can save at least 20-
30% of the energy spent by legacy data centers. Our models show
that each gateway requres just modest memory upgrades.
A number of questions need to be answered before NaDa can be
effectively deployed. First, users may be concerned that the ISP
takes advantage of their own electricity to offer services to others,
and may need to be “incentivized” to accept NaDa. Simple rewards
schemes could be used to this end, motivating users through aware-
ness programs or rewarding them with service credit equivalent to
their energy use increase.
A second issue is that of uplink bandwidth and Gateway capacity.
NaDa critically relies on these resources, and it requires substan-
tial investment from an ISP to secure the necessary uplink band-
width, and to deploy powerful enough Gateways. However, from
discussions with ISPs, these investments could indeed be made: the
Gateway is strategically located as the hub to the Internet, and ISPs
may want to leverage this strong position to offer home networking
services. ISPs currently providing triple-play services are in fact
considering deployment of powerful Gateways with memory and
multicore processors in order to provide more services. In addi-
tion, FTTH deployment is making progress and removes the uplink
bandwidth limitation, as well as simplifies the network architecture
(while DSLAMs can theoretically perform routing tasks, most of
the gateway-to-gateway communication as of today still has to go
up to the router to which the DSLAM is connected).
We therefore believe that NaDa not only can, but also should
happen, given its potential. We are currently working on prototypes
of advanced Gateways and on fine-tuning the tracker and placement
mechanisms to deliver VoD services.
[3] Data Center Report to Congress.
[4] Google Data Center Power Usage Efficiency.
[5] Thomson’s sapphire server data sheet.
CDT-3039D- 2.pdf.
[6] Uptime Institute.
[7] L. A. Barroso and U. Hölzle. The case for
energy-proportional computing. Computer, 40(12):33–37,
[8] M. Cha, H. Kwak, P. Rodriguez, Y.-Y. Ahn, and S. Moon. I
tube, you tube, everybody tubes: analyzing the world’s
largest user generated content video system. In IMC ’07:
Proceedings of the 7th ACM SIGCOMM conference on
Internet measurement, pages 1–14, New York, NY, USA,
2007. ACM.
[9] M. Cha, P. Rodriguez, J. Crowcroft, S. Moon, and
X. Amatriain. Watching television over an ip network. In
IMC ’08: Proceedings of the 8th ACM SIGCOMM
conference on Internet measurement, pages 71–84, New
York, NY, USA, 2008. ACM.
[10] J.Chabarek,J.Sommers,P.Barford,C.Estan,D.Tsiang,
and S. Wright. Power awareness in network design and
routing. In INFOCOM 2008. The 27th Conference on
Computer Communications. IEEE, pages 457–465, 2008.
[11] J. S. Chase, D. C. Anderson, P. N. Thakar, A. M. Vahdat, and
R. P. Doyle. Managing energy and server resources in hosting
centers. In Proc. of ACM SOSP ’01, pages 103–116, 2001.
[12] K. Church, A. Greenberg, and J. Hamilton. On delivering
embarrassingly distributed cloud services. In Proc. of
HotNets-VII, 2008.
[13] Cisco Visual Networking Index. http://newsroom.
[14] M. Dischinger, A. Haeberlen, K. P. Gummadi, and S. Saroiu.
Characterizing residential broadband networks. In IMC ’07:
Proceedings of the 7th ACM SIGCOMM conference on
Internet measurement, pages 43–56, New York, NY, USA,
2007. ACM.
[15] X. Fan, C. Ellis, and A. Lebeck. Memory controller policies
for dram power management. In ISLPED ’01: Proceedings
of the 2001 international symposium on Low power
electronics and design, pages 129–134, New York, NY,
USA, 2001. ACM.
[16] X. Fan, W.-D. Weber, and L. A. Barroso. Power provisioning
for a warehouse-sized computer. In ISCA ’07: Proceedings
of the 34th annual international symposium on Computer
architecture, pages 13–23, New York, NY, USA, 2007.
[17] C. Gunaratne, K. Christensen, and B. Nordman. Managing
energy consumption costs in desktop pcs and lan switches
with proxying, split tcp connections, and scaling of link
speed. Int. J. Netw. Manag., 15(5):297–310, 2005.
[18] M. Gupta and S. Singh. Greening of the internet. In
SIGCOMM ’03: Proceedings of the 2003 conference on
Applications, technologies, architectures, and protocols for
computer communications, pages 19–26, New York, NY,
USA, 2003. ACM.
[19] J. Hamilton. Where does power go in high-scale data
centers? Tutorial, ACM Sigmetrics, 2009.
[20] N. Laoutaris, P. Rodriguez, and L. Massoulie. Echos: edge
capacity hosting overlays of nano data centers. ACM
SIGCOMM Comput. Commun. Rev., 38(1):51–54, 2008.
[21] J. R. Lorch. Operating systems techniques for reducing
processor energy consumption. PhD thesis, 2001.
Chair-Smith, Alan Jay.
[22] S. McAllister, V. P. Carey, A. Shah, C. Bash, and C. Patel.
Strategies for effective use of exergy-based modeling of data
center thermal management systems. Microelectron. J.,
39(7):1023–1029, 2008.
[23] J. C. Mogul. Improving Energy Efficiency for Networked
Applications, Dec. 2007. Keynote Address at ACM/IEEE
ANCS 2007.
[24] S. Nedevschi, L. Popa, G. Iannaccone, S. Ratnasamy, and
D. Wetherall. Reducing network energy consumption via
sleeping and rate-adaptation. In NSDI’08: Proceedings of the
5th USENIX Symposium on Networked Systems Design and
Implementation, pages 323–336, Berkeley, CA, USA, 2008.
USENIX Association.
[25] S. Nedevschi, L. Popa, G. Iannaccone, S. Ratnasamy, and
D. Wetherall. Reducing network energy consumption via
sleeping and rate-adaptation. In NSDI’08: Proceedings of the
5th USENIX Symposium on Networked Systems Design and
Implementation, pages 323–336, Berkeley, CA, USA, 2008.
USENIX Association.
[26] A. Sinha and A. P. Chandrakasan. Energy efficient real-time
scheduling. In ICCAD ’01: Proceedings of the 2001
IEEE/ACM international conference on Computer-aided
design, pages 458–463, Piscataway, NJ, USA, 2001. IEEE
[27] K. Suh, J. F. Kurose, L. Massoulie, D. Towsley, M. Varvello,
C. Diot, and C. Neumann. Push-to-peer video-ondemand
system: design and evaluation. Technical report, In UMass
Computer Science Techincal Report 2006â ˘
S59, 2006.
[28] Z. Wen, S. Triukose, and M. Rabinovich. Facilitating focused
internet measurements. In SIGMETRICS ’07: Proceedings of
the 2007 ACM SIGMETRICS international conference on
Measurement and modeling of computer systems, pages
49–60, New York, NY, USA, 2007. ACM.
[29] B. Zhai, D. Blaauw, D. Sylvester, and K. Flautner.
Theoretical and practical limits of dynamic voltage scaling.
In DAC ’04: Proceedings of the 41st annual conference on
Design automation, pages 868–873, New York, NY, USA,
2004. ACM.
... According to [6] a data center's average usage is 20 to 30 %. Also, resources are using energy even if there is no traffic flow [7]. Many solutions for energy efficiency in DCNs are available [8][9][10][11][12]. Due to the large amount of energy consumption in data center networks has become abundant [13][14][15]. ...
... The rest of the devices are put to power off to save energy. Energy-aware routing algorithm [13], Energy optimization on-device interface approach [30], Elastic tree [9], Artificial intelligence abstraction model for Energy efficiency in DCN [2] and traffic merge [31] approaches are some examples in this domain. ...
... One of the energy aware solutions is elastic tree [9]. In this approach, the authors have focused to find a minimum number of communication devices that can be used with a lesser impact on network performance. ...
Full-text available
The world is moving towards cost-effective and time-constrained solutions. The uses of applications and automated devices have been growing day by day. In computing, resources available in personal computers are limited due to less storage capacity and lower computation speeds. Using all applications on personal systems may not be cost-effective. Therefore, the trends of online storage and computing have become popular. On the other hand, there must be some serving end for these users. One of the major issues, due to the growth of data centers is the increase in power usage of a larger number of servers and network devices. These devices are power-hungry and consume energy even during idle hours even if there are no network traffic loads. The cost of energy used and dissipated is increased in this situation. In this paper, we have given a solution for efficient usage of energy efficiency in data center networks based on traffic loads. We have proposed a model to use traffic intensity to decide the number of machines inactive conditions so that we can save the energy consumption in data center networks. We have implemented this proposed model and simulated it to validate it.
... University (CMU), where the virtualized data center was very small, and in distributed fashion, serves users near the edge (Bilal, Khalid, Erbad, & Khan, 2018;Ren, Zhang, He, Zhang, & Li, 2019;Satyanarayanan, Bahl, Caceres, & Davies, 2009). These terms are also related to other similar objects and concepts like Nano-data centers (Valancius, Laoutaris, Massoulié, Diot, & Rodriguez, 2009). Edge technologies can work together and cooperate, even though various edge technologies are defined independently (Borcoci, 2016). ...
There is a long history of engagement between computing and healthcare, adopting telemedicine is slow due to political willingness, limited infrastructure development frameworks, and availability. The Internet of Medical Things (IoMT) is going to be one of the fastest developments, and it’s anticipated to bring in the largest delivery of technology in existence. Hence, human-to-machine (H2M), machine-to-machine (M2M), and person-to-person (P2P) interactions can be fully updated along with telemedicine for the betterment of society in general. The use of IoMT-based sensors helps in the real-time detection of diseases and has significantly reduced the mortality rate. To meet the requirements of low latency and energy efficiency during medical data analysis, edge computing combined with 5G speeds is the answer. Edge computing has contributed greatly in the areas of low latency, massive connectivity of devices, and higher data rates in a network for ultra-reliable communication in the smart healthcare system. Therefore, this chapter discusses the areas of applicability and several extraordinary opportunities brought by the edge-enabled IoMT-based system in the healthcare system. The chapter also discusses the research challenges of the deployment of Edge IoMT-based system in the healthcare system and proposes an edge-enabled IoMT-based system framework. The application of the edge-enabled IoMT system brings end users and data sources running at a distance closer to the network nodes. The proposed framework can be used for real-time data capture, diagnosis, and monitoring patients’ health conditions like body motion, speech signals, body temperature, blood pressure, blood glucose, and heart rate, among others using various devices and sensors. Besides, the system can be useful in the emergency cases such as heart attacks, hysteria, anxiety, and epilepsy, thus saving the life of any patient.
... It provides a self-contained and secure computing environment to run customer applications and generally fits into one or two racks. [43] compared the energy consumption of an NDC enabled system with traditional cloud data centers and asserted that using NDCs brings a reduction of around 20-30% in the total energy consumption. While comparing, authors cited several reasons to reduce energy consumption such as exploiting underutilized gateways, reduction in cooling cost, and network energy consumption, etc. ...
Full-text available
The advancements in sensing technologies, smart devices, wearable gadgets, and communication paradigm enable the vision of the internet of things, smart city, virtual and augmented reality, pervasive healthcare, to name a few. These applications have strict requirements of low latency delivery, high data rate, and instant response. To support this, various new technologies, such as fog computing, mobile edge computing, cloudlet, Micro, and Nano centers, mini and micro clouds, etc., have emerged. The entire set of emerging edge computing paradigms are commonly referred as "edge technologies" in which computational resources and storage are closer to the user/terminal devices somewhere between the device and the cloud data center. The edge technologies aim to deliver computing services with minimal delay by reducing the downward and upward time and data traffic volume. Like cloud service providers, edge service providers are emerging, and a market of edge computing resources has been created. Therefore, Auction theory, a subfield of Economics, is being widely applied for the allocation of resources in emerging edge technologies. This work presents a comprehensive survey on auction-based resource allocation and pricing approaches in emerging edge technologies. An overview of edge technologies and auction theory is given, followed by a thorough review and comparison of the existing auction-based approaches applied in edge technologies for resource allocation and pricing in terms of economic properties. Various open research issues have been deliberated to set the future research direction at the end.
... Edge computing deals with the services hosted on nodes that are held by Internet Service Providers (ISPs), and other smaller servers which are available near users in the network. The Nano Datacenters (NaDa) [66] model is one of the examples of the current trend of disruption of the Edge of complex network operations that were traditionally performed "Network-centric". ...
Full-text available
In recent years, there has been a dramatic change in attitude towards computers and the use of computer resources in general. Cloud and Edge computing have emerged as the most widely used technologies, including fog computing and the Internet of Things (IoT). There are several benefits in exploiting Cloud and Edge computing paradigms, such as lower costs and higher efficiency. It Provides data computation and storage where data is processed, enables better data control, faster understanding and actions, and continuous operation. However, these benefits seem to be appealing, and their effects on computer forensics are somewhat undesirable. The complexity of the Cloud and Edge environment and their key features present many technical challenges from multiple stakeholders. This paper seeks to establish an in-depth understanding of the impact of Cloud and Edge computing-based environmental factors. Software and hardware tools used in the digital forensic process, forensic methods for handling tampered sound files, hidden files, image files, or images with steganography etc. The technical/ legal challenges and the open design problems (like distributed maintenance, multi-tasking and practicality) highlight the various challenges for the digital forensics process.
Full-text available
Global warming is one of the most compelling environmental threats today, as the rise in energy consumption and CO2 emission caused a dreadful impact on our environment. The data centers, computing devices, network equipment, etc., consume vast amounts of energy that the thermal power plants mainly generate. Primarily fossil fuels like coal and oils are used for energy generation in these power plants that induce various environmental problems such as global warming ozone layer depletion, which can even become the cause of premature deaths of living beings. The recent research trend has shifted towards optimizing energy consumption and green fields since the world recognized the importance of these concepts. This paper aims to conduct a complete systematic mapping analysis on the impact of high energy consumption in cloud data centers and its effect on the environment. To answer the research questions identified in this paper, one hundred nineteen primary studies published until February 2022 were considered and further categorized. Some new developments in green cloud computing and the taxonomy of various energy efficiency techniques used in data centers have also been discussed. It includes techniques like VM Virtualization and Consolidation, Power-aware, Bio-inspired methods, Thermal-management techniques, and an effort to evaluate the cloud data center’s role in reducing energy consumption and CO2 footprints. Most of the researchers proposed software level techniques as with these techniques, massive infrastructures are not required as compared with hardware techniques, and it is less prone to failure and faults. Also, we disclose some dominant problems and provide suggestions for future enhancements in green computing.
The previous chapter explored the Cloud computing ecosystem (Cloud-to-thing continuum) as the mainstay of current and future technological solutions for eHealth systems. In this chapter, we delve into existing architecture models in which stream analytic scenarios can be deployed. Criteria for architecture selection are also discussed. The chapter illustrates different communication protocols most commonly used in stream applications, focusing on infrastructure and application protocols.
The importance of content delivery networks (CDN) continues to rise with the exponential increase in generation and consumption of electronic media. In order to ensure high quality of experience, CDNs often deploy cache servers that capable of storing some of the popular files close to users. Such edge caching solutions not only increase the content availability, but also result in higher download rates and lower latency. Different from the classical eviction-based algorithms, the present work formulates the content placement problem from an optimization perspective and puts forth an online algorithm for the same. In contrast to the existing optimization-based solutions, the proposed algorithm is incremental and incurs very low computation cost, while yielding storage allocations that are provably near-optimal. The proposed algorithm can handle time-varying content popularity, thereby obviating the need for periodically estimating demand distribution. Using synthetic and real IPTV data, we show that the proposed policies outperform the traditional eviction based techniques in terms of various metrics.
Computation offloading is a prominent exposition for the mobile devices that lack the computational power to execute applications that require a high computational cost. There are several criteria on which computational offloading can be performed. The common measures’ being load harmonizing at the servers on which task is to be computed, energy management, security and privacy of tasks to be offloaded and the most important being the computational requirement of the task. That being said more and more solutions for offloading use various machine learning (ML) and deep learning (DL) algorithms for predicting the best nodes off to which task is to be offloaded improving the performance of offloading by reducing the delay in computing the tasks. We present various computational offloading techniques which use ML and DL. Also, we describe numerous middleware technologies and the criteria's that are crucial for offloading in specific developments.
Full-text available
Internet hosting centers serve multiple service sites from a common hardware base. This paper presents the design and implementation of an architecture for resource management in a hosting center operating system, with an emphasis on energy as a driving resource management issue for large server clusters. The goals are to provision server resources for co-hosted services in a way that automatically adapts to offered load, improve the energy efficiency of server clusters by dynamically resizing the active server set, and respond to power supply disruptions or thermal events by degrading service in accordance with negotiated Service Level Agreements (SLAs).Our system is based on an economic approach to managing shared server resources, in which services "bid" for resources as a function of delivered performance. The system continuously monitors load and plans resource allotments by estimating the value of their effects on service performance. A greedy resource allocation algorithm adjusts resource prices to balance supply and demand, allocating resources to their most efficient use. A reconfigurable server switching infrastructure directs request traffic to the servers assigned to each service. Experimental results from a prototype confirm that the system adapts to offered load and resource availability, and can reduce server energy usage by 29% or more for a typical Web workload.
Full-text available
Very large data centers are very expensive (servers, power/cool-ing, networking, physical plant.) Newer, geo-diverse, distributed or containerized designs offer a more economical alternative. We argue that a significant portion of cloud services are embarras-singly distributed – meaning there are high performance realiza-tions that do not require massive internal communication among large server pools. We argue further that these embarrassingly distributed applications are a good match for realization in small distributed data center designs. We consider email delivery as an illustrative example. Geo-diversity in the design not only im-proves costs, scale and reliability, but also realizes advantages stemming from edge processing; in applications such as spam fil-tering, unwanted traffic can be blocked near the source to reduce transport costs.
Conference Paper
Full-text available
A large and rapidly growing proportion of users connect to the Internet via residential broadband networks such as Digital Subscriber Lines (DSL) and cable. Residential networks are often the bottleneck in the last mile of today's Internet. Their characteristics critically affect Internet applications, including voice-over-IP, online games, and peer-to-peer content sharing/delivery systems. However, to date, few studies have investigated commercial broadband deployments, and rigorous measurement data that characterize these networks at scale are lacking. In this paper, we present the first large-scale measurement study of major cable and DSL providers in North America and Europe. We describe and evaluate the measurement tools we developed for this purpose. Our study characterizes several properties of broadband networks, including link capacities, packet round-trip times and jitter, packet loss rates, queue lengths, and queue drop policies. Our analysis reveals important ways in which residential networks differ from how the Internet is conventionally thought to operate. We also discuss the implications of our findings for many emerging protocols and systems, including delay-based congestion control (e.g., PCP) and network coordinate systems (e.g., Vivaldi).
Conference Paper
Full-text available
For half a century, television has been a dominant and per- vasive mass media, driving many technological advances. Despite its widespread usage and importance to emerging applications, the ingrained TV viewing habits are not com- pletely understood. This was primarily due to the difficulty of instrumenting monitoring devices at individual homes at a large scale. The recent boom of Internet TV (IPTV) has enabled us to monitor the user behavior and network usage of an entire network. Such analysis can provide a clearer picture of how people watch TV and how the underlying networks and systems can better adapt to future challenges. In this paper, we present the first analysis of IPTV work- loads based on network traces from one of the world's largest IPTV systems. Our dataset captures the channel change ac- tivities of 250,000 households over a six month period. We characterize the properties of viewing sessions, channel pop- ularity dynamics, geographical locality, and channel switch- ing behaviors. We discuss implications of our findings on networks and systems, including the support needed for fast channel changes. Our data analysis of an operational IPTV system has important implications on not only existing and future IPTV systems, but also the design of the open In- ternet TV distribution systems such as Joost and BBC's iPlayer that distribute television on the wider Internet.
Conference Paper
Full-text available
User Generated Content (UGC) is re-shaping the way people watch video and TV, with millions of video producers and consumers. In particular, UGC sites are creating new view- ing patterns and social interactions, empowering users to be more creative, and developing new business opportunities. To better understand the impact of UGC systems, we have analyzed YouTube, the world's largest UGC VoD system. Based on a large amount of data collected, we provide an in-depth study of YouTube and other similar UGC systems. In particular, we study the popularity life-cycle of videos, the intrinsic statistical properties of requests and their re- lationship with video age, and the level of content aliasing or of illegal content in the system. We also provide insights on the potential for more efficient UGC VoD systems (e.g. utilizing P2P techniques or making better use of caching). Finally, we discuss the opportunities to leverage the latent demand for niche videos that are not reached today due to information filtering effects or other system scarcity distor- tions. Overall, we believe that the results presented in this paper are crucial in understanding UGC systems and can provide valuable information to ISPs, site administrators, and content owners with major commercial and technical implications.
The IT equipment comprising the Internet in the USA uses about $6 billion of electricity every year. Much of this electricity use is wasted on idle, but fully powered-up, desktop PCs and network links. We show how to recover a large portion of the wasted electricity with improved power management methods that are focused on network issues. Copyright © 2005 John Wiley & Sons, Ltd.
Conference Paper
Real-time scheduling on processors that support dynamic voltage and frequency scaling is analyzed. The Slacked Earliest Deadling First (SEDF) algorithm is proposed and it is shown that the algorithm is optimal in minimizing processor energy consumption and maximum lateness. An upper bound on the processor energy savings is also derived. Real-time scheduling of periodic tasks is also analyzed and optimal voltage and frequency allocation for a given task set is determined that guarantees schedulability and minimizes energy consumption.