Towards energy-awareness in managing wireless LAN applications
ABSTRACT We have investigated the scope for enabling WLAN applications to manage the trade-off between performance and energy usage. We have conducted measurements of energy usage and performance in our 802.11n WLAN testbed, which operates in the 5 GHz ISM band. We have defined an effective energy usage envelope with respect to application-level packet transmission, and we demonstrate how performance as well as the effective energy usage envelope is effected by various configurations of IEEE 802.11n, including transmission power levels and channel width. Our findings show that the packet size and packet rate of the application flow have the greatest impact on application-level energy usage, compared to transmission power and channel width. As well as testing across a range of packet sizes and packet rates, we emulate a Skype flow, a YouTube flow and file transfers (HTTP over Internet and local server) to place our results in context. Based on our measurements we discuss approaches and potential improvements of management in effective energy usage for the tested applications.
-
Conference Paper: The Effect of the 802.11 Power Save Mechanism (PSM) on Energy Efficiency and Performance During System Activity
[Show abstract] [Hide abstract]
ABSTRACT: 802.11 WLAN is a popular choice for wireless access on a range of ICT devices. A growing concern is the increased energy usage of ICT, for reasons of cost and environmental protection. The Power Save Mode (PSM) in 802.11 deactivates the wireless network interface during periods of inactivity. However, applications increasingly use push models, and so devices may be active much of the time. We have investigated the effectiveness of PSM, and considered its impact on performance when a device is active. Rather than concentrate on the NIC, we have taken a system-wide approach, to gauge the impact of the PSM from an application perspective. We experimentally evaluated performance at the packet level and system-wide power usage under various offered loads, controlled by packet size and data rate, on our 802.11n testbed. We have measured the system- wide power consumption corresponding to the individual traffic profiles and have derived application-specific effective energy-usage. We have found that in our scenarios, no significant benefit can be gained from using PSM.GreenCom 2012 - IEEE Intl. Conf. Green Computing and Communications; 11/2012 -
Conference Paper: Low RSSI in WLANs: Impact on Application-Level Performance
[Show abstract] [Hide abstract]
ABSTRACT: Widespread use of wireless LAN (WLAN) may soon cause an over-crowding problem in use of the ISM spectrum. One way in which this manifests itself is the low Received Signal Strength Indication (RSSI) at the WLAN stations, impacting performance. Meanwhile, the IEEE 802.11 standard is being evolved and extended, for example with new coding schemes and the 802.11n standard, which makes use of 5GHz and 2.4GHz. We report on measurements of the upper and lower bounds of performance with good and poor RSSI in 802.11g and 802.11n. We find that in operation under poor (low) RSSI, performance is indeed impacted. In some cases the impact is such that there may be little benefit in using the newer 802.11n over the mature 802.11g.ICNC 2013 - IEEE Intl. Conf. Computing, Networking and Communications; 01/2013 - SourceAvailable from: Saleem Bhatti
Conference Paper: The Case for Heterogeneous WLAN Environments for Converged Networks
[Show abstract] [Hide abstract]
ABSTRACT: We demonstrate that in a future converged network scenario, it may be beneficial to allow selection of 802.11 variant based on application requirements. We analyse traces from the campus network from the University of Twente, comprising ∼5000 users. We have evaluated a performance envelope derived from testbed experiments for individual IEEE 802.11 variants and compare these with the traffic patterns from the campus network. From our comparison, we find that specific IEEE 802.11 variants (e.g. 802.11g or 802.11n) may be better suited to specific applications, such as video streaming, rather than using a single WLAN standard for all traffic.ICNC 2013 - IEEE Intl. Conf. Computing, Networking and Communications; 01/2013
Page 1
Towards Energy-Awareness in Managing Wireless
LAN Applications
Markus Tauber, Saleem N. Bhatti, Yi Yu
School of Computer Science, University of St Andrews, UK
{markus,saleem,yi}@cs.st-andrews.ac.uk
Abstract—We have investigated the scope for enabling WLAN
applications to manage the trade-off between performance and
energy usage. We have conducted measurements of energy usage
and performance in our 802.11n WLAN testbed, which operates
in the 5 GHz ISM band. We have defined an effective energy usage
envelope with respect to application-level packet transmission,
and we demonstrate how performance as well as the effective
energy usage envelope is effected by various configurations of
IEEE 802.11n, including transmission power levels and channel
width. Our findings show that the packet size and packet rate
of the application flow have the greatest impact on application-
level energy usage, compared to transmission power and channel
width. As well as testing across a range of packet sizes and packet
rates, we emulate a Skype flow, a YouTube flow and file transfers
(HTTP over Internet and local server) to place our results in
context. Based on our measurements we discuss approaches and
potential improvements of management in effective energy usage
for the tested applications.
I. INTRODUCTION
The convenience of use of WLAN leads to its use in many
devices even those that are not mobile, e.g. new Internet-
enabled television sets. At the same time, there exists a very
large existing deployment of WLAN systems, spanning a
range of evolutions of the IEEE 802.11 standard. Various
IEEE 802.11 standards are widely implemented in many
consumer devices. This includes devices such as hand-held
games consoles, smartphones and televisions, as well as the
more ‘traditional’ uses in laptops/netbooks and desktops.
Meanwhile, there is world-wide concern for the increasing
carbon footprint of ICT systems, with current carbon emissions
similar to the aviation industry, and emissions set to triple
between 2002 to 2012 [1]. So, widespread and growing use of
IEEE 802.11 suggests that it is prudent to consider the energy
usage of systems incorporating the various 802.11 standards.
A. Motivation
Several power saving mechanisms which aim to enhance
energy efficiency have been proposed and are partially imple-
mented: e.g. 802.11 Power Save Mode (PSM) [2], Unsched-
uled Automatic Power Save Delivery (U-APSD) [3], WMM
Power Save (WMM-PS) [4], Dynamic MIMO Power Save
[5], and Wake-on-Wireless [6]. However, such features are
not widely implemented by all vendors, or are not widely de-
ployed, or are not easily accessible for use by non-expert users.
Also, there is a large base of legacy (i.e. old) equipment which
cannot support such features due to hardware constraints.
Additionally, in situations in which modern WLAN hard-
ware is used to provide key infrastructure (e.g. [7]), providers
may be reluctant to jeopardise performance by enabling new
energy-saving features [8], [9]. For instance, the generic IEEE
802.11 Power Save Mode (PSM), the 802.11n specific Spatial
Multiplexing Power Save (SMPS) and the Power Save Multi-
Poll (PSMP) mode implement algorithms which either put
the (redundant) WLAN interface into sleep mode or increase
buffers. Other approaches to use the WLAN resources more
efficiently include packet aggregation. [8], [9].
Another factor impacting the use of power-save modes with
sleep mechanisms is the growing popularity of cloud-based
services and the use of ‘push’ service paradigms, giving little
opportunity for the user’s system to enter sleep mode.
Even as WLAN equipment is upgraded, legacy kit will
remain, and the energy efficiencies that might be gained by
considering only the NIC, even when expertly configured, are
small compared to considering the system as a whole, e.g.
transmission power is 1mW-50mW and NIC power usage is
low, so NIC-only savings are also low. Thus, our motivation is
to determine the scope for application-level (self-)management
of energy usage on a system-side basis. Our eventual aim is
to identify management mechanisms that could be applied to
existing (legacy) deployments of 802.11, as well as to new
deployments, and is complementary to and independent of,
hardware power-saving schemes, such as those listed above.
B. Research focus
We take the position that it is possible to improve the energy
usage of applications by actions taken at the application level,
even if 802.11 power-saving enhancements are not available.
So, one of our objectives is to understand the performance
and energy usage dynamics of existing, commonly-used IEEE
802.11 equipment, with the overall goal that application-level
self-adaptation could provide energy efficiencies in 802.11 sys-
tems, for example by adjusting packet flow construction and
packet transmission [10]. As IEEE 802.11 is so widely used,
by applying functionality retrospectively to existing systems,
e.g. through software updates and/or patches, we believe that
considerable energy efficiencies might be achieved: even small
savings for individual devices could have a large impact if
applied universally.
Our previous work defines an application-level effective
energy usage envelope showing the energy usage of IEEE
802.11a and IEEE802.11n at 5 GHz [11]. In this study, we
978-1-4673-0269-2/12/$31.00 c ? 2012 IEEE
IEEE/IFIP NOMS 2012 - Network Operations and Management Symposium, Hawaii, USA, 16-20 Apr 2012
Page 2
focus on IEEE 802.11n at 5 GHz, to identify the dynamics
of the performance and the effective energy usage envelope in
response to changes in transmission power and channel width.
Note that we use off-the-shelf hardware with default configura-
tion and consider system-wide usage for a node, and not only
the NIC within that node. We adopt an empirical approach,
taking measurements of power usage for deducing effective
energy usage, and measuring system wide performance in our
testbed comprising off-the-shelf consumer equipment. While
we are concerned primarily with energy usage, analyses only
make sense when placed in the context of system performance
under different system workloads.
C. Contributions
The contributions of this paper are:
• An empirical study of the energy usage and performance
trade-off for a typical WLAN configuration of 802.11n.
• We identify the scope of adaptation that is possible for
an application. This acts to define boundaries for what is
achievable through management actions or interventions.
This is captured in an effective energy usage envelope,
which indicates the upper and lower bounds of what is
possible.
• An analyses of how a (self-)management system for
different classes of applications (or how different end-
system platforms) might make use of such information in
order to provide appropriate application-level adaptation
decisions. These could be policy-driven or might be
realised as an autonomic system.
We observe that performance and effective energy usage
vary greatly between WLAN configurations and traffic pro-
files, i.e. for different classes of applications. We find that
the kind of (self-)management adaptation that is possible may
need to be per-user, per-application and context specific. So,
MAC layer solutions, such as packet aggregation, that are
applied unilaterally might not necessarily provide the best
trade-off between energy usage and performance.
D. Structure of this paper
In Section II, we explain our methodology, describe our
testbed, our observables and define our energy usage metric. In
Section III, we present our observations and discuss what they
show. A discussion on how our findings can be exploited to
make applications energy-aware and enable self-management
is provided in Section IV. We present a summary of related
work in Section V, and conclude in Section VI.
II. APPROACH, TESTBED AND METRICS
Our eventual aim is to allow applications to be self-
managing. We design our experiments to investigate the scope
for self-adaptation of application-level flow characteristics.
To do that we adapt the testbed and methodology from our
previous work, further details of which can be found in
[11]. We consider that most deployed systems are used in
‘out-of-the-box’ configurations, without specific tuning for
energy efficiency. Many WLAN NIC drivers permit various
controls of the hardware features, but these might not be easily
accessible or comprehensible for modification by most users.
A practical constraint we have used is that of a 5 GHz-only
testbed. This was because in our local environment, we have
exclusive usage of 5 GHz and so the probability for biasses
of our measurements by interference was small.
A. Testbed
We have experimentally evaluated energy usage and per-
formance in our 5 GHz testbed. We generated packet flows
of offered loads with various bit-rates and packets sizes, and
measured power usage during the packet transmission. Our
testbed (Figure 1) consisted of a single client host, a host
running a wireless access-point (AP) and experimental control
units (only one shown in Figure 1) for monitoring the WLAN
environment, providing storage for measurement data, ntp1
services and system configuration. The WLAN hosts were
setup in a teaching lab in the University of St Andrews with
a distance of ∼24±0.5 m between the 2 dBi antennas.
Fig. 1. Schematic of test-bed showing physical connectivity. We used 5 GHz
only, and the testbed was configured separately for 802.11n (20 MHz and
40 MHz channels, as well as transmission (TX) powers of 0 dBm and 17 dBm)
experiments. The experiment controller uses Ethernet for control messages
and shared file-system access. Power meter readings are logged by experiment
controllers. The separation between the 2 dBi antennas of the client and access
point/server is 24±0.5 m. Data packets generated by iperf were transferred
across the WLAN link.
We wished to test 802.11n at 0 dBm (1 mW, minimum
RF power), 17 dBm (50 mW, typical maximum indoor RF
power), and with both a 20 MHz (default) channel and a
40 MHz channel, a total of 4 different combinations. This
means that all our experimental workloads in Table I are
executed 4 times, once with each of these combinations. Our
WLAN card uses the popular Atheros2chipset, now bought
by Qualcomm so likely to be used even more widely3. All
machines used Ubuntu 10.04 a minimal server distribution (no
desktop service daemons or GUI overhead), with the default
kernel 2.6.32-24-generic-pae, and updated WLAN modules
(compat-wireless-2011-05-02), which will soon be part of the
standard distribution.
1http://www.ntp.org/
2http://www.atheros.com/
3http://www.qualcomm.com/news/releases/2011/01/05/qualcomm-acquire-
atheros-leader-connectivity-networking-solutions
Page 3
B. Experiments
Packet generation and performance measurement for UDP
traffic was conducted using iperf4for which the AP was used
as the server. A wrapper script at the client executed iperf
and extracted throughput and loss for individual UDP flows
using iperf server reports. The specific packet sizes and bit
rates of the UDP workload are given in Table I. Motivation for
using UDP is its popularity for Voice and Video over IP (VoIP
and ViIP) applications and because it allows better control of
application specific offered load bit rates compared to TCP,
which is modulated by its congestion control behaviour.
TABLE I
GENERIC UDP WORKLOAD.
Packet size
Bit rate of the
offered load
64; 1460 bytes
32; 50; 100; 256; 512 Kbps
1; 5; 10; 15; 20; 25; 30; 35; 40; 45;
50; 60; 70; 80; 90 Mbps
Combining packet size and bit-rate gives 40 configurations; 5 measurements
for each gives 200 flows; 4 power/channel combinations gives 800 flows; each
flow had a duration of 4 minutes, giving over 53 hours of measurements.
We have used emulated flows for Skype, YouTube and
HTTP (Internet and Intranet), as summarised in Table II.
Traffic emulating a Skype (VoIP) flow was based on previous
studies [12], [13], as was traffic emulating a YouTube (ViIP)
flow [14], [15]. We do not incur the audio/video codec
overhead in our experiments. We have deduced HTTP-specific
downstream traffic profiles from preliminary experiments us-
ing wget5to generate HTTP flows from a local server and from
http://mirror.ox.ac.uk/ for downloading of an Ubuntu ISO CD
image file. For each of the above application specific traffic
profiles we have emulated 5 sequential UDP flows with iperf.
We have used a flow duration of 4 minutes based on previous
studies of VoIP and ViIP traffic stated above, but used this
duration for all traffic workloads for comparability.
TABLE II
APPLICATION UDP WORKLOAD EMULATION.
Skype
YouTube
HTTP (Internet)
HTTP (Intranet)
300 byte packets, 65 Kbps
1431 byte packets, 639 Kbps
1420 byte packets, 11 Mbps
1460 byte packets, 90 Mbps
Combining packet size and bit rate gives 20 configurations; with 4
power/channel combinations gives 80 flows; a flow duration of 4 minutes
gives a total of 320 minutes of measurements. Workload based on [12]–[15],
as well as on preliminary measurements.
C. Observed variables and metrics
We have measured the observables as described below:
• Performance: throughput and loss, as recorded by iperf’s
server reports on the client for each UDP flow.
• Power: every 30 seconds we have recorded the current
power consumption in Watts at the AP and client.
• WLAN spectrum: the signal strength (showing channel
utilisation) as recorded every 30 seconds via the USB-
connected spectrum analyser WiSpy6at one of the exper-
4https://sourceforge.net/projects/iperf/
5http://www.gnu.org/software/wget/
6http://www.metageek.net/products/wi-spy/
iment controllers. This was for initial calibrations; for
further on demand analysis; and also to confirm that
during the measurements, only our test-bed was operating
at 5 GHz, i.e. to spot possible interference from other
sources.
The monitoring intervals for all of the above observables
were deduced from preliminary experiments. Motivation for
doing this was to avoid excessive disk usage, but having
sufficient monitoring sample sets for determining significant
differences between experiments.
We measured power consumption on the client and the AP
at 30 second intervals. For power measurements, we used a
CC128 power meter7.
For assessing energy usage, we define effective energy usage
(EA) as follows:
EA=mean power used during transmission of flow
mean throughput of flow
EAhas units Joules/Mega-bit (J/Mb):
power in Watts
throughput in Mbps=
J/s
Mb/s= J/Mb
and the lower the value of EA, the better in terms of energy
usage. To generate values for EA, for each individual flow, we
use the following measurements:
EA=PA− PI
TA
(1)
PA
Mean power consumption measured during the trans-
mission of flow [Watts].
Mean power consumption measured for an idling
node [Watts] (was measured to be 48 Watts).
Mean throughput measured (using iperf) during flow
transmission [Mbps].
PI
TA
III. RESULTS AND DISCUSSION
We find that a higher application-specific bit rate results
in a lower effective energy usage, i.e low values of EA.
Conversely, low application-specific bit rates result in higher
effective energy usage, i.e. high values of EA. Of course, lower
values of EAare better.
For a given bit rate of offered load, increasing packet size
reduces the effective energy usage. This is shown in Figure 2,
which plots the mean value of EA over all experiments
(to show trends in effective energy usage), including the
operational points of the emulated applications. This is due
to amortisation of the transmission overhead and system-wide
energy usage across a greater number of transmitted bytes.
That is, we have derived EAbased on power measurements for
the system has a whole and then evaluated the effective energy
usage across the transmitted packets. This is in contrast to
other studies that have considered only the NIC, for example.
By considering the system as a whole, we are also in a
7http://www.currentcost.com/product-cc128.html
Page 4
good position to then suggest system-wide (self-)management
actions that can be applied at the application level and should
then yield reductions in effective energy usage across the
system as a whole.
0.1
1
10
100
500
0.1 1 10 100
avg. EA [J/Mb]
offered load [Mbps]
Skype
(300 B, 64 Kbps)
YouTube
(1431 B, 639 Kbps)
Threshhold 1, 1 Mbps
Threshhold 2, 5 Mbps
HTTP, Intranet
(1420 B, 11 Mbps)
HTTP, Internet
(1460 B, 90 Mbps)
64 B packets
1460 B packets
Fig. 2.
points of emulated application flows. Threshold 1 – up to which little/no
loss occurs; Threshold 2 – up to which large and small packets result in the
approximately the same throughput.
Mean values of EAfor small and large packets, with operational
We have identified a ∼1 Mbps threshold (Threshold 1 in
Figure 2) for bit rates of offered load up to which no significant
loss nor any difference with respect to throughput can be
identified due to different packet sizes. We also observe that
up to ∼5 Mbps, there is still little difference in throughput, but
there are bursts of loss, up to ∼20 % (Threshold 2 in Figure 2).
The figure also shows that EAranges from ∼500 J/Mb, for
low data rates (< 1 Mbps) with both large and small packets,
to ∼3 J/Mb, for small packets with high data rates, and
∼0.5 J/Mb for large packets with high data rates (90 Mbps).
We show in Table III how the highest mean throughput
relates to loss and EA in all of our experiments. Table III
shows that high throughput correlates to low effective energy
usage EA(maximum EA∼ 500 J/Mb), but also results in high
loss in each of our experiments with high TX power (17 dBm).
TABLE III
HIGHEST MEAN THROUGHPUT, WITH LOSS AND EAMEASUREMENTS FOR
64B AND 1460B PACKETS, IN ALL EXPERIMENTS, ARRANGED BY
CHANNEL WIDTH (CW) AND TX POWER (TX). POWER.
64 B packets
tmax
[Mbps]
3.8
6.7
5.7
6.6
1460 B packets
tmax
[Mbps]
19.3
72.3
35.5
73.4
CW
[MHz]
20
20
40
40
TX
[dBm]
0
17
0
17
loss
[%]
3.2
47.4
23.8
37.2
EA
loss
[%]
0.0
15.4
0.8
19.8
EA
[J/Mb][J/Mb]
10
2.3
2.9
6.2
0.7
0.5
1.1
0.5
A. Detailed Analyses
In all our graphs in Figures 4–7, we have: (i) offered load
on the horizontal axis (the configured values of offered bit
rate, from our generated workload); (ii) used standard error
bars, but in some cases, the error bars may be too small to
be visible. In Figure 4, we show throughput, loss and EA,
for a 20 MHz channel, with the left column showing results
measured at 0 dBm (1 mW) and the right column measured at
17 dBm (50 mW). Figure 5 shows the corresponding results
for a 40 MHz channel. The graphs for EAin Figures 4 and 5
show clearly the effective energy usage envelopes, the region
between the lines plotted for small packets and large packets.
For throughput, upto 1 Mbps, we see very little differ-
ence between the different systems configurations (TX power,
channel width), or workload (packet size, offered bit rate).
However, beyond this, we see the expected changes: maximum
TX power (17 dBm) and wider channel (40 MHz) give better
throughput than minimum TX power (0 dBm) and the normal
channel width (20 MHz). A counterintuitive observation is
that in all our experiments with high TX power (which also
equates to higher RSSI in our setup), higher throughput comes
at the cost of a higher loss rate. However, for our energy
usage metric, EA, we see the largest impact is made by
the adjustment of the application-level workload: increasing
packet size and increasing data rate yield better energy usage.
We have made a comparative analysis of the 802.11n config-
urations with respect to both transmission power and channel
width, which we will call a delta (∆) analyses. For transmis-
sion power, in Figure 6: (i) ∆ throughput was computed as the
normalised relation of throughput0dBm/throughput17dBm;
(ii) as the loss is already a normalised value, we have simply
computed ∆ loss as the difference of loss0dBm−loss17dBm;
(iii) for energy usage, ∆EA, was computed as the normalised
value of EA0dBm/EA17dBm. For channel width, in Figure 7:
(i) ∆ throughput was computed as the normalised relation of
throughput20MHz/throughput40MHz; (ii) as loss is already
a normalised value ∆ is the difference of loss20MHz −
loss40MHz. For energy usage, ∆EA was derived from the
normalised relation of EA20MHz/EA40MHz. Here, again, we
see that; (i) there is not much difference observed below 1
Mbps; (ii) the main differences occur due to the packet size
and data rates of the offered loads.
Overall, below ∼1 Mbps, we see that there is little dif-
ference in performance with respect to loss and throughput
across all the experiments. This applies to 20
40 MHz channels; small packets and large packets; and to low
and high transmission power. So, where application flows are
below 1 Mbps, there may not be much benefit in performance
by applying (self-)management actions in adjusting the packet
size or data rate. However, we do see a significant difference
(an order of magnitude) in effective energy usage for data rates
values from low data rates up to 1 Mbps. This means that, for
applications developers, the main incentive for applying (self-
)management actions is going to be energy efficient as there
is no loss (or gain) in performance.
MHz and
Above 1 Mbps, we again see significant differences in
effective energy usage (another order of magnitude from
∼1 Mbps to ∼25 Mbps), but the other performance parameters
(loss and throughput in our measurements) vary greatly. In this
case, a more complex (self-)management system, policies or
mechanisms may be required to achieve lower effective energy
usage without compromising performance.
So, from a management viewpoint, the system-wide effec-
tive energy usage and other performance parameters must be
considered within any (self-)management functionality. In the
next Section, we discuss how such trade-offs could be made
for different classes of applications, based on our experiments.
Page 5
IV. MANAGING ENERGY-AWARENESS IN APPLICATIONS
Based on our analyses in Section III, we can now determine
management actions and interventions that are possible in
order to enable energy efficient operation on a system-wide
basis by actions at the application-level. Our results also tell
us what is achievable: in Figure2, we have the mean effective
energy usage envelope across all our experiments.
A vertical ‘profile’ of this graph, for example, gives us
the margin of change that is possible for the same data rate
but with different packet sizes: the lower line shows the best
effective energy usage we can obtain.
A horizontal ‘profile’ of this graph shows us that, what
adjustment is possible to the maximum data without increasing
effective energy usage. Of course, this is not the complete
picture: other performance issues, such as loss, may need to
be considered, as shown in Figures 4 and 5.
Some applications will have greater scope for improving ef-
fective energy usage than others. Different approaches may be
required. For instance, VoIP applications require real time data
exchange and may be able to adapt their flow characteristics
in when the network conditions permit. The same applies to
ViIP applications, but as they normally do not require real-time
data transfers they may be able to apply a greater magnitude
of change to their flow characteristics and also use increased
buffering/caching at the application layer. We see that EAin
file transfer applications also benefits from higher data rates.
A. Energy-aware self-adaptation for VoIP flows
VoIP applications use small packets to minimise impact
of packet loss and to reduce end-to-end delay. However, we
observe that use of small packets is not energy efficient. If
network conditions permit, a VoIP application could alter its
packet size, without changing its offered load, in order to
improve EA. For example, consider Figure 3, which depicts
the delay budget for a VoIP application.
Fig. 3. Trading off delay for packet size in a VoIP transmitter. The maximum
delay tolerable, dM, is the upper bound. The application can estimate the
current end-to-end delay, de, using reports from the receiver. Assuming some
safety margin, dS(includes receiver-side delay), the application can delay the
packet transmission by da to create larger packets.
For a VoIP application to produce larger packets, it has to
delay packet transmission more than for a smaller packet size,
increasing the per-packet, end-to-end delay. Given an upper
bound on the maximum delay for the application, dM, there
may be an acceptable delay, da, that could be introduced by
the application. This delay, da, could be evaluated by taking
measurements of the current end-to-end delay (de, variable),
using existing means (e.g. RTCP Reports), and allowing some
safety margin (dS, for receiver delays such as decoding and
playout buffering for jitter smoothing). Additionally, the loss
of a larger packet could result in greater impact at the
receiver. Of course, such adaptation would also depend on the
capabilities of the audio codec in use, as different codecs have
different encoding/decoding delay and tolerance to loss. e.g.
a G.711/PCM-based codec may be more amenable to such
techniques than a CELP-based codec. The simple example
above shows only delay, but other factors may need to be
considered. Overall, the application would need to assess
current network and end-system conditions against its own
operating modes in order to make appropriate adjustments,
e.g. assess current packet loss experienced by the application,
and current power usage – different codecs may have different
impact on power consumption in an end-system, for example
as described in [10]. Such adaptation could also be integrated
with congestion control mechanisms (e.g. DCCP8), or as part
of an autonomic management policy (e.g. with the manage-
ment framework introduced in [16]). Similar considerations
would apply to real-time video, with the added complexity of
video coding/framing.
B. Energy-efficient video streaming
For ViIP (streamed, non-real-time video), if client-side
buffering is used to compensate for loss, it may be possible
to use a similar approach as in Section IV-A, but with less
aggressive constraints, as non-real time transfers are more
tolerable to end-to-end delay and loss. In this case, dynamic
codec selection may be used instead of packet size adaptation
in order to change data rates, or, where possible the use of
modern scaleable codecs such as H.264 AVC. In the limit, with
large buffering at the client-side, the streaming of a video file
can be approximated to the case of file transfer (see below). If
we consider the results of the EAvalues in Figures 4 and 5,
we see that the emulated youtube traffic has EAvalues that are
much higher than our emulations of file transfers using HTTP.
Again, the energy cost of the video codec must be considered.
C. Green caching for file transfer applications
File transfer applications (e.g. HTTP) already operate in
fairly energy efficient manner, according to our measurements.
So, there is a lower margin for improving effective energy
usage by adapting their flow characteristics. However, we
also observe that increasing data rates also reduces effective
energy usage. Our measurements for a HTTP download locally
(Intranet) and for a remote server (Internet) show that the local
download at a higher data rate was more energy efficient –
by an order of magnitude in our particular measurements, as
shown in Section III. So, if content is cached locally, this may
improve download data rates and energy usage. Of course,
locally cached content also avoids the energy overhead of
using the network resources for fetching the content from
the remote source. Although this may be hard to assess
quantitatively, the intuition is that it will be more energy
8http://www.erg.abdn.ac.uk/∼gerrit/dccp/apps/
Page 6
efficient. As site-wide caches are already available and used in
widely today, this offers an easy path to more energy efficient
file transfers.
Now, this may be seen as a cost trade-off: could monetary
savings in energy usage and lower volumes of downstream
traffic due to local caching be sufficient to offset the cost
of the installation and maintenance of the cache? Developing
appropriate models to investigate this is future work. Such
caching may not be appropriate for all content or all content
providers. For example, the youtube business model requires
visits to a youtube server so that the number of views of
content can be recorded. This is an engineering challenge,
again, suitable for future work.
V. RELATED WORK
The authors’ previous work [11] compares energy efficiency
in IEEE 802.11a and IEEE 802.11n at 5 GHz. It presents
an initial evaluation of the effective energy usage envelope
and provides the basic model for this study. The key new
contributions in this paper are to consider the dynamics of
the envelope with: (i) lower transmission power, 0 dBm (only
17 dBm previously); (ii) a 40 MHz channel (only 20 MHz
previously); (iii) higher data rates, up to 90 Mbps (only up to
25 Mbps previously); (iv) emulated data transfer with HTTP
via Internet and Intranet (previously only Skype and YouTube).
As well as work already listed in the Introduction, we provide
below a non-exhaustive summary of other relevant work.
In [17], the authors examine 802.11n energy efficiency and
conclude that transmissions with larger packets and higher data
rates are more energy efficient than those using smaller packet
sizes and lower data rates. However, their study measures the
power consumption directly at the wireless NIC, so they do not
capture system-wide effects. Additionally, their consideration
of bit rates is by looking at the modulation and coding scheme
(MCS) that is selected by the WLAN driver, while we consider
the data rate that is measured at the application-level.
In [18], the authors study energy efficiency and propose
their own extension to the 802.11 MAC protocol in order to
improve energy efficiency. Alternatively, in [19], the authors
consider an analytical model to evaluate energy efficiency.
They also propose an enhancement to 802.11, not at the
protocol level, but by allowing non-transmitting or receiving
NICs to remain in sleep mode for the duration of any ongoing
transmissions that are not of interest to those nodes. We
have not considered such work in our study, as our objective
was to understand typical energy efficiency for off-the-shelf
equipment with out-of-the-box configuration.
As we wished to consider off-the-shelf equipment, with out-
of-the-box configurations, we have for this study excluded
use of 802.11n features which are not normally turned on by
default, might be optional for implementation by vendors, are
vendor specific, or would require specific expert knowledge
by the user in order to configure those features for use.
Such features include 802.11 Power Save Mode (PSM) [2],
Unscheduled Automatic Power Save Delivery (U-APSD) [3],
WMM Power Save (WMM-PS) [4], Dynamic MIMO Power
Save [5], and Wake-on-Wireless [6]. A survey of energy
efficiency through MAC layer techniques is presented in [20].
There are recent studies on performance, e.g. [21], [22],
[23], but none of these consider energy usage. Our perfor-
mance results are in general agreement with those studies.
VI. CONCLUSION
We have performed an empirical investigation of IEEE
802.11n at 5 GHz in order to investigate its energy usage
dynamics, including how the effective energy usage envelope is
effected by changes in transmission power and channel usage.
We find that, as might be expected, transmission power and
channel width do effect performance, such as throughput and
loss. However, they have relatively little impact on application-
level energy usage. Our key observation is that we can gain
a better appreciation of energy usage by considering the
amortisation of energy usage across the system as a whole.
The greatest impact on effective energy usage, EA, results
from the use of large packets and higher transmission rates:
the energy usage of the system is amortised over the greater
number of bytes that are transmitted. We see changes of two
orders of magnitude in effective energy usage with changes
in data rates, and significant changes in energy usage for the
same data rate but with different packet sizes.
This means that there is great potential for application-
level (self-)adaptation to achieve greater energy efficiencies
than is currently achieved for IEEE 802.11n at 5 GHz, by
adjusting packet flow construction and packet transmission.
We also provide examples of how our findings can be exploited
in order to allow applications to trade off performance against
energy usage, but different mechanisms may apply to different
classes of applications. Although our experiments were carried
out for a WLAN cell, as we are observing the amortisation
of the energy/transmission costs of the end-system, the results
are also found to be similar for wired Ethernet [11].
This study is a first step towards understanding the dynamics
of real IEEE 802.11 systems with respect to building energy-
aware (self-)adaptive applications. Future extensions of this
work include (but are not limited to) examining multiple
clients in a cell; considering 2.4 GHz, e.g. 802.11b and
802.11g; comparing application-specific and OS-specific en-
ergy issues; and of course, considering the nature of the
application-level self-adaptation that can be realised based
on our new understanding of IEEE 8021.11 energy usage
dynamics. User studies would also be required in order to
determine the efficacy and impact on quality of experience
(QoE) of such adaptive techniques for multimedia applica-
tions. Such adaptation of traffic flows would also impact
network management techniques based on packet analysis.
ACKNOWLEDGEMENTS
Lito Kriara (University of Edinburgh, UK) provided feed-
back on this work. Prof Krishna M. Sivalingam and Narendran
Krishnan (IIT Madras, India) provided discussion and moti-
vational background. This work has been supported by the
EPSRC-funded project IU-ATC (http://www.iu-atc.com/).
Page 7
0
10
20
30
40
50
60
70
80
0.1 1 10 100
throughput [Mbps]
offered load [Mbps]
64 B packets
1460 B packets
Skype
YouTube
http/Internet
http/Intranet
0
10
20
30
40
50
60
70
80
0.1 1 10 100
throughput [Mbps]
offered load [Mbps]
64 B packets
1460 B packets
Skype
YouTube
http/Internet
http/Intranet
0
10
20
30
40
50
60
70
80
0.1 1 10 100
packet-loss [%]
offered load [Mbps]
64 B packets
1460 B packets
Skype
YouTube
http/Internet
http/Intranet
0
10
20
30
40
50
60
70
80
0.1 1 10 100
packet-loss [%]
offered load [Mbps]
64 B packets
1460 B packets
Skype
YouTube
http/Internet
http/Intranet
0.1
1
10
100
0.1 1 10 100
EA [J/Mb]
offered load [Mbps]
Skype
YouTube
http/Internet
http/Intranet
64 B packets
1460 B packets
0.1
1
10
100
0.1 1 10 100
EA [J/Mb]
offered load [Mbps]
Skype
YouTube
http/Internet
http/Intranet
64 B packets
1460 B packets
Fig. 4.IEEE 802.11n 20 MHz Channel - 0 dBm (left column) and 17 dBm (right column) TX power.
0
10
20
30
40
50
60
70
80
0.1 1 10 100
throughput [Mbps]
offered load [Mbps]
64 B packets
1460 B packets
Skype
YouTube
http/Internet
http/Intranet
0
10
20
30
40
50
60
70
80
0.1 1 10 100
throughput [Mbps]
offered load [Mbps]
64 B packets
1460 B packets
Skype
YouTube
http/Internet
http/Intranet
0
10
20
30
40
50
60
70
80
0.1 1 10 100
packet-loss [%]
offered load [Mbps]
64 B packets
1460 B packets
Skype
YouTube
http/Internet
http/Intranet
0
10
20
30
40
50
60
70
80
0.1 1 10 100
packet-loss [%]
offered load [Mbps]
64 B packets
1460 B packets
Skype
YouTube
http/Internet
http/Intranet
0.1
1
10
100
0.1 1 10 100
EA [J/Mb]
offered load [Mbps]
Skype
YouTube
http/Internet
http/Intranet
64 B packets
1460 B packets
0.1
1
10
100
0.1 1 10 100
EA [J/Mb]
offered load [Mbps]
Skype
YouTube
http/Internet
http/Intranet
64 B packets
1460 B packets
Fig. 5.IEEE 802.11n 40 MHz Channel. 0 dBm (left column) and 17 dBm (right column) TX power.
Page 8
-80
-40
0
40
80
0.1 1 10 100
∆ throughput [%]
offered load [Mbps]
64 B packets
1460 B packets
-80
-40
0
40
80
0.1 1 10 100
∆ throughput [%]
offered load [Mbps]
64 B packets
1460 B packets
-80
-40
0
40
80
0.1 1 10 100
∆ loss [%]
offered load [Mbps]
64 B packets
1460 B packets
-80
-40
0
40
80
0.1 1 10 100
∆ loss [%]
offered load [Mbps]
64 B packets
1460 B packets
-80
-40
0
40
80
0.1 1 10 100
∆ EA [%]
offered load [Mbps]
64 B packets
1460 B packets
-80
-40
0
40
80
0.1 1 10 100
∆ EA [%]
offered load [Mbps]
64 B packets
1460 B packets
Fig. 6.
Horizontal zero line is a visual aid. Positive values indicate where 0dBm TX power has higher values. Application-specific workloads omitted for readability
.
Differences in the main observables depending on TX power (0dBm - 17 dBm). 20 MHz channel (left column) and 40 MHz channel (right column).
-80
-40
0
40
80
0.1 1 10 100
∆ throughput [%]
offered load [Mbps]
64 B packets
1460 B packets
-80
-40
0
40
80
0.1 1 10 100
∆ throughput [%]
offered load [Mbps]
64 B packets
1460 B packets
-80
-40
0
40
80
0.1 1 10 100
∆ loss [%]
offered load [Mbps]
64 B packets
1460 B packets
-80
-40
0
40
80
0.1 1 10 100
∆ loss [%]
offered load [Mbps]
64 B packets
1460 B packets
-80
-40
0
40
80
0.1 1 10 100
∆ EA [%]
offered load [Mbps]
64 B packets
1460 B packets
-80
-40
0
40
80
0.1 1 10 100
∆ EA [%]
offered load [Mbps]
64 B packets
1460 B packets
Fig. 7.
Horizontal zero line is a visual aid. Positive values indicate where 20MHz channel has higher values. Application-specific workloads omitted for readability
.
Differences in the main observables depending on channel width (20Mhz - 40MHz). 0 dBm TX power (left column) and 17 dBm (right column).
Page 9
REFERENCES
[1] “SMART2020: Enabling the low carbon economy in the information
age.” The Climate Group, Tech. Rep., 2008. [Online]. Available:
http://www.smart2020.org/
[2] D. Qiao and K. Shin, “Smart power-saving mode for IEEE 802.11
wireless LANs,” in Proc. INFOCOM 2005, vol. 3, Mar 2005, pp. 1573–
1583.
[3] H. Liu and Y. Wang, “An Implementation Scheme of Power Saving
Mechanism for IEEE 802.11e,” in Computer Science and Software
Engineering, 2008 Intl. Conf. on, vol. 1, Dec 2008, pp. 1178–1181.
[4] B. Alfonsi, “In brief: Wi-Fi certification program aims to boost battery
life,” IEEE Distributed Systems Online, vol. 7, no. 1, Jan 2006.
[5] H. Kim, C.-B. Chae, G. de Veciana, and R. Heath, “Energy-efficient
adaptive MIMO systems leveraging dynamic spare capacity,” in Infor-
mation Sciences and Systems, 2008. CISS 2008. 42nd Annual Conference
on, Mar 2008, pp. 68–73.
[6] N. Mishra, K. Chebrolu, B. Raman, and A. Pathak, “Wake-on-WLAN,”
in Proceedings of the 15th international conference on World Wide Web.
New York, NY, USA: ACM, 2006, pp. 761–769.
[7] G. Bernardi, P. Buneman, and M. K. Marina, “Tegola Tiered Mesh
Network Testbed in Rural Scotland,” in Proc. ACM Mobicom 2008,
Workshop on ”Wireless Networks and Systems for Developing Regions”
(WiNS-DR), 2008.
[8] S. Baek and B. D. Choi, “Performance analysis of power save mode in
IEEE 802.11 infrastructure WLAN,” in Telecommunications, 2008. ICT
2008. Intl Conf. on, Jun 2008, pp. 1–4.
[9] T. Jun, Z. Ying-jiang, W. An-qing, Y. Zhi-wei, and J. Ming-bo, “A
research of wlan packet aggregation,” in Education Technology and
Computer Science (ETCS), 2010 Second International Workshop on,
vol. 2, Mar 2010, pp. 480–483.
[10] S. N. Bhatti and G. Knight, “Enabling QoS adaptation decisions for
Internet applications,” Computer Networks, vol. 31, pp. 669–692, 1999.
[11] M. Tauber, S. N. Bhatti, and Y. Yu, “Application Level Energy and Per-
formance Measurements in a Wireless LAN,” in Proc. GreenCom2011:
IEEE/ACM Intl. Conf. on Green Computing & Comms., Aug 2011.
[12] K.-T. Chen, C.-Y. Huang, P. Huang, and C.-L. Lei, “Quantifying Skype
User Satisfaction,” in Proc. SIGCOMM 2006, Aug 2006.
[13] D. Bonfiglio, M. Mellia, and M. Meo, “Revealing Skype Traffic: When
Randomness Plays with You,” in Proc. SIGCOMM 2007, Aug 2007.
[14] M. Zink, K. Suh, Y. Gu, and J. Kurose, “Watch Global, Cache Local:
YouTube Network Traffic at a Campus Network - Measurements and
Implications,” in Proc. IEEE MMCN 2008, 2008.
[15] P. Gill, M. Arlitt, Z. Li, and A. Mahanti, “Youtube traffic characteriza-
tion: a view from the edge,” in Proc. IMC 2007, 2007.
[16] M. Tauber, G. Kirby, and A. Dearle, “Self-Adaptation Applied to
Peer-Set Maintenance in Chord via a Generic Autonomic Management
Framework,” in Self-Adaptive Networks (SAN) Workshop, IEEE Intl.
Conf. on Self-Adaptive and Self-Organizing Systems (SASO2010), 2010.
[17] D. Halperin, B. Greensteiny, A. Shethy, and D. Wetherall, “Demystifying
802.11n power consumption,” in Proc. 2010 International Conference on
Power Aware Computing and Systems.
Association, 2010.
[18] S.-L. Wu and P.-C. Tseng, “An Energy Efficient MAC Protocol for IEEE
802.11 WLANs,” Communication Networks and Services Research,
Annual Conference on, vol. 0, pp. 137–145, 2004.
[19] M. Ergen and P. Varaiya, “Decomposition of Energy Consumption in
IEEE 802.11,” in IEEE Intl. Conference on Communications, 2007. ICC
’07, 2007, pp. 403–408.
[20] S.-L. Tsao and C.-H. Huang, “A survey of energy efficient MAC
protocols for IEEE 802.11 WLAN,” Computer Communications, vol. 34,
no. 1, pp. 54 – 67, 2011.
[21] S. Sendra, P. Fernandez, C. Turro, and J. Lloret, “IEEE 802.11a/b/g/n
Indoor Coverage and Performance Comparison,” in 6th International
Conference on Wireless and Mobile Communications (ICWMC), 2010.
[22] J. Pacheco de Carvalho, H. Veiga, N. Marques, C. Pacheco, and A. Reis,
“A contribution to performance measurements of IEEE 802.11 a, b,
g laboratory WEP point-to-point links using TCP, UDP and FTP,” in
International Conference on Applied Electronics (AE 2010), 2010.
[23] E. Gelal, K. Pelechrinis, I. Broustis, S. Krishnamurhty, S. Mohammed,
A. Chockalingam, and S. Kasera, “On the Impact of MIMO Diversity on
Higher Layer Performance,” in Distributed Computing Systems (ICDCS),
2010 IEEE 30th International Conference on, 2010, pp. 764–773.
Berkeley, CA, USA: USENIX
Supplementary resources (1)
-
noms2012-tby2012